Datasets:

dongqi-me

/

SciNews

like 3

Medicine

New study shows how cells can be led down non-cancer path

PKM2 methylation by CARM1 activates aerobic glycolysis to promote tumorigenesis, Nature Cell Biology. DOI: 10.1038/ncb3630 Journal information: Nature Cell Biology

http://dx.doi.org/10.1038/ncb3630

https://medicalxpress.com/news/2017-10-cells-non-cancer-path.html

Abstract Metabolic reprogramming is a hallmark of cancer. Herein we discover that the key glycolytic enzyme pyruvate kinase M2 isoform (PKM2), but not the related isoform PKM1, is methylated by co-activator-associated arginine methyltransferase 1 (CARM1). PKM2 methylation reversibly shifts the balance of metabolism from oxidative phosphorylation to aerobic glycolysis in breast cancer cells. Oxidative phosphorylation depends on mitochondrial calcium concentration, which becomes critical for cancer cell survival when PKM2 methylation is blocked. By interacting with and suppressing the expression of inositol-1,4,5-trisphosphate receptors (InsP 3 Rs), methylated PKM2 inhibits the influx of calcium from the endoplasmic reticulum to mitochondria. Inhibiting PKM2 methylation with a competitive peptide delivered by nanoparticles perturbs the metabolic energy balance in cancer cells, leading to a decrease in cell proliferation, migration and metastasis. Collectively, the CARM1–PKM2 axis serves as a metabolic reprogramming mechanism in tumorigenesis, and inhibiting PKM2 methylation generates metabolic vulnerability to InsP 3 R-dependent mitochondrial functions. Main One hallmark of cancer 1 , 2 is the Warburg effect, whereby tumour cells rely mainly on aerobic glycolysis for adenosine-5′-triphosphate (ATP) production, even with sufficient oxygen 3 . However, metabolic adaptation in tumours extends beyond the Warburg effect, including balancing energy needs with equally important needs for macromolecular synthesis and redox homeostasis 1 , 2 , 4 . Emerging evidence suggests that mitochondrial respiration is crucial for tumorigenesis and presents a target for cancer therapy 5 , 6 , 7 , 8 . Pyruvate kinase (PK) catalyses the final step in glycolysis, converting phosphoenolpyruvate to pyruvate while phosphorylating ADP to produce ATP. The M1 and M2 isoforms of PK are produced by mutually exclusive alternative splicing of PKM pre-messenger RNA 9 . Although PKM1 and PKM2 differ by only 22 amino acids, PKM1 is not allosterically regulated and exists in a tetrameric form with high PK activity. PKM2 shifts between inactive dimeric and active tetrameric forms, modulated by phosphotyrosine signalling 10 , metabolic intermediates (for example, FBP, serine and SAICAR) 11 , 12 and post-translational modifications 13 . Switching PKM2 to PKM1 reverses aerobic glycolysis to oxidative phosphorylation and reduces tumour formation in nude mice 14 , identifying PKM2 as a potential cancer therapy target. However, a recent report challenged the PKM2-catalysed reaction as a rate-limiting step in cancer cell glycolysis 15 and a possible protein kinase activity of PKM2 remains controversial 16 . Co-activator-associated arginine methyltransferase 1 (CARM1), also known as PRMT4, is a type I protein arginine methyltransferase (PRMT) that asymmetrically dimethylates protein substrates including histones, transcriptional factors and co-regulators, splicing factors and RNA polymerase II 17 , 18 , 19 , 20 . CARM1 is overexpressed in breast cancer to promote cancer growth 21 , and elevated CARM1 expression correlates with poor prognosis 22 . Recently, we discovered that chromatin remodelling factor BAF155 methylation by CARM1 promotes breast cancer progression and metastasis 23 . However, whether CARM1 regulates energy metabolism in cancer cells remains unknown. Here, we discovered CARM1–PKM2 interaction as a major contributor to metabolic reprogramming in cancer. CARM1 methylates the dimeric form of PKM2 at Arg445/447/455. Methylated PKM2 promotes tumour cell proliferation, migration and lung metastasis by reprogramming oxidative phosphorylation to aerobic glycolysis, and this effect was reversed by a competitive PKM2 peptide delivered using nanoparticles. We showed that methylated PKM2 localized in mitochondria-associated endoplasmic reticulum membrane, through interaction with inositol 1,4,5-trisphosphate receptors (InsP 3 Rs), decreasing mitochondrial membrane potential (Δ Ψ m ) and Ca 2+ uptake, which is essential for activating pyruvate dehydrogenase (PDH) to support oxidative phosphorylation 24 . Blocking PKM2 methylation elevates InsP 3 R expression, increasing mitochondrial Ca 2+ uptake, PDH activation and oxidative phosphorylation. Thus, PKM2 methylation represents an important regulator of switching between oxidative phosphorylation to aerobic glycolysis in cancer cells. RESULTS CARM1 interacts with and methylates PKM2 Consistent with CARM1 promotion of tumour development and progression 21 , 23 , knocking out CARM1 decreased DNA synthesis in MCF7 cells ( Supplementary Fig. 1a ). CARM1 knockout (KO) also increased mitochondrial oxygen consumption rate (OCR) but decreased lactate production without affecting glucose uptake ( Supplementary Fig. 1b–f ). These results prompted us to test whether CARM1 modulates energy metabolism in breast cancer cells. We identified PKM2 as a putative CARM1-interacting protein by mass spectrometry when Halo-tagged CARM1 was overexpressed in HEK293T cells ( Supplementary Table 1 ). Endogenous CARM1–PKM2 interaction was confirmed by reciprocal co-immunoprecipitation in MCF7 cells ( Fig. 1a ). To determine whether CARM1 directly interacts with PKM2, we performed glutathione S -transferase (GST) pulldown using recombinant GST-tagged CARM1, Flag-tagged PKM2 and Flag-PABP1 (positive control) 25 and observed that GST-CARM1, but not GST alone, bound PKM2, indicating direct CARM1–PKM2 interaction ( Fig. 1b ). Interestingly, recombinant PKM1 also associated with CARM1 in vitro ( Fig. 1b ). To map the PKM2 region that binds CARM1, we expressed truncated Flag-PKM2 using in vitro transcription and translation and performed in vitro interaction assays with GST-CARM1. Deleting PKM2 domain C (Δ1) or N/A1 (Δ4) significantly decreased the interaction of PKM2 with CARM1, and truncation of both domains (Δ8) completely abolished the interaction, suggesting that the C and N/A1 domains are responsible for the interaction ( Fig. 1c, d ). Although these domains are identical between PKM1 and PKM2, intriguingly only PKM2, but not PKM1, can be methylated by CARM1 in vitro ( Fig. 1e ). Figure 1: Only the dimeric form of PKM2 is methylated by CARM1. ( a ) Reciprocal co-immunoprecipitation of PKM2 and CARM1 in MCF7 cells followed by western blot analysis. ( b ) Direct interaction of recombinant PKM1, PKM2 and PABP1 (positive control) proteins with CARM1 in GST pulldown assays. GST and GST-CARM1 were detected by anti-GST antibody, and Flag-tagged PKM1, PKM2 and PABP1 were detected by anti-Flag antibody, respectively. ( c , d ) Schematic showing truncations of PKM2 and in vitro interaction assays with CARM1 using full-length PKM and truncated proteins expressed by in vitro transcriptional and translational systems. ( e ) In vitro methylation assays using recombinant CARM1, GST-PKM1 or PKM2 protein in the presence of 3 H-SAM. ( f ) In vitro methylation assays of PKM2 by CARM1 in the presence of FBP (100 μM) or TEPP-46 (1 μM). Histone proteins were used as negative controls. ( g ) In vitro methylation assays of PKM2 by PRMTs. Data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image Since non-methylatable PKM1 forms only tetramers while methylatable PKM2 forms tetramers or dimers, we hypothesized that tetramer formation may prevent CARM1-mediated methylation. TEPP-46 (ref. 26 ), a PKM2 activator, or FBP stimulated PKM2 tetramer formation ( Supplementary Fig. 2a, b ), inhibiting PKM2 methylation by CARM1 ( Fig. 1f ). Neither TEPP-46 nor FBP affected histone H3 methylation by CARM1 ( Fig. 1f ), excluding the possibility that TEPP-46 and FBP interfere with CARM1 methyltransferase activity. Thus, only the dimeric form of PKM2 is methylated by CARM1. Mammalian genomes encode nine PRMTs that sometimes share the same substrates 17 . Using purified Halo-tagged 23 PRMTs (1–8), we assayed methylation of recombinant PKM2 by PRMTs and showed that PKM2 is uniquely methylated by CARM1 ( Fig. 1g ). CARM1 methylates PKM2 at Arg445/447/455 To narrow down the PKM2 methylation site(s), we assayed methylation in vitro using purified GST-tagged full-length (FL) or truncated PKM2 ( Fig. 2a ). Among the three truncated fragments, only the C domain was significantly methylated ( Fig. 2b ), suggesting that it might contain the methylation site(s) of PKM2. Three arginine residues (Arg445, Arg447 and Arg455) were identified in the in vitro methylated GST-PKM2 (390–531 amino acid) fragment using liquid chromatography coupled tandem mass spectrometry (LC–MS/MS) ( Fig. 2c ). To further discern the major methylation site(s) of PKM2, we substituted lysine for the three methylated arginines to preserve their positive charge, individually or in combination, in GST-PKM2 ( Fig. 2d ). While individually mutating each arginine site slightly affected PKM2 methylation ( Fig. 2e ), mutating any two sites dramatically decreased methylation, and mutating all three sites abolished methylation ( Fig. 2e ). While all three arginine residues reside in the C domain that fosters tetramer formation, none is at the tetrameric interface ( Supplementary Fig. 2a ). Size-exclusion chromatography using purified recombinant His-tagged proteins showed that neither PKM2 ( Supplementary Fig. 2b ) nor PKM1 ( Supplementary Fig. 2c ) tetramer formation was affected by mutating Arg445, Arg447 and Arg455 to lysine. Thus, while CARM1 predominantly methylates dimeric PKM2, PKM2 tetramer formation is not affected by PKM2 methylation. PKM1 is not methylated by CARM1 possibly because the corresponding arginines in tetrameric PKM1 are inaccessible to CARM1. Figure 2: CARM1 methylates PKM2 at Arg445, Arg447 and Arg455. ( a ) Schematic diagram of full-length PKM2 and its truncated derivatives. ( b ) Mapping of the methylation domain of PKM2 by CARM1 in in vitro methylation assays. ( c ) Identification of dimethylated Arg445, Arg447 and Arg455 of the in vitro methylated GST-PKM2 protein. ( d ) Schematic showing arginine to lysine mutations on the GST-PKM2 protein. ( e ) In vitro methylation assays of PKM2 mutants relative to the wild-type PKM2. In b , e , data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image Inhibiting PKM2 methylation decreases breast cancer cell proliferation and migration and tumour growth To investigate the function of PKM2 methylation in breast cancer cell lines, we employed CRISPR/Cas9 technology to knock out endogenous PKM2. Genomic DNA sequencing of two representative PKM2-KO clones of MCF7 or MDA-MB-231 revealed non-homologous end-joining-induced random insertions or deletions at the targeted site ( Supplementary Fig. 3a ), causing mRNA degradation and PKM2 protein loss ( Fig. 3a, b ). Specific PKM2 KO led to compensatory PKM1 induction ( Fig. 3a, b ), as in primary mouse embryonic fibroblast (MEF) cells 27 . Whole-proteome analysis showed that, among >4,000 proteins detected, 158 proteins significantly increased (fold >= 1.5, P < 0.05), and 261 proteins decreased (fold <= 0.7, P < 0.05) in PKM2 KO versus parental cells ( Fig. 3c and Supplementary Table 2 ). In agreement with specific PKM2 KO, PKM2 protein levels decreased by >5-fold and PKM1 increased by >1.6-fold. The total PK activity in PKM2-KO cells decreased as compared with the parental cells ( Supplementary Fig. 3b ), possibly because PKM1 restoration in PKM2-KO cells is insufficient to compensate PKM2 loss. No PK activity change was observed in paired parental and CARM1-KO MCF7 and MDA-MB-231 cell lines ( Supplementary Fig. 3c ), indicating that CARM1-mediated PKM2 methylation does not affect PKM2 PK activity. Figure 3: Inhibition of PKM2 methylation decreases cell proliferation and migration. ( a ) Quantitative PCR analyses of mRNA levels of PKM1 and PKM2 in parental MCF7 and MDA-MB-231 cells and their respective PKM2-KO clones ( n = 3 independent experiments). ( b ) Western blot analyses of PKM1 and PKM2 in parental MCF7 and MDA-MB-231 cells and their respective PKM2-KO clones. ( c ) Mass spectrometry analysis of global protein changes between parental MCF7 and PKM2-KO cells ( n = 3 independent experiments). ( d ) EdU incorporation assays of parental MCF7 and PKM2-KO clones ( n = 3 independent experiments). ( e ) Cell cycle analyses of parental MCF7 and MCF7 PKM2-KO clones ( n = 3 independent experiments). ( f ) Western blot analysis of PKM2, methyl-PKM2 and PKM1 in parental and genetically engineered MCF7 and MDA-MB-231 cells, where shPKM1 represents PKM1 KD. ( g ) Measurement of cell proliferation by MTT assays in parental MCF7, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells ( n = 10 independent experiments). ( h ) Measurement of cell migration by Transwell assays in parental MDA-MB-231, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells ( n = 3 independent experiments). Scale bars, 50 μm. ( i ) The growth curves of MDA-MB-231 PKM2 WT /shPKM1 and PKM2 mut /shPKM1 xenografts ( n = 6 biologically independent mice per group). ( j ) Representative images of the xenograft tumours. Statistical significance was assessed using ANOVA. In a , d , g , i , data are shown as mean ± s.d. and statistics source data are available in Supplementary Table 7 . Statistical significance was assessed using two-tailed t -test ( a , d ) and ANOVA ( g , i ), ∗ ∗ ∗ P < 0.001. In b , f , h , data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image In MEF cells, deleting PKM2 from one or both allele(s) stimulated PKM1 expression and arrested proliferation, with reduced DNA synthesis due to PKM1 expression, rather than PKM2 loss 27 . Similarly, EdU incorporation and S-phase accumulation revealed dramatically reduced DNA synthesis when PKM2 was knocked out in MCF7 cells ( Fig. 3d, e ). However, the mechanism of reduced DNA synthesis in MCF7 cells appears to differ from that of MEF cells. PKM2 KO induced massive reduction of nucleotides in MEF cells 27 , but not in MCF7 and MDA-MB-231 cells ( Supplementary Table 3 ). Thus, reduced DNA synthesis in MCF7 cells was not due to lack of nucleotides, as in MEFs. Also in contrast to MEFs 27 , PKM1 overexpression in MCF7 cells did not inhibit cell proliferation ( Supplementary Fig. 3d ) in the presence of PKM2 expression. To delineate the functions of PKM2 methylation on energy metabolism without interference from PKM1, we restored PKM2 WT or PKM2 mut (where mut = mut1,2,3 from Fig. 2d ) in PKM2-KO cell lines, followed by knocking down PKM1 ( Fig. 3f ). While knocking down PKM1 60–80% in PKM2-KO cells did not affect cell proliferation and oxidative phosphorylation ( Supplementary Fig. 3e–g ), cell viability was dramatically reduced when PKM1 knockdown (KD) reached nearly 100%, possibly because these cells have insufficient PK to support survival. To ensure that PKM2 mut was defective in PKM2 methylation, we generated an antibody against an asymmetrically dimethylated Arg445 and Arg447 peptide of PKM2, referred to as the methyl-PKM2 antibody. While PKM2 WT and PKM2 mut were restored to similar levels, methylated PKM2 was detected only in parental and PKM2 WT -expressing cells, but not in PKM2-KO or PKM2 mut -expressing cells ( Fig. 3f ) or in CARM1-KO cells ( Supplementary Fig. 3h ), demonstrating the antibody specificity. We used these cell lines to investigate the effects of methyl-PKM2 on cell proliferation and colony formation in MCF7 cells and cell migration in MDA-MB-231. PKM2 methylation-defective cells (for example, PKM2 KO and PKM2 mut ) elicited reduced cell proliferation and colony formation ( Fig. 3g and Supplementary Fig. 3i ) compared with parental and PKM2 WT MCF7 cells; however, these effects appeared not to be caused by apoptosis ( Supplementary Fig. 3j ). PKM2 methylation status also affected cell migration ( Fig. 3h ) and tumour growth of MDA-MB-231 xenografts ( Fig. 3i, j ). Therefore, PKM2 methylation is required for promoting cell proliferation, migration and tumour growth in various breast cancer cell models. Blocking PKM2 methylation results in elevated mitochondrial respiration in breast cancer cells We next examined whether PKM2 methylation regulates metabolic fluxes. The switch from PKM2 to PKM1 in MCF7 and MDA-MB-231 PKM2-KO cells significantly increased OCR and reduced lactate production ( Fig. 4a–c ). Remarkably, the balance of OCR and lactate production could be reversed by restoring PKM2 WT , but not PKM2 mut ( Fig. 4b, c ). TEPP-46 treatment, which triggers PKM2 tetramerization, thus blocking its methylation (Figs 1f and 4d ), also significantly increased OCR while decreasing lactate production ( Fig. 4e–g ). Thus, PKM2 methylation modulates energy metabolism in cancer cells. Figure 4: Inhibiting PKM2 methylation increased mitochondrial oxidative phosphorylation. ( a ) The OCR curves in parental MCF7, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells treated with oligomycin, FCCP and rotenone/antimycin A ( n = 6 independent experiments). ( b ) Basal OCR and lactate production normalized to the cell numbers in parental MCF7, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells ( n = 6 independent experiments). ( c ) Basal OCR and lactate production normalized to cell numbers in parental MDA-MB-231, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells ( n = 6 independent experiments). ( d ) Western blot analysis of methyl-PKM2 in MCF7 cells treated with DMSO or TEPP-46. ( e ) The OCR curves in parental MCF7 cells treated with DMSO or TEPP-46 ( n = 6 independent experiments). ( f , g ) Basal OCR and lactate production normalized to cell numbers in MCF7 cells ( f ) or MDA-MB-231 ( g ) treated with DMSO or TEPP-46 ( n = 6 independent experiments). In a – c , e – g , data are shown as mean ± s.d. and statistics source data are available in Supplementary Table 7 . Statistical significance was assessed using two-tailed t -test ( f , g ) and ANOVA ( b , c ), ∗ ∗ P < 0.01, ∗ ∗ ∗ P < 0.001. In d , data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image Reactive oxygen species (ROS), an inevitable by-product of mitochondrial oxidative phosphorylation, are often scavenged by nicotinamide adenine dinucleotide phosphate (NADPH) and glutathione (GSH) 1 , 28 . As blocking PKM2 methylation increases oxidative phosphorylation, it may elevate ROS and deplete NADPH and GSH. Indeed, ROS levels were elevated by knocking out PKM2 ( Supplementary Fig. 4a ), mutating PKM2 methylation sites ( Supplementary Fig. 4b ) or knocking out CARM1 ( Supplementary Fig. 4c ) in MCF7 cells relative to corresponding controls, and increased ROS was accompanied by a decreased NADPH/NADP + ratio ( Supplementary Fig. 4d, e ) and GSH concentration ( Supplementary Fig. 4f, g ). Similarly, the NADPH/NADP + ratio ( Supplementary Fig. 4h ) and GSH concentration ( Supplementary Fig. 4i ) were higher in parental and PKM2 WT -expressing cells than PKM2-KO and PKM2 mut -expressing MDA-MB-231 cells. However, mitoTEMPO, a specific scavenger of mitochondrial superoxide, and glutathione did not alter cell proliferation and migration in cells producing high levels of ROS (that is, PKM2-KO expressing PKM2 mut and CARM1-KO cells) ( Supplementary Fig. 4j–o ). Thus, glycolytic metabolism and growth effects regulated by PKM2 methylation are largely independent of ROS. Inhibiting PKM2 methylation increases Ca 2+ uptake and mitochondrial membrane potential To investigate the mechanism by which mitochondrial respiration was elevated in PKM2 methylation-defective cells relative to PKM2 methylation-competent cells, we evaluated the effects of PKM2 methylation on mitochondrial membrane potential (Δ Ψ m ), an indicator of oxidative energy metabolism. Incorporation of mitochondrial-specific JC-1 dye followed by flow cytometry showed that Δ Ψ m increased after knocking out PKM2 in MCF7 cells, and that restoring PKM2 WT , but not PKM2 mut , in PKM2-KO cells abrogated increased Δ Ψ m ( Fig. 5a ). To validate this, we used tetramethylrhodamine ethyl ester (TMRE) to measure Δ Ψ m . Similar results were obtained in both MCF7 ( Fig. 5b ) and MDA-MB-231 cells ( Supplementary Fig. 5a ). In accordance with the PKM2 methylation-dependent Δ Ψ m change, CARM1 KO also increased Δ Ψ m in MCF7 cells ( Fig. 5c ). Increased mitochondrial DNA copy number is another indicator of increased mitochondrial activity. Ablating PKM2 expression or PKM2 methylation resulted in increased mitochondrial DNA content in MCF7 cells ( Supplementary Fig. 5b ). Therefore, PKM2 methylation suppresses mitochondrial function and loss of PKM2 methylation releases this suppression and elevates Δ Ψ m . Figure 5: Inhibiting PKM2 methylation increases mitochondrial membrane potential and [Ca 2+ ] mito . ( a – c ) Measurement of mitochondrial membrane potential (Δ Ψ ) by incorporation of JC-1 dye ( a ) or TMRE dye ( b , c ) followed by flow cytometry. The Δ Ψ was measured by incorporation of JC-1 ( a ) or TMRE dyes ( b ) in parental MCF7, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells ( n = 3 independent experiments). Alternatively, the Δ Ψ was measured by TMRE dye incorporation in parental MCF7, PKM2-KO and CARM1-KO cells ( c ) ( n = 3 independent experiments). ( d ) Representative images of co-localized mitochondrial tracker and Rhod-2 in MCF7 PKM2-KO cells. Scale bars, 5 μm. ( e ) Representative images of Rhod-2-labelled mitochondria in parental MCF7, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells. Scale bars, 10 μm. ( f – i ) Relative basal [Ca 2+ ] mito in Rhod-2-labelled parental, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 MCF7 ( f ) ( n = 3 independent experiments) or corresponding MDA-MB-231 cells ( g ) ( n = 3 independent experiments); or parental MEF and PKM2-KO cells ( h ) ( n = 3 independent experiments); or parental MCF7, CARM1-KO and PKM2-KO cells ( i ) ( n = 3 independent experiments). ( j – m ) Western blotting of phosphorylated PDH and total PDH in the indicated MCF7 ( j ) or corresponding MDA-MB-231 cells ( k ); or parental MEF and PKM2-KO ( l ); or parental MCF7, CARM1-KO and PKM2-KO cells ( m ). ( n , o ) Western blot analysis of phosphorylated PDH and total PDH in MCF7 ( n ) or MDA-MB-231 ( o ) cells treated with DCA. ( p , q ) Basal OCR ( p ) and lactate production ( q ) normalized to the cell numbers in MCF7 or MDA-MB-231 cells treated with DCA ( n = 6 independent experiments). In b , c , f – h , p – q , data are shown as mean ± s.d. and statistics source data are available in Supplementary Table 7 . Statistical significance was assessed using two-tailed t -test ( h , p , q ) and ANOVA ( b , c , f , g , i ), ∗ P < 0.05, ∗ ∗ P < 0.01, ∗ ∗ ∗ P < 0.001. NS, not significant. In d , e , data represent one of two independent experiments with similar results. In a , j – o , data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image Mitochondrial Ca 2+ , primarily driven by Δ Ψ m , stimulates oxidative phosphorylation to maintain cellular energy homeostasis 24 , 29 . Δ Ψ m sensitivity to PKM2 methylation prompted us to assess basal mitochondrial Ca 2+ concentration ([Ca 2+ ] mito ) using a Ca 2+ -sensitive Rhod-2 AM dye and confocal imaging ( Fig. 5d ). PKM2 KO profoundly increased basal [Ca 2+ ] mito in MCF7 ( Fig. 5e, f ) and MDA-MB-231 ( Fig. 5g ) cells. The effect is not cancer-cell-specific, as [Ca 2+ ] mito also increased in PKM2-KO MEF cells (PKM2 fl/fl , Cre-oestrogen receptor) 27 ( Fig. 5h ). Restoring PKM2 WT , but not PKM2 mut , abrogated the elevated basal [Ca 2+ ] mito caused by PKM2 KO in MCF7 and MDA-MB-231 cells ( Fig. 5f, g ). Basal [Ca 2+ ] mito was also elevated in CARM1-KO MCF7 cells ( Fig. 5i ). Thus, methylated PKM2 suppresses mitochondrial Ca 2+ uptake. Mitochondrial matrix calcium regulates oxidative phosphorylation through activating several dehydrogenases, including PDH, which couples glycolysis to the tricarboxylic acid (TCA) cycle by pyruvate decarboxylation 24 . PDH activity is suppressed by phosphorylation by PDH kinase and enhanced by dephosphorylation by Ca 2+ -dependent pyruvate phosphatases 24 , 30 . To investigate whether altered [Ca 2+ ] mito levels change PDH activity, we measured phosphorylated PDH levels in PKM2-KO, CARM1-KO and PKM2 WT - or PKM2 mut -restored cell lines by western blot. PKM2 KO dramatically decreased PDH phosphorylation in MCF7 ( Fig. 5j ), MDA-MB-231 ( Fig. 5k ) and MEF ( Fig. 5l ) cells, implying increased PDH activity by [Ca 2+ ] mito influx. Restoring PKM2 WT , but not PKM2 mut , increased phosphorylated PDH in MCF7 and MDA-MB-231 cells ( Fig. 5j, k ), implying attenuated PDH activity. CARM1 KO similarly decreased PDH phosphorylation in MCF7 ( Fig. 5m ). As expected, treating cells with dichloroacetate (DCA), a PDH kinase inhibitor, also decreased PDH phosphorylation and lactate production while increasing oxidative phosphorylation ( Fig. 5n–q ). Thus, PKM2 methylation is critical for restraining mitochondrial oxidative phosphorylation via decreasing mitochondrial Δ Ψ m and Ca 2+ uptake, and increasing PDH phosphorylation. Methylated PKM2 decreases InsP 3 R expression A previous study reported PKM2 in mitochondria 31 . We confirmed PKM2 in the mitochondrial fraction by subcellular fractionation ( Supplementary Fig. 6a ) and observed a portion of PKM2 co-localizing with HSPA9 in mitochondrial outer membrane ( Supplementary Fig. 6b ). However, the mitochondrial localization of PKM2 appeared not to be affected by its methylation ( Supplementary Fig. 6c ). To elucidate how methylated PKM2 modulates mitochondrial oxidative phosphorylation, we overexpressed Flag-tagged PKM2 WT or PKM2 mut in HEK293T cells, and performed MS analyses on Flag-PKM2 co-immunoprecipitated proteins to identify differentially interacting proteins. Many interacting proteins were identical between PKM2 WT and PKM2 mut ( Supplementary Tables 4 and 5 ); however, the endoplasmic reticulum (ER) calcium-releasing proteins InsP 3 R1, 2 and 3 showed the most notable difference between PKM2 WT and PKM2 mut ( Supplementary Fig. 6d ). Interestingly, InsP 3 R1 and InsP 3 R3 are not only putative PKM2-interacting proteins but their levels also increased in PKM2-KO MCF7 cells ( Fig. 3c and Supplementary Fig. 6e ). To confirm the proteomics results, we examined InsP 3 R1 and InsP 3 R3 expression levels by western blotting in different PKM2-KO clones from MCF7 and MDA-MB-231 cells. Both InsP 3 R1 and InsP 3 R3 were significantly elevated in three different PKM2-KO clones of two cell lines ( Fig. 6a ). As a negative control, levels of HSPA9, another PKM2-interacting protein, were insensitive to PKM2 KO. To test whether increased InsP 3 Rs resulted from increased PKM1 in PKM2-KO cells, we measured the kinetics of protein changes in immortalized, tamoxifen-inducible PKM2-KO (PKM2 fl/fl , Cre-oestrogen receptor) MEFs. As reported previously 27 , PKM1 expression was elevated after a two-day 4-hydroxytamoxifen (4-OHT) treatment and plateaued after four-day treatment ( Fig. 6b ). However, increased InsP 3 R1 and InsP 3 R3 levels were detected only later when PKM2 was substantially lower ( Fig. 6b ). Moreover, overexpressing PKM1 failed to increase InsP 3 R expression ( Supplementary Fig. 6f ), reinforcing that InsP 3 Rs are regulated by PKM2 not PKM1. Co-immunoprecipitation showed that PKM2, but not PKM1, interacted with the endogenous InsP 3 R1 and InsP 3 R3 in breast cancer cells ( Fig. 6c and Supplementary Fig. 6g ). Thus, elevated InsP 3 R levels in PKM2-KO cells probably result from PKM2 loss rather than PKM1 gain. Inverse expression changes of InsP 3 Rs and PKM2 were also observed in The Cancer Genome Atlas (TCGA) breast tumour specimens 32 . Across 1,093 primary breast tumours in TCGA, CARM1 and PKM2 mRNA levels were positively correlated but negatively correlated with InsP 3 R1 and InsP 3 R2 expression ( Fig. 6d ). Similarly, in the CPTAC 77 breast tumour proteogenomics database 33 , InsP 3 R2 protein level was negatively correlated with CARM1 or PKM ( Fig. 6e ). Figure 6: Methylated PKM2 decreases InsP 3 R expression. ( a ) Western blotting of InsP 3 R1, InsP 3 R3, p53 and HSPA9 in parental and PKM2-KO MCF7 or MDA-MB-231 cells. ( b ) Western blotting of InsP 3 R1, InsP 3 R3, p53, PKM1 and PKM2 in MEF (PKM2 fl/fl , Cre-ER) cells treated with 4-OHT for the indicated time. ( c ) Co-immunoprecipitation of InsP 3 Rs and HSPA9 with PKM2 but not PKM1. Flag-tagged PKM1 or PKM2 is immunoprecipitated from cell lysates derived from parental MCF7 or PKM2-KO cells using anti-Flag antibody followed by detection of PKM1, PKM2, InsP 3 R1, InsP 3 R3 and HSPA9 by western blot. ( d ) mRNA correlation of CARM1, InsP 3 R1, InsP 3 R2, InsP 3 R3 and PKM2 in 1,093 primary breast tumours of TCGA ( n = 1,093 biologically independent patient samples). ( e ) Protein correlation of CARM1, InsP 3 R1, InsP 3 R2, InsP 3 R3 and PKM in 77 breast tumours of the CPTAC collection 33 ( n = 77 biologically independent patient samples). ( f ) Co-immunoprecipitation of InsP 3 R1 and InsP 3 R3 with PKM2 WT but not PKM2 mut in MCF7 (top panel) and overexpressed PKM1 in HEK293T PKM2-KO cells (bottom panel). InsP 3 R1, InsP 3 R3, PKM1, PKM2 and methyl-PKM2 were detected by western blot in Flag-PKM1/2 immunoprecipitates using corresponding antibodies. ( g ) Western blotting of InsP 3 Rs and p53 in parental MCF7 (or MDA-MB-231), PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells. ( h ) Western blotting of p53, InsP 3 R1, InsP 3 R3 and PKM1 in MCF7 PKM2-KO and MDA-MB-231 PKM2-KO cells expressing control shRNA and p53 shRNA. ( i ) Western blotting of p53 in parental MCF7 and CARM1-KO cells. In a – c , f – i , data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image To determine whether PKM2–InsP 3 R interaction is affected by PKM2 methylation, we precipitated PKM2 from MCF7 cells stably expressing Flag-tagged PKM2 WT or PKM2 mut using anti-Flag M2 resin. PKM2–InsP 3 R interaction was detected in PKM2 WT - but not PKM2 mut -expressing cells ( Fig. 6f , upper panel). Even when Flag-PKM1 was overexpressed in HEK293T PKM2-KO cells, no PKM1–InsP 3 R interaction was detected ( Fig. 6f , lower panel). Interestingly, InsP 3 R expression appears sensitive to PKM2 methylation since restoring PKM2 WT , but not PKM2 mut , abrogated elevated InsP 3 R expression in PKM2-KO cells ( Fig. 6g ). Furthermore, blocking PKM2 methylation by knocking out CARM1 or TEPP-46 treatment elevated InsP 3 R3 expression ( Supplementary Fig. 6h ). Thus, InsP 3 R expression is inversely correlated with methylated PKM2. ChIP-seq 34 revealed p53 binding to InsP 3 R promoters, indicating that InsP 3 Rs could be direct p53 target genes. InsP 3 R mRNA levels were increased by PKM2 KO, particularly in MDA-MB-231 cells ( Supplementary Fig. 6i ), consistent with increased p53 protein levels in PKM2-KO clones of MCF7, MDA-MB-231 and MEF cells ( Fig. 6a, b ). Knocking down p53 using short hairpin RNAs (shRNAs) significantly decreased InsP 3 R protein levels ( Fig. 6h ). p53 protein levels were sensitive to PKM2 methylation, as restoring PKM2 WT but not PKM2 mut reduced p53 expression ( Fig. 6g ). Conversely, CARM1 KO induced p53 expression ( Fig. 6i ). Accordingly, methylated PKM2, via downregulating p53, is one means to control InsP 3 R levels. Methylated PKM2 restrains mitochondrial addiction to Ca 2+ through InsP 3 Rs To test whether methylated PKM2 regulates mitochondrial functions through modulating InsP 3 Rs, which sustain mitochondrial functions, we stably knocked down InsP 3 R3 in MCF7 PKM2-KO and MDA-MB-231 cell lines, which highly express InsP 3 R3 ( Supplementary Fig. 7a, b ). InsP 3 R3 KD reduced the basal [Ca 2+ ] mito level ( Fig. 7a, b ) and Δ Ψ m ( Fig. 7c and Supplementary Fig. 7c ). Similarly, OCR was decreased by knocking down InsP 3 R3 in MCF7 PKM2-KO ( Fig. 7d ) and MDA-MB-231 cells ( Fig. 7e ). To delineate the roles of InsP 3 Rs in PKM2-modulated mitochondrial activity, we knocked down InsP 3 R3 in PKM2 WT - or PKM2 mut -expressing cells. InsP 3 R3 KD significantly reduced OCR in PKM2 WT or PKM2 mut cells. Basal OCR in PKM2 mut /shInsP 3 R3 cells was higher than in PKM2 WT /shInsP 3 R3 cells, possibly due to other InsP 3 Rs (InsP 3 R1 and 2) that remain abundant in PKM2 mut cells ( Fig. 7f, g ). Calcium transport between the ER and mitochondria is essential for Ca 2+ homeostasis and cell survival 29 . To assess the importance of calcium haemostasis and oxidative phosphorylation to cell survival in PKM2-KO or methylation-defective cells, we treated cells with xestospongin B (XeB), a specific InsP 3 R inhibitor, to inhibit IP3R-mediated ER Ca 2+ release. PKM2-KO or PKM2 mut cells were more vulnerable to XeB than parental and PKM2 WT cells ( Fig. 7h–j and Supplementary Fig. 7d ), indicating that addiction to oxidative phosphorylation following loss of PKM2 methylation plays essential roles in cell survival. Thus, methylated PKM2 repressed mitochondrial addiction to Ca 2+ via interacting with and suppressing the expression of InsP 3 Rs. Figure 7: Methylated PKM2 restrains mitochondrial addiction to Ca 2+ through InsP 3 Rs. ( a , b ) Relative basal [Ca 2+ ] mito levels in MCF7 PKM2-KO and MDA-MB-231 cells with control shRNA or InsP 3 R3 shRNA KD ( n = 3 independent experiments). ( c ) Δ Ψ measurement in MCF7 PKM2-KO and MDA-MB-231 cells with control shRNA or InsP 3 R3 shRNA KD. ( d – g ) Normalized basal OCR values in MCF7 PKM2-KO and MDA-MB-231 cells ( d , e ) and MCF7 (or MDA-MB-231) PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells ( f , g ) with control shRNA or InsP 3 R3 shRNA KD ( n = 6 independent experiments). ( h ) Representative images of parental MCF7, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells after 3.5 μM InsP 3 R inhibitor XeB treatment for 24 h. Scale bars, 50 μm. ( i , j ) Cell death measured by PI staining in parental MCF7, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells after treatment with 3.5 μM XeB for 24 h ( i ) or in parental MDA-MB-231, PKM2-KO, PKM2 WT /shPKM1 and PKM2 mut /shPKM1 cells after treatment with 5 μM XeB for 24 h ( n = 3 independent experiments). In a , b , d – g , i , j , data are shown as mean ± s.d. and statistics source data are available in Supplementary Table 7 . Statistical significance was assessed using two-tailed t -test ( a , b , d , e ) and ANOVA ( f , g , i , j ), ∗ P < 0.05, ∗ ∗ P < 0.01, ∗ ∗ ∗ P < 0.001. In f (lower panel) and g (lower panel), data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image Inhibiting PKM2 methylation with a nanoparticle-delivered competitive peptide blocks cancer cell metastasis in vivo Having established that PKM2 methylation controls ER-mitochondrial Ca 2+ signalling and promotes breast cancer cell proliferation and migration, we investigated whether PKM2 methylation can be therapeutically targeted. To assess PKM2 methylation dynamics, we estimated the extent of endogenous PKM2 methylation in cancer cells. We used excess methyl-PKM2 antibody for immunoprecipitation and measured the proportion of PKM2 in the supernatant and pellets (that is, in the methylated form) in MCF7 cells. Approximately 40% of endogenous PKM2 was methylated ( Supplementary Fig. 8a ). This partial PKM2 methylation in cancer cells implies that PKM2 methylation is dynamic and regulatable. Peptide drugs have made huge impacts on cancer treatment 35 . We evaluated whether a competitive, non-methylated PKM2 peptide encompassing the methylation sites could inhibit endogenous PKM2 methylation and reverse aerobic glycolysis to oxidative phosphorylation. As a negative control, we also synthesized a corresponding peptide with Arg445 and Arg447 asymmetrically dimethylated. In vitro , the non-methyl-peptide, but not the control methyl-peptide, abrogated CARM1-mediated methylation of PKM2 ( Fig. 8a ). In contrast, the peptides only partially inhibited methylation of histone H3, a control CARM1 substrate, suggesting that PKM2 is the primary target of inhibition by this peptide. Figure 8: Inhibiting PKM2 methylation using a competitive PKM2 peptide reduces proliferation, migration and lung metastasis of cancer cells due to increased oxidative phosphorylation. ( a ) In vitro methylation assays showing the inhibitory effects of the methyl- or non-methyl-PKM2 peptides on the methylation of PKM2 or a control histone H3 protein. ( b ) In vitro methylation assays showing the inhibitory effects of the methyl- or non-methyl-PKM2 peptides encapsulated in the UMNPs on the methylation of PKM2 or a control histone H3 protein. ( c ) Flow cytometric measurement of FAM-labelled peptide uptake delivered by UMNP in MDA-MB-231 cells. ( d ) Western blotting of endogenous PKM2 methylation and the InsP 3 R3 protein levels following cellular uptake of UMNP–methyl-peptide or UMNP–non-methyl-peptide. ( e , f ) Normalized basal OCR values in MCF7 ( e ) or MDA-MB-231 ( f ) cells treated with UMNP–methyl-peptide or UMNP–non-methyl-peptide ( n = 6 independent experiments). ( g ) MTT assays of MCF7 cells under the indicated treatment conditions ( n = 10 independent experiments). ( h ) The relative migratory cell numbers of MDA-MB-231 cells treated with UMNP–methyl-peptide or UMNP–non-methyl-peptide. Scale bars, 50 μm. ( n = 3 independent experiments.) ( i ) Bioluminescence in lungs of mice treated with UMNP–methyl-peptide or UMNP–non-methyl-PKM2 peptide ( n = 6 biologically independent mice per group). ( j ) Representative bioluminescence images of mice after 28 days of treatment. The colour scale depicts the photon flux (photons per second) emitted from the lung. In e – i , data are shown as mean ± s.d. and statistics source data are available in Supplementary Table 7 . Significance was assessed using two-tailed t -test ( e , f , h ) and ANOVA ( g , i ), ∗ ∗ P < 0.01, ∗ ∗ ∗ P < 0.001. In a , b , d , data represent one of three independent experiments with similar results. Unprocessed original scans of blots are shown in Supplementary Fig. 9 . Full size image We employed unimolecular nanoparticles (UMNPs) to deliver a peptide to inhibit methylation of PKM2 in vivo ( Supplementary Fig. 8b–f ). To ensure that UMNP encapsulation did not alter the PKM2 peptide’s inhibitory activity, we performed PKM2 in vitro methylation assay in the presence of non-methyl-peptide-loaded UMNP or methyl-peptide-loaded UMNP. The UMNP–non-methyl-peptide, but not UMNP–methyl-peptide, inhibited CARM1-mediated PKM2 methylation ( Fig. 8b ). Under this condition, histone H3 methylation was not inhibited. The results imply that UMNP–non-methyl-peptide selectively inhibited PKM2 methylation, an effect similar to expressing PKM2 mut and TEPP-46 treatment. MDA-MB-231 cells efficiently took up the 6-carboxyfluorescein-conjugated peptide (FAM-peptide)-loaded UMNPs in a dose-dependent manner ( Fig. 8c ). In addition, only UMNP–non-methyl-peptide inhibited endogenous PKM2 methylation, increasing InsP 3 R3 protein levels ( Fig. 8d ). Treatment with UMNP–non-methyl-peptide significantly increased OCR in MCF7 and MDA-MB-231 cells relative to the UMNP–methyl-peptide control ( Fig. 8e, f ). Just as CARM1 KO did not affect PK activity ( Supplementary Fig. 3c ), inhibiting PKM2 methylation by UMNP–non-methyl-peptide did not alter PKM2 PK activity ( Supplementary Fig. 8g ). Moreover, the non-methyl-peptide, but not methyl-peptide, inhibited MCF7 cell proliferation ( Fig. 8g ) and MDA-MB-231 cell migration ( Fig. 8h ). To test whether the competitive PKM2 peptide inhibits breast cancer lung metastasis in vivo , LM2 cells 36 , a metastatic MDA-MB-231 derivative clone, were tail-vein-injected into nude mice. While the majority of cells did not survive one day after injection, the remaining cells colonized in the lungs and reached 30–40% of the initial cell numbers by day 7. The mice were treated with UMNPs loaded with methyl-peptide or non-methyl-peptide on day 8 and treatment continued for three weeks. LM2 colonization and outgrowth in the lungs of the two cohorts were monitored by bioluminescence imaging 23 . Bioluminescence intensities in the UMNP-non-methyl-peptide-treated group were significantly decreased compared with those in the UMNP–methyl-peptide-treated group, indicating that non-methyl-peptide inhibited LM2 lung colonization ( Fig. 8i, j ). Thus, targeting PKM2 methylation is a feasible therapeutic strategy to reverse oncogenic processes. DISCUSSION We show here that reversible PKM2 methylation reprogrammes cancer metabolism from oxidative phosphorylation to aerobic glycolysis. PKM2 methylation by CARM1 inhibits Ca 2+ influx from ER to mitochondria. In breast cancer cells, mitochondrial oxidative phosphorylation dramatically increased following the loss of CARM1, PKM2 or PKM2 methylation, increasing basal mitochondrial [Ca 2+ ] and Δ Ψ m ( Supplementary Fig. 8h ). These findings provide mechanistic insights into the metabolic reprogramming controlled by the CARM1–PKM2 axis in breast cancer cells and show that inhibiting PKM2 methylation has therapeutic applications. PKM2 plays an important role in aerobic glycolysis by distributing glycolytic intermediates for anabolic and catabolic purposes in cancer cells 11 . Several post-translational modifications of PKM2 (refs 37 , 38 ) have been reported to modulate PKM2 function by inhibiting PK activity. However, a recent study challenged the PKM2-catalysed reaction as a rate-limiting step in cancer cell glycolysis 15 . Consistent with this, PKM2 PK activity was not affected by knocking out CARM1 ( Supplementary Fig. 3c ), or by inhibiting PKM2 methylation using non-methyl-PKM2 peptide ( Supplementary Fig. 8g ), suggesting that PKM2 methylation has little effect on its PK activity. Mitochondrial oxidative phosphorylation dramatically increased in CARM1-KO, PKM2-KO or PKM2 mut -expressing breast cancer cells, suggesting that non-glycolytic function of PKM2 regulates aerobic glycolysis rather than PK activity. Indeed, we found that PKM2 methylation elicits profound effects on energy production by altering mitochondrial oxidative phosphorylation. Notably, mitochondria have well-recognized roles in producing ATP and intermediates for macromolecule biosynthesis in normal and cancerous cells, and are promising chemotherapeutic targets 8 . In breast cancer cells, knock down of mitochondrial p32, a critical regulator of tumour metabolism via maintenance of oxidative phosphorylation, shifted metabolism from oxidative phosphorylation to glycolysis, yet tumorigenesis was impaired 7 , suggesting that high levels of glycolysis without adequate oxidative phosphorylation do not always benefit tumour growth. Thus, our results support the notion that the balance between aerobic glycolysis and mitochondrial respiration is essential for tumour progression. Cancer cells rely on mitochondria for TCA cycle intermediates to fuel lipid, nucleic acid and protein biosynthesis essential for growth 8 . The TCA cycle is regulated by mitochondrial Ca 2+ , which activates matrix dehydrogenases, including pyruvate-, α-ketoglutarate- and isocitrate-dehydrogenases 29 , to promote oxidative phosphorylation and ATP production 39 . Mitochondrial Ca 2+ is primarily taken from ER at mitochondria-associated ER membrane contacts. A minor fraction is from cytosol through low-affinity mitochondrial calcium uniporters. Both processes are tightly controlled by InsP 3 Rs, the ubiquitous family of ER Ca 2+ -release channels 40 . Interestingly, we found that InsP 3 R expression levels are inversely associated with PKM2 expression, and are sensitive to PKM2 methylation, that is, high InsP 3 Rs in PKM2 methylation-defective cells. Accordingly, mitochondrial Ca 2+ uptake increases in PKM2 methylation-defective cells, activating PDH and increasing oxidative phosphorylation. PKM2 methylation, on the contrary, decreases InsP 3 R expression and [Ca 2+ ] mito , increasing PDH phosphorylation and inactivation, decreasing Δ Ψ m , and switching energy homeostasis from mitochondrial respiration to aerobic glycolysis. We found that methylated PKM2 suppresses the expression of InsP 3 Rs via negatively regulating p53, the transcription factor regulating InsP 3 R expression. In addition to controlling InsP 3 R expression, methylated PKM2 co-precipitates with InsP 3 R1 and InsP 3 R3 ( Fig. 6f ). Thus, methylated PKM2, through regulating InsP 3 R expression and interaction, delicately controls Ca 2+ uptake by mitochondria. Mitochondrial Ca 2+ addiction was recently identified as a feature of cancer cells 41 . While inhibiting ER-to-mitochondria Ca 2+ transfer creates a bioenergetic crisis in normal and tumour cells, normal cells trigger autophagy to sustain survival, whereas the same autophagic response in tumour cells is insufficient for survival. Tumour cell survival uniquely depends on InsP 3 R-regulated, constitutive ER-to-mitochondria Ca 2+ transfer, since inhibiting InsP 3 R activity reduces cancer cell line proliferative potential in vitro and impairs tumour growth in vivo 41 . Accordingly, increased InsP 3 R expression and/or activity are associated with cancer cell proliferation, survival and invasiveness. All three InsP 3 R subtypes are expressed in breast cancer cells at various levels to regulate intracellular Ca 2+ release, which is essential for growth control of these cells 41 . Although inhibiting PKM2 methylation appears to reduce tumour cell growth, migration and metastasis in various breast cancer cell lines, it is insufficient to alter cell survival ( Supplementary Fig. 3j ), whereas inhibiting both PKM2 methylation and InsP 3 R activity is lethal to cancer cells ( Fig. 7h–j ). The results indicate a gain of dependence on mitochondrial Ca 2+ by cancer cells in order to maintain cell viability. The acquired mitochondrial addiction to Ca 2+ renders cancer cells susceptible to therapies based on inhibiting InsP 3 R (for example, XeB). Thus, combinatory inhibition of InsP 3 R activity and PKM2 methylation may elicit synergistic therapeutic effects. Targeting cancer-specific metabolism pathways (that is, aerobic glycolysis and ER-to-mitochondria Ca 2+ transfer) should provide new therapeutic avenues for cancer treatment, as exemplified by the UMNP peptide delivery system here.□ Methods Materials. The antibodies, reagents, shRNAs and primers are listed in Supplementary Table 6 . Cell culture and generation of PKM2-knockout cells. MCF7, MDA-MB-231 and HEK293T cell lines were purchased from ATCC, LM2 was provided by J. Massagué (Howard Hughes Medical Institute), and immortalized MEFs (PKM2 fl/fl , Cre-ER) were kindly provided by M. G. Vander Heiden (Koch Institute for Integrative Cancer Research, Massachusetts Institute of Technology) and were maintained in DMEM supplemented with 10% fetal bovine serum (FBS) purchased from Gibco. None of the cell lines used in this study was found in the database of commonly misidentified cell lines that are maintained by ICLAC and NCBI Biosample. Cell lines were authenticated by short tandem repeat profiling and were routinely tested for mycoplasma contamination. PKM2-KO cell lines were generated using a PKM2-specific CRISPR/Cas9/eGFP plasmid. For PKM2 KO in MEF cells, MEF cells were treated with 1 μM 4-hydroxytamoxifen (4-OHT) at least for 8–10 days to allow complete knockout of PKM2. Virus packaging and stable cell line generation. For lentivirus packaging, three plasmids (PAX2-, VSVG- and pLKO-shRNA-expressing plasmid) were transfected into HEK293T cells. For retrovirus packaging, the three plasmids PHIT60-, VSVG- and pLNCX-PKM2-expressing plasmid were employed. Supernatant containing the virus was harvested for cell line infection after 48 h. To generate stable cell lines expressing PKM2 wild type or methylation-defective mutant in accompanying with PKM1 KD, 2 × 10 5 PKM2-KO cells were seeded into six-well plates. For infection the next day, 1 ml of retrovirus was mixed with 1 ml fresh cell culture medium; Polybrene was added at a final concentration of 5 μg ml −1 to increase the infection efficiency. On the second day, the cells were cultured with fresh medium containing 400 μg ml −1 G418 (or 200 μg ml −1 G418 for MDA-MB-231) for at least four weeks to obtain cell lines stably expressing the PKM2 wild type or the methylation-defective mutant To knock down PKM1, the above cell lines were infected with pLKO-PKM1 shRNA lentivirus and selected with 2 μg ml −1 puromycin for at least 2 weeks. The expression of PKM2 in stable cell lines was individually examined by western blotting. Co-immunoprecipitation. Co-immunoprecipitation was performed as previously described 23 . In vitro methylation assay. In vitro methylation assay was performed as previously described 23 . Quantitative real-time PCR. Quantitative real-time PCR was performed as described previously 23 . Gel filtration. Recombinant His-tagged PKM2 protein was incubated with TEPP-46 (10 μM) for 1 h on ice, and then separated in the Superdex 200 Increase 10/300 GL column (GE Healthcare) in 0.01 M phosphate buffer and 0.14 M NaCl at pH 7.4. The flow speed rate is 0.5 ml min −1 . Fractions (0.3 ml) were collected and analysed by UV absorbance or SDS–PAGE and western blot. Pyruvate kinase and lactate dehydrogenase assays. Pyruvate activity was measured as described previously 42 . Briefly, 2 μg whole-cell lysate was incubated in 1 ml buffer (Tris pH 7.5 (50 mM), KCl (100 mM), MgCl 2 (5 mM), ADP (0.6 mM), phosphoenolpyruvate (0.5 mM), NADH (180 μM) and LDH (8 units)). The change in absorbance at 340 nm owing to the oxidation of NADH was measured using a Nanodrop ND-2000 1-position spectrophotometer (Thermo). LDH activity was determined by measuring the decreased fluorescence intensity at 340 nm from the NADH oxidation in buffer (Tris pH 7.5 (50 mM), KCl (100 mM), MgCl 2 (5 mM), pyruvate (20 mM), NADH (180 μM)). Subcellular fractionation. The isolation of cytosol and mitochondria was conducted as described previously 43 . Briefly, the cell pellets were resuspended in 11 ml ice-cold RSB hypo buffer (10 mM NaCl 1.5 mM MgCl 2 10 mM Tris-HCl (pH 7.5)) and the cells were allowed to swell for 5–10 min; then the swollen cells were broken open with several strokes in the presence of 8 ml of 2.5× MS homogenization buffer (525 mM mannitol, 175 mM sucrose, 12.5 mM Tris-HCl, 2.5 mM EDTA, pH 7.5) to give a final concentration of 1× MS homogenization buffer. The homogenate was centrifuged at 1,300 g for 5 min, repeated several times. After centrifugation at 7,000 g for 15 min, the supernatant (cytosolic fraction) and the crude mitochondria fraction were separated. The pellet was resuspended with 1× MS homogenization buffer followed by 7,000 g sedimentation several times. Confocal imaging. MAD-MB-231 cells were fixed with 4% paraformaldehyde in culture media for 15 min at 37 °C and permeabilized with 0.2% Triton X-100 for 10 min at room temperature. The nonspecific binding was blocked by incubation with 4% BSA in PBS for 60 min, and cells were subsequently stained with primary PKM2 and HSPA9 antibodies overnight at 4 °C. The slides were washed in PBS three times (5 min each time) and were incubated for 1 h with the following secondary antibodies: FITC-conjugated goat anti-mouse IgG and Dylight 594-conjugated goat anti-rabbit IgG. After being washed three times in PBS and air-dried, the coverslips were mounted in ProLong Gold anti-fade reagent with DAPI (Invitrogen). Fluorescence was examined using a Leica SP8 3X STED Super-resolution microscope equipped with a 63× objective lens with laser excitation at 405 nm, 488 nm or 592 nm. For z -stack analysis, optical sections were obtained along the z axis at 0.5-μm intervals. Images were analysed with ImageJ software. Expression and purification of recombinant proteins. Human PRMT1-8 cDNAs were cloned into pFN21K HaloTag CMV Flexi Vector (Promega) and the corresponding proteins were purified as described previously 23 . GST pulldown assays. GST and GST-CARM1 proteins were expressed in E. coli BL-21 competent cells and purified by glutathione Sepharose 4B resin (GE Healthcare Life Sciences). Recombinant, Flag-tagged PKM2 proteins were purified from HEK 293T cells. The GST pulldown experiment was conducted as described previously 23 . In vitro protein–protein interaction assay. Flag-tagged full-length PKM2 and its truncation constructs were translated by the T7 Quick Coupled Translation/Transcription system (Promega). Interaction with GST-CARM1 fusion protein (1 μg ml −1 ) was conducted as described previously 23 . Cell proliferation and DNA synthesis assays. The MTT assay was conducted as described previously 23 . For 5-ethynyl-2′-deoxyuridine (EdU) incorporation assay, 3 × 10 5 cells were seeded into a six-well plate and incubated with 10 μM EdU for 1 h, followed by procedures described in the Click-iT EdU cytometry assay kit on a BD LSRII. For the clonogenicity assay, 1,000 viable transfected cells were cultured in six-well plates for two weeks. Colonies were washed with PBS and fixed with 3.7% formaldehyde at room temperature for 15 min, and then stained for 20 min with 0.05% crystal violet. Transwell cell migration assays. The Transwell cell migration assay was performed as described previously 23 . Briefly, 1 × 10 5 cells in 100 μl serum-free media were added into the upper chamber; 500 μl culture media with 20% FBS was in the lower well. After 12-h culture, cells on the upper surface of the membrane were removed and migrant cells on the lower surface were fixed with 3.7% formaldehyde in culture media at 37 °C for 15 min, and then stained with 0.05% crystal violet for 20 min. Generation of methylated PKM2 (methyl-PKM2) specific antibody. Methyl-PKM2 specific anti-peptide antibody was generated by Genemed Synthesis. The KLH-conjugated PKM2 peptide RYR(as)PR(as)APIIAVTC, with Arg445 and Arg447 asymmetrically dimethylated, was synthesized. This peptide corresponding to human PKM2 (amino acids 443–454) was used to immunize rabbits. Purification of antibody was conducted as described previously 23 . Measurement of oxygen consumption rate. The OCR was measured in an XF96 extracellular analyser (Seahorse Bioscience). A total of 2 × 10 4 cells per well were seeded into a 96-well plate and incubated in DMEM media with 10% FBS at 37 °C. The next day, the medium was changed to analysis media containing 10 mM glucose. The cells were incubated in a CO 2 -free incubator at 37 °C for 1 h. Cells were sequentially exposed to oligomycin (1 μM), FCCP (1 μM) and rotenone (0.5 μM). Each point in the traces represents the average measurement from six different wells. Measurement of glucose uptake. Cells were washed and subsequently studied in a modified balanced salt solution (MBSS) containing (in mM): 140 NaCl, 5.4 KCl, 0.5 MgCl 2 , 0.4 MgSO 4 , 3.3 NaHCO 3 , 2.0 CaCl 2 , 10 HEPES, 5.5 glucose, pH 7.4; 2-NBDG (0.1–0.3 mM) was added to the bathing media, and after 20 min incubation and several washes, uptake of 2-NBDG was measured by fluorescence spectrometry. Mass spectrometry analysis of arginine methylation. In-gel digestions. Experiments were performed with a previously described protocol with modifications 1 . Briefly, the gel was destained twice with 100 mM (NH 4 )HCO 3 /50% methanol, followed by dehydration with 25 mM (NH 4 )HCO 3 /50% acetonitrile and 100% acetonitrile. The gel particles were rehydrated with freshly prepared 25 mM dithiothreitol at 56 °C. The proteins were then alkylated with 55 mM iodoacetamide in the dark. The gel pieces were incubated with 50 ng of trypsin in 25 mM (NH 4 )HCO 3 /0.01% ProteasMAX (Promega) at 42 °C. The tryptic peptides were extracted by adding 2.5% TFA solution. Samples were analysed on a Waters nanoAcquity UPLC system coupled to a Q-Exactive quadrupole Orbitrap mass spectrometer (Thermo). Tryptic peptides were loaded onto the Waters Symmetry C18 trap column (180 μm × 20 mm, 5 μm) and separated over a 70 min gradient on the fabricated column (150 mm × 75 μm, 1.7 μm). The data were acquired under the data-dependent acquisition mode (DDA, top 20). Mass spectrometric conditions were as follows: normalized high-energy collision dissociation (HCD) 30%, 70 K and 17.5 K resolution for full scan and MS/MS scan, automatic gain control (AGC) of 2 × 10 5 , maximum ion injection time of 100 ms, isolation window of 2 m / z , and fixed first mass of 100 m / z . Mass spectral data were searched in PEAKS Studio (Bioinformatics) against a home-made FASTA database containing the PKM2 protein sequence. Trypsin was selected and two missed cleavage sites were allowed. PEAKS searches were performed with a precursor mass tolerance of 20 ppm and a fragment mass tolerance of 0.05 Da. The fixed modifications consisted of carbamidomethylation (57.0215 Da) of cysteine residues. The variable modifications consisted of dimethylation (28.0313 Da) of arginine residues and oxidation (15.9949 Da) of methionine. Peptide spectral matches were validated on the basis of P value of 1% false discovery rate. Quantitative proteomics analysis. Cells were lysed in lysis buffer containing 8 M urea, 50 mM Tris-HCl, 30 mM NaCl, and a protease inhibitor tablet. The lysates were homogenized and processed by centrifugation. Protein content was determined by bicinchoninic acid (BCA) assay (Thermo). Protein extract was diluted to a urea concentration of 0.9 M and subjected to digestion with trypsin. The samples were quenched with 10% TFA and desalted with a Sep-Pak C18 cartridge (Waters). DiLeu labelling was performed according to previous studies 2 . Briefly, each DiLeu tag was dissolved in anhydrous DMF and activated with DMTMM and NMM. The supernatant was used immediately to label the peptides and the reaction was quenched with hydroxylamine. Samples were pooled at a 1:1 ratio across all of the channels and dried in vacuo . Strong cation exchange fractionation was performed on a Waters Alliance e2695 HPLC (Milford, MA) with a flow rate of 0.2 ml min −1 . Tryptic peptides were loaded onto a polySULFOETHYL A (200 mm × 2.1 mm, 5 μm) column. The samples were offline separated over a 90 min gradient, collected every 1.5 min and concatenated into 10 fractions. All samples were reconstituted in 0.1% FA, 3% ACN and loaded onto the fabricated column (150 mm × 75 μm, 1.7 μm). Peptides were online separated with a Dionex UltiMate 3000 LC system before entering the Orbitrap Fusion Lumos tribrid mass spectrometer (Thermo). Survey scans (300–1,500 m / z ) were acquired at a resolution of 60K, followed by top 15 HCD fragmentation (normalized collision energy 30, isolation window 1 m / z ). Values of 2 × 10 5 and 5 × 10 5 were selected as the AGC target for MS and MS/MS scans, respectively. The maximum injection time was set to 100 ms. The OMSSA Proteomic Analysis Software Suite (COMPASS) was used for peptide identification. Raw files were searched against the Homo sapiens Uniprot database (December 2015). Trypsin was selected as the enzyme and a maximum of two missed cleavages were allowed. Precursor and fragment ion tolerance were set to 25 ppm and 0.02 Da. DiLeu labelling on peptide N termini and lysine residues (+145.1267748 Da), and carbamidomethylation of cysteine residues (+57.02146 Da) were chosen as static modifications. Methionine oxidation (+15.99492 Da) and DiLeu labelling on tyrosine residues (+145.1267748) were selected as variable modifications. Search results were filtered to 1% false discovery rate at both the peptide and protein levels. Quantification was performed using an in-house software called the DiLeu tool. Reporter ion abundances are corrected for isotope impurities with python script. Cellular metabolite extraction and analysis. Global metabolomic analysis of nucleotides was performed by Metabolon as described previously 44 . Measurement of mitochondrial membrane potential (Δ Ψ ) and DNA content. Cells were loaded with JC-1 (10 μg ml −1 ) for 10 min or tetramethylrhodamine ethyl ester (TMRE) for 20 min in phenol-red-free media at 37 °C followed by three washes with PBS. Trypsinized cells were subjected to flow cytometry analysis. Relative quantification of mitochondrial DNA levels was performed using the ratio of mitochondrial tRNA Leu to the nuclear-encoded B2-microglobulin 45 . Measurement of calcium flux. Cells were loaded with Rhod-2 dye (MCF7 (1 μM), MDA-MB-231 and MEFs (0.45 μM)) and 0.002% Pluronic F-127 and incubated for 10 min at room temperature, and then the buffer was changed to Rhod-2-free buffer and the cells were incubated for a further 30 min. Calcium imaging was carried out by confocal microscopy at 561 nm excitation using a ×63 oil objective. Images were analysed and quantified using ImageJ (NIH). Cell apoptosis and death analyses. Cell apoptosis and death assays were conducted as described previously 41 . Cell apoptosis was determined by annexin V/APC and propidium iodide (PI) staining. Cell death was determined by PI staining using flow cytometry. Unimolecular nanoparticle (UMNP) synthesis. β-benzyl L -aspartate N -carboxyanhydride (BLA-NCA) monomer, poly (β-benzyl L -aspartate)-poly(ethylene glycol) (PBLA-PEG) block copolymers (that is, PBLA-mPEG and PBLA-PEG-Maleimide (Mal)), and poly(amidoamine)-poly(β-benzyl L -aspartate)-poly(ethylene glycol)-OCH 3 /Mal (PAMAM-PBLA-PEG-Mal) were all synthesized following the methods previously reported 46 , 47 . 1 H NMR (400 MHz, CDCl 3 ). BLA-NCA: 7.42–7.28 (5H, m, Ar- H ), 6.30 (1H, s, N H ), 5.20 (2H, s, C H 2 -Ar), 4.6 (1H, t, C H ), and 2.9 (2H, t, COCHC H 2 ) ppm. 1 H NMR (400 MHz, DMSO-D 6 ). PBLA-mPEG: 7.26–7.38 (102H, m, Ar- H ), 5.00–5.10 (40H, s, C H 2 -Ar), 4.55–4.68 (20H, m, COC H CH 2 ), 3.35–3.53 (450H, m, C H 2 C H 2 O from PEG), and 2.48–2.90 (41H, m, COCHC H 2 ) ppm. PBLA-PEG-Mal: 7.26–7.38 (102H, m, Ar- H ), 6.95 (2H, s, Mal), 5.00–5.10 (40H, s, C H 2 -Ar), 4.55–4.68 (20H, m, COC H CH 2 ), 3.35–3.53 (450H, m, C H 2 C H 2 O from PEG), and 2.48–2.90 (41H, m, COCHCH 2 ) ppm. PAMAM-PBLA-PEG-Mal: 7.30–7.10 (105H, m, Ar- H ), 6.90 (0.5H, s, Mal), 4.90–5.00 (40H, s, C H 2 -Ar), 4.60–4.50 (20H, m, COC H CH 2 ), 3.22–3.50 (450H, m, C H 2 C H 2 O from PEG), and 2.50–2.80 (41H, m, COCHC H 2 ) ppm. To synthesize poly(amidoamine)-poly(aspartate diethyltriamine)-poly(ethylene glycol)-OCH 3 /Mal (PAMAM-PAsp(DET)-PEG-Mal), PAMAM-PBLA-PEG-Mal (20 mg) was dissolved in 5 ml DMF. Diethyltriamine (224 μl) was added to the solution dropwise at 4 °C and the resulting mixture was stirred at room temperature for 4 h. Thereafter, it was added to 10 ml deionized water, neutralized by 1 M HCl, and dialysed against deionized water (MWCO = 15 kDa) for 48 h. 1 H NMR (400 MHz, D 2 O): 6.90 (0.5H, s, Mal), 4.80–4.50 (20H, s, COCHC H 2 ), 4.00–3.60 (450H, m, C H 2 C H 2 O from PEG), 3.30 (40H, s, CONHC H 2 ), 3.10–2.98 (81H, m, C H 2 NHC H 2 ), 2.75 (41H, C H 2 NH 2 ), and 2.40–2.25 (m, 42H, COCHC H 2 ) ppm. To synthesize poly(amidoamine)-poly(aspartate diethyltriamine- r -imidazole)-poly(ethylene glycol)-OCH 3 /Mal (PAMAM-PAsp(DET- r -Im)-PEG-Mal), PAMAM-PAsp(DET)-PEG-Mal (20 mg), 4-imidazolecarboxylic acid (1.13 mg), 1,3-dicyclohexylcarbodiimide (2.06 mg) and N -hydroxysuccinimide (1.15 mg) were dissolved in 5 ml DMF and stirred at room temperature for 24 h at pH 6.5 followed by dialysis against deionized water (MWCO 15 kDa) for 48 h. 1 H NMR (400 MHz, D 2 O): 8.24–8.22 (5H, s, Im), 7.53–7.52 (5H, d, Im), 6.90 (0.5 H, s, Mal), 4.80–4.50 (20H, s, COC H CH 2 ), 4.00–3.50 (450H, m, C H 2 C H 2 O from PEG), 3.30 (40H, s, CONHC H 2 ), 3.10–2.98 (80H, m, C H 2 NHC H 2 ), 2.75 (40H, C H 2 NH 2 ), and 2.70–2.50 (m, 41H, COCHC H 2 ) ppm. Poly(amidoamine)-poly(aspartate diethyltriamine-aconitic acid- r -imidazole)-poly(ethylene glycol)-OCH 3 /Mal (PAMAM-PAsp(DET-Aco- r -Im)-PEG-Mal) was synthesized using PAMAM-PAsp(DET- r -Im)-PEG/Mal following a previously reported method 48 . 1 H NMR (400 MHz, D 2 O): 8.24–8.22 (5H, s, Im), 7.53–7.52 (5H, s, Im), 6.90 (0.5H, s, Mal), 5.82–5.72 (15H, s, COC H C(COOH)CH 2 COOH), 4.80–4.50 (20H, s, COC H CH 2 ), 4.00–3.50 (450H, m, C H 2 C H 2 O from PEG), 3.30 (40H, s, CONHC H 2 ), 3.10–2.98 (82H, m, C H 2 NHC H 2 ), 2.75 (41H, C H 2 NH 2 ), 2.70–2.50 (m, 41H, COCHC H 2 ), 1.80 (30H, s, COCHC(COOH)C H 2 COOH) ppm. To synthesize PAMAM-PAsp(DET-Aco- r -Im)-PEG-TAT, PAMAM-PAsp(DET-Aco- r -Im)-PEG (20 mg), TAT (sequence CYGRKKRRQRRR, 0.3 mg) and tris(2-carboxyethyl)phosphine hydrochloride (0.57 mg) were dissolved in PBS buffer (pH 7.4, 5 ml). After 24 h stirring, the solution was purified by dialysis against deionized water (MWCO 15 kDa). 1 H NMR (400 MHz, D 2 O). 8.24–8.22 (5H, s, Im), 7.55–7.30 (15H, m, Im and TAT), 5.82–5.72 (15H, s, COC H C(COOH)CH 2 COOH), 4.80–4.50 (20H, s, COC H CH 2 ), 4.00–3.50 (450H, m, C H 2 C H 2 O from PEG), 3.30 (40H, s, CONHC H 2 ), 3.10–2.98 (82H, m, C H 2 NHC H 2 ), 2.75 (41H, CH 2 N H 2 ), 2.70–2.50 (m, 41H, COCHC H 2 ), 1.80 (30H, s, COCHC(COOH)C H 2 COOH) ppm. To prepare PKM2 peptide-loaded UMNPs, 6-carboxyfluorescein (FAM)-conjugated PKM2 peptide (FAM-PKM2 peptide, 0.5 mg) was dissolved in 0.5 ml deionized water under stirring, while the PAMAM-PAsp(DET-Aco- r -Im)-PEG-TAT polymer (UMNP, 2 mg) was dissolved in 1 ml deionized water at pH 7. The UMNP solution was slowly added to the peptide solution under stirring at room temperature for 4 h and dialysed against deionized water (MWCO = 100 kDa). The peptide loading level was determined by a UV–Vis spectrometer (Cary 5000 UV–Vis–NIR, Agilent Technologies) with the absorbance of FAM at 495 nm. UMNP–non-methyl-peptide and UMNP–methyl-peptide without FAM conjugation were prepared using the same method. Characterization. The 1 H NMR spectra were collected on a Bruker Advance 400 NMR spectrometer. The hydrodynamic size distribution and zeta-potential of the UMNPs were characterized using a dynamic light scattering (DLS) spectrometer (Malvern Zetasizer Nano ZS) at a polymer concentration of 0.1 mg ml −1 . Principle of PKM2 peptide delivery using UMNP. Positively charged PKM2 peptides were loaded onto the charge conversional polyanionic (PAsp(DET-Aco- r -Im)) segments through electrostatic interactions 49 . Under neutral pH (for example, the blood stream), the PAsp (DET-Aco- r -Im) segments carried negative charges, allowing for the complexation of the PKM2 peptides. Once the UMNPs were endocytosed by target cancer cells, the polyanionic PAsp(DET-Aco- r -Im) segments were converted to polycationic PAsp(DET- r -Im) segments in the acidic endocytic compartments due to the acid-induced cleavage of the aconitic acid side groups, thereby facilitating the release of PKM2 peptides. The imidazole functional groups were incorporated into the PAsp(DET-Aco- r -Im) segments to enhance the endosomal escape capability, thereby preventing potential damage of the peptides in the acidic endosomes/lysosomes 50 . A cell-penetrating peptide TAT was conjugated onto the surfaces of the nanocarriers to enhance their cellular uptake 51 . Animal experiments. All animal work was performed in accordance with protocols approved by the Research Animal Resource Center of UW-Madison and the study was compliant with ethical regulations regarding animal research. Balb/c nude female mice at 4–6 weeks old were used for all xenograft experiments (Harlan). For xenograft tumours assays, 1 × 10 6 cells (MDA-MB-231 PKM2 WT /shPKM1 or PKM2 mut /shPKM1 cells) were injected into the inguinal mammary fat pads of nude mice ( n = 6 per group). Tumour size was determined using calliper measurement and the tumour volume was calculated using the formula 1/2 × L × W 2 . For lung metastasis assays, 1 × 10 5 LM2, a metastatic MDA-MB-231 derivative clone, was resuspended in 0.1 ml PBS and tail-vein-injected into mice. Mice were imaged for luciferase activity immediately after injection (day 0) to exclude any mice that were not successfully xenografted. Luciferase-based non-invasive bioluminescent imaging and analysis were performed as described previously 36 using an IVIS Imaging System (Caliper Life Sciences). Briefly, mice were anaesthetized and injected intraperitoneally with 2 mg D -luciferin (20 mg ml −1 in PBS) (Gold Biotechnology). Imaging was completed between 5 to 15 min after injection. For bioluminescence plots, total photon flux was calculated for each mouse in the gated areas. Then, the mice were retro-orbitally injected with UMNP–methyl-peptide or UMNP–non-methyl-peptide (100 μl, 1 g l −1 ) at the indicated time interval. Imaging was performed every week and endpoint assays were conducted four weeks after injection. Statistics and reproducibility. Statistical testing was performed using the unpaired two-tailed Student’s t -test and/or ANOVA analysis. All experiments were repeated at least three times unless otherwise indicated. N numbers are indicated in the figure legends. P value <0.05 ( ∗ ) was considered as statistically significant. Data availability. Mass spectrometry data have been deposited in ProteomeXchange with the primary accession code PXD007671 ( ) 52 . Global metabolomic nucleotide data have been deposited in MetaboLights with the primary accession code MTBLS533 ( ) 53 . The RNA-seq data of 1,093 human primary solid breast tumour samples were derived from the TCGA Research Network: and they are available in FireBrowse ( ). The proteomics data of 77 human breast tumour samples were derived from the CPTAC Research Network ( ) and they are available in ( ). Data related to CARM1- or PKM2-interacting proteins are provided in Supplementary Tables 1, 4 and 5 . Source data for Figs 3a, d, g, i , 4a–c, e–g , 5b, c, f–h, p, q , 7a, b, d–g, i, j and 8e–i and Supplementary Figs 1a–f , 2b–d, f, g , 4a–l , 5a, b , 6i and 8g have been provided in Supplementary Table 7 . All other data supporting the findings of this study are available from the corresponding author on reasonable request. Additional Information Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Change history 13 November 2017 In the version of this Article originally published, an amino acid (aa) range in Fig. 2a incorrectly read 390–53 aa. The correct range is 390–531 aa. In addition, two labels from Fig. 8j were displaced during production and instead appeared over Fig. 8g. These errors have now been corrected in the online version of the Article.

As cells with a propensity for cancer break down food for energy, they reach a fork in the road: They can either continue energy production as healthy cells, or shift to the energy production profile of cancer cells. In a new study published Monday in the journal Nature Cell Biology, University of Wisconsin-Madison researchers map out the molecular events that direct cells' energy metabolism down the cancerous path. The findings could lead to ways to interrupt the process. "Cancer cells often change their nutrient utilization and energy production, so many efforts are being made to develop drug inhibitors of cancer cell metabolism to starve them," says senior author Wei Xu, the Marian A. Messerschmidt Professor in Cancer Research at the UW Carbone Cancer Center and McArdle Laboratory for Cancer Research. "We have found that inhibiting a chemical modification of a cancer-associated metabolism protein is enough to inhibit the aggressive nature of cancer cells." Cancer biologists have identified nearly a dozen "hallmarks of cancer," or large-scale changes that send a precancerous cell over the tipping point to become a cancerous one. One hallmark of cancer is the loss of properly regulated energy metabolism, a process referred to as the "Warburg effect" after the Nobel laureate, Otto Warburg, who identified it. Other hallmarks of cancer include continuous activation of growth pathways, the inability to respond to signals that put the brakes on cell growth, and a gain of invasion and spread to distant organs. "My lab studies a protein, CARM1, which is associated with worse outcomes in breast cancer patients, though it has also been found expressed in many other cancer types," Xu says. "CARM1 chemically modifies its target proteins to alter their function, and in doing so directly leads to the activation of several hallmarks of cancer." In the study, Xu and her colleagues found that CARM1 protein modifies a cell metabolism protein, PKM2, and changes its function. This drives the Warburg effect, activating a hallmark of cancer. Nearly a decade ago, researchers found that PKM2 was expressed at high levels in cancer cells, but how these levels translated to more aggressive cancers was not known. So, Xu and colleagues performed a protein interaction assay in a breast cancer cell line and found that CARM1 interacts with and chemically modifies PKM2. They also assessed whether CARM1-directed modifications of PKM2 might be responsible for leading cells down a cancerous pathway. By engineering cells to express "normal" PKM2 or a mutated form that was not modifiable, the researchers learned that PKM2 appears to be the deciding factor in picking the direction cell metabolism takes at that fork in the road. The CARM1-modified PKM2 shifted cells toward the cancer cell metabolism path while cells with PKM2 that could not be modified took the metabolic path associated with noncancerous cells. With a clearer picture of how cancer cells shift their metabolism, the researchers next used a mouse model of breast cancer and a competitor drug that prevents CARM1 from effectively modifying PKM2 to test what would happen. "When we block PKM2 modification by CARM1, the metabolic energy balance in cancer cells is reversed, and we see a decrease of cell growth and cell spreading potential," Xu says. "This study, then, identifies another therapeutic target to help reverse several hallmarks of cancer." In addition to targeting PKM2 modification by CARM1, Xu's lab is investigating how CARM1 recognizes all of its many target proteins, with the goal of disrupting those protein modifications from driving aggressive cancers.

10.1038/ncb3630

Medicine

Cancer cells send out 'drones' to battle immune system from afar

Exosomal PD-L1 contributes to immunosuppression and is associated with anti-PD-1 response, Nature (2018). DOI: 10.1038/s41586-018-0392-8 , www.nature.com/articles/s41586-018-0392-8 Journal information: Nature

http://dx.doi.org/10.1038/s41586-018-0392-8

https://medicalxpress.com/news/2018-08-cancer-cells-drones-immune-afar.html

Abstract Tumour cells evade immune surveillance by upregulating the surface expression of programmed death-ligand 1 (PD-L1), which interacts with programmed death-1 (PD-1) receptor on T cells to elicit the immune checkpoint response 1 , 2 . Anti-PD-1 antibodies have shown remarkable promise in treating tumours, including metastatic melanoma 2 , 3 , 4 . However, the patient response rate is low 4 , 5 . A better understanding of PD-L1-mediated immune evasion is needed to predict patient response and improve treatment efficacy. Here we report that metastatic melanomas release extracellular vesicles, mostly in the form of exosomes, that carry PD-L1 on their surface. Stimulation with interferon-γ (IFN-γ) increases the amount of PD-L1 on these vesicles, which suppresses the function of CD8 T cells and facilitates tumour growth. In patients with metastatic melanoma, the level of circulating exosomal PD-L1 positively correlates with that of IFN-γ, and varies during the course of anti-PD-1 therapy. The magnitudes of the increase in circulating exosomal PD-L1 during early stages of treatment, as an indicator of the adaptive response of the tumour cells to T cell reinvigoration, stratifies clinical responders from non-responders. Our study unveils a mechanism by which tumour cells systemically suppress the immune system, and provides a rationale for the application of exosomal PD-L1 as a predictor for anti-PD-1 therapy. Main Extracellular vesicles, such as exosomes and microvesicles (also known as shedding vesicles), carry bioactive molecules that influence the extracellular environment and the immune system 6 , 7 , 8 . We purified exosomes from a panel of human primary and metastatic melanoma cell lines by differential centrifugation 9 , 10 , 11 , 12 , and verified them by transmission electron microscopy (TEM) and nanoparticle tracking analysis (NTA) (Fig. 1a, b ). Proteins associated with the exosomes were then analysed by reverse phase protein array (RPPA), a large-scale antibody-based quantitative proteomics technology 13 . Analysis by RPPA and western blot revealed the presence of PD-L1 in exosomes, and its level was significantly higher in exosomes derived from metastatic melanoma cells compared to those from primary melanoma cells (Fig. 1c, d , Extended Data Fig. 1a ) . Iodixanol density gradient centrifugation further confirmed the association of PD-L1 with the exosomes (Extended Data Fig. 1b ). PD-L1 was also detected in microvesicles, but at a lower level (Extended Data Fig. 1c–e ). PD-L1 was also detected in extracellular vesicles generated from mouse metastatic melanoma B16-F10 cells (Extended Data Fig. 1f ). Fig. 1: Extrafacial expression of PD-L1 on melanoma cell-derived exosomes and its regulation by IFN-γ. a , A representative TEM image of purified exosomes from WM9 cells. Scale bar, 50 nm. b , Characterization of purified exosomes using nanoparticle tracking. c , RPPA data showing the levels of PD-L1 in exosomes secreted by primary or metastatic melanoma cell lines ( n = 3 for WM1552C, WM902B, A375, WM164; n = 4 for WM35, WM793, UACC-903, WM9). See Extended Data Fig. 1a for statistical analysis. d , Immunoblots for PD-L1 in the whole cell lysate (W) and purified exosomes (E) from different metastatic melanoma cell lines. All lanes were loaded with the same amount of total protein. e , A representative TEM image of WM9 cell-derived exosomes immunogold-labelled with anti-PD-L1 antibodies. Arrowheads indicate 5-nm gold particles. Scale bar, 50 nm. f , Schematic (left) of ELISA to measure PD-L1 concentration (right) on the surface of exosomes isolated from indicated cell types. TMB, 3, 3′, 5, 5′-tetramethylbenzidine; SA-HRP, streptavidin–horseradish peroxidase. g , ELISA of PD-L1 on exosomes from melanoma cells, with or without IFN-γ treatment. h , PD-l binding of exosomes with IFN-γ or blocking PD-L1 antibody (PD-L1 Ab) (see Methods ). i , Western blot analysis of PD-L1 in whole cells and exosomes from IFN-γ-treated cells and control cells. All lanes were loaded with the same amount of total protein (left). Quantification of exosomal PD-L1 by western blotting (right). Results shown represent three ( a , b ) or two ( d , e ) independent experiments. Data are mean ± s.d. of three ( f , h , i ) or four ( g ) independent biological replicates. P values are from a two-sided unpaired t -test ( g , i ). Full gel source data ( d , i ) are shown in Supplementary Fig. 1 . Source Data Full size image Tumour cell surface PD-L1 can be upregulated in response to IFN-γ secreted by activated T cells, and PD-L1 binds to PD-1 through its extracellular domain to inactivate T cells 2 , 3 , 14 . Using immuno-electron microscopy and enzyme-linked immunosorbent assay (ELISA) (Fig. 1e–g ), we found that exosomal PD-L1 has the same membrane topology as cell surface PD-L1, with its extracellular domain exposed on the surface of the exosomes. Exosomal PD-L1 binds PD-1 in a concentration-dependent manner, and this interaction can be disrupted by PD-L1-blocking antibodies (Fig. 1h ). Furthermore, the level of exosomal PD-L1 secreted by melanoma cells increased markedly upon IFN-γ treatment (Fig. 1f, g, i ), and correspondingly, these exosomes displayed increased binding to PD-1 (Fig. 1h ). Exosomes are generated and released through a defined intracellular trafficking route 7 , 9 , 10 . Genetic knockdown of the ESCRT subunit Hrs, which mediates the recognition and sorting of exosomal cargos 15 , using short hairpin (sh)RNA led to a decrease in the level of PD-L1 in the exosomes and an increase of PD-L1 in the cell (Extended Data Fig. 1g, h ). In addition, PD-L1 co-immunoprecipitated with Hrs from cell lysates (Extended Data Fig. 1i ). PD-L1 co-localized with Hrs and CD63, an exosome marker, in melanoma cells (Extended Data Fig. 1j, k ). Knockdown of Rab27A, which mediates exosome release 16 , also blocked PD-L1 secretion via the exosomes (Extended Data Fig. 1l ). To investigate the secretion of exosomal PD-L1 by melanoma cells in vivo, we established human melanoma xenografts in nude mice. Blood from these mice was collected for exosome purification and subsequent detection of human PD-L1 proteins by ELISA (Fig. 2a ). Antibodies against human PD-L1 specifically identified human PD-L1 on the circulating exosomes from mice bearing human melanoma xenografts but not the control mice (Fig. 2b , Extended Data Fig. 2a, b ). Moreover, the level of circulating exosomal PD-L1 positively correlated with tumour size (Fig. 2c ). Fig. 2: The level of PD-L1 on circulating exosomes distinguishes patients with metastatic melanoma from healthy donors. a , ELISA of human PD-L1 on exosomes in plasma samples from mice with human melanoma xenograft. b , Levels of PD-L1 on exosomes isolated from the plasma samples of control nude mice or mice bearing human WM9 melanoma xenograft, measured by ELISA ( n = 10). c , Pearson correlation between the exosomal PD-L1 in plasma and tumour burden in xenograft-bearing nude mice ( n = 10). d – f , ELISA of circulating exosomal PD-L1 ( d ), total PD-L1 ( e ) or extracellular vesicle (EV)-excluded PD-L1 ( f ) in healthy donors (HD, n = 11) and melanoma patients (MP, n = 44). The exosomes were purified using the exosome isolation kit. g , ROC curve analysis for the indicated parameters in patients with metastatic melanoma compared to healthy donors. Data are mean ± s.d. P values are from a two-sided unpaired t -test ( b , d – f ). Source Data Full size image PD-L1 has been found in blood samples derived from melanoma patients 17 . Recent studies suggest the presence of PD-L1 in extracellular vesicles isolated from blood samples of patients with cancer, and the level of PD-L1 correlates with pathological features of these patients 18 , 19 , 20 . We purified extracellular vesicles from the plasma of melanoma patients (Extended Data Fig. 2c–g ). The level of PD-L1 on the circulating exosomes was significantly higher in patients with metastatic melanoma than in healthy donors (Fig. 2d , Extended Data Figs. 2f , 3a , 3b ), whereas there was no or only marginal difference in the number of circulating exosomes or the total protein level on these exosomes (Extended Data Fig. 3c, d ). There was less difference in PD-L1 levels in circulating microvesicles compared to the circulating exosomes (Extended Data Fig. 3e ). The data analysis and receiver operating characteristic (ROC) curve show that, among all the parameters tested, the level of circulating exosomal PD-L1 best distinguished melanoma patients from healthy donors (Fig. 2d–g , Extended Data Fig. 3e, f ). The current model for PD-L1-mediated immunosuppression is based on the interaction between PD-L1 on the tumour cell surface and PD-1 on CD8 T cells. Here we tested whether exosomal PD-L1 inhibits CD8 T cells. First, we used confocal microscopy to show a physical interaction between tumour exosomes and CD8 T cells purified from human peripheral blood (Extended Data Fig. 4a, b ). Flow cytometry analyses further indicated that the level of interaction was higher for activated CD8 T cells than for non-activated counterparts (Extended Data Fig. 4c ). Moreover, exosomes derived from melanoma cells treated with IFN-γ exhibited a higher level of binding to CD8 T cells (Extended Data Fig. 4d ). Next, we tested the effect of exosomal PD-L1 on CD8 T cells, taking advantage of MEL624 cells, which do not express endogenous PD-L1 (Extended Data Fig. 5a–d ) and other immunosuppressive proteins such as FasL and TRAIL 1 . Exosomes derived from MEL624 cells expressing exogenous PD-L1 inhibited the proliferation, cytokine production and cytotoxicity of CD8 T cells, as demonstrated by the decreased proportion of cells containing diluted carboxyfluorescein succinimidyl ester (CFSE, a cell division-tracking dye), reduced expression of Ki-67 and granzyme B (GzmB), and the inhibited production of IFN-γ, IL-2, and TNF (Fig. 3a , Extended Data Fig. 5e, f ). Pre-treatment of the exosomes with anti-PD-L1 antibodies nearly abolished these effects. Similar effects were observed using exosomes secreted from WM9 cells, which express endogenous PD-L1 (Extended Data Fig. 5e–h ). Exosomes derived from mouse melanoma B16-F10 cells also inhibited the proliferation and cytotoxicity of mouse splenic CD8 T cells (Extended Data Fig. 6a–d ). Pre-treating OT-I T cells (which specifically recognize OVA peptide) (Extended Data Fig. 6e ) with B16-F10 cell-derived exosomes inhibited their ability to kill their target cells (Extended Data Fig. 6f ). Extracellular vesicles from human lung and breast cancer cells also contain immunosuppressive PD-L1, mostly of which is in exosomes, and PD-L1 expression is also upregulated by IFN-γ in some of these cell lines (Extended Data Fig. 7a–e ). Fig. 3: Exosomal PD-L1 inhibits CD8 T cells and facilitates the progression of melanoma in vitro and in vivo. a , Representative histogram of CFSE-labelled human peripheral CD8 T cells (top left) and representative contour plots of human peripheral CD8 T cells examined for the expression of Ki-67 (middle left) and granzyme B (GzmB) (bottom left) after indicated treatments. The proportions of cells with diluted CFSE dye, or positive Ki-67 or GzmB expression are shown on the right ( n = 3 independent biological experiments). b , Growth curve of PD-L1(KD) B16-F10 tumours with indicated treatments ( n = 7 mice per group). c , The proportions of Ki-67 + PD-1 + CD8 TILs or splenic or lymph node CD8 T cells after indicated treatments ( n = 6 for tumour samples of the EXO-IgG group, and n = 7 for all the other groups). See Extended Data Fig. 8d for representative contour plots. Data are mean ± s.d. ( a – c ). P values are from a two-sided unpaired t -test ( a , c ) or two-way ANOVA ( b ). Source Data Full size image To examine the effects of exosomal PD-L1 in vivo, we established a syngeneic mouse melanoma model in C57BL/6 mice using B16-F10 cells in which PD-L1 expression had been knocked down (PD-L1(KD) B16-F10) (Extended Data Fig. 8a ). Injection of exosomes derived from parental B16-F10 cells promoted the growth of tumours derived from PD-L1(KD) B16-F10 cells, whereas pre-treatment of the exosomes with anti-PD-L1 antibodies, but not with IgG isotype or CD63-blocking antibodies, inhibited the effect (Fig. 3b , Extended Data Fig. 8b, c ). The number of tumour-infiltrating CD8 T lymphocytes (TILs) decreased significantly after the injection of exosomes (Fig. 3c , Extended Data Fig. 8d, e ). B16-F10 exosomes also decreased the proportion of proliferating PD-1 + CD8 T cells in both spleen and lymph nodes (Fig. 3c , Extended Data Fig. 8d ), suggesting that exosomal PD-L1 suppresses anti-tumour immunity systemically. We then examined the level of PD-L1 on circulating extracellular vesicles in melanoma patients during anti-PD-1 therapy. The pre-treatment level of circulating exosomal PD-L1 was significantly higher in patients who failed to respond to the anti-PD-1 treatment with pembrolizumab (Fig. 4a ). The difference was, however, not significant for total circulating PD-L1, and undetectable for PD-L1 on circulating microvesicles, or extracellular vesicle-excluded PD-L1 (Fig. 4b–d ). A higher level of circulating exosomal PD-L1 before the treatment was associated with poorer clinical outcomes (Fig. 4e ). IFN-γ upregulates exosomal PD-L1 and the pre-treatment levels of IFN-γ were significantly higher in patients who did not respond to pembrolizumab 21 . The level of circulating exosomal PD-L1 positively correlated with the level of circulating IFN-γ and overall tumour burden (Fig. 4f, g ), which were shown to be indicative of poor prognosis 21 . Fig. 4: The level of circulating exosomal PD-L1 stratifies clinical responders to pembrolizumab and non-responders. a – d , Comparison of the pre-treatment levels of circulating exosomal PD-L1 ( a ), total PD-L1 ( b ), microvesicle PD-L1 ( c ), or extracellular vesicle-excluded PD-L1 ( d ) between melanoma patients with or without clinical response to pembrolizumab. R, responders; n = 21; NR, non-responders; n = 23. e , Objective response rate (ORR) for patients with high and low pre-treatment levels of circulating exosomal PD-L1. f , g , Pearson correlation of the IFN-γ level ( f , n = 27) or overall tumour burden ( g , n = 39) to the exosomal PD-L1 level in the plasma of patients with melanoma. h , Circulating exosomal PD-L1 at serial time points pre-treatment and on-treatment ( n = 39). i , Circulating exosomal PD-L1 in clinical responders ( n = 19) and non-responders ( n = 20) at serial time points pre- and on-treatment. j , Comparison of the maximum fold change of circulating exosomal PD-L1 at week 3–6 between clinical responders and non-responders. k , ROC curve analysis for the maximum fold change of circulating exosomal PD-L1 at week 3–6 in clinical responders compared to non-responders. AUC, area under curve. l – o , ORR for patients with high and low fold changes of circulating exosomal PD-L1 ( l ), total PD-L1 ( m ), microvesicle PD-L1 ( n ), or extracellular vesicle-excluded PD-L1 ( o ), at weeks 3–6 of treatment. Data are mean ± s.d. * P < 0.05, two-sided unpaired t -test ( a – d , j ), two-sided paired t - test ( h , i) , or two-sided Fisher’s exact test ( e , l – o ). Source Data Full size image Next, we examined the level of circulating exosomal PD-L1 in patients undergoing pembrolizumab therapy. In clinical responders, there were increased levels of PD-L1 on circulating exosomes, mostly within 6 weeks of therapy (Fig. 4h, i ). The level of PD-L1 on microvesicles also increased in the same cohort of patients, but to a lesser extent in comparison to exosomes (Extended Data Fig. 9a ). Proliferation and reinvigoration of CD8 T cells peaked at week 3 of treatment and preceded the peaking of exosomal PD-L1 at week 6 (Extended Data Fig. 9b ). Moreover, in pembrolizumab-responsive patients, both the absolute value and maximal fold change of Ki-67 in PD-1 + CD8 T cells after 3–6 weeks of treatment positively correlated with those of circulating exosomal PD-L1 (Extended Data Fig. 9c, d ). The responders displayed a larger increase in the level of circulating exosomal PD-L1 as early as 3–6 weeks following the initial treatment (Fig. 4j ). ROC analysis determined that a fold change of 2.43 in exosomal PD-L1 at week 3–6 stratified patients by clinical response to pembrolizumab (Fig. 4k ); a fold change in circulating exosomal PD-L1 greater than 2.43 at week 3–6 was associated with a better response to anti-PD-1 therapy by objective response rate (ORR), progression-free and overall survival (Fig. 4l , Extended Data Fig. 9e ). The fold increase of total circulating PD-L1, microvesicle PD-L1, and extracellular vesicle-excluded PD-L1 was inferior to that of exosomal PD-L1 for distinguishing responders from non-responders (Fig. 4k, m–o , Extended Data Fig. 9f–h ). Our studies suggest that melanoma cells release PD-L1-positive extracellular vesicles into the tumour microenvironment and circulation to counter the anti-tumour immunity systemically. Since exosomal PD-L1-mediated T cell inhibition can be blocked by antibodies against either PD-L1 or PD-1, our results raise the possibility that disrupting the interaction between exosomal PD-L1 and PD-1 on T cells is a previously unrecognized mechanism in PD-L1/PD-1 blockade-based therapies. The level of PD-L1 on extracellular vesicles is upregulated by IFN-γ, and PD-L1 on extracellular vesicles primarily targets PD-1 + CD8 T cells, which represent the antigen-experienced T cells that secrete IFN-γ. Exosomal PD-L1 may therefore reflect the dynamic interplay between tumour and immune cells. Besides PD-L1, other extracellular vesicle proteins such as FasL may also contribute to immunosuppressive effects 19 , 22 , 23 , 24 . However, PD-L1 enables exosomes to target predominantly PD-1 + CD8 T cells, allowing tumour cells to counteract the immune pressure at the effector stage. In addition to the interaction between exosomal-PD-L1 and PD-1, the involvement of other molecules including B7 and CD28 25 , 26 in this process also warrant investigation. Our study suggests that circulating exosomal PD-L1 prior to and during pembrolizumab treatment may reflect distinct states of anti-tumour immunity. The pre-treatment PD-L1 level may correlate with a role of exosomal PD-L1 in immune dysfunction. High levels of exosomal PD-L1 may reflect the ‘exhaustion’ of T cells to a stage at which they can no longer be reinvigorated by anti-PD-1 treatment. In on-treatment patients, however, an increase in the level of exosomal PD-L1, following and correlating positively with T cell reinvigoration, would reflect the presence of a successful anti-tumour immunity elicited by the anti-PD-1 therapy. Although the increase in exosomal PD-L1 in response to IFN-γ could enable tumour cells to adaptively inactivate CD8 T cells, this is futile because the interaction between PD-L1 and PD-1 is blocked by pembrolizumab. We observed no marked increase in exosomal PD-L1 in non-responders. This could be a result of failure to elicit an adequate T cell response, or a resistance mechanism to IFN-γ from tumours. Tumour cells in non-responders may have adaptively downregulated their response to IFN-γ to avoid the detrimental increase in antigen presentation and to escape the anti-proliferative effects induced by IFN-γ 5 , 27 . Our study offers a rationale for developing circulating exosomal PD-L1 as a predictor for the clinical outcomes of anti-PD-1 therapy, and sheds light on possible causes for the failure of anti-PD-1 therapies experienced by many patients (Extended Data Fig. 10 ). Tumour PD-L1 has been used as a predictive biomarker for clinical responses to anti-PD-1 therapy 28 , 29 , 30 . Considering the heterogeneity and dynamic changes of PD-L1 expression in tumours, and the invasive nature of tumour biopsy, developing exosomal PD-L1 as a blood-based biomarker could be an attractive option. Methods Cell culture The A375 human melanoma and B16-F10 mouse melanoma cells were purchased from ATCC. The control and PD-L1-overexpressing human melanoma MEL624 cells were provided by H. Dong (Mayo Clinic). Mouse melanoma B16 cells stably expressing chicken OVA (B16-OVA) were provided by H. C. J. Ertl (The Wistar Institute). The UACC-903 human melanoma cells were provided by M. Powell (Stanford University). The melanoma cell lines WM1552C, WM35, WM793, WM902B, WM9 and WM164 presented in this study were established in M. Herlyn’s laboratory (The Wistar Institute). All cell lines were authenticated by DNA fingerprinting, and were tested routinely before use to avoid mycoplasma contamination. Human melanoma cell lines MEL624, PD-L1/MEL624, WM1552C, WM35, WM902B, WM793, UACC-903, WM9, A375 and WM164 were cultured in RPMI 1640 medium (Invitrogen) supplemented with 10% (v/v) fetal bovine serum (FBS) (Invitrogen). B16-F10 and B16-OVA cells were cultured in DMEM (Sigma) supplemented with 10% (v/v) FBS. For stimulation with IFN-γ, cells were incubated with 100 ng/ml of recombinant human or mouse IFN-γ (Peprotech) for 48 h. Generation of stable Hrs, Rab27a or PD-L1 knockdown melanoma cells Short hairpin RNAs (shRNAs) against human Hrs (also known as HGS ) (NM_004712, GCACGTCTTTCCAGAATTCAA, GCATGAAGAGTAACCACAGC), human RAB27A (NM_004850, GCTGCCAATGGGACAAACATA, CAGGAGAGGTTTCGTAGCTA) (gift from A. Weaver, Vanderbilt University), mouse PD-L1 (also known as Cd274 ) (NM_021893, GCGTTGAAGATACAAGCTCAA) or scrambled shRNA-control (Addgene) were packaged into lentiviral particles using 293T cells co-transfected with the viral packaging plasmids. Lentiviral supernatants were harvested 48–72 h after transfection. Cells were infected with filtered lentivirus and selected by 2 μg/ml puromycin. Patients and specimen collection Patients with stage III to IV melanoma (Supplementary Table 1 ) were enrolled for treatment with pembrolizumab (2 mg/kg by infusion every 3 weeks) under an Expanded Access Program at Penn ( identifier NCT02083484) or with commercial Keytruda. Patients gave consent in writing for blood collection under the University of Pennsylvania Abramson Cancer Center’s melanoma research program tissue collection protocol UPCC 08607 in accordance with the ethics committee and The Institutional Review Board of the University of Pennsylvania. Peripheral blood was obtained in sodium heparin tubes before each pembrolizumab infusion every 3 weeks for 12 weeks. Clinical response was determined as best response based on immune-related RECIST (irRECIST) using unidimensional measurements 31 . The assessment of clinical responses for patients was performed independently in a double-blind fashion. Blood samples from healthy donors were collected at The Wistar Institute after approval by the ethics committee and Institutional Review Board of The Wistar Institute. Written consent was obtained from each healthy donor before blood collection. All experiments involving blood samples from healthy donors were performed in accordance with relevant ethical regulations. Flow cytometry of patients’ PBMCs Peripheral blood mononuclear cells (PBMCs) were isolated using Ficoll gradient and stored using standard protocols. Cryopreserved PBMC samples from pretreatment, cycles 1–4 (weeks 3–12) were thawed and analysed by flow cytometry as previously described 21 . In brief, live or dead cell discrimination was performed using Live/Dead Fixable Aqua Dead Cell Stain Kit (Life Technologies). Cell surface staining was performed for 30 min at 4 °C. Intracellular staining was performed for 60 min on ice after using a fixation/permeabilization kit (eBioscience). Purification of extracellular vesicles For exosome purification from cell culture supernatants, cells were cultured in media supplemented with 10% exosome-depleted FBS. Bovine exosomes were depleted by overnight centrifugation at 100,000 g . Supernatants were collected from 48–72 h cell cultures and extracellular vesicles were purified by a standard differential centrifugation protocol 9 , 10 , 11 , 12 . In brief, culture supernatants were centrifuged at 2,000 g for 20 min to remove cell debris and dead cells (Beckman Coulter, Allegra X-14R). Microvesicles were pelleted after centrifugation at 16,500 g for 45 min (Beckman Coulter, J2-HS) and resuspended in PBS. Supernatants were then centrifuged at 100,000 g for 2 h at 4 °C (Beckman Coulter, Optima XPN-100). The pelleted exosomes were suspended in PBS and collected by ultracentrifugation at 100,000 g for 2 h. For purification of circulating extracellular vesicles by differential centrifugation, venous citrated blood from melanoma patients or healthy donors was centrifuged at 1,550 g for 30 min to obtain cell-free plasma (Beckman Coulter, Allegra X-14R). Then, 1 ml of the obtained plasma was centrifuged at 16,500 g for 45 min (Eppendorf, 5418R). The pelleted microvesicles were suspended in PBS. The collected supernatants were then centrifuged at 100,000 g for 2 h at 4 °C (Beckman Coulter, Optima TM MAX-XP) to pellet the exosomes. For purification of circulating exosomes using the exosome isolation kit, cell-free plasma was first centrifuged at 16,500 g for 45 min (Eppendorf, 5418R) to pellet large membrane vesicles. Exosomes were then purified from the supernatants using the exosome isolation kit (Invitrogen, Cat# 4484450). Characterization of purified exosomes For verification of purified exosomes using electron microscopy, purified exosomes suspended in PBS were dropped on formvar carbon-coated nickel grids. After staining with 2% uranyl acetate, grids were air-dried and visualized using a JEM-1011 transmission electron microscope. For immunogold labelling, purified exosomes suspended in PBS were placed on formvar carbon-coated nickel grids, blocked, and incubated with mouse anti-human monoclonal antibody that recognizes the extracellular domain of PD-L1 (clone 5H1-A3) 1 , followed by incubation with the anti-mouse secondary antibody conjugated with protein A-gold particles (5 nm). Each staining step was followed by five PBS washes and ten ddH 2 O washes before contrast staining with 2% uranyl acetate. The size and concentration of exosomes purified from cell culture supernatants or patients’ plasma were determined using a NanoSight NS300 (Malvern Instruments), which is equipped with fast video capture and particle-tracking software. For iodixanol density gradient centrifugation, exosomes harvested by differential centrifugation were loaded on top of a discontinuous iodixanol gradient (5%, 10%, 20% and 40%, made by diluting 60% OptiPrep aqueous iodixanol with 0.2 5M sucrose in 10 mM Tris) and centrifuged at 100,000 g for 18 h at 4 °C (Beckman Coulter, Optima MAX-XP). Twelve fractions of equal volume were collected from the top of the gradients, with the exosomes distributed at the density range between 1.13 and 1.19 g/ml, as previously demonstrated 8 , 9 , 11 , 12 . The exosomes were further pelleted by ultracentrifugation at 100,000 g for 2 h at 4 °C. Immunoprecipitation To analyse the role of ESCRT machinery in exosomal secretion of PD-L1 in melanoma cells, PD-L1/MEL624 cells were transfected with Flag-Hrs plasmid or vector and then lysed. The cleared lysate was incubated with Anti-FLAG Affinity Gel (Sigma-Aldrich) overnight at 4 °C. The immunoprecipitated proteins were resolved by SDS-polyacrylamide gel electrophoresis and transferred to nitrocellulose membranes. PD-L1 and Flag (Hrs) were determined by western blot using specific antibodies. ELISA For detection of PD-L1 on extracellular vesicles, cell supernatants or patients’ plasma, ELISA plates (96-well) (Biolegend) were coated with 0.25 μg per well (100 μl) of monoclonal antibody against PD-L1 (clone 5H1-A3) overnight at 4 °C. Free binding sites were blocked with 200 μl of blocking buffer (Pierce) for 1 h at room temperature. Then, 100 μl of plasma samples with or without extracellular vesicle removal, or extracellular vesicle samples purified from plasma or cell culture supernatants, were added to each well. The exosome or microvesicle samples purified from cell culture supernatants were prepared by serial dilution according to the total protein level to analyse the enrichment of PD-L1 on exosomes and microvesicles. The concentration of PD-L1 on the surface of exosomes isolated from indicated cells was calculated based on the linear range of the ELISA assay data. The exosome or microvesicle samples derived from the plasma samples of healthy donors or melanoma patients were prepared using the same volume of PBS as the plasma as they were originally derived from. The plasma samples with (extracellular vesicle-excluded) or without (total) extracellular vesicle removal were diluted with PBS in a 1:0.75 volume ratio. After overnight incubation at 4 °C, biotinylated monoclonal PD-L1 antibody (clone MIH1, eBioscience) was added to each well and incubated for 1 h at room temperature. A total of 100 μl per well of horseradish peroxidase-conjugated streptavidin (BD Biosciences) diluted in PBS containing 0.1% BSA was then added and incubated for 1 h at room temperature. Plates were developed with tetramethylbenzidine (Pierce) and stopped with 0.5N H 2 SO 4 . The plates were read at 450 nm with a BioTek plate reader. Recombinant human PD-L1 protein (R&D Systems, Cat# 156-B7) was used to make a standard curve. Recombinant P-selectin protein (R&D Systems, Cat# 137-PS) was used as negative control to verify the detection specificity. The result of standard curve demonstrated that the established ELISA exhibited a reliable linear detection range from 0.2 to 12 ng/ml. For detection of IFN-γ, TNF and IL-2, the supernatant of human CD8 T cells was harvested and measured according to the kit manufacturer’s instructions (Biolegend). PD-1 – PD-L1 binding assay To test the binding of exosomal PD-L1 to PD-1, 100 μl of exosome samples of different concentrations were captured onto PD-L1 antibody (clone 5H1-A3)-coated 96-well ELISA plates by overnight incubation at 4 °C. Then 100 μl of 4 μg/ml biotin-labelled human PD-1 protein (BPS Bioscience, Cat# 71109) was added and incubated for 2 h at room temperature. A total of 100 μl per well of horseradish peroxidase-conjugated streptavidin (BD Biosciences) diluted in PBS containing 0.1% BSA was then added and incubated for 1 h at room temperature. Plates were developed with tetramethylbenzidine (Pierce) and stopped using 0.5N H 2 SO 4 . The plates were read at 450 nm with a BioTek plate reader. Recombinant human PD-L1 protein directly coated onto the plates was used as the positive control. Treatment of CD8 T cells with the exosomes To block PD-L1 on the exosome surface, purified exosomes (200 μg) were incubated with PD-L1 blocking antibodies (10 μg/ml) or IgG isotype antibodies (10 μg/ml) in 100 μl PBS, and then washed with 30 ml PBS and pelleted by ultracentrifugation to remove the non-bound free antibodies. Human CD8 T cells purified from peripheral blood using immunodepletion on a Ficoll-Hypaque gradient (RosetteSep, StemSep Technologies) or mouse CD8 T cells purified from splenocytes using Dynabeads Untouched Mouse CD8 Cells Kit (Invitrogen) were stimulated with anti-CD3 (2 μg/ml) and anti-CD28 (2 μg/ml) antibodies for 24 h and then incubated with human melanoma cell-derived exosomes or mouse B16-F10 cell-derived exosomes with or without PD-L1 blocking for 48 h in the presence of anti-CD3/CD28 antibodies. For human CD8 T cells (2 × 10 5 cells/well in a 96-well plate), 25 μg/ml of human WM9 cell-derived exosomes (carrying surface PD-L1 at a level of 0.05 ng per μg of exosomes as determined by ELISA, Fig. 1i ) were used as the circulating exosomal PD-L1 level in melanoma patients is around 1.25 ng/ml (Fig. 2h ). For mice CD8 T cells (2 × 10 5 cells/well in a 96-well plate), 100 μg/ml of mouse B16-F10 cell-derived exosomes (carrying surface PD-L1 at a level of 0.016 ng per μg of exosomes as determined by ELISA) were used as the circulating exosomal PD-L1 level in mice bearing B16-F10 tumours is around 1.63 ng/ml. The treated cells were then collected, stained, and analysed by flow cytometry. Information about the primary antibodies is included in Supplementary Table 2 . To assay for the proliferation of CD8 T cells, CFSE, a dye for the tracking of cell division (Molecular Probes) was used. A total of 1 × 10 6 CD8 T cells were stained with CFSE at 5 μM. The cells were then incubated at 37 °C for 20 min and the reaction was stopped by adding 5 volumes of cold medium with 10% FBS, and treated as above. Unstimulated CFSE-labelled cells served as a non-dividing control. The exosome – T cell binding assay To verify the physical interactions between melanoma cell-derived exosomes and CD8 T cells, purified exosomes were stained with CFSE in 100 μl PBS, and then washed with 10 ml PBS and pelleted by ultracentrifugation. Unstimulated or stimulated human CD8 T cells (2 × 10 5 cells/well in 96-well plates) were treated with CFSE-labelled exosomes (25 μg/ml) for 2 h, and then fixed for flow cytometry or confocal microscopy after immunostaining for CD8 T cells. Generation of dendritic cells from bone marrow Dendritic cells (DCs) were generated from bone marrow of C57BL/6 mice and cultured in RPMI 1640 with 10% (v/v) FBS, 20 mM l -glutamine, 50 μM β-mercapoethanol, 20 ng/ml IL-4 and 20 ng/ml GM-CSF. After 3 days, half of the culture medium was replaced by fresh medium containing 40 ng/ml IL-4 and 40 ng/ml GM-CSF. To prime antigen-specific OT-I CD8 T cells, DCs were subsequently loaded with 2 μg/ml SIINFEKL (OVA 257–264 ) peptide overnight. CD8 T cell-mediated tumour cell killing assay To determine the effects of melanoma cell-derived exosomes on the ability of CD8 T cells to kill tumour cells, CD8 T cells were purified from the splenocytes of OT-I mice expressing a transgene encoding a T cell receptor that specifically recognized SIINFEKL peptide bound to MHC-I H-2k b 32 . OT-I CD8 T cells (4 × 10 5 cells/well in a 48-well plate) were then activated by incubation with SIINFEK-loaded (2 μg/ml) bone marrow-derived DCs (2 × 10 5 cells/well). The activated OT-I CD8 T cells (4 × 10 5 cells/well in 48-well plate) were treated with PBS (as a control) or B16-F10-derived exosomes (100 μg/ml for 48 h) with or without IgG isotype or PD-L1 antibody blocking (10 μg/ml), and then co-cultured with CFSE-labelled melanoma PD-L1 (KD) B16/OVA cells (4 × 10 5 ) in 6-well plates for 48 h at an effector to target (E:T) ratio of 1:1. Cells were then harvested, intracellularly stained with BV650-conjugated antibody against cleaved-caspase-3 (BD Biosciences) and analysed by flow cytometry. Information about the primary antibodies is included in Supplementary Table 2 . Immunofluorescence staining Immunofluorescence staining was performed on fixed cells or formalin-fixed, paraffin-embedded (FFPE) sections. For fixed cells, permeabilization with 0.1% Triton X-100 was performed before blocking with bovine serum albumin (BSA) buffer for 1 h. For FFPE sections, antigen retrieval by steaming in citrate buffer (pH = 6.0) was performed before blocking. The fixed cells or FFPE sections were incubated with primary antibodies overnight at 4 °C, followed by incubation with fluorophore-conjugated secondary antibodies for 1 h. Nuclei were stained with DAPI. Samples were observed using a Nikon confocal microscope at 100× magnification. Western blot analysis Whole cell lysates or exosomal proteins were separated using 12% SDS–PAGE and transferred onto nitrocellulose membranes. The blots were blocked with 5% non-fat dry milk at room temperature for 1 h, and incubated overnight at 4 °C with the corresponding primary antibodies at dilutions recommended by the suppliers, followed by incubation with HRP-conjugated secondary antibodies (Cell Signaling Technology) at room temperature for 1 h. The blots on the membranes were developed with ECL detection reagents (Pierce). CD63, Hrs, Alix, and TSG101 were used as exosome markers. TYRP-1 and TYRP-2 were used as melanoma-specific markers. GAPDH was used as a loading control. Information about the primary antibodies was included in Supplementary Table 2 . Quantitative PCR (qPCR) Total RNA was isolated from CD8 T cells using TRIzol Reagent (Invitrogen), and reverse transcribed into first-strand complementary DNA (cDNA) with random primer with RevertAid First Strand cDNA Synthesis Kit (ThermoFisher Scientific). The samples were then analysed in an Applied Biosystems QuantStudio 3 Real-Time PCR system. GAPDH was used as an internal control. Information about the primers is included in Supplementary Table 3 . In vivo mice study All animal experiments were performed according to protocols approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Pennsylvania. For establishing human melanoma xenograft model in nude mice, WM9 cells (5 × 10 6 cells in 100 μl medium) were injected into flanks of 8-week-old female athymic nude mice. Tumours were measured using a digital caliper and the tumour volume was calculated by the formula: (width) 2 × length/2. Mice were euthanized 30 days after cell inoculation or if the longest dimension of the tumours reached 2.0 cm before 30 days. Immediately following euthanasia, blood samples were harvested by cardiac puncture, and exosomes were purified and detected by ELISA using the aforementioned method. Exosomes purified from sex-, age- and weight-matched healthy nude mice without xenograft were used as the control. For establishing syngeneic mouse melanoma model in C57BL/6 mice, B16-F10 cells or B16-F10 PD-L1 (KD) cells (5 × 10 5 cells in 100 μl medium) were subcutaneously injected into immunocompetent C57BL/6 mice. Based on the difference in the level of circulating exosomal PD-L1 between mice bearing parental B16-F10 and PD-L1 (KD) B16-F10 tumours (1.63 ng/ml vs 0.70 ng/ml), a total of 100 μg of parental B16-F10 cell-derived exosomes (carrying surface PD-L1 at a level of 0.016 ng per μg of exosomes) with or without IgG isotype, CD63 or PD-L1 blocking (10 μg/ml) were injected into mice after inoculation of PD-L1 (KD) B16-F10 cells to examine the functional significance of PD-L1. The dose of 100 μg exosomes used for our in vivo study was equivalent to approximately 30% of the physiological level of circulating exosomes in mice, and was also comparable to those from a palpable tumour in mice according to our data. Tail vein injections of exosomes (100 μg in 100 μl PBS) were performed every 3 days. Mice were weighed every 3 days. Tumours were measured using a digital caliper and the tumour volume was calculated by the formula: (width) 2 × length/2. The mice were euthanized before the longest dimension of the tumours reached 2.0 cm. Mice were allocated randomly to each treatment group. Downstream analyses of mouse samples (immunofluorescence staining, flow cytometry and ELISA) were performed in a blinded fashion. For flow cytometry, the spleen and tumour samples were harvested, and single cell suspensions were prepared and red blood cells were lysed using ACK Lysis Buffer. Information about the primary antibodies is included in Supplementary Table 2 . Reverse phase protein array (RPPA) RPPA was performed at the MD Anderson Cancer Center core facility using 50 µg protein per sample. All of the antibodies were validated by western blot 13 . Methods for data analysis are described below. Statistical analyses RPPA data analysis was performed according to the protocol from the MD Anderson Cancer Center. Specifically, relative protein levels for each sample were determined by interpolation of each dilution curve from the ‘standard curve’ (supercurve) of the slide (antibody). Supercurve is constructed by a script in R written by the RPPA core facility. The package binaries of SuperCurve and SuperCurveGUI are available in R-Forge ( ). These values are defined as supercurve log 2 value. All the data points were normalized for protein loading and transformed to linear value, designated as ‘normalized linear’. Normalized linear value was transformed to the log 2 value, and then median-centred for further analysis. Median-centred values were obtained by subtracting the median of all samples in a given protein. All of the above-mentioned procedures were performed by the RPPA core facility. The normalized data provided by the RPPA core facility were analysed by Cluster 3.0 ( ) and visualized using the Java TreeView 1.0.5 ( ). All other statistical analyses were performed using GraphPad Prism v.6.0. Normality of distribution was determined by D’Agostino–Pearson omnibus normality test and variance between groups was assessed by the F -test. For normally distributed data, significance of mean differences was determined using two-tailed paired or unpaired Student’s t -tests; for groups that differed in variance, unpaired t -test with Welch’s correction was performed. For data that were not normally distributed, non-parametric Mann–Whitney U -tests or Wilcoxon matched-pairs tests were used for unpaired and paired analysis, respectively. Correlations were determined by Pearson’s r coefficient. Two-way ANOVA was used to compare mouse tumour volume data among different groups. log-rank and Wilcoxon tests were used to analyse the mouse survival data. Error bars shown in graphical data represent mean ± s.d. A two-tailed value of P < 0.05 was considered statistically significant. Reporting summary Further information on experimental design is available in the Nature Research Reporting Summary linked to this paper. Data availability All data and materials are available from the authors upon reasonable request.

Cancer cells are more than a lump of cells growing out of control; they participate in active combat with the immune system for their own survival. Being able to evade the immune system is a hallmark of cancer. Cancer cells release biological "drones" to assist in that fight—small vesicles called exosomes circulating in the blood and armed with proteins called PD-L1 that cause T cells to tire before they have a chance to reach the tumor and do battle, according to researchers from the University of Pennsylvania. The work, published in the journal Nature, is a collaboration between Wei Guo, Ph.D., a professor of Biology in the School of Arts and Sciences, and Xiaowei Xu, MD, Ph.D., a professor of Pathology and Laboratory Medicine in the Perelman School of Medicine. While primarily focused on metastatic melanoma, the team found that breast and lung cancer also release the PD-L1-carrying exosomes. The research offers a paradigm-shifting picture of how cancers take a systemic approach to suppressing the immune system. In addition, it also points to a new way to predict which cancer patients will respond to anti-PD1 therapy that disrupts immune suppression to fight tumors and a means of tracking the effectiveness of such therapies. "Immunotherapies are life-saving for many patients with metastatic melanoma, but about 70 percent of these patients don't respond," said Guo. "These treatments are costly and have toxic side effects so it would be very helpful to know which patients are going to respond. Identification of a biomarker in the bloodstream could potentially help make early predictions about which patients will respond, and, later on, could offer patients and their doctors a way to monitor how well their treatment is working." "Exosomes are tiny lipid-encapsulated vesicles with a diameter less than 1/100 of a red blood cell. What we have found with these circulating exosomes, is truly remarkable," said Xu. "We collected blood samples from melanoma patients treated with anti-PD1 therapy. This type of liquid biopsy assay allows us to monitor tumor-related immune suppression with time. " One of the most successful innovations in cancer therapy has been the use of checkpoint inhibitor drugs, which are designed to block attempts by cancer cells to suppress the immune system to allow tumors to thrive and spread. One of the primary targets for this class of drugs is PD-1, a protein on the surface of T cells. On tumor cells, they express a counterpart molecule called PD-L1, which interacts with the PD-1 protein on T cells, effectively turning off that cell's anti-cancer response. Blocking that interaction using checkpoint inhibitors reinvigorates T cells, allowing them to unleash their cancer-killing power on the tumor. While it was known that cancer cells carried PD-L1 on their surface, in this new work, the team found that exosomes from human melanoma cells also carried PD-L1 on their surface. Exosomal PD-L1 can directly bind to and inhibit T cell functions. Identification of the exosomal PD-L1 secreted by tumor cells provides a major update to the immune checkpoint mechanism, and offers novel insight into tumor immune evasion. "Essentially exosomes secreted by melanoma cells are immunosuppressive." Guo said. "We propose a model in which these exosomes act like drones to fight against T cells in circulation, even before the T cells get near to the tumor." Since a single tumor cell is able to secrete many copies of exosomes, the interaction between the PD-L1 exosomes and T cells provides a systemic and highly effective means to suppress anti-tumor immunity in the whole body. This may explain why cancer patients might have weakened immune system. Because exosomes circulate in the bloodstream, they present an accessible way of monitoring the cancer/T cell battle through a blood test, compared to the traditional more-invasive biopsy of tumors. After an acute phase of treatment, the researchers envision such a test as a way to monitor how well the drugs are keeping cancer cells in check. By measuring pre-treatment levels of PD-L1, oncologists may be able to predict the extent of tumor burden in a patient and associate that with treatment outcome. In addition, a blood test could measure the effectiveness of a treatment, for example, levels of exosomal PD-L1 could indicate the level of T cell invigoration by immune checkpoint inhibitors. "In the future, I think we will begin to think about cancers as a chronic disease, like diabetes," says Guo. "And just as diabetes patients use glucometers to measure their sugar levels, it's possible that monitoring PD-L1 and other biomarkers on the circulating exosomes could be a way for clinicians and cancer patients to keep tabs on the treatments. It's another step toward precision and personalized medicine."

10.1038/s41586-018-0392-8

Medicine

New vaccine promising for tackling sleeping sickness in Africa

Delphine Autheman et al, An invariant Trypanosoma vivax vaccine antigen induces protective immunity, Nature (2021). DOI: 10.1038/s41586-021-03597-x Journal information: Nature

http://dx.doi.org/10.1038/s41586-021-03597-x

https://medicalxpress.com/news/2021-05-vaccine-tackling-sickness-africa.html

Abstract Trypanosomes are protozoan parasites that cause infectious diseases, including African trypanosomiasis (sleeping sickness) in humans and nagana in economically important livestock 1 , 2 . An effective vaccine against trypanosomes would be an important control tool, but the parasite has evolved sophisticated immunoprotective mechanisms—including antigenic variation 3 —that present an apparently insurmountable barrier to vaccination. Here we show, using a systematic genome-led vaccinology approach and a mouse model of Trypanosoma vivax infection 4 , that protective invariant subunit vaccine antigens can be identified. Vaccination with a single recombinant protein comprising the extracellular region of a conserved cell-surface protein that is localized to the flagellum membrane (which we term ‘invariant flagellum antigen from T. vivax ’) induced long-lasting protection. Immunity was passively transferred with immune serum, and recombinant monoclonal antibodies to this protein could induce sterile protection and revealed several mechanisms of antibody-mediated immunity, including a major role for complement. Our discovery identifies a vaccine candidate for an important parasitic disease that has constrained socioeconomic development in countries in sub-Saharan Africa 5 , and provides evidence that highly protective vaccines against trypanosome infections can be achieved. Main African trypanosomiasis is an infectious disease that is caused by unicellular parasites of the genus Trypanosoma , which are transmitted by the bite of an infected tsetse fly. In humans, trypanosome infections cause sleeping sickness: a deadly disease that threatens the lives of millions of people who live in over 30 countries in sub-Saharan Africa 6 . Some species of trypanosome also infect important livestock animals (including cattle, goats and pigs) and cause the wasting disease nagana (animal African trypanosomiasis), which affects the livelihoods of people who rely on these animals for milk, food and draught power 1 . Approximately three million cattle die from this disease every year, which results in an estimated direct annual economic cost of many hundreds of millions of dollars 7 and represents a major barrier for the socioeconomic advancement of many countries in Africa. Nagana is primarily caused by T. vivax and Trypanosoma congolense and is currently managed with drugs, but resistance is increasing 8 . Previous attempts to develop subunit vaccines against African trypanosome infections have highlighted the difficulties in overcoming the immune evasion strategies that have been evolved by these parasites to enable them to survive in host blood 9 . These strategies include antigenic variation: the serial expression of an abundant allelically excluded variable surface glycoprotein (VSG), and the rapid removal of surface-bound antibodies by hydrodynamic sorting 3 , 10 . It is widely thought that the VSG forms a constantly changing, impenetrable surface coating that sterically shields other surface proteins from host antibodies, which leads to chronic infections characterized by oscillating parasitaemia; however, a careful analysis of recent structural information suggests that this model may not fully explain the protective role of the VSGs 11 . We therefore hypothesized that the VSGs may also subvert natural immunity by preventing the acquisition of high-titre antibody responses to protective antigens, suggesting that eliciting prophylactic unnatural host immunity by vaccination could be achieved. Here we report the identification of a conserved cell-surface protein that—when used as a subunit vaccine in a mouse model of T. vivax infection—is capable of eliciting highly protective immunity. A subunit vaccine induces immunity to T. vivax To identify subunit vaccine candidates for T. vivax , we established a genome-led vaccinology approach using a bioluminescent mouse model of infection 4 (Fig. 1a ). We determined that adoptive transfer of about 100 parasites intravenously into BALB/c mouse hosts resulted in an acute, reproducible infection that permitted sensitive and accurate quantification of parasitaemia using light-based imaging 12 . We selected subunit vaccine candidates by searching the T. vivax genome 13 for genes that encode predicted cell-surface and secreted proteins that are likely to be accessible to vaccine-elicited antibodies. We selected 60 candidates using the following criteria: (1) they did not belong to paralogous gene families (to minimize the risk of functional redundancy in different mammalian hosts) 14 ; (2) contained more than 300 amino acids in their predicted extracellular region (and so are likely to be accessible at the cell surface); and (3) had evidence of expression in the blood stages 15 (Supplementary Table 1 ). We synthesized gene sequences encoding the entire predicted extracellular region, and cloned them into a mammalian protein expression plasmid that contained an exogenous secretion peptide and purification tags. We expressed the candidates as soluble recombinant proteins in HEK293 cells to increase the chance that structurally critical posttranslational modifications were appropriately added. Of the 60 expression plasmids we tested, 39 yielded sufficient protein after purification for vaccination trials (Extended Data Fig. 1a ). For vaccination, we used a prime and two-boost regime that used alum as an adjuvant to bias host responses towards humoral immunity. To reduce any systemic adjuvant-elicited effects on disease progression, we rested the vaccinated mice for a minimum of four weeks following the final boost before parasite challenge (Extended Data Fig. 1b, c ). Trypanosoma vivax loses virulence once removed from donor mice, and so, to avoid confounding effects due to the loss of parasite viability during the infection procedure, we ensured that infections were comparable in control mice challenged before and after the vaccinated mice (Extended Data Fig. 1b ). Fig. 1: Candidate V23 (IFX) induces protective immunity in a mouse model of T. vivax infection. a , Schematic of the genome-led vaccinology approach. b , Summary of mean parasitaemia ( n ≥ 3 mice) quantified by bioluminescence in cohorts of vaccinated mice challenged with T. vivax . Although the V58 candidate had lower mean parasitaemia relative to controls on day 8, it rebounded on day 9 and so was not considered further. c , d , Bioluminescent images of adjuvant-only control (left panels) and mice vaccinated with V31 ( c right) and V23 ( d right); images were taken five (top panels) and eight (bottom panels) days after infection. Source data Full size image We determined the elicited antibody titres to each antigen, and the vast majority (90%) had mean half-maximal responses at serum dilutions greater than 1:10,000 (Extended Data Fig. 1d ). We found that, of the 39 antigens tested, 34 had no effect on the infection parameters relative to controls (Fig. 1b , Extended Data Fig. 2 ). We observed statistically significant effects on parasite growth with four antigens (Fig. 1b ): two candidates (V2 and V8) exhibited a slight delay, and one candidate (V31) showed a longer delay, to the ascending phase of parasitaemia (Fig. 1b, c ), and one candidate (V23) showed no detectable parasites in all five vaccinated mice (Fig. 1b, d ). For each of the four candidates, we repeated the experiments using independent protein preparations and larger cohorts of mice. The two candidates that induced a slight delay (V2 and V8) did not replicate, and so were not pursued further (Extended Data Fig. 3a ). Candidate V31 reduced the rate of parasite multiplication once more and induced improved protection compared to the initial screen so that 9 out of 15 mice survived until day 16 after infection (Extended Data Fig. 3a ). V23 vaccination again elicited robust protection, and longitudinal sampling of these mice showed that 10 out of the 15 mice were protected beyond at least day 170 (Extended Data Fig. 3a, b ). We dissected protected mice several months after infection, which revealed no detectable extravascular reservoirs of parasites (Extended Data Fig. 4 ). On the basis of these and subsequent findings, we propose the name ‘invariant flagellum antigen from T. vivax ’ (IFX) for the V23 candidate (TvY486_0807240). IFX localizes to the flagellum IFX is a previously uncharacterized type-I cell-surface glycoprotein that contains a short (18 amino acid) cytoplasmic region that does not include any known protein domains and has no paralogues (protein sequence identity greater than 25%) within T. vivax , nor homologues in other sequenced genomes of Trypanosoma species. To begin the functional characterization of IFX, we asked whether it had a specific localization in blood-stage parasites. Immunocytochemistry showed that staining was localized along the length of the flagellum and loosely concentrated in discrete puncta (Fig. 2a ). Using immunogold electron microscopy, we found that IFX was enriched at the boundaries of where the flagellum is attached to the cell body; in different sections, these clusters were either unilaterally or bilaterally located (Fig. 2b ). In mid-sagittal sections, IFX was located along the length of the flagellum membrane and concentrated in discrete clusters at the points at which the flagellum was in close apposition to the cell membrane; specifically, the gold particles were located between the flagellum and cell body membranes (Extended Data Fig. 5a–f ). These data demonstrated that IFX was localized to the flagellum membrane and particularly enriched as continuous or punctuated bilateral stripes along the flagellum, and bordering the region where the flagellum is attached to the parasite cell body, which suggests a structural role in maintaining flagellar function. Fig. 2: IFX is expressed on the flagellum membrane of T. vivax and concentrated at the periphery of the flagellum–cell-body contact. a , Immunofluorescence staining of T. vivax with rabbit anti-IFX antiserum (red) (left) or control pre-immune serum (right) counterstained with DAPI (blue) demonstrates localization of IFX to the flagellum. Scale bars, 5 μm. b , Immunogold electron microscopy using an anti-IFX mouse monoclonal antibody localized IFX to the borders of where the flagellum is in contact with the parasite cell body in transverse sections (black arrows in left and middle images) compared to isotype-matched control (right). Scale bars, 100 nm. Representative images of at least two independent experiments are shown. Full size image Passive protection by antibodies to IFX To determine the immunological mechanisms of IFX-mediated protection, we first demonstrated that antibodies contributed to immunity by transferring immune serum from IFX-vaccinated mice to naive recipient mice, which inhibited parasite growth in a dose-dependent manner (Fig. 3a ). Depletion of CD4- and CD8-positive T lymphocytes and NK1.1-positive natural killer cells in IFX-vaccinated mice did not affect protective efficacy, demonstrating that these cell types were not direct executors of immunity once it was established (Extended Data Fig. 6 ). To further investigate the role of antibodies in immunity using an independent approach, we selected six hybridomas that secreted monoclonal antibodies to IFX. Of the six monoclonal antibodies we selected, three affected parasite growth when used in passive protection experiments (Fig. 3b ). We determined the approximate location of the monoclonal-antibody-binding sites on IFX and quantified their binding affinities, but did not observe a simple positive correlation between their protective efficacy and either the location of their epitope or binding affinity (Extended Data Fig. 7 ). The inhibitory effects of one antibody (designated 8E12) titrated with dose (Fig. 3c ). Fig. 3: Passive transfer of immunity to T. vivax infections with anti-IFX antibodies. a , Dose-dependent inhibition of T. vivax by adoptive transfer of sera from IFX-vaccinated mice relative to sera from unimmunized control mice. Groups of five mice were compared by one-way analysis of variance (ANOVA) with Sidak post hoc test; * P ≤ 0.05, **** P ≤ 0.0001. b , Three out of six anti-IFX IgG1-isotype monoclonal antibodies (6B3, 8E12 and 3D12), each given at a dose of 3 × 100 μg, passively protect against T. vivax infection relative to an isotype-matched control. Groups of five mice were compared by one-way ANOVA with Sidak post hoc test; **** P ≤ 0.0001. c , Passive protection of the 8E12–IgG1 monoclonal antibody is dose-dependent. Parasitaemia was quantified at day 5 using bioluminescence. Data points represent individual mice; bars indicate mean ± s.d., groups of five mice were compared by one-way ANOVA with Sidak post hoc test; ** P ≤ 0.01, **** P ≤ 0.0001. The background bioluminescence threshold is indicated by grey shading. Source data Full size image Several mechanisms of anti-IFX protection Isotyping the monoclonal antibodies to IFX revealed that they were all of the IgG1 subclass, which, in mice, do not effectively recruit immune effector functions such as complement or bind activating Fc receptors with high affinity 16 ; this suggests that direct antibody binding to IFX affected parasite viability. To establish the role of Fc-mediated immune effectors in anti-IFX antibody protection, we selected 8E12 (which gave intermediate protective effects) and, by cloning the rearranged antibody variable regions, switched the monoclonal antibody isotype from IgG1 to IgG2a (Extended Data Fig. 8 ). We observed that the 8E12–IgG2a monoclonal antibodies had a substantially increased potency compared to the 8E12–IgG1 when used in passive transfer experiments (Fig. 4a ), and titrating this antibody showed that 3 doses of 50 micrograms or more conferred sterile protection (Fig. 4b , Extended Data Fig. 9a ). This demonstrated that the recruitment of antibody-mediated immune effectors was important for parasite neutralization, and to quantify their relative contributions we engineered three further monoclonal antibodies, each of which lacked the binding sites for C1q (ΔC1q), FcRs (ΔFcR) or both (ΔC1qΔFcR) 17 (Extended Data Fig. 8c ). When used in passive protection experiments, we observed that mutation of the C1q binding site almost completely reversed the inhibition of parasite growth, which demonstrates that C1q-mediated complement recruitment was a major protective mechanism (Fig. 4c , Extended Data Fig. 9b ). Mutating the FcR binding site also relieved the inhibition of parasite growth (but to a lesser extent), whereas mutation of both C1q and FcR sites inhibited growth with a potency similar to that of the IgG1 isotype (Fig. 4c , Extended Data Fig. 9b ). These experiments revealed that anti-IFX antibodies inhibited parasite multiplication by several immune mechanisms, dominated by the recruitment of complement. Fig. 4: There are several mechanisms of antibody-mediated anti-IFX immunological protection, dominated by complement recruitment. a , Anti-IFX 8E12–IgG2a monoclonal antibodies passively protect against T. vivax infection more potently than do 8E12–IgG1 monoclonal antibodies. b , Dose titration of 8E12–IgG2a monoclonal antibody compared to isotype-matched control. c , Passive transfer of 8E12–IgG2a monoclonal antibodies containing mutations that prevent binding to C1q (ΔC1q), FcRs (ΔFcR) or both (ΔC1qΔFcR) relieved the inhibition of parasite multiplication to differing degrees, demonstrating that there are several mechanisms of antibody-based immunological protection, including a major role for complement. Parasitaemia was quantified at day 5 using bioluminescence. Data points represent individual mice. Grey shading indicates background bioluminescence. Bars indicate mean ± s.d. One of two independent experiments with very similar outcomes is shown. Source data Full size image IFX is highly conserved across isolates To further assess and develop IFX as a potential vaccine target, we tested appropriate routes of administration and other adjuvants that would bias antibody responses towards more-protective isotypes. Of two selected adjuvants that have previously been used in veterinary vaccines and can be delivered subcutaneously, we found that the saponin-based adjuvant Quil-A elicited consistent antibody titres equivalent to the protective responses induced by alum, of which a large proportion were of the IgG2 isotypes (Extended Data Fig. 10a, b ); these mice were potently protected against parasite challenge (Fig. 5a , Extended Data Fig. 10c ). One potential challenge with subunit vaccines is that the genes encoding antigens that elicit protective immune responses in natural infections can be subject to diversifying selection, which potentially leads to strain-specific immunity and limits the usefulness of the vaccine 18 . We therefore analysed the IFX gene sequence in 29 cosmopolitan T. vivax genomes and showed that it is highly conserved by comparison to other surface antigens. We observed only a single non-synonymous polymorphism in 2 out of 29 strains (Fig. 5b ) and both of these were heterozygotes; the frequency of the mutation across all strains was therefore very low (0.058), demonstrating that IFX is almost completely invariant within the parasite population. This high level of sequence conservation across isolates suggests that the IFX protein is not a target of host immune responses and, consistent with this, sera from naturally infected cattle were not immunoreactive to IFX (Fig. 5c ). Finally, a successful vaccine must be able to elicit long-lasting protection: we therefore repeatedly challenged IFX-vaccinated mice over 100 days after they received their final immunization. We observed that mice remained fully protected, including when parasites were delivered subcutaneously (Fig. 5d ). Fig. 5: IFX is highly conserved and can elicit long-lasting immunity to T. vivax infections. a , Comparing veterinary adjuvants using subcutaneous delivery demonstrates that Quil-A is as effective as the positive control (IFX adjuvanted in alum and delivered intraperitoneally). Parasitaemia was quantified at day 5 using bioluminescence. Data points represent individual mice. Grey shading indicates background bioluminescence. Bars indicate mean ± s.d. Mont., montanide ISA 201 VG. b , Parasite population genetic analysis shows that IFX is highly conserved compared to other genes; mean synonymous (Ds) and nonsynonymous (Dn) substitution densities are shown (+s.d. where appropriate). c , IFX is not immunogenic in the context of a natural infection. Immunoreactivity to the indicated proteins in sera from Cameroonian (C) ( n = 10) or Kenyan (K) ( n = 10) cattle, or uninfected control cattle from the UK (U) ( n = 6). Bar indicates median. d , Mice vaccinated with IFX adjuvanted in alum were protected from two intravenous (i.v.) and one final subcutaneous (s.c.) T. vivax challenge given over 100 days after the final booster immunization. Source data Full size image Discussion We have shown that it is possible to elicit apparently sterile protection to an experimental trypanosome infection with a subunit vaccine that corresponds to the ectodomain of the invariant cell-surface parasite protein IFX. The localization of IFX to the boundaries of where the flagellum is in contact with the parasite cell body suggests it performs a role in flagellar structure and function. Our demonstration that antibodies are required for immunity raises questions about the immunoprotective mechanisms used by trypanosomes, and, importantly, that their vulnerabilities that can be exploited to develop vaccines. The inhibition of parasite growth by antibodies to IFX suggest that the VSG surface coat of T. vivax cannot fully shield it from antibody binding, and that anti-IFX antibodies are not removed by endocytosis within the flagellar pocket from the parasite surface with sufficient rapidity to prevent immune effector recruitment mediated by antibodies. The findings that the IFX gene sequence was highly conserved across parasite isolates, and that sera from infected cattle living in nagana-endemic regions were not immunoreactive to IFX, suggest that natural parasite infections in some species can subvert host immunity to avoid eliciting protective antibody responses. These mechanisms could include perturbations of the B cell compartment, which have previously been described in experimental models of other trypanosome species 19 , 20 , 21 , or that the IFX protein may not be suitably presented to the host immune system in the context of a natural infection. Preliminary experiments to translate these findings to goats did not show protection 22 and highlighted the need to develop infection models that are suitable for vaccine testing, and a greater understanding of which antibody isotypes and adjuvants elicit the necessary mechanisms of immune effector recruitment. The discovery of an antigen that can elicit protection to a trypanosome infection provides optimism and a technical road map that could be followed to identify vaccine antigens not only for other trypanosome species, but also parasites that have thus far proven intractable to vaccine development. Finally, IFX represents a very attractive vaccine candidate for an important livestock disease that has been a major barrier to socioeconomic development in sub-Saharan Africa. Methods A group size of five mice was selected for the initial vaccine screening on the basis of the highly reproducible nature of the T. vivax infection between individual mice as quantified by bioluminescence, and the requirement for a strong effect size for an effective vaccine. In a typical vaccine test experiment, we calculate a mean and s.d. of 6 × 10 7 ± 1 × 10 7 photons per second ( n = 5) in an adjuvant control cage on day five after infection. A sample size of five mice at 90% statistical power would be sufficient to allow us to detect a reduction in parasitaemia of ≥35% using a one-sided t -test at P ≤ 0.05. Larger group sizes of up to 15 mice were used in replication studies. The experiments were not randomized, and investigators were not blinded to allocation during experiments and outcome assessment, although vaccination and parasite quantification were performed by independent researchers. Mouse strains and ethical approvals All mouse experiments were performed under UK Home Office governmental regulations (project licence numbers PD3DA8D1F and P98FFE489) and European directive 2010/63/EU. Research was ethically approved by the Sanger Institute Animal Welfare and Ethical Review Board. Mice were maintained under a 12-h light/dark cycle at a temperature of 19–24 °C and humidity between 40 and 65%. The mice used in this study were 6–14-week-old female Mus musculus strain BALB/c, which were obtained from a breeding colony at the Research Support Facility, Wellcome Sanger Institute. Cell lines and antibodies Recombinant proteins and antibodies used in this study were expressed in HEK293E 23 or HEK293-6E 24 provided by Y. Durocher. Neither cell line was authenticated but they were regularly tested for mycoplasma (Surrey Diagnostics) and found to be negative. The antibodies used in this study were as follows. Primary antibodies were: six anti-IFX mouse monoclonal antibodies were selected and validated in this study: 6B3, 8C9, 3D12, 2H3, 10E2 and 8E12 from hybridomas that all secreted IgG1 isotypes. A rabbit polyclonal antibody to the entire ectodomain of IFX was generated by Cambridge Research Biochemicals and validated by enzyme-linked immunosorbent assay (ELISA) against the recombinant IFX ectodomain. The 8E12 antibody was cloned and expressed recombinantly as a mouse IgG2a isotype as described in ‘Antibody cloning, isotype switching, mutagenesis and purification’. Mouse isotype control antibodies were: IgG1 (MOPC-21, BE0083, BioXcell) and IgG2a (C1.18.4, BE0085, BioXcell). Antibodies used for in vivo leukocyte cell depletion were: anti-mouse CD4 (clone GK1.5, BP0003-1, BioXcell), anti-mouse CD8 (clone 2.43, BP0061, BioXcell), anti-mouse NK1.1 (clone PK136, BE0036, BioXcell), and control anti-keyhole limpet haemocyanin (clone LTF-2, BP0090, BioXcell). Antibodies used for protein quantification for ELISAs were mouse monoclonal anti-His (His-Tag monoclonal antibody, 70796, EMD-Millipore), and biotinylated mouse anti-rat CD4 (clone OX68). OX68 was purified from the spent tissue culture medium from the hybridoma, which was a gift from N. Barclay. Secondary antibodies used were: goat anti-mouse alkaline phosphatase conjugated secondary (A3562, Sigma-Aldrich) and rabbit anti-bovine alkaline phosphatase conjugated secondary (A0705 Sigma-Aldrich). Mouse antibody isotypes were determined using the mouse monoclonal antibody isotyping kit (ISO2-KT Sigma-Aldrich). Vaccine target identification and expression The T. vivax genome was searched for genes encoding predicted type I, GPI-anchored and secreted proteins using protein feature searching in TriTrypDB 25 . The regions corresponding to the entire predicted extracellular domains of T. vivax cell-surface and secreted proteins from the Y486 strain were determined by using transmembrane 26 and GPI-anchor 27 or signal peptide 28 prediction software. Protein sequences encoding the predicted extracellular domain and lacking their signal peptide were codon-optimized for expression in human cells and made by gene synthesis (Geneart and Twist Bioscience). The sequences were flanked by unique NotI and AscI restriction enzyme sites and cloned into a pTT3-based mammalian expression vector 23 between an N-terminal signal peptide to direct protein secretion and a C-terminal tag that included a protein sequence that could be enzymatically biotinylated by the BirA protein–biotin ligase 29 and a 6-His tag for purification 30 . The ectodomains were expressed as soluble recombinant proteins in HEK293 cells as previously described 31 , 32 . To prepare purified proteins for immunization, between 50 ml and 1.2 l (depending on the level at which the protein was expressed) of spent culture medium containing the secreted ectodomain was collected from transfected cells, filtered and purified by Ni 2+ immobilized metal-ion affinity chromatography using HisTRAP columns using an AKTAPure instrument (GEHealthcare). Proteins were eluted in 400 mM imidazole as previously described 33 , and extensively dialysed into HEPES-buffered saline (HBS) before being quantified by spectrophotometry at 280 nm. Protein purity was determined by resolving one to two micrograms of purified protein by SDS–PAGE using NuPAGE 4–12% Bis Tris precast gels (ThermoFisher) for 50 min at 200 V. Where reducing conditions were required, NuPAGE reducing agent and anti-oxidant (Invitrogen) were added to the sample and the running buffer, respectively. The gels were stained with InstantBlue (Expedeon) and imaged using a c600 Ultimate Western System (Azure biosystems). Purified proteins were aliquoted and stored frozen at −20 °C until used. Where enzymatically monobiotinylated proteins were required to determine antibody titres by ELISA, proteins were co-transfected with a secreted version of the protein biotin ligase (BirA) as previously described 32 , and extensively dialysed against HBS and their level of expression determined by ELISA using a mouse monoclonal anti-His antibody (His-Tag monoclonal antibody, 70796, EMD Millipore) as primary antibody and a goat anti-mouse alkaline phosphatase-conjugated secondary (A3562, Sigma-Aldrich). Vaccine formulation and administration For the initial screening of antigens, aliquots of purified protein for immunization were thawed, diluted and mixed 50% v/v with alhydrogel adjuvant 2% (InvivoGen) for two hours at room temperature. For each antigen, groups of five 6–8-week-old female BALB/c mice were immunized intraperitoneally using a prime and two-boost strategy using the amounts of protein documented in Supplementary Table 1 . For retesting those antigens that had shown some effect in the preliminary screen, one group of 15 mice received three intraperitoneal immunizations of the query protein adjuvanted in alum using similar amounts as used in the initial screen (Supplementary Table 1 ); a control group, also of 15 mice, received the adjuvant alone. For evaluating different IFX vaccine–adjuvant formulations, groups of five mice received three immunizations of 50 μg IFX adjuvanted with either alhydrogel, montanide ISA 201 VG or Quil-A in a total volume of 200 μl. IFX was formulated with montanide ISA 201 VG according to the manufacturer’s instructions using a stirrer to create the water-in-oil emulsion. IFX was mixed in a 1:1 (v/v) ratio with Quil-A adjuvant using a 0.5 mg ml −1 solution. IFX adjuvanted in montanide, or Quil-A were administered subcutaneously at two different injection sites (100 μl per site), whereas IFX adjuvanted in alhydrogel was administered intraperitoneally. Quantification of serum antibody titres by ELISA To determine the serum antibody responses to immunized proteins, blood biopsies were collected between ten to twelve days after the final immunization from the tail of each mouse and clotted for two hours at room temperature. Cells were removed by centrifugation, the serum collected, supplemented with sodium azide to a final concentration of 2 mM as a preservative and stored at −20 °C until use. Cattle sera were donated from archived material at the University of Liverpool, originally collected from natural T. vivax infections in Cameroon (northwest state), and Kenya (western state) where the infection was positively identified by thick blood smear and the parasite identified as T. vivax using the VerY Diag field test. To determine the antibody titre against an antigen of interest, individual sera were initially diluted 1:1,000 and then six fourfold serial dilutions in PBST and 2% BSA were prepared. These dilutions were pre-incubated overnight at room temperature with 100 μg ml −1 of purified rat CD4d3 + 4-BLH protein to adsorb any anti-biotin or His tag antibodies. Sera were transferred to streptavidin-coated ELISA plates on which the biotinylated target antigen was immobilized. To ensure that all anti-tag antibodies were adsorbed, binding of the lowest dilution of antisera was also tested against biotinylated rat CD4d3 + 4-BLH protein similarly immobilized on the ELISA plate to confirm the absence of any anti-tag immunoreactivity 34 . Sera were incubated for one hour at room temperature followed by three washes with PBST before incubating with an anti-mouse IgG secondary antibody conjugated to alkaline phosphatase (Sigma-Aldrich) used as a 1:5,000 dilution for one hour. Following three further washes with PBST, 100 μl of 1 mg ml −1 Sigma 104 phosphatase substrate was added and substrate hydrolysis quantified at 405 nm using a plate reader (Spark, Tecan). To quantify immunoreactivity to T. vivax antigens in the context of natural infections, cattle sera were diluted 1:800 in PBST and 2% BSA and incubated for two hours at room temperature with biotinylated ectodomains of V2, V53, IFX or control rat CD200, adsorbed on the microtitre plate. Following three washes with PBST, a secondary rabbit anti-bovine IgG antibody (A0705, Sigma-Aldrich) diluted 1:20,000 was incubated for one hour and washed three times with PBST before adding colorimetric phosphatase substrate and acquiring absorbance readings. Antibody isotyping Isotyping of the monoclonal antibodies and polyclonal sera responses was performed using the Mouse Monoclonal Antibody Isotyping Kit (ISO2-KT, Sigma-Aldrich), according to the manufacturer’s instructions. In brief, the biotinylated ectodomain of the IFX protein was immobilized on a streptavidin-coated plate, incubated with sera diluted 1:1,000 in PBST and 2% BSA or hybridoma supernatants, washed in PBST before adding isotype-specific goat anti-mouse secondary antibodies diluted 1:1,000. Binding was quantified with an alkaline-phosphatase-conjugated rabbit anti-goat tertiary antibody (1:5,000, Sigma-Aldrich) followed by a colorimetric phosphatase substrate, and hydrolysis products quantified by absorbance readings at 405 nm. Trypanosoma parasite strain and maintenance A transgenic form of T. vivax genetically engineered to ubiquitously express the firefly luciferase enzyme 35 was provided by P. Minoprio. The parental strain of this parasite is the IL1392 line derived from the Y486 strain used for genome sequencing 13 and has previously been fully documented 12 . Parasites were initially recovered from a frozen stabilate by intraperitoneal administration into two BALB/c female mice. Parasites were maintained by weekly serial blood passage in wild-type female BALB/c mice by taking a blood biopsy, quantifying living parasites in PBS and 20 mM d -glucose by microscopy and infecting four naive mice intravenously. During the course of the project, two further aliquots of frozen parasites were thawed and then used for infection challenges: no significant differences in the kinetics of infection were observed. Luciferase-expressing T. congolense parasites were a gift from B. Wickstead and C. Gadelha, and were maintained by weekly serial intravenous blood passage in wild-type female BALB/c mice. Trypanosoma vivax infections For infection challenges, bloodstream forms of T. vivax parasites were obtained from the blood of an infected donor mouse at the peak of parasitaemia, diluted in PBS and 20 mM d -glucose, quantified by microscopy and used to infect mice by intravenous injection. While establishing the infection model in our facility, we observed that the T. vivax parasite was labile and gradually lost virulence once removed from living mice. To reduce the possibility of any artefactual protective effects being due to the loss of parasite virulence during the challenge procedure, we screened the protective effects of antigens in a cohort design. Each cohort contained six cages of five mice: four cages contained mice immunized with a different query subunit-vaccine candidate, and the other two cages contained control mice immunized with adjuvant alone. Vaccinated mice were rested for four to eight weeks after the final immunization to mitigate any possible non-specific protective effects elicited by the adjuvant. During the infection procedure, the mice in the control cages were challenged first and last, and the data from the cohort used only if the infections in the control mice from the two cages were comparable. During the infection procedures, parasites were outside of a living mouse for no more than 40 min. Mice were normally challenged by intravenous delivery of 10 2 (cohorts 1–7, 10 and 11) to 10 3 (cohorts 8 and 9) parasites for the initial screening and passive transfer protection experiments, but were also challenged intraperitoneally during the establishment of the model and subcutaneously when investigating the duration of protection. The mice were not randomized between cages and the operator was not blinded to the group condition. Occasionally, individual infected mice within a group unexpectedly exhibited only background levels of bioluminescence, which was attributed to the injected luciferin substrate not distributing from the site of delivery (possibly due to mislocalization of the injection bolus); in these instances, the mice were excluded from the analysis. This occurred 12 times out of 1,650 injections (0.7%) in screening cohorts 3, 5, 6 and 9. Groups were compared using bioluminescence quantification as a proxy for parasitaemia and one-way ANOVA with Dunnett’s post hoc test unless specified. Quantification of T. vivax infections by bioluminescent in vivo imaging The luciferase substrate d -luciferin (potassium salt, Source BioScience) was reconstituted to 30 mg ml −1 in Dulbecco’s PBS (Hyclone), filter-sterilized (0.22 μm) and stored in aliquots at −20 °C. Aliquots were thawed and administered to animals at a dose of 200 mg kg −1 , by intraperitoneal injection ten minutes before bioluminescence acquisitions. The mice were given three minutes of free movement before being anaesthetized with 2.5% isoflurane and placed in the imaging chamber where anaesthesia was maintained for acquisition. An average background bioluminescence measurement was determined by luciferin administration in five female BALB/c mice and calculating the mean whole-body bioluminescence; where appropriate, this value is indicated as a light grey shading on bioluminescence plots. To determine long-term persistence of the parasites in different organs of infected mice, mice were administered with luciferin, imaged and then euthanized with an overdose of anaesthetic. Mice were then perfused with PBS until the perfusion fluid ran clear, the organs dissected, arranged on a Petri dish and bathed in PBS containing 20 mM glucose and 3.3 mg ml −1 luciferin for imaging. Emitted photons were acquired by a charge coupled device (CCD) camera (IVIS Spectrum Imaging System, Perkin Elmer). Regions of interest (ROIs) were drawn and total photons emitted from the image of each mouse were quantified using Living Image software version 4.7.4 (Xenogen), the results were expressed as the number of photons s −1 . Bioluminescence values were exported and plotted in Prism GraphPad version 8.0.2, which was also used for testing statistical significance where needed. Where necessary, peripheral parasitaemia was quantified by direct microscopic observation as previously described 12 . In brief, five microlitres of blood obtained from the tail vein were appropriately diluted in PBS containing 20 mM glucose and parasite counts were expressed as number of parasites per blood millilitre. Passive transfer of immunity To obtain sufficient sera for adoptive transfer experiments, fifty 6–8-week-old female BALB/c mice were immunized intraperitoneally three times with 20 μg of purified IFX adjuvanted in alum, with each immunization separated by two weeks. Nine days after the final immunization, sera were collected as above, aliquoted and stored at −20 °C until use. For passive transfer experiments, groups of 10-to-14-week-old female BALB/c mice were dosed three times with either sera or purified monoclonal antibodies on three consecutive days; three hours after the second dosing, mice were challenged intravenously with 10 2 T. vivax parasites. When using immune serum for passive transfer protection experiments, doses of 100 and 200 μl of sera from either IFX-vaccinated mice or non-immunized control mice were administered. For monoclonal antibodies, the purified antibody was diluted to the required dose in PBS and 200 μl administered intravenously. Control isotypes antibodies used were MOPC-21 for the IgG1 isotype and C1.18.4 for the IgG2a isotype (both from BioXcell). The serum half-life for mouse IgG1 and IgG2a are known to be between 6 and 8 days 36 . In vivo cell depletion Groups of five mice were immunized three times with 50-μg doses of purified IFX to induce protective immunity to T. vivax . To deplete immune mice of defined leucocyte lineages, mice within each group were depleted by intraperitoneal administration of lineage-specific monoclonal antibodies using standard procedures. In brief, natural killer cells were depleted by four injections of 500 μg of the PK136 monoclonal antibody that targets the NK1.1 glycoprotein at days −5, −1, 0 and 2 relative to T. vivax challenge. Mouse CD4 and CD8 T lymphocytes were depleted by one intraperitoneal 750-μg injection of the monoclonal antibodies targeting CD4 (clone GK1.5) or CD8 (clone 2.43) receptors, respectively, the day before the infection. The LFT-2 monoclonal antibody (750 μg) was used as an isotype-matched control antibody. Mice were challenged with 10 2 T. vivax parasites and parasitaemia quantified was using bioluminescent imaging as described in ‘Quantification of T. vivax infections by bioluminescent in vivo imaging’. Trypanosome genomic sequence analysis To identify whether IFX had any homologues in other Trypanosome species, the entire IFX sequence was analysed with Interproscan; this showed that it does not contain any known protein domains other than the predicted N-terminal signal peptide and transmembrane helix. Comparison of the predicted IFX protein sequence with all the other sequenced Trypanosoma species genomes in TriTrypDB 25 ( ) using tBLASTx returned no reliable matches; moreover, comparison of a hidden Markov model of the IFX protein sequence with all Trypanosoma brucei, T. congolense and Trypanosoma cruzi proteins using HMMER also produced no matches, demonstrating IFX is unique to T. vivax . To confirm that IFX is present in a single copy, the IFX protein sequence was compared with a six-frame translation of the genome sequence using tBLASTn to identify any sequence copies (annotated or not) with >98% amino acid identity typical of allelic variation. Illumina sequencing reads from 29 clinical strains isolated from Nigeria, Togo, Burkina Faso, The Gambia, Ivory Coast, Uganda and Brazil were mapped to the T. vivax Y486 reference sequence using BWA 37 before single-nucleotide polymorphisms (SNPs) were called using the GATK4 analysis toolkit 38 . Insertion and/or deletions (indels) and variant positions with QD < 2.0, FS > 60.0, MQ < 40.0, MQRankSum < −12.5 or ReadPosRankSum < −8.0 were excluded to produce a final list of 403,190 SNPs. Individuals were classified as missing if allele calls were supported by fewer than three reads. Coding SNPs and synonymous or non-synonymous codon alterations were identified by comparison to the reference annotation using a custom Biopython script; pi-values were calculated on a per site basis using vcftools 39 . Genes selected for comparison were V31 (TvY486_0003730), two VSGs (TvY486_0031620 and TvY486_0040490), ISG (TvY486_0503980), HpHbR (TvY486_0040690) and the ‘housekeeping’ gene GAPDH (TvY486_1006840). Electron microscopy Trypanosoma vivax parasites were resuspended in 1% formalin prepared using freshly dissolved paraformaldehyde in PBS for 30 min (all steps at room temperature), washed three times in PBS, blocked with PBS and glycine followed by 5% fetal calf serum for 30 min and then incubated with a mouse monoclonal antibody to IFX (clone 8E12) for 1 h. After rinsing, the parasites were incubated with goat anti-mouse IgG preadsorbed to 10 nm gold particles (ab27241 Abcam) for 30 min, washed and fixed in a mixture of 2% formalin and 2.5% glutaraldehyde in 0.1 M sodium cacodylate buffer for 30 min. After washing again, the parasites were post-fixed in 1% osmium tetroxide for 30 min, dehydrated in an ethanol series, embedded in epoxy resin and 60-nm ultrathin sections were cut on a Leica UC6 ultramicrotome, contrasted with uranyl acetate and lead citrate and examined on a 120 kV FEI Spirit Biotwin using a Tietz F4.16 CCD camera. The density of anti-IFX gold particle staining was determined by counting the number of gold particles per μm of membrane on both sagittal and transverse sections. Membrane lengths were determined using the segmented line function in ImageJ version 1.45s and a known image scaling factor. To assign dorsal and ventral sectors, a line was drawn across (transverse) or along (sagittal) the flagellum midpoint. Anti-IFX antibody selection and characterization To raise polyclonal antisera against IFX, the entire ectodomain of IFX was expressed and purified and injected into rabbits (Cambridge Research Biochemicals). The sera were purified on Hi-Trap Protein G HP columns (GE Healthcare) according to the manufacturer’s instructions. Hybridomas secreting monoclonal antibodies to IFX were selected using standard protocols, as previously described 40 . In brief, the SP2/0 myeloma cell line was grown in advanced DMEM/F12 medium (Invitrogen) supplemented with 20% fetal bovine serum, penicillin (100 U ml −1 ), streptomycin (100 μg ml −1 ) and l -glutamine (2 mM). Following spleen dissection and dissociation, 10 8 splenocytes were fused to 10 7 SP2/0 myeloma in 50% PEG (PEG 1500, Roche), using standard procedures. The resulting hybridomas were plated over ten 96-well plates and initially grown in advanced DMEM and F12 medium (Invitrogen) supplemented with 20% fetal bovine serum, penicillin (100 U ml −1 ), streptomycin (100 μg ml −1 ) and l -glutamine (2 mM) before addition of hypoxanthine–aminopterin–thymidine (HAT) selection medium 24 h after the fusion. After 11 days, hybridoma supernatants were collect to determine the presence of antibodies reacting to the IFX protein using an ELISA-based method, as previously described 40 . Seven wells (2H3, 3D12, 6B3, 8C9, 8E12, 8F10 and 10E2) containing hybridoma colonies secreting antibodies that reacted with IFX, but not to a control protein containing the same purification tags, were identified and cultured for a further four days in HAT-selection medium. Hybridoma cells from each of the positive wells were cloned by limiting dilution over two 96-well plates at a density of 0.5 cells per well and grown in HAT-free SP2/0 conditioned medium. Eleven days later, twelve wells corresponding to each of the seven clones were selected and tested again by ELISA for reactivity to the IFX protein. Hybridoma 8F10 was not successfully cloned but three positive wells from the remaining hybridomas were chosen for a second round of dilution cloning in the conditions described above. After a final test for reactivity to IFX, a single well from each of the positive clones was expanded and adapted to grow in Hybridoma-SFM serum-free medium (Thermo Fisher). To determine the location of the anti-IFX monoclonal antibody epitopes, subfragments of the IFX ectodomain corresponding to the boundaries of predicted secondary structure (M1–T251, M1–S472, S135–T251 and N442–S535) were designed, produced by gene synthesis and cloned into a mammalian expression plasmid with an enzymatically biotinylated C-terminal tag (Twist Biosciences). Biotinylated proteins were expressed as secreted recombinant proteins in HEK293 cells as described in ‘Vaccine target identification and expression’ and dialysed to remove free d -biotin. Biotinylated IFX fragments were immobilized on a streptavidin-coated plate and binding of the six mouse monoclonal antibodies was tested by ELISA and detected with an alkaline-phosphatase-conjugated anti-mouse secondary antibody (Sigma-Aldrich) as previously described 40 . Binding of a rabbit polyclonal antibody raised to the entire ectodomains of IFX (Cambridge Research Biochemicals) was used as a positive control for each of the subdomains, and detected with an alkaline-phosphatase-conjugated anti-rabbit secondary antibody (Jackson Immunoresearch). For affinity-purification of monoclonal antibodies from hybridoma culture supernatants, spent supernatants were supplemented with 0.1 M sodium acetate, pH 5.0 immediately before purification on a HiTrap Protein G HP 1 mL column (GE Healthcare) using an AKTA pure instrument. Elution was performed in 0.1 M glycine, pH 2.7 followed by immediate neutralization with 1 M Tris-HCl, pH 9.0. Purified antibodies were extensively dialysed against PBS and stored at 4 °C until use. To capture antibodies on streptavidin-coated sensor chips for biophysical interaction analysis, 300 μg of purified monoclonal antibodies were chemically biotinylated using a 20-fold molar excess of sulfo-NHS-biotin (ThermoFisher) for two hours at room temperature; to remove excess biotin the solutions were dialysed against 5 l PBS for 16 h. Antibody affinity by surface plasmon resonance Antibody affinities were determined by surface plasmon resonance (SPR) essentially as previously described 41 using a Biacore 8K instrument (GE Healthcare). To measure antibody interaction affinity rather than avidity, between 400 to 600 RU of biotinylated anti-IFX monoclonal antibodies were immobilized on a streptavidin-coated sensor chip prepared using the Biotin CAPture kit (GE Healthcare); a biotinylated mouse monoclonal antibody (OX68) was used as a non-binding control in the reference flow cell. The entire ectodomain of IFX was used as the analyte, which was first purified and resolved by size-exclusion chromatography on a Superdex 200 Increase 10/300 column (GE Healthcare) in HBS-EP (10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.05% v/v P20 surfactant) immediately before use in SPR experiments to remove any protein aggregates that might influence kinetic measurements. Increasing concentrations of twofold dilutions of the entire ectodomain of IFX as a soluble analyte were injected at 30 μl min −1 for a contact time of 120 s and dissociation of 600 s. Both kinetic and equilibrium binding data were analysed in the manufacturer’s Biacore 8K evaluation software version 1.1 (GE Healthcare) and plotted in Prism GraphPad version 8.0.2. All experiments were performed at 37 °C in HBS-EP. Antibody cloning, isotype switching, mutagenesis and purification To switch the isotype of the 8E12 anti-IFX monoclonal antibody from IgG1, it was first necessary to amplify the genes encoding the rearranged light and heavy variable regions from the hybridoma; this was performed essentially as previously described 40 . In brief, total RNAs were extracted from the cloned 8E12 hybridoma using the RNAqueous-micro total RNA isolation kit (Ambion) followed by reverse transcription with Superscript III (Thermo Fisher). PCR products encoding the rearranged heavy and light chain regions were individually amplified using sets of degenerate oligonucleotides and then assembled in a subsequent fusion PCR using a linker fragment to create a single PCR product containing both the rearranged light and heavy chains, as previously described 42 . The fusion PCR product was ligated using the NotI and AscI restriction sites into an expression plasmid obtained from Addgene (plasmid no. 114561) in frame with the mouse constant IgG2a heavy chain 43 . Competent Escherichia coli were transformed and purified plasmids used in small-scale transfections of HEK293 cells to identify those plasmids encoding functional antibodies as previously described 44 . To perturb the recruitment of immune effectors in the mouse IgG2a recombinant antibody and retain serum half-life, we mutated the C1q and FcR binding sites in the IgG2a constant heavy chain by site-directed mutagenesis as previusly described 17 . Mutation to the binding site of Fcγ receptors (ΔFcR) was achieved by introducing the L234A and L235A substitutions using primers FcR sense 5′-GCACCTAACGCTGCAGGTGGACCATCCG-3′ and FcR anti-sense 5′-TGGTCCACCTGCAGCGTTAGGTGCTGGGC-3′. To abrogate C1q binding (ΔC1q), a single amino-acid change P329A was introduced using primers C1q sense 5′-CAAAGACCTCGCTGCGCCCATCGAGAGAACC-3′ and C1q anti-sense 5′-GATGGCGCAGCGAGGTCTTTGTTGTTGACC-3′. In both cases, antibody mutagenesis was achieved by first amplifying 20 ng of an expression vector containing the mouse constant IgG2a heavy chain with each oligonucleotide separately for nine cycles (denaturation for 45 s at 94 °C; annealing for 40 s at 58 °C; elongation for 7 min and 30 s at 72 °C), using the KOD Hot Start DNA polymerase (Merck). Amplification reactions performed with complementary oligonucleotides were then mixed, 0.5 μl KOD Hot Start DNA polymerase was added to the reaction, and the amplification was resumed for a further 18 cycles. At the end of the reaction, half of the PCR reaction was digested with 20 U DpnI enzyme (New England Biolabs), which specifically cleaves methylated strands from the parental plasmid, for 3 h at 37 °C before transforming 5 μl into TOP 10 chemically competent bacteria (Invitrogen). Mutations were confirmed in selected clones by DNA sequencing. To generate a double mutant lacking both the C1q and FcR binding sites (ΔC1qΔFcR), site-directed mutagenesis was performed on an expression plasmid containing the FcR mutation, using the set of oligonucleotides designed for C1q mutagenesis. Both single mutants and the double-mutant backbones were doubly digested with NotI and AscI restriction enzymes and the fusion PCR product encoding the variable regions of the 8E12 recombinant antibody cloned into them, plasmids purified and verified by sequencing. Antibodies were produced by transfecting HEK293 cells with plasmids encoding the recombinant 8E12–IgG2a monoclonal antibody with the wild-type IgG2a heavy chain, single mutants that lacked C1q and FcR binding, and the double mutant. Six days after transfection, the cell culture supernatant was collected and the recombinant antibodies were purified on a HiTrap Protein G HP 1-ml column, according to the manufacturer’s instructions as previously described 40 . Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this paper. Data availability Annotated T. vivax genome data were obtained from TriTrypDB ( ). All data generated or analysed during this study are included in this Article, and/or are available from the corresponding author (G.J.W., who can also be contacted at [email protected]) on reasonable request. Source data are provided with this paper.

Scientists have identified a promising vaccine candidate for a parasitic disease that causes a chronic wasting disease in livestock animals across sub Saharan Africa. The disease, called Nagana, threatens the livelihoods of millions of people who depend on animals for milk, food and draft power. It's hoped the research—from the University of York and the Wellcome Sanger Institute—will help alleviate the socioeconomic hardship of people living in some of the most deprived regions in the world. Deadly The disease is caused by trypanosomes: insect-borne parasites that live in the blood which—in humans—are responsible for a deadly disease called African sleeping sickness. Millions of people living in sub-Saharan countries are at risk of trypanosome infections and the economic impact of the livestock disease is vast, with around three million cattle dying every year costing billions of dollars in lost productivity. Trypanosomes have evolved sophisticated protective mechanisms that allow them to thrive in blood and have so far presented a barrier to develop vaccines—which would be an important tool to control these diseases. The researchers identified a vaccine candidate that can elicit sterile immunity to T. vivax—one of the trypanosome parasites responsible for nagana. Parasites Researchers used an approach called "reverse vaccinology" to find new potential vaccines. Professor Gavin Wright from the Department of Biology and the York Biomedical Research Institute said: "Vaccinating against trypanosomes has long been considered a low value option to tackle this awful disease which affects the lives and livelihoods of people living in some of the most deprived regions on the planet. Because of their protective molecular armory, developing a vaccine against these parasites seemed like trying to stop a tank with a peashooter. "Our discovery identifies a vaccine candidate for an important parasitic disease that has constrained the socioeconomic development of sub-Saharan African countries and provides evidence that highly protective vaccines against trypanosome infections can be achieved." Infection The research was in collaboration with Dr. Andrew Jackson at the University of Liverpool. Dr. Jackson said: "Vaccines against parasites are notoriously difficult to produce. For 40 years, researchers have used a trial-and-error approach to look for vaccine candidates that protect against infection, but there are thousands of possibilities. "The real strength of this research is its reverse vaccinology method. It considers all the options and hones in on the protective antigens, taking the luck out of the search. Discovery of this vaccine candidate is a fantastic breakthrough, it removes a long-standing obstacle to progress and will fundamentally change our approach to preventing livestock trypanosomiasis."

10.1038/s41586-021-03597-x

Physics

Novel algorithm proposed for inversion of aerosol optical depth

Yizhe Fan et al, Aerosol Retrieval Study from a Particulate Observing Scanning Polarimeter Onboard Gao-Fen 5B without Prior Surface Knowledge, Based on the Optimal Estimation Method, Remote Sensing (2023). DOI: 10.3390/rs15020385

https://dx.doi.org/10.3390/rs15020385

https://phys.org/news/2023-02-algorithm-inversion-aerosol-optical-depth.html

Abstract Guidelines Hypothesis Interesting Images Letter New Book Received Obituary Opinion Perspective Proceeding Paper Project Report Protocol Registered Report Reply Retraction Short Note Study Protocol Systematic Review Technical Note Tutorial Viewpoint All Article Types Advanced Search Section All Sections AI Remote Sensing Atmospheric Remote Sensing Biogeosciences Remote Sensing Coral Reefs Remote Sensing Discoveries in Remote Sensing Earth Observation Data Earth Observation for Emergency Management Ecological Remote Sensing Engineering Remote Sensing Environmental Remote Sensing Forest Remote Sensing Ocean Remote Sensing Remote Sensing and Geo-Spatial Science Remote Sensing Communications Remote Sensing Image Processing Remote Sensing in Agriculture and Vegetation Remote Sensing in Geology, Geomorphology and Hydrology Remote Sensing of the Water Cycle Satellite Missions for Earth and Planetary Exploration Urban Remote Sensing All Sections Special Issue All Special Issues Remote Sensing: 15th Anniversary 100 Years ISPRS - Advancing Remote Sensing Science 2nd Edition Advances in Remote Sensing for Archaeological Heritage 3D Point Clouds in Rock Mechanics Applications Accuracy Assessment of UAS Lidar Active-Passive Microwave Sensing for Earth System Parameters Advanced Communication and Networking Techniques for Remote Sensing Advanced Machine Learning and Big Data Analytics in Remote Sensing for Natural Hazards Management Advanced Machine Learning Approaches for Hyperspectral Data Analysis Advanced Radar Signal Processing and Applications Advanced Remote Sensing Methods for 3D Vegetation Mapping and Characterization Advanced Satellite-Terrestrial Networks Advanced Topics in Remote Sensing Advancement of Remote Sensing and UAS in Cartography and Visualisation Advancements in Remote Sensing and GIS in Mobile Computing and Location-Based Services Advances and Innovative Applications of Unmanned Aerial Vehicles Advances in Applications of Volunteered Geographic Information Advances in Climate and Geophysics Advances in Geographic Object-Based Image Analysis (GEOBIA) Advances in Mobile Laser Scanning and Mobile Mapping Advances in Object and Activity Detection in Remote Sensing Imagery Advances in Object and Activity Detection in Remote Sensing Imagery II Advances in Object-Based Image Analysis—Linking with Computer Vision and Machine Learning Advances in Quantitative Remote Sensing in China – In Memory of Prof. Xiaowen Li Advances in Real Aperture and Synthetic Aperture Ground-Based Interferometry Advances in Remote Sensing and Geographic Information Science and Their Uses in Geointelligence Advances in Remote Sensing Applications for the Detection of Biological Invasions Advances in Remote Sensing for Archaeological Heritage Advances in Remote Sensing in Coastal Geomorphology Advances in Remote Sensing of Agriculture Advances in Remote Sensing of Crop Water Use Estimation Advances in Remote Sensing of Forestry Advances in Remote Sensing of Wildland Fires Advances in Remote Sensing-based Disaster Monitoring and Assessment Advances in SAR: Sensors, Methodologies, and Applications Advances in Satellite Altimetry Advances in Satellite Altimetry and Its Application Advances in Synthetic Aperture Radar Remote Sensing Advances in the Remote Sensing of Terrestrial Evaporation Advances in VIIRS Data Advances of Remote Sensing in Environmental Geoscience Aerosol and Cloud Remote Sensing Airborne Laser Scanning Analysis of Remote Sensing Image Data Application of Machine Learning in Volcano Monitoring Application of Multi-Sensor Fusion Technology in Target Detection and Recognition Application of Remote Sensing in Hydrological Modeling and Watershed Management Applications and New Trends in Metrology for Radar/LiDAR-Based Systems Applications of Deep Learning in Smart Agriculture Applications of Full Waveform Lidar Applications of GNSS Reflectometry for Earth Observation Applications of Micro- and Nano-Satellites for Earth Observation Applications of Remote Sensing in Landscape Ecology in Latin America Applications of Remote Sensing in Rangelands Research Applications of Synthetic Aperture Radar (SAR) for Land Cover Analysis Approaches for Monitoring Land Degradation with Remote Sensing Archaeological Prospecting and Remote Sensing Artifacts in Remote Sensing Data Analysis: An Ecological Perspective Artificial Intelligence and Automation in Sustainable Smart Farming Artificial Neural Networks and Evolutionary Computation in Remote Sensing Assessment of Renewable Energy Resources with Remote Sensing Assimilation of Remote Sensing Data into Earth System Models Atmospheric Remote Sensing Baltic Sea Remote Sensing Big Data Analytics for Secure and Smart Environmental Services Big Earth Observation Data: From Cloud Technologies to Insights and Foresight Bistatic HF Radar Calibration and Validation of Synthetic Aperture Radar Calibration and Verification of Remote Sensing Instruments and Observations Carbon Cycle, Global Change, and Multi-Sensor Remote Sensing Cartography of the Solar System: Remote Sensing beyond Earth Celebrating the 50th Anniversary of the Moon Landing with Views of Earth from Satellite and Apollo Missions Citizen Science and Earth Observation Climate Modelling and Monitoring Using GNSS Close Range Remote Sensing Close-Range Remote Sensing by Ground Penetrating Radar Compact Polarimetric SAR Concurrent Positioning, Mapping and Perception of Multi-source Data Fusion for Smart Applications Contemporary Microwave and Radar Techniques in Remote Sensing—MIKON 2022, IRS 2022 CORINE Land Cover System: Limits and Challenges for Territorial Studies and Planning Cross-Calibration and Interoperability of Remote Sensing Instruments Cryospheric Remote Sensing Data Mining in Multi-Platform Remote Sensing Data Science in Remote Sensing Data Science, Artificial Intelligence and Remote Sensing Deep Learning for Intelligent Synthetic Aperture Radar Systems Deep Learning for Target Object Detection and Identification in Remote Sensing Data Deep Learning in Remote Sensing: Sample Datasets, Algorithms and Applications Dense Image Time Series Analysis for Ecosystem Monitoring Design and Calibration of Microwave Radiometers and Scatterometers for Remote Sensing of the Earth Designing and Managing the Next Generation of Transportation Infrastructure Digital Forest Resource Monitoring and Uncertainty Analysis Discovering A More Diverse Remote Sensing Discipline Drone-Based Ecological Conservation Earth Environment Monitoring with Advanced Spaceborne Synthetic Aperture Radar: New Architectures, Operational Modes, and Processing Techniques Earth Monitoring from A New Generation of Geostationary Satellites Earth Observation and Sustainable Development in Marine and Freshwater Systems Earth Observation for Ecosystems Monitoring in Space and Time Earth Observation for Water Resource Management in Africa Earth Observation in Planning for Sustainable Urban Development Earth Observation Technology Cluster: Innovative Sensor Systems for Advanced Land Surface Studies Earth Observation to Support Disaster Preparedness and Disaster Risk Management Earth Observation with AVHRR Data and Interconsistency Earth Observations for a Better Future Earth Earth Observations for Addressing Global Challenges Earth Observations for Geohazards Earth Observations for the Sustainable Development Ecogeomorphological Research Using Satellite Images Ecological Status and Change by Remote Sensing Environmental Research with Global Navigation Satellite System (GNSS) EO Solutions to Support Countries Implementing the SDGs ESA - NRSCC Cooperation Dragon 4 Final Results Estimation of Forest Biomass from SAR European Remote Sensing-New Solutions for Science and Practice Feature Papers Field Spectroscopy and Radiometry Fire Remote Sensing: Capabilities, Innovations, Opportunities and Challenges Frontiers in Spectral Imaging and 3D Technologies for Geospatial Solutions Fusion of High-Level Remote Sensing Products Fusion of LiDAR Point Clouds and Optical Images Fusion of Multi-Satellite and Multi-Sensor SARs Data for Investigating Long-Term Surface Ground Movements and Processes GEOBIA in the Era of Big EO Data Geological Remote Sensing Geomorphological Mapping and Process Monitoring Using Remote Sensing Geomorphological Processes and Natural Hazards Geospatial Statistics and Spatial/Spatiotemporal Analysis in Remote Sensing GIS and Remote Sensing advances in Land Change Science Global Croplands Global Navigation Satellite Systems for Earth Observing System Global Positioning Systems (GPS) and Applications GNSS High Rate Data for Research of the Ionosphere Google Earth Engine and Cloud Computing Platforms: Methods and Applications in Big Geo Data Science Google Earth Engine Applications Google Earth Engine: Cloud-Based Platform for Earth Observation Data and Analysis GPS/GNSS Contemporary Applications GPS/GNSS for Earth Science and Applications High Performance Computing in Remote Sensing High-precision GNSS: Methods, Open Problems and Geoscience Applications Human-Induced Global Change Hydrological Remote Sensing Hydrometeorological Prediction and Mapping Hyperspectral Imaging for Precision Farming Hyperspectral Remote Sensing Image Compression and Coding in Remote Sensing Indoor Localization Innovative Application of AI in Remote Sensing Innovative Remote Sensing for Monitoring and Assessment of Natural Resources InSAR for Earth Observation InSAR in Remote Sensing Instrumenting Smart City Applications with Big Sensing and Earth Observatory Data: Tools, Methods and Techniques Integrated Applications of Geo-Information in Environmental Monitoring Integrating Remote Sensing and Social Sensing Intelligence Computing Paradigms with Remote Sensing Networks in Water Hydrology Joint Artificial Intelligence and Computer Vision Applications in Remote Sensing Land Change Assessment Using Remote Sensing and Geographic Information Science Land Cover Classification Using Multispectral LiDAR Data Land Cover/Land Use Change (LC/LUC) – Causes, Consequences and Environmental Impacts in South/Southeast Asia Land Surface Fluxes Land Surface Global Monitoring from PROBA-V to Sentinel-3 Land Surface Processes and Interactions—From HCMM to Sentinel Missions and beyond Land Use / Land Cover Monitoring using SAR and Complementarity with Optical Sensors Landsat-8 Sensor Characterization and Calibration Laser Scanning in Forests Latest Advances in Remote Sensing-Based Environmental Dynamic Models Latest Developments in Clustering Algorithms for Hyperspectral Images Lessons Learned from the SPOT4 (Take5): Experiment in Preparation for Sentinel-2 LiDAR Lidar/Laser Scanning in Urban Environments Machine and Deep Learning for Earth Observation Data Analysis Machine Learning Applications in Earth Science Big Data Analysis Machine Learning Models from Regression to Deep Learning Neural Networks for Assessment, Prediction, Mitigation, and Control of Geospatial, Socioeconomic, and Environmental Impacts of Climate Change Machine Learning Techniques for Remote Sensing and Electromagnetic Applications Mapping Ecosystem Services Flows and Dynamics Using Remote Sensing Mapping the Dynamics of Forest Plantations in Tropical and Subtropical Regions from Multi-Source Remote Sensing Mapping, Monitoring and Impact Assessment of Land Cover/Land Use Changes in South and South East Asia Mathematical Models for Remote Sensing Image and Data Processing Microwave Indices from Active and Passive Sensors for Remote Sensing Applications Microwave Remote Sensing Mobile Laser Scanning Mobile Mapping Technologies Mobile Remote Monitoring Technology for Mental Health Modern Advances in Electromagnetic Imaging and Remote Sensing: Enabling Hardware, Computational Techniques and Machine Learning Monitoring and Modelling of Dynamics in Tropical Coastal Systems Monitoring Global Vegetation with AVHRR NDVI3g Data (1981-2011) Monitoring Land Surface Dynamic with AVHRR Data Monitoring of Land Changes Mountain Remote Sensing Multi-Constellation Global Navigation Satellite Systems: Methods and Applications Multi-Sensor and Multi-Data Integration in Remote Sensing Multi-Temporal Remote Sensing Multiscale and Multitemporal High Resolution Remote Sensing for Archaeology and Heritage: From Research to Preservation Multistatic and Bistatic SAR: State of Affairs and Way Forward Multitemporal Land Cover and Land Use Mapping Natural Hazard Assessment and Disaster Management Using Remote Sensing New Frontiers of Multiscale Monitoring, Analysis and Modeling of Environmental Systems via Integration of Remote and Proximal Sensing Measurements New Insights in InSAR and GNSS Measurements New Outstanding Results over Land from the SMOS Mission New Perspectives of Remote Sensing for Archaeology New Sensors, Multi-Sensor Integration, Large Volumes: New Opportunities and Challenges in Forest Fire Research New Technologies for Earth Remote Sensing New Trends in Forest Fire Research Incorporating Big Data and Climate Change Modeling Novel Bistatic SAR Scattering Theory, Imaging Algorithms, and Applications Object Detection, Recognition and Identification Using Remote Sensing Technique Object-Based Image Analysis Observation of Lake–Air Interaction and Atmospheric Boundary Layer Process, Remote Sensing Inversion over the Pan-Third Pole Observations, Modeling, and Impacts of Climate Extremes Observing the Ocean’s Interior from Satellite Remote Sensing Ocean Remote Sensing Open Resources in Remote Sensing Opportunities and Challenges for Medium Resolution (hecta- to kilometric) Earth Observation of Land Surfaces in the Advent of European Proba-V and Sentinel-3 Missions Optical Remote Sensing of the Atmosphere OPTIMISE: Innovative Optical Tools for Proximal Sensing of Ecophysiological Processes Photogrammetry and Image Analysis in Remote Sensing Point Cloud Processing and Analysis in Remote Sensing Polar Sea Ice: Detection, Monitoring and Modeling Positioning and Navigation in Remote Sensing Precise Orbit Determination with GNSS Processing of Bi-static, Geo-Synchronous and Multi-Satellite SAR Constellation Data Progress on the Use of UAS Techniques for Environmental Monitoring Proximal and Remote Sensing in the MWIR and LWIR Spectral Range Quantifying and Validating Remote Sensing Measurements of Chlorophyll Fluorescence Quantifying the Environmental Impact of Forest Fires Quantitative Inversion and Validation of Satellite Remote Sensing Products Quantitative Remote Sensing of Land Surface Variables Quantitative Remote Sensing Product and Validation Technology Quantitative Volcanic Hazard Assessment and Uncertainty Analysis in Satellite Remote Sensing and Modeling Quantitative Volcanic Hazard Assessment and Uncertainty Analysis in Satellite Remote Sensing and Modeling: Part II Radar Applications in Cultural Heritage Radar High-Speed Target Detection, Tracking, Imaging and Recognition - 2nd Edition Radar Remote Sensing for Agriculture Radar Remote Sensing of Cloud and Precipitation Radar Remote Sensing on Life Activities Radar Systems for the Societal Challenges Radiative Transfer Modelling and Applications in Remote Sensing Radio Frequency Interference (RFI) in Microwave Remote Sensing Rapid Processing and Analysis for Drone Applications Real-time GNSS Precise Positioning Service and its Augmentation Technology Real-Time Radar Imaging and Sensing Recent Advances in Geophysical Exploration and Monitoring on Coastal Areas Recent Advances in GPR Imaging Recent Advances in Pattern Recognition and Analysis in Landscape Ecology Recent Advances in Polarimetric SAR Interferometry Recent Advances in Remote Sensing for Crop Growth Monitoring Recent Advances in Subsurface Sensing Technologies Recent Advances in Thermal Infrared Remote Sensing Recent Developments in Remote Sensing for Physical Geography Recent Progress and Developments in Imaging Spectroscopy Recent Progress in Ground Penetrating Radar Remote Sensing Recent Trends in UAV Remote Sensing Regional and Global Land Cover Mapping Remote and Proximal Assessment of Plant Traits Remote Sensed Data and Processing Methodologies for 3D Virtual Reconstruction and Visualization of Complex Architectures Remote Sensing and Cyber Situational Awareness Remote Sensing and Decision Support for Precision Orchard Production Remote Sensing and GIS for Habitat Quality Monitoring Remote Sensing and Machine Learning Applications in Atmospheric Physics, Weather, and Air Quality Remote Sensing and Oil Spill Response: Leveraging New Technologies to Safeguard the Environment Remote Sensing and Vegetation Mapping Remote Sensing Application for Promoting Ecosystem Services and Land Degradation Management in Mid-Latitude Ecotone (MLE) Remote Sensing Applications in Monitoring of Protected Areas Remote Sensing Applications to Human Health Remote Sensing Applied to Soils: From Ground to Space Remote Sensing as Tool in Geofluids Dynamics and Related Risks Remote Sensing Big Data: Theory, Methods and Applications Remote Sensing by Synthetic Aperture Radar Technology Remote Sensing Data Application, Data Reanalysis and Advances for Mesoscale Numerical Weather Models Remote Sensing Data Fusion and Applications Remote Sensing Dedicated to Geographical Conditions Monitoring Remote Sensing for 3D Urban Morphology Remote Sensing for Archaeological Heritage Preservation Remote Sensing for Coral Reef Monitoring Remote Sensing for Cultural Heritage Remote Sensing for Environment and Disaster Remote Sensing for EU Habitats Directive Application Remote Sensing for Improved Understanding of Land Surface, Hydrology, and Water Quality Remote sensing for Intelligent Transportation Systems Remote Sensing for Land Cover/Land Use Mapping at Local and Regional Scales Remote Sensing for Land Cover/Land Use Mapping at Local and Regional Scales: Part II Remote Sensing for Land Surface Temperature (LST) Estimation, Generation, and Analysis Remote Sensing for Landslide Monitoring, Mapping and Modeling Remote Sensing for Landslides Investigation: From Research into Practice Remote Sensing for Maritime Safety and Security Remote Sensing for Monitoring Wildlife and Habitat in a Changing World Remote Sensing for Near-Real-Time Disaster Monitoring Remote Sensing for Public Health Remote Sensing for Sustainable Energy Systems Remote Sensing for the Definition and Near-Real Time Monitoring of Meteorological Extremes Remote Sensing for Understanding Coral Reef Dynamics and Processes: Photo-Systems to Coral Reef Systems Remote Sensing for Water Productivity Assessments in Agriculture, Ecosystems and Water Resources Remote Sensing for Wetland Inventory, Mapping and Change Analysis Remote Sensing from Unmanned Aerial Vehicles (UAVs) Remote Sensing in 2021 Remote Sensing in 2023 Remote Sensing in Applications of Geoinformation Remote Sensing in Climate Modeling Remote Sensing in Climate Monitoring and Analysis Remote Sensing in Coastal Ecosystem Remote Sensing in Coastal Environments Remote Sensing in Dryland Assessment and Monitoring Remote Sensing in Ecosystem Modelling Remote Sensing in Flood Monitoring and Management Remote Sensing in Food Production and Food Security Remote Sensing in Geology Remote Sensing in Geomorphology Remote Sensing in Natural and Cultural Heritage Remote Sensing in Precision Agriculture Remote Sensing in Public Health Remote Sensing in Sea Ice Remote Sensing in Seismology Remote Sensing in Support of Aeolian Research Remote Sensing in Support of Environmental Governance Remote Sensing in Support of Environmental Policy Remote Sensing in the Age of Electronic Ecology Remote Sensing in Tibet and Siberia Remote Sensing in University of Warsaw: Celebrating 60th Anniversary on Remote Sensing Remote Sensing Measurements for Monitoring Achievement of the Sustainable Development Goals (SDGs) Remote Sensing Monitoring Aerosols and Its Effects on Atmospheric Radiation Remote Sensing Observations of the Giant Planets Remote Sensing of Aerosol - Cloud Interactions Remote Sensing of Arctic Tundra Remote Sensing of Atmosphere and Underlying Surface Using OLCI and SLSTR on Board Sentinel-3: Calibration, Algorithms, Geophysical Products and Validation Remote Sensing of Biodiversity Remote Sensing of Biodiversity Monitoring Remote Sensing of Biogeochemical Cycles Remote Sensing of Biological Diversity Remote Sensing of Biomass Burning Remote Sensing of Biophysical Parameters Remote Sensing of Changing Northern High Latitude Ecosystems Remote Sensing of Desert Landscapes to Monitor Impacts of Renewable Energy Developments Remote Sensing of Desertification Remote Sensing of Drought Monitoring Remote Sensing of Dryland Environment Remote Sensing of Ecosystem Structure and Function Dynamics Due to Climate Change and Human Activities Remote Sensing of Environmental Changes in Cold Regions Remote Sensing of Essential Climate Variables and Their Applications Remote Sensing of Floodpath Lakes and Wetlands Remote Sensing of Forest Health Remote Sensing of Geo-Hydrological Process in an Arid Region Remote Sensing of Geopolitics Remote Sensing of Glaciers Remote Sensing of Grassland Ecosystem Remote Sensing of Human-Environment Interactions along the Urban-Rural Gradient Remote Sensing of Hydro-Meteorology Remote Sensing of Image Pansharpening Remote Sensing of Islands Remote Sensing of Land Degradation and Drivers of Change Remote Sensing of Land Degradation in Drylands Remote Sensing of Land Surface Radiation Budget Remote Sensing of Mangroves: Observation and Monitoring Remote Sensing of Night Lights – Beyond DMSP Remote Sensing of Peatlands I Remote Sensing of Phytoplankton Remote Sensing of Rainfall and Snowfall - Recent Advances Remote Sensing of Regional Soil Moisture Remote Sensing of Solar Surface Radiation Remote Sensing of the Oceans: Blue Economy and Marine Pollution Remote Sensing of Tropical Environmental Change Remote Sensing of Urban Ecology Remote Sensing of Vegetation Proportion, Attribute, Condition, and Change Remote Sensing of Vegetation Structure and Dynamics Remote Sensing of Water Quality Remote Sensing of Water Resources Remote Sensing of Wildfire Remote Sensing of Wind Energy Remote Sensing on Earth Observation and Ecosystem Services Remote Sensing on Land Surface Albedo Remote Sensing with Nighttime Lights Remote Sensing-Based Proxies to Predict Socio-Economic and Demographic Data Remote Sensing: 10th Anniversary Remote Sensing: A Themed Issue in Honor of Professor John R. Jensen Remotely Sensed Estimates of Fire Radiative Energy Reproducibility and Replicability in Remote Sensing Workflows Retrieval and Validation of Trace Gases Using Remote Sensing Measurements Retrieval, Validation and Application of Satellite Soil Moisture Data Root Dynamics Tracking Using Remote Sensing SAR for Forest Mapping SAR Imagery for Landslide Detection and Prediction Satellite Altimetry for Earth Sciences Satellite Climate Data Records and Applications Satellite Derived Bathymetry Satellite Hydrological Data Products and Their Applications Satellite Mapping Technology and Application Satellite Monitoring of Water Quality and Water Environment Satellite Remote Sensing Applications for Fire Management Satellite Remote Sensing for Water Resources in a Changing Climate Satellite Remote Sensing of High-Temperature Thermal Anomalies Satellite Remote Sensing of High-Temperature Thermal Anomalies, Volume II Satellite-Derived Wind Observations Science of Landsat Analysis Ready Data Sea Ice Remote Sensing and Analysis Selected Papers from IGL-1 2018 — First International Workshop on Innovating GNSS and LEO Occultations & Reflections for Weather, Climate and Space Weather Selected Papers from Remote Sensing Science Workshop Selected Papers from the "2019 International Symposium on Remote Sensing" Selected Papers from the “International Symposium on Remote Sensing 2018” Selected Papers from the “International Symposium on Remote Sensing 2021” Selected Papers from the 1st International Electronic Conference on Remote Sensing (ECRS-1) Selected Papers from the 2nd International Electronic Conference on Remote Sensing (ECRS-2) Selected Papers from the 5th International Electronic Conference on Remote Sensing Selected Papers from the First "Symposia of Remote Sensing Applied to Soil Science", as part of the "21st World Congress of Soil Science, 2018" Selected Papers from the Polarimetric Interferometric SAR Workshop 2019 Semantic Segmentation Algorithms for 3D Point Clouds She Maps Silk-Road Disaster Monitoring and Risk Assessment Using Remote Sensing and GIS Societal and Economic Benefits of Earth Observation Technologies Space Geodesy and Ionosphere Space-Borne Gravimetric Measurements for Quantifying Earth System Mass Change Spatial Data Infrastructures for Big Geospatial Sensing Data Spatial Demography and Health – The 1st Internaitonal Symposium on Lifecourse Epidemiology and Spatial Science (ISLES) Spatial Enhancement of Hyperspectral Data and Applications Spatial Modelling of Natural Hazards and Water Resources through Remote Sensing, GIS and Machine Learning Methods Spatio-Temporal Mobility Data State-of-the-Art Remote Sensing in North America 2019 State-of-the-Art Remote Sensing in North America 2023–2024 State-of-the-Art Remote Sensing in South America State-of-the-Art Technology of Remote Sensing in Russia Synthetic Aperture Radar (SAR) Technology Advance and for Cultural Heritage Applications Synthetic Aperture Radar (SAR)—New Techniques, Missions and Applications Systems and Technologies for Remote Sensing Application through Unmanned Aerial Systems (STRATUS) Teaching and Learning in Remote Sensing Technological Developments of Unmanned Aerial Systems (UAS) for Remote Sensing Applications Ten Years of Remote Sensing at Barcelona Expert Center Ten Years of TerraSAR-X—Scientific Results Terrestrial Laser Scanning The Application of Thermal Urban Remote Sensing to Understand and Monitor Urban Climates The Development and Validation of Remote Sensing Products for Terrestrial, Hydrological, and Ecological Applications at the Regional Scale The Environmental Mapping and Analysis Program (EnMAP) Mission: Preparing for Its Scientific Exploitation The Internet of Things (IoT) in Remote Sensing: Opportunities and Challenges The Kyoto and Carbon Initiative—Environmental Applications by ALOS-2 PALSAR-2 The Role of Remote Sensing in Sustainable Renewable Energy Exploitation The Use of Earth Observations for Exposure Assessment in Epidemiological Studies The Use of UAV Platforms for Cultural Heritage Monitoring and Surveying Themed Issue: Practical or Economic Applications of Optical/Thermal Remote Sensing: A Themed Issue in Honor of Professor Toby N. Carlson Thermal Remote Sensing Applications for the Atmospheric Surface Layer Thermal Remote Sensing Applications: Present Status and Future Possibilities Time Series Analysis in Remote Sensing: Algorithm Development and Applications Time-of-Flight Range-Imaging Cameras Towards Practical Application of Artificial Intelligence in Remote Sensing Towards Remote Long-Term Monitoring of Wetland Landscapes Trends in UAV Remote Sensing Applications Trends in UAV Remote Sensing Applications: Part II Two Decades of MODIS Data for Land Surface Monitoring: Exploitation of Existing Products and Development of New Algorithms UAV and IoT-based Measurement of Vegetation Indices for Environmental Monitoring UAV Imagery for Precision Agriculture UAV Photogrammetry and Remote Sensing UAV-Based Remote Sensing Methods for Modeling, Mapping, and Monitoring Vegetation and Agricultural Crops Uncertainties in Remote Sensing Understanding Animal-Ecosystem Interactions with Remote Sensing Underwater 3D Recording & Modelling Underwater Acoustic Remote Sensing Unmanned Aerial Vehicle Applications in Cryospheric Sciences Unmanned Aerial Vehicles (UAVs) based Remote Sensing Urban Remote Sensing Use of LiDAR and 3D point clouds in Geohazards Validation and Inter-Comparison of Land Cover and Land Use Data Validation on Global Land Cover Datasets Visible Infrared Imaging Radiometers and Applications Volcanic Processes Monitoring and Hazard Assessment Using Integration of Remote Sensing and Ground-Based Techniques Volcano Remote Sensing Water Optics and Water Colour Remote Sensing What can Remote Sensing Do for the Conservation of Wetlands? All Special Issues Volume Issue Number Page Logical Operator Operator AND OR Search Text Search Type All fields Title Abstract Keywords Authors Affiliations Doi Full Text

A research team led by Prof. Sun Xiaobing from the Anhui Institute of Optics and Fine Mechanics, Hefei Institutes of Physical Science (HFIPS), Chinese Academy of Sciences (CAS), has proposed an optimal inversion algorithm based on combined utilization of multi-band intensity and polarization information. This algorithm can meet the requirements of single-angle and multi-band polarization aerosol detection. The study was published in Remote Sensing. Aerosol optical depth (AOD) is used to characterize the extinction effect of aerosol on solar radiation, which plays an important role in remote sensing atmospheric correction and fine particulate pollution assessment. The proposed algorithm does not need prior information of the ground. It uses polarization information of short-wave infrared band to separate ground and atmospheric information, and then uses scalar information to obtain the final result. "By decoupling the surface and atmosphere, our method can avoid inversion error and spatio-temporal matching error caused by the late updating of the surface reflectance database," said Prof. Sun. The researchers used the observation data of a high-precision polarization scanner (POSP) onboard the hyperspectral observation satellite (GF-5B) to verify the algorithm. "Compared with AOD products of aerosol robotic network stations in different regions, the algorithm can retrieve AOD over different surfaces," said Prof. Sun. "Another advantage is that the effectiveness of the algorithm under different pollution conditions is verified with Moderate-Resolution Imaging Spectroradiomete (MODIS) AOD products."

10.3390/rs15020385

Medicine

Researchers discover DNA variants significantly influence body fat distribution

Protein-coding variants implicate novel genes related to lipid homeostasis contributing to body-fat distribution, Nature Genetics (2019). DOI: 10.1038/s41588-018-0334-2 Journal information: Nature Genetics

http://dx.doi.org/10.1038/s41588-018-0334-2

https://medicalxpress.com/news/2019-02-dna-variants-significantly-body-fat.html

Abstract Body-fat distribution is a risk factor for adverse cardiovascular health consequences. We analyzed the association of body-fat distribution, assessed by waist-to-hip ratio adjusted for body mass index, with 228,985 predicted coding and splice site variants available on exome arrays in up to 344,369 individuals from five major ancestries (discovery) and 132,177 European-ancestry individuals (validation). We identified 15 common (minor allele frequency, MAF ≥5%) and nine low-frequency or rare (MAF <5%) coding novel variants. Pathway/gene set enrichment analyses identified lipid particle, adiponectin, abnormal white adipose tissue physiology and bone development and morphology as important contributors to fat distribution, while cross-trait associations highlight cardiometabolic traits. In functional follow-up analyses, specifically in Drosophila RNAi-knockdowns, we observed a significant increase in the total body triglyceride levels for two genes ( DNAH10 and PLXND1 ). We implicate novel genes in fat distribution, stressing the importance of interrogating low-frequency and protein-coding variants. Main Central body-fat distribution, as assessed by waist-to-hip ratio (WHR), is a heritable and a well-established risk factor for adverse metabolic outcomes 1 , 2 , 3 , 4 , 5 , 6 . Lower values of WHR are associated with lower risk of cardiometabolic diseases such as type 2 diabetes 7 , 8 , or differences in bone structure and gluteal muscle mass 9 . These epidemiological associations are consistent with our previously reported genome-wide association study (GWAS) results of 49 loci associated with WHR (after adjusting for body mass index, WHRadjBMI) 10 . Notably, genetic predisposition to higher WHRadjBMI is associated with increased risk of type 2 diabetes and coronary heart disease (CHD), which seems to be causal 9 . Recently, large-scale studies have identified roughly 125 common loci for multiple measures of central obesity, primarily non-coding variants of relatively modest effect 10 , 11 , 12 , 13 , 14 , 15 , 16 . Large-scale interrogation of coding and splice site variants, including both common (minor allele frequency, MAF ≥5%) and low-frequency or rare variants (LF/RV: MAF <5%), may lead to additional insights into the etiology of central obesity. Here, we identify and characterize such variants associated with WHRadjBMI using ExomeChip array genotypes. Results Protein-coding and splice site variation associations We conducted a two-stage fixed-effects meta-analysis testing additive and recessive models to detect protein-coding genetic variants that influence WHRadjBMI ( Methods and Fig. 1 ). Stage 1 included up to 228,985 variants (218,195 low-frequency or rare) in up to 344,369 individuals from 74 studies of European, South and East Asian, African and Hispanic/Latino descent individuals (Supplementary Data 1 – 3 ). Stage 2 assessed 70 suggestive ( P < 2 × 10 −6 ) stage 1 variants in two cohorts, UK Biobank (UKBB) and deCODE for a total stage 1 and 2 sample size of 476,546 (88% European). Of the 70 variants considered, two common and five low-frequency or rare were not available in stage 2 (Tables 1 and 2 and Supplementary Data 4 – 6 ). Variants are considered novel and statistically significant if they were greater than one megabase (Mb) from a previously identified WHRadjBMI SNP 10 , 11 , 12 , 13 , 14 , 15 , 16 and achieve array-wide significance ( P < 2 × 10 −7 , stage 1 and 2). Fig. 1: Summary of meta-analysis study design and workflow. Abbreviations: EUR, European, AFR, African, SAS, South Asian, EAS, East Asian and HIS, Hispanic/Latino ancestry. Full size image Table 1 Association results for combined sexes. Association results based on an additive or recessive model for coding variants that met array-wide significance ( P < 2 × 10 −7 ) in the sex-combined meta-analyses Full size table Table 2 Association results for sex-stratified analyses. Association results based on an additive or recessive model for coding variants that met array-wide significance ( P < 2 × 10 −7 ) in the sex-specific meta-analyses and reach Bonferroni corrected P value for sex heterogeneity ( P sexhet < 7.14 × 10 −4 ) Full size table In our primary meta-analysis, including all stage 1 and 2 samples, we identified 48 coding variants (16 novel) across 43 genes, 47 assuming an additive model and one under a recessive model (Table 1 and Supplementary Figs. 1 – 4 ). Due to possible heterogeneity, we also performed European-only meta-analysis. Here, four additional coding variants were significant (three novel) assuming an additive model (Table 1 and Supplementary Figs. 5 – 8 ). Of these 52 significant variants, 11 were low-frequency or rare and displayed larger effect estimates than many previously reported common variants 10 , including seven novel variants in RAPGEF3 , FGFR2, R3HDML, HIST1H1T, PCNXL3, ACVR1C and DARS2 . Variants with MAF ≤1% had effect sizes approximately three times greater than those of common variants (MAF ≥5%). Despite large sample size, we cannot rule out the possibility that additional low-frequency or rare variants with smaller effects exist (see estimated 80% power in Fig. 2 ). However, in the absence of common variants with similarly large effects, our results point to the importance of investigating LF/RVs. Fig. 2: Minor allele frequency compared to estimated effect. This scatter plot displays the relationship between MAF and the estimated effect (β) for each significant coding variant in our meta-analyses. All novel WHRadjBMI variants are highlighted in orange, and variants identified only in models that assume recessive inheritance are denoted by diamonds and only in sex-specific analyses by triangles. Eighty percent power was calculated based on the total sample size in the stage 1 and 2 meta-analysis and P = 2 × 10 −7 . Estimated effects are shown in original units (cm cm −1 ) calculated by using effect sizes in s.d. units times SD of WHR in the ARIC study (sexes combined = 0.067, men = 0.052, women = 0.080). Full size image Given established sex differences in the genetic underpinnings of WHRadjBMI 10 , 11 , we also performed sex-stratified analyses. We detected four additional novel variants that exhibit significant sex-specific effects ( P sexhet < 7.14 × 10 −4 , Methods ) in UGGT2 and MMP14 for men; and DSTYK and ANGPTL4 for women (Table 2 and Supplementary Figs. 9 – 15 ); including low-frequency or rare in UGGT2 and ANGPTL4 (MAF men = 0.6% and MAF women = 1.9%, respectively). Additionally, 14 variants from the sex-combined meta-analyses displayed significantly stronger effects in women, including the novel, low-frequency ACVR1C variant (rs55920843, MAF = 1.1%). Overall, 19 of the 56 variants (32%) identified across all meta-analyses (48 from all ancestry, 4 from European-only and 4 from sex-stratified analyses) showed significant sex-specific effects on WHRadjBMI: 16 variants with significantly stronger effects in women and three in men (Fig. 1 ). In summary, we identified 56 array-wide significant coding variants ( P < 2.0 × 10 −7 ); 43 common (14 novel) and 13 low-frequency or rare (9 novel). For the 55 significant variants from the additive model, we examined potential collider bias 17 , 18 ( Methods , Supplementary Table 1 and Supplementary Note ). Overall, 51 of 55 variants were robust to collider bias 17 , 18 . Of these, 25 variants were nominally associated with BMI ( P BMI < 0.05), yet effect sizes changed little after correction for potential biases (15% change in effect estimate on average). For four of the 55 SNPs (rs141845046, rs1034405, rs3617, rs9469913), attenuation following correction was noted ( P corrected > 9 × 10 −4 , 0.05/55), including one novel variant, rs1034405 in C3orf18 , demonstrating a possible overestimation of these effects in the current analysis. Using stage 1 results, we then aggregated LF/RVs across genes and tested their joint effect with SKAT and burden tests 19 (Supplementary Table 2 and Methods ). None of the five genes that reached array-wide significance ( P < 2.5 × 10 −6 , 0.05/16,222 genes tested: RAPGEF3 , ACVR1C , ANGPTL4 , DNAI1 and NOP2 ) remained significant after conditioning on the most significant single variant. Conditional analyses We next implemented conditional analyses to determine (1) the total number of independent signals identified and (2) whether the 33 variants near known GWAS signals (<+/−1 Mb) represent independent novel associations. We used approximate joint conditional analyses to test for independence in stage 1 ( Methods and Supplementary Table 3 ) 19 . Only the RSPO3-KIAA0408 locus contains two independent variants 291 kilo bases apart, rs1892172 in RSPO3 (MAF = 46.1%, P conditional = 4.37 × 10 −23 in the combined sexes and P conditional = 2.4 × 10 −20 in women) and rs139745911 in KIAA0408 (MAF = 0.9%, P conditional = 3.68 × 10 −11 in combined sexes and P conditional = 1.46 × 10 −11 in women, Supplementary Table 3 and Fig. 3a ). For the 33 variants within 1 Mb of previously identified WHRadjBMI SNPs, sex-combined conditional analyses identified one coding variant representing a novel independent signal in a known locus ( RREB1 ; stage 1 meta-analysis, rs1334576, MAF = 44%, P conditional = 3.06 × 10 −7 , Supplementary Table 4 and Fig. 3b ; UKBB analysis, P conditional = 1.24 × 10 −8 , Supplementary Data 7 ). Fig. 3: Regional association plots for known loci with novel coding signals identified by conditional analyses. Point color reflects r 2 calculated from the ARIC dataset. a , There are two independent variants in RSPO3 and KIAA0408 , based on results from the stage 1 all ancestry women ( N = 180,131 for RSPO3 and 139,056 for KIAA0408 ). b , We have a variant in RREB1 that is independent of the GWAS variant rs1294410, based on results from the stage 1 all ancestry sex-combined individuals ( N = 319,090; GWAS signal tagging variant rs1294421, rs1294410-rs1294421, r 2 =0.85, 1000 Genomes Phase 3 version 5 all ancestries reference set). Full size image In summary, we identified 56 WHRadjBMI-associated coding variants in 41 independent association signals, 24 of which are new or independent of known GWAS-identified tag SNPs (either >+/−1 Mb or array-wide significant following conditional analyses) (Fig. 1 ). Thus, we identified 15 common and 9 low-frequency or rare novel and independent variants following conditional analyses. Gene set and pathway enrichment analysis To determine whether significant coding variants highlight novel or previously identified biological pathways, we applied two complementary methods, EC-DEPICT (ExomeChip Data-driven Expression-Prioritized Integration for Complex Traits) 20 , 21 and PASCAL 22 ( Methods ). For PASCAL all variants were used, for EC-DEPICT we examined only 361 variants with suggestive significance ( P < 5 × 10 −4 ) 10 , 23 from the all ancestries combined sexes analysis (which after clumping and filtering became 101 lead variants in 101 genes). We separately analyzed variants that exhibited significant sex-specific effects ( P sexhet < 5 × 10 −4 ). The sex-combined analyses identified 49 significantly enriched gene sets (false discovery rate, FDR < 0.05) that grouped into 25 meta-gene sets ( Supplementary Note and Supplementary Data 8 , 9 ). We noted a cluster of meta-gene sets with direct relevance to metabolic aspects of obesity (‘enhanced lipolysis’, ‘abnormal glucose homeostasis’, ‘increased circulating insulin level’ and ‘decreased susceptibility to diet-induced obesity’); we observed two significant adiponectin-related gene sets in these meta-gene sets. While these pathway groups had previously been identified in the GWAS DEPICT analysis (Fig. 4 ), many of the individual gene sets in these meta-gene sets were not significant in the previous GWAS analysis, such as ‘insulin resistance,’ ‘abnormal white adipose tissue physiology’ and ‘abnormal fat cell morphology’ (Supplementary Data 8 , Fig. 4 and Supplementary Fig. 16a ), but represent similar biological underpinnings implied by the shared meta-gene sets. Despite their overlap with the GWAS results, these analyses highlight novel genes that fall outside known GWAS loci and with strong contributions to the substantially enriched gene sets related to adipocyte and insulin biology (for example, MLXIPL, ACVR1C and ITIH5) (Fig. 4 ). Fig. 4: Heat maps showing DEPICT gene set enrichment results from the stage 1 all ancestry sex-combined individuals ( N = 344,369). For any given square, the color indicates how strongly the corresponding gene ( x axis) is predicted to belong to the reconstituted gene set ( y axis). This value is based on the gene’s z -score for gene set inclusion in DEPICT’s reconstituted gene sets, where red indicates a higher and blue a lower z -score. To visually reduce redundancy and increase clarity, we chose one representative ‘meta-gene set’ for each group of highly correlated gene sets based on affinity propagation clustering ( Methods and Supplementary Note ). Heatmap intensity and DEPICT P values (Supplementary Data 8 – 9 ) correspond to the most substantially enriched gene set in the meta-gene set. Annotations for the genes indicate (1) the minor allele frequency of the significant ExomeChip (EC) variant (blue; if multiple variants, the lowest-frequency variant was kept), (2) whether the variant’s P value reached array-wide significance (<2 × 10 −7 ) or suggestive significance (<5 × 10 −4 ) (shades of purple), (3) whether the variant was novel, overlapping ‘relaxed’ GWAS signals from Shungin et al. 10 (GWAS P < 5 × 10 −4 ), or overlapping ‘stringent’ GWAS signals (GWAS P < 5 × 10 −8 ) (pink) and (4) whether the gene was included in the gene set enrichment analysis or excluded by filters (shades of brown/orange) ( Methods and Supplementary Note ). Annotations for the gene sets indicate if the meta-gene set was found significant (shades of green; FDR < 0.01, < 0.05, or not significant) in the DEPICT analysis of GWAS results from Shungin et al. 10 . Full size image Also, we conducted pathway analyses after excluding variants from previous WHRadjBMI analyses 10 ( Supplemental Note ). Seventy-five loci/genes were included in the EC-DEPICT analysis, and we identified 26 substantially enriched gene sets (13 meta-gene sets). Here, all but one gene set, ‘lipid particle size’, were related to skeletal biology, probably reflecting an effect on the pelvic skeleton (hip circumference), shared signaling pathways between bone and fat (such as TGF-beta) and shared developmental origin 24 (Supplementary Data 9 and Supplementary Fig. 16b ). These previously identified GWAS DEPICT significant findings provide a fully independent replication of their biological relevance for WHRadjBMI. We used PASCAL ( Methods ) to further distinguish between enrichment based on coding-only variant associations (this study) and regulatory-only variant associations (up to 20 kb upstream of the gene from a previous GIANT study 10 ), finding 116 significantly enriched coding pathways (FDR < 0.05; Supplementary Data 10 ). We also compared the coding pathways to those identified in the total previous GWAS effort (using both coding and regulatory variants) identifying a total of 158 gene sets. Forty-two gene sets were enriched in both analyses, and we found high concordance in the -log10 ( P values) between ExomeChip and GWAS gene set enrichment (Pearson’s r (coding versus regulatory only) = 0.38, P < 10 −300 ; Pearson’s r (coding versus coding + regulatory) = 0.51, P < 10 −300 )). Nonetheless, some gene sets were enriched specifically for variants in coding regions (for example, decreased susceptibility to diet-induced obesity, abnormal skeletal morphology) or unique to variants in regulatory regions (for example, transcriptional regulation of white adipocytes) (Supplementary Fig. 17 ). The EC-DEPICT and PASCAL results showed a moderate but strongly significant correlation (for EC-DEPICT and the PASCAL max statistic, r = .28, P = 9.8 × 10 −253 ; for EC-DEPICT and the PASCAL sum statistic, r = 0.287, P = 5.42 × 10 −272 ). Common gene sets strongly implicate a role for skeletal biology, glucose homeostasis/insulin signaling and adipocyte biology (Supplementary Fig. 18 ). Cross-trait associations To assess the clinical relevance of our identified variants with cardiometabolic, anthropometric and reproductive traits, we conducted association lookups from existing ExomeChip studies of 15 traits (Supplementary Data 11 and Supplementary Fig. 19 ) 21 , 25 , 26 , 27 , 28 , 29 . Variants in STAB1 and PLCB3 displayed the greatest number of significant associations with seven different traits ( P < 9.8 × 10 −4 , 0.05/51 variants tested). Also, these two genes cluster together with RSPO3 , DNAH10 , MNS1 , COBLL1 , CCDC92 and ITIH3 . The WHR-increasing alleles in this cluster exhibit a previously described pattern of increased cardiometabolic risk (for example, increased fasting insulin, two-hour glucose (TwoHGlu)) and triglycerides and decreased high-density lipoprotein cholesterol (HDL)), but also decreased BMI 30 , 31 , 32 , 33 , 34 , 35 , 36 . The impact of central obesity may be causal, as a 1 s.d. increase in genetic risk of central adiposity was previously associated with higher total cholesterol, triglycerides, fasting insulin and TwoHGlu and lower HDL 9 . We conducted a search in the NHGRI-EBI GWAS Catalog 37 , 38 to determine whether our variants are in high linkage disequilibrium ( R 2 > 0.7) with variants associated with traits or diseases not covered by our cross-trait lookups (Supplementary Data 12 ). We identified several cardiometabolic traits (adiponectin, CHD and so on), diet/behavioral traits potentially related to obesity (carbohydrate, fat intake and so on), behavioral and neurological traits (schizophrenia, bipolar disorder and so on) and inflammatory or autoimmune diseases (Crohn’s Disease, multiple sclerosis and so on). Given the established correlation between total body-fat percentage and WHR of up to 0.483 39 , 40 , 41 , we examined the association of our top exome variants with both total body-fat percentage and truncal fat percentage available in a sub-sample of UKBB ( N = 118,160) (Supplementary Tables 5 and 6 ). Seven of the common novel variants were significantly associated ( P < 0.001, 0.05/48 variants examined) with both body-fat percentage and truncal fat percentage in the sexes-combined analysis ( COBLL1 , UHRF1BP1 , WSCD2 , CCDC92 , IFI30 , MPV17L2 , IZUMO1 ) and two with truncal fat percentage in women only ( EFCAB12 , GDF5) . Only rs7607980 in COBLL1 is near a known body-fat percentage GWAS locus (rs6738627; R 2 = 0.1989, distance, 6,751 base pairs, with our tag SNP) 42 . Of the nine SNPs associated with at least one of these two traits, all variants displayed much greater magnitude of effect on truncal fat percentage compared to body-fat percentage (Supplementary Fig. 20 ). Previous studies have demonstrated the importance of examining common and LF/RVs in genes with mutations known to cause monogenic diseases 43 , 44 . Thus, we assessed enrichment of WHRadjBMI variants in monogenic lipodystrophy and/or insulin resistance genes 43 , 44 (Supplementary Data 13 ). No significant enrichment was observed, possibly due in part to the small number of implicated genes and the relatively small number of variants in monogenic disease-causing genes (Supplementary Fig. 21 ). Genetic architecture of WHRadjBMI coding variants We used summary statistics from our stage 1 primary meta-analysis results to estimate the phenotypic variance explained by subsets of variants across various significance thresholds ( P < 2 × 10 −7 to 0.2) and conservatively using only independent SNPs (Supplementary Table 7 , Methods and Supplementary Fig. 22 ). For only independent coding variants that reached suggestive significance in stage 1 ( P < 2 × 10 −6 ), 33 SNPs explain 0.38% of the variation. The 1,786 independent SNPs with a liberal threshold of P < 0.02 explain 13 times more variation (5.12%), however, these large effect estimates may be subject to winner’s curse. When considering all coding variants on the ExomeChip in combined sexes, 46 SNPs with a P < 2 × 10 −6 and 5,917 SNPs with a P < 0.02 explain 0.51 and 13.75% of the variance in WHRadjBMI, respectively. As expected given the design of the ExomeChip, most of the variance explained is attributable to rare and low-frequency coding variants. However, for LF/RVs, those that passed significance in stage 1 explain only 0.10% of the variance in WHRadjBMI. We also estimated variance explained for the same SNPs in women and men separately and observed a greater variance explained in women compared to men ( P RsqDiff < 0.002 = 0.05/21, Bonferroni-corrected threshold) at each significance threshold considered (differences ranged from 0.24 to 0.91%). We conducted penetrance analysis using the UKBB (both sexes combined, and men- and women only) to determine whether there is a significant accumulation of the minor allele in either the centrally obese or non-obese groups ( Methods ). Three rare variants (MAF ≤1%) with larger effect sizes (effect size >0.90) were included in the penetrance analysis using World Health Organization cut-offs for central obesity. Of these, one SNP (rs55920843- ACVR1C ; P sex-combined = 9.25 × 10 −5 ; P women = 4.85 × 10 −5 ) showed a statistically significant difference in the number of carriers and non-carriers of the minor allele in the combined and female-only analysis (sex-combined obese carriers, 2.2%; non-obese carriers, 2.6%; women obese carriers, 2.1%; non-obese women carriers, 2.6%; Supplementary Table 8 and Supplementary Fig. 23 ). Drosophila knockdown Considering the genetic evidence of adipose and insulin biology in determining body-fat distribution 10 , and the lipid signature of the variants described herein, we examined whole-body triglyceride levels in adult Drosophila , a model organism in which the fat body is an organ functionally analogous to mammalian liver and adipose tissue as triglycerides are the major source of fat storage 45 . Of the 51 genes harboring our 56 significantly associated variants, we identified 27 Drosophila orthologs for functional follow-up analyses. We selected genes with large changes in triglyceride levels (>20% increase or >40% decrease, as chance alone is not a probable cause for changes of this magnitude) from an existing large-scale screen with ≤2 replicates per knockdown strain 45 . Two orthologs, for PLXND1 and DNAH10 , met these criteria and were subjected to additional knockdown experiments with ≥5 replicates using tissue-specific drivers (fat body (cg-Gal4) and neuronal (elav-Gal4) specific RNAi-knockdowns) (Supplementary Table 9 ). A significant ( P < 0.025, 0.05/2 orthologs) increase in the total body triglyceride levels was observed in DNAH10 ortholog knockdown strains for both the fat body and neuronal drivers. Only the neuronal driver knockdown for PLXND1 produced a significant change in triglyceride storage. DNAH10 and PLXND1 both lie in previous GWAS-identified regions. Adjacent genes have been highlighted as probable candidates for the DNAH10 association region, including CCDC92 and ZNF664 based on expression quantitative trait locus (eQTL) evidence. Of note, rs11057353 in DNAH10 showed suggestive significance after conditioning on the known GWAS variants in nearby CCDC92 (sex-combined P conditional = 7.56 × 10 −7 ; women-only rs11057353 P conditional = 5.86 × 10 −7 , Supplementary Table 4 ) thus providing some evidence of multiple causal variants/genes underlying this signal. Further analyses are needed to determine whether the implicated coding variants from the current analysis are the putatively functional variants. eQTL lookups We examined the cis-association of variants with expression level of nearby genes in subcutaneous and visceral omental adipose, skeletal muscle and pancreas tissue from the Genotype-Tissue Expression (GTEx) 46 project, and assessed whether exome and eQTL associations implicated the same signal ( Methods and Supplementary Data 14 – 15 ). The lead exome variant was associated with expression level of the gene itself for DAGLB, MLXIPL, CCDC92, MAPKBP1, LRRC36 and UQCC1 . However, for MLXIPL, MAPKBP1 and LRRC36 , the lead variant is also associated with expression of additional nearby genes. At three additional loci, the lead exome variant is only associated with expression level of nearby genes ( HEMK1 at C3orf18; NT5DC2, SMIM4 and TMEM110 at STAB1/ITIH3 and C6orf106 at UHRF1BP1 ). Thus, although detected with a missense variant, these results are also consistent with a regulatory mechanism of effect, and the association signal may well be due to linkage disequilibrium with nearby regulatory variants. Some of the coding genes implicated by eQTL analyses are known to be involved in adipocyte differentiation or insulin sensitivity: for example, for MLXIPL , the encoded carbohydrate responsive element binding protein is a transcription factor, regulating glucose-mediated induction of de novo lipogenesis in adipose tissue and expression of its beta -isoform in adipose tissue is positively correlated with adipose insulin sensitivity 47 , 48 . For CCDC92 , the reduced adipocyte lipid accumulation on knockdown confirmed the involvement of its encoded protein in adipose differentiation 49 . Biological curation To investigate the possible functional role of the identified variants, we conducted thorough searches of the literature and publicly available bioinformatics databases (Supplementary Data 16 – 17 , Box 1 and Methods ). Many of our novel LF/RVs are in genes that are intolerant of non-synonymous mutations (for example, ACVR1C , DARS2 , FGFR2 ; ExAC Constraint Scores >0.5). Other coding variants lie in genes that are involved in glucose homeostasis (for example, ACVR1C , UGGT2 , ANGPTL4 ), angiogenesis ( RASIP1 ), adipogenesis ( RAPGEF3 ) and lipid biology ( ANGPTL4 , DAGLB ). Box 1 Genes of biological interest harboring WHR-associated variants PLXND1 (3:129284818, rs2625973, known locus). The major allele of a common non-synonymous variant in Plexin D1 (L1412V, MAF 26.7%) is associated with increased WHRadjBMI (β (s.e.m.) = 0.0156 (0.0024), P = 9.16 × 10 −11 ). PLXND1 encodes a semaphorin class 3 and 4 receptor gene, and therefore, is involved in cell to cell signaling and regulation of growth and development for a number of different cell and tissue types, including those in the cardiovascular system, skeleton, kidneys and the central nervous system 77 , 78 , 79 , 80 , 81 . Mutations in this gene are associated with Moebius syndrome 82 , 83 , 84 , 85 and persistent truncus arteriosus 79 , 86 . PLXND1 is involved in angiogenesis as part of the SEMA and VEGF signaling pathways 87 , 88 , 89 , 90 . PLXND1 was implicated in the development of type 2 diabetes through its interaction with SEMA3E in mice. SEMA3E and PLXND1 are upregulated in adipose tissue in response to diet-induced obesity, creating a cascade of adipose inflammation, insulin resistance and diabetes mellitus 81 . PLXND1 is highly expressed in adipose (both subcutaneous and visceral) (GTeX). PLXND1 is highly intolerant of mutations and therefore highly conserved (Supplementary Data 16 ). Last, our lead variant is predicted as damaging or possibly damaging for all algorithms examined (SIFT, Polyphen2/HDIV, Polyphen2/HVAR, LRT, MutationTaster). ACVR1C (2:158412701, rs55920843, novel locus). The major allele of a low-frequency non-synonymous variant in activin A receptor type 1C (rs55920843, N150H, MAF 1.1%) is associated with increased WHRadjBMI (β (s.e.m.) = 0.0652 (0.0105), P = 4.81 × 10 −10 ). ACVR1C , also called Activin receptor-like kinase 7 ( ALK7 ), encodes a type I receptor for TGFB (Transforming Growth Factor, Beta-1), and is integral for the activation of SMAD transcription factors; therefore, ACVR1C plays an important role in cellular growth and differentiation 64 , 65 , 66 , 67 , 68 , including adipocytes 68 . Mouse Acvr1c decreases secretion of insulin and is involved in lipid storage 69 , 72 , 73 , 91 . ACVR1C exhibits the highest expression in adipose tissue, but is also highly expressed in the brain (GTEx) 69 , 70 , 71 . Expression is associated with body fat, carbohydrate metabolism and lipids in both obese and lean individuals 70 . ACVR1C is moderately tolerant of mutations (ExAC constraint scores: synonymous, −0.86; non-synonymous, 1.25; LoF, 0.04 and Supplementary Data 16 ). Last, our lead variant is predicted as damaging for two of five algorithms examined (LRT and MutationTaster). FGFR2 (10:123279643, rs138315382, novel locus). The minor allele of a rare synonymous variant in Fibroblast Growth Factor Receptor 2 (rs138315382, MAF 0.09%) is associated with increased WHRadjBMI (β (s.e.m.) = 0.258 (0.049), P = 1.38 × 10 −07 ). The extracellular portion of the FGFR2 protein binds with fibroblast growth factors, influencing mitogenesis and differentiation. Mutations in this gene have been associated with many rare monogenic disorders, including skeletal deformities, craniosynostosis, eye abnormalities and LADD syndrome, as well as several cancers including breast, lung and gastric cancer. Methylation of FGFR2 is associated with high birth weight percentile 92 . FGFR2 is tolerant of synonymous mutations, but highly intolerant of missense and loss-of-function mutations (ExAC constraint scores: synonymous, −0.9; missense, 2.74; LoF, 1.0 and Supplementary Data 16 ). Last, this variant is not predicted to be damaging on the basis of any of the five algorithms tested. ANGPTL4 (19:8429323, rs116843064, novel locus). The major allele of a non-synonymous low-frequency variant in Angiopoietin Like 4 (rs116843064, E40K, EAF 98.1%) is associated with increased WHRadjBMI (β (s.e.m.) = 0.064 (0.011) P = 1.20 × 10 −09 ). ANGPTL4 encodes a glycosylated, secreted protein containing a C-terminal fibrinogen domain. The encoded protein is induced by peroxisome proliferation activators and functions as a serum hormone that regulates glucose homeostasis, triglyceride metabolism 93 , 94 and insulin sensitivity 95 . Angptl4-deficient mice have hypotriglyceridemia and increased lipoprotein lipase activity, while transgenic mice overexpressing Angplt4 in the liver have higher plasma triglyceride levels and decreased ipoprotein lipase activity 96 . The major allele of rs116843064 has been previously associated with increased risk of CHD and increased triglycerides 63 . ANGPTL4 is moderately tolerant of mutations (ExAC constraint scores: synonymous, 1.18; missense, 0.21; LoF,0.0 and Supplementary Data 16 ). Last, our lead variant is predicted damaging for four of five algorithms (SIFT, Polyphen 2/HDIV, Polyphen2/HVAR and MutationTaster). RREB1 (6:7211818, rs1334576, novel association signal). The major allele of a common non-synonymous variant in the Ras responsive element binding protein 1 (rs1334576, G195R, MAF = 56%) is associated with increased WHRadjBMI (β (s.e.m.) = 0.017 (0.002), P = 3.9 × 10 −15 ). This variant is independent of the previously reported GWAS signal in the RREB1 region (rs1294410; 6:6738752 10 ). The protein encoded by this gene is a zinc finger transcription factor that binds to RAS-responsive elements (RREs) of gene promoters. It has been shown that the calcitonin gene promoter contains an RRE and that the encoded protein binds there and increases expression of calcitonin, which may be involved in Ras/Raf-mediated cell differentiation 97 , 98 , 99 . The ras-responsive transcription factor RREB1 is a candidate gene for type 2 diabetes associated end-stage kidney disease 98 . This variant is highly intolerant to loss-of function (ExAC constraint score LoF, 1, Supplementary Data 16 ). DAGLB (7:6449496, rs2303361, novel locus). The minor allele of a common non-synonymous variant (rs2303361, Q664R, MAF 22%) in DAGLB (Diacylglycerol lipase beta) is associated with increased WHRadjBMI (β (s.e.m.) = 0.0136 (0.0025), P value = 6.24 × 10 −8 ). DAGLB encodes a diacylglycerol (DAG) lipase that catalyzes the hydrolysis of DAG to 2-arachidonoyl-glycerol, the most abundant endocannabinoid in tissues. In the brain, DAGL activity is required for axonal growth during development and for retrograde synaptic signaling at mature synapses (2-AG) 100 . The DAGLB variant, rs702485 (7:6449272, r 2 = 0.306 and D ’ = 1 with rs2303361) has been previously associated with HDL. Pathway analyses indicate a role in the triglyceride lipase activity pathway 101 . DAGLB is tolerant of synonymous mutations, but intolerant of missense and loss-of function mutations (ExAC constraint scores: synonymous, −0.76; missense, 1.07; LoF, 0.94 and Supplementary Data 16 ). Last, this variant is not predicted to be damaging by any of the algorithms tested. MLXIPL (7:73012042, rs35332062 and 7:73020337, rs3812316, known locus). The major alleles of two common non-synonymous variants (A358V, MAF = 12%; Q241H, MAF = 12%) in MLXIPL (MLX interacting protein like) are associated with increased WHRadjBMI (β (s.e.m.) = 0.02 (0.0033), P = 1.78 × 10 −9 ; β (s.e.m.) = 0.0213 (0.0034), P = 1.98 × 10 −10 ). These variants are in strong linkage disequilibrium ( r 2 = 1.00, D ’ = 1.00, 1000 Genomes CEU). This gene encodes a basic helix-loop-helix leucine zipper transcription factor of the Myc/Max/Mad superfamily. This protein forms a heterodimeric complex and binds and activates carbohydrate response element (ChoRE) motifs in the promoters of triglyceride synthesis genes in a glucose-dependent manner 74 , 75 . This gene is possibly involved in the growth hormone signaling pathway and lipid metabolism. The WHRadjBMI-associated variant rs3812316 in this gene has been associated with the risk of non-alcoholic fatty liver disease and coronary artery disease 74 , 102 , 103 . Furthermore, Williams–Beuren syndrome (an autosomal dominant disorder characterized by short stature, abnormal weight gain, various cardiovascular defects and mental retardation) is caused by a deletion of about 26 genes from the long arm of chromosome 7 including MLXIPL . MLXIPL is generally intolerant to variation, and therefore conserved (ExAC constraint scores: synonymous = 0.48, missense = 1.16, LoF = 0.68, Supplementary Data 16 ). Last, both variants reported here are predicted as possible or probably damaging by one of the algorithms tested (PolyPhen). RAPGEF3 (12:48143315, rs145878042, novel locus). The major allele of a low-frequency non-synonymous variant in Rap Guanine-Nucleotide-Exchange Factor (GEF) 3 (rs145878042, L300P, MAF = 1.1%) is associated with increased WHRadjBMI (β (s.e.m) = 0.085 (0.010), P = 7.15 × 10 −17 ). RAPGEF3 codes for an intracellular cAMP sensor, also known as Epac (the Exchange Protein directly Activated by Cyclic AMP). Among its many known functions, RAPGEF3 regulates the ATP sensitivity of the KATP channel involved in insulin secretion 104 , may be important in regulating adipocyte differentiation 105 , 106 , 107 , plays an important role in regulating adiposity and energy balance 108 . RAPGEF3 is tolerant of mutations (ExAC constraint scores: synonymous, −0.47; non-synonymous, 0.32; LoF, 0 and Supplementary Data 16 ). Last, our lead variant is predicted as damaging or possibly damaging for all five algorithms examined (SIFT, Polyphen2/HDIV, Polyphen2/HVAR, LRT, MutationTaster). TBX15 (1:119427467, rs61730011, known locus). The major allele of a low-frequency non-synonymous variant in T-box 15 (rs61730011, M460R, MAF = 4.3%) is associated with increased WHRadjBMI (β (s.e.m.) = 0.041(0.005)). T-box 15 ( TBX15 ) encodes a developmental transcription factor expressed in adipose tissue, but with higher expression in visceral adipose tissue than in subcutaneous adipose tissue, and is strongly downregulated in overweight and obese individuals 109 . TBX15 negatively controls depot-specific adipocyte differentiation and function 110 and regulates glycolytic myofiber identity and muscle metabolism 111 . TBX15 is moderately intolerant of mutations and therefore conserved (ExAC constraint scores: synonymous, 0.42; non-synonymous, 0.65; LoF, 0.88 and Supplementary Data 16 ). Last, our lead variant is predicted as damaging or possibly damaging for four of five algorithms (Polyphen2/HDIV, Polyphen2/HVAR, LRT, MutationTaster). Show more Discussion Our analysis of coding variants from ExomeChip data in up to 476,546 individuals identified a total of 56 array-wide significant WHRadjBMI-associated variants in 41 independent association signals, including 24 newly identified (23 novel and one independent of known GWAS signals). Nine of these variants were low-frequency or rare, indicating an important role for such variants in the polygenic architecture of fat distribution. While, due to their rarity, these coding variants explain a small proportion of the trait variance at a population level, they may be more functionally tractable than non-coding variants and have a critical impact at the individual level. For instance, the association between a LF/RV (rs11209026; R381Q; MAF <5% in ExAC) located in the IL23R gene and multiple inflammatory diseases 50 , 51 , 52 , 53 led to development of new therapies targeting IL23 and IL12 in the same pathway 54 , 55 , 56 . Thus, we are encouraged that our LF/RVs displayed large effect sizes; all but one of the nine novel LF/RVs display larger effects than the 49 SNPs reported in Shungin et al. 10 , and some of these effects were up to seven-fold larger than those previously reported for GWAS. This finding mirrors results for other cardiometabolic traits 57 , and suggests variants of possible clinical significance with even larger effect and rarer variants will probably be detected with greater sample sizes. We continue to observe sexual dimorphism in the genetic architecture of WHRadjBMI 11 . We identified 19 coding variants with significant sex differences, of which 16 (84%) display larger effects in women compared to men. Of the variants outside GWAS loci, we reported three (two LF/RVs) that show a significantly stronger effect in women and two (one LF/RV) that show a stronger effect in men. Genetic variants continue to explain a higher proportion of the phenotypic variation in body-fat distribution in women compared to men 10 , 11 . Of the novel female ( DSTYK and ANGPTL4 ) and male ( UGGT2 and MMP14 ) specific signals, only ANGPTL4 implicated fat distribution related biology associated with both lipid biology and cardiovascular traits (Box 1 ). Sexual dimorphism in fat distribution is apparent 58 , 59 , 60 and at sexually dimorphic loci, hormones with different levels in men and women may interact with genomic and epigenomic factors to regulate gene activity, although this remains to be tested. Dissecting the underlying molecular mechanisms of the sexual dimorphism in body-fat distribution and how it is correlated with—and causing—important comorbidities such as cardiometabolic diseases will be crucial for improved understanding of disease pathogenesis. Overall, we observe fewer significant associations, pathways and cross-trait associations between WHRadjBMI and coding variants on the ExomeChip than Turcot et al. for BMI 25 . One reason for this may be smaller sample size ( N WHRadjBMI = 476,546, N BMI = 718,639), and thus, lower statistical power. Power is probably not the only contributing factor, as trait architecture, heritability (possibly overestimated in some phenotypes) and phenotype precision all probably contribute to our study’s capacity to identify LF/RVs with large effects. Further, it is possible that the comparative lack of significant findings for WHRadjBMI may be a result of higher selective pressure against genetic predisposition to cardiometabolic phenotypes, thus rarer risk variants 61 . The ExomeChip is limited by the variants present on the chip, which was largely dictated by sequencing studies in European-ancestry populations and MAF detection criteria of ~0.012%. It is probable that through increased sample size, use of chips designed to detect variation across a range of continental ancestries, and high quality, deep imputation with large reference samples, future studies will detect additional variation from the entire allele frequency range that contributes to fat distribution. The collected genetic and epidemiologic evidence has demonstrated that increased central adiposity is correlated with risk of type 2 diabetes and CVD, and that this association is probably causal with potential mediation through blood pressure, triglyceride-rich lipoproteins, glucose and insulin 9 . This observation yields an immediate follow-up question: which mechanisms regulate depot-specific fat accumulation and are risks for disease driven by increased visceral and/or decreased subcutaneous adipose tissue mass? Pathway analysis identified several novel pathways and gene sets related to metabolism and adipose regulation, bone growth and development and adiponectin, a hormone that has been linked to ‘healthy’ expansion of adipose tissue and insulin sensitivity 62 . Similarly, expression/eQTL results support the relevance of adipogenesis, adipocyte biology, and insulin signaling, supporting our previous findings for WHRadjBMI 10 . We also provide evidence suggesting known biological functions and pathways contributing to body-fat distribution (for example, diet-induced obesity, angiogenesis, bone growth/morphology and lipolysis). The ultimate aim of genetic investigations of obesity-related traits is to identify dysregulated genomic pathways leading to obesity pathogenesis that may result in a myriad of downstream illnesses. Thus, our findings may enhance the understanding of central obesity and identify new molecular targets to avert its negative health consequences. Significant cross-trait associations are consistent with expected direction of effect for several traits; that is, the WHR-increasing allele is associated with higher values of triglycerides, DBP, fasting insulin, total cholesterol, LDL and type 2 diabetes across many significant variants. However, it is worth noting that there are some exceptions. For example, rs9469913-A in UHRF1BP1 is associated with both increased WHRadjBMI and increased HDL. Also, we identified two variants in MLXIPL (rs3812316 and rs35332062), a well-known lipids-associated locus, in which the WHRadjBMI-increasing allele also increases all lipid levels, risk for hypertriglyceridemia, SBP and DBP. However, our findings show a significant and negative association with HbA1C, and nominally significant and negative associations with two-hour glucose, fasting glucose, and Type 2 diabetes, and potential negative associations with biomarkers for liver disease (for example, gamma glutamyl transpeptidase). Other notable exceptions include ITIH3 (negatively associated with BMI, HbA1C, LDL and SBP), DAGLB (positively associated with HDL), and STAB1 (negatively associated with total cholesterol, LDL and SBP). Therefore, caution in selecting pathways for therapeutic targets is warranted; we must look beyond the effects on central adiposity to the potential cascading effects of related diseases. A major finding from this study is the importance of lipid metabolism for body-fat distribution. In fact, pathway analyses that highlight enhanced lipolysis, cross-trait associations with circulating lipid levels, existing biological evidence from the literature, and knockdown experiments in Drosophila , point to novel candidate genes ( ANGPTL4 , ACVR1C , DAGLB , MGA , RASIP1 and IZUMO1 ) and new candidates in known regions ( DNAH10 10 and MLXIPL 14 ) related to lipid biology and their role in fat storage. ACVR1C , MLXIPL and ANGPTL4 , all of which are involved in lipid homeostasis, all are excellent candidate genes for central adiposity. Carriers of inactivating mutations in ANGPTL4 ( Angiopoietin Like 4 ), for example, display low triglycerides and low risk of coronary artery disease 63 . ACVR1C encodes the activin receptor-like kinase 7 protein (ALK7), a receptor for the transcription factor TGFB-1, well-known for its central role in general growth and development 64 , 65 , 66 , 67 , 68 and adipocyte development in particular 68 . ACVR1C exhibits the highest expression in adipose tissue, but is also highly expressed in the brain 69 , 70 , 71 . In mice, decreased activity of ACVR1C upregulates PPARγ and C/EBPα pathways and increases lipolysis in adipocytes, thus decreasing weight and diabetes 69 , 72 , 73 . Such activity suggests a role for ALK7 in adipose tissue signaling and a possible therapeutic target. MLXIPL , also important for lipid metabolism and postnatal cellular growth, encodes a transcription factor that activates triglyceride synthesis genes in a glucose-dependent manner 74 , 75 . The lead exome variant in MLXIPL is highly conserved, most probably damaging and associated with reduced MLXIPL expression in adipose tissue. Furthermore, in a recent longitudinal, in vitro transcriptome analysis of adipogenesis in human adipose-derived stromal cells, gene expression of MLXIPL was upregulated during the maturation of adipocytes, suggesting a critical role in the regulation of adipocyte size and accumulation 76 . However, given our cross-trait associations with variants in MLXIPL and diabetes-related traits, development of therapeutic targets must be approached cautiously. Our 24 novel variants for WHRadjBMI highlight the importance of lipid metabolism in the genetic underpinnings of body-fat distribution. We continue to demonstrate the critical role of adipocyte biology and insulin resistance for central obesity and offer support for potentially causal genes underlying previously identified fat distribution loci. Notably, our findings offer potential new therapeutic targets for intervention in the risks associated with abdominal fat accumulation and represents a major advance in our understanding of the underlying biology and genetic architecture of central adiposity. URLs Ensembl ortholog database, ; RAREMETALWORKER, ; RVTEST, ; EC-DEPICT, code available at: ; gplots, . Methods Studies Stage 1 included 74 studies (12 case/control, 59 population-based and five family) comprising 344,369 adults of European ( N = 288,492), African ( N = 15,687), South Asian ( N = 29,315), East Asian ( N = 6,800) and Hispanic ( N = 4,075) descent. Stage 1 meta-analyses were conducted in each ancestry and in all ancestries together, for both sex-combined and sex-specific analyses. Follow-up analyses were performed in 132,177 individuals of European ancestry from deCODE and the UK Biobank, Release 1 112 (UKBB) (Supplementary Data 1 – 3 ). Informed consent was obtained by the parent study and protocols approved by each study’s institutional review boards. Phenotypes For each study, WHR (waist circumference divided by hip circumference) was corrected for age, BMI and genomic principal components (derived from GWAS data, the variants with MAF >1% on the ExomeChip, and ancestry informative markers available on the ExomeChip), as well as any additional study-specific covariates (for example, recruiting center), in a linear regression model. For studies with unrelated individuals, residuals were calculated separately by sex, whereas for family based studies sex was included as a covariate in models with both men and women. Residuals for case/control studies were calculated separately. Finally, residuals were inverse normal transformed and used as the outcome in association analyses. Phenotype descriptives by study are shown in Supplementary Data 3 . Genotypes and quality control Most studies followed a standardized protocol and performed genotype calling using the algorithms indicated in Supplementary Data 2 , which typically included zCall 3 . For 10 studies from the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) Consortium, raw intensity data for samples from seven genotyping centers were combined for joint calling 4 . Study-specific quality control of the genotyped variants was implemented before association analysis (Supplementary Data 1 , 2 ). To assess whether any significant associations with rare and low-frequency variants could be due to allele calling in smaller studies, we performed a sensitivity meta-analysis of all large studies (>5,000 participants) compared to all studies. We found very high concordance for effect sizes, suggesting that smaller studies do not bias our results (Supplementary Fig. 24 ). Study-level statistical analyses Each cohort performed single-variant analyses for both additive and recessive models in each ancestry, for sexes combined and sex-specific groups, with either RAREMETALWORKER (see URLs) or RVTESTs (see URLs). to associate inverse normal transformed WHRadjBMI with genotype accounting for cryptic relatedness (kinship matrix) in a linear mixed model. Both programs perform score-statistic rare-variant association analysis, accommodate unrelated and related individuals and provide single-variant results and variance-covariance matrices. The covariance matrix captures linkage disequilibrium between markers within 1 Mb, which is used for gene-level meta-analyses and conditional analyses 113 , 114 . Centralized quality control Individual cohorts identified ancestry outliers based on 1000 Genomes Phase 1 reference populations. A centralized quality control procedure implemented in EasyQC 115 was applied to individual cohort summary statistics to identify cohort-specific problems: (1) possible errors in phenotype residual transformation; (2) strand issues and (3) inflation due to population stratification, cryptic relatedness and genotype biases. Meta-analyses Meta-analyses were carried out in parallel by two analysts at two sites using RAREMETAL 113 . We excluded variants if they had call rate <95%, Hardy–Weinberg equilibrium P < 1 × 10 −7 , or large allele frequency deviations from reference populations (>0.6 for all ancestries analyses and >0.3 for ancestry-specific population analyses). We also excluded markers not present on the Illumina ExomeChip array 1.0, Y-chromosome and mitochondrial variants, indels, multiallelic markers and problematic variants on the basis of Blat-based sequence alignment. Significance for single-variant analyses was defined at an array-wide level ( P < 2 × 10 −7 ). For all suggestive significant variants ( P < 2 × 10 −6 ) from stage 1, we calculated P sexhet for each SNP, testing for differences between women-specific and men-specific beta estimates and standard errors using EasyStrata 116 . Each SNP that reached P sexhet < 0.05 per number of variants tested (70 variants brought forward from stage 1, P sexhet < 7.14 × 10 −4 ) was considered significant. Additionally, while each individual study was asked to perform association analyses stratified by ancestry and adjusted for population stratification, all study-specific summary statistics were combined in our all ancestry meta-analyses. To investigate potential heterogeneity across ancestries, we examined ancestry-specific meta-analysis results for our top 70 variants from stage 1 and found no evidence of significant across-ancestry heterogeneity for any of our top variants ( I 2 values noted in Supplementary Data 4 – 6 ). For the gene-based analyses, we applied two sets of criteria to select variants with a MAF <5% in each ancestry based on coding variant annotation from five prediction algorithms (PolyPhen2, HumDiv and HumVar, LRT, MutationTaster and SIFT) 117 . Our broad gene-based tests included nonsense, stop-loss, splice site and missense variants annotated as damaging by at least one algorithm mentioned above. Our strict gene-based tests included only nonsense, stop-loss, splice site and missense variants annotated as damaging by all five algorithms. These analyses were performed using the sequence kernel association test (SKAT) and variable threshold methods in RAREMETAL 113 . Statistical significance for gene-based tests was set at a Bonferroni-corrected threshold of P < 2.5 × 10 −6 (0.05/~20,000 genes). Genomic inflation We observed marked genomic inflation of the test statistics even after controlling for population stratification arising mainly from common markers; λ GC in the primary meta-analysis (combined ancestries and combined sexes) was 1.06 for all variants and 1.37 for common coding and splice site markers, respectively (Supplementary Figs. 3 , 7 and 13 and Supplementary Table 10 ). Such inflation is expected for a highly polygenic trait such as WHRadjBMI, for studies using a non-random set of variants across the genome, and is consistent with our very large sample size 115 , 118 , 119 . Conditional analyses The RAREMETAL R-package 113 was used to identify independent WHRadjBMI association signals across all ancestries and European meta-analysis results. RAREMETAL performs conditional analyses using covariance matrices to distinguish true signals from shadows of adjacent significant variants in linkage disequilibrium. First, we identified lead variants ( P < 2 × 10 −7 ) based on a 1 Mb window centered on the most significant variant. We then conditioned on the lead variants in RAREMETAL and kept new lead signals at P < 2 × 10 −7 for conditioning in a second round of analysis. The process was repeated until no additional signal emerged below the pre-specified P value threshold ( P < 2 × 10 −7 ). To test if the associations detected were independent of previously published WHRadjBMI variants 10 , 14 , 16 , we used RAREMETAL to perform conditional analyses in the stage 1 discovery set if the GWAS variant or its proxy ( r 2 ≥ 0.8) was on the ExomeChip. All variants identified in our meta-analysis and the previously published variants were available in the UKBB dataset 112 , which was used as a replacement dataset if a good proxy was not on the ExomeChip. All conditional analyses in the UKBB were performed using SNPTEST 120 , 121 , 122 . The conditional analyses were carried out reciprocally, conditioning on the ExomeChip variant and then the previously published variant. An association was considered independent if it was significant before conditional analysis ( P < 2 × 10 −7 ) with both the exome chip variant and the previously published variant, and the observed association with our variant remained significant on conditional analysis. Conditional P values between 9 × 10 −6 and 0.05 were considered inconclusive, while those <9 × 10 −6 were considered suggestive. Stage 2 meta-analyses In stage 2, we sought to validate 70 stage 1 variants ( P < 2 × 10 −6 ) in two independent studies, UKBB ( N = 119,572) and deCODE ( N = 12,605), using the same quality control and analytical methodology. Genotyping, study descriptions and phenotype descriptives are provided in Supplementary Data 1 – 3 . Stage 1 and 2 meta-analysis was performed using the inverse-variance weighted fixed-effects method. Significant associations were defined as those nominally significant ( P < 0.05) in stage 2 when available in stage 2, and array-wide significance for stage 1 and 2 at P < 2 × 10 −7 (0.05/~250,000 246,328 variants tested). Variants are considered novel and statistically significant if they were greater than 1 Mb from a previously identified WHRadjBMI lead SNP 10 , 11 , 12 , 13 , 14 , 15 , 16 and achieved a significance threshold of P < 2 × 10 −7 . Pathway enrichment analyses: EC-DEPICT We adapted DEPICT, a gene set enrichment analysis method for GWAS data, for use with the ExomeChip (‘EC-DEPICT’) described further in a companion manuscript 21 . DEPICT uses ‘reconstituted’ gene sets, where different types of gene set (for example, canonical pathways, protein-protein interaction networks, and mouse phenotypes) were extended through large-scale microarray data (see Pers et al. 20 for details). EC-DEPICT computes P values based on Swedish ExomeChip data (Malmö Diet and Cancer (MDC), All New Diabetics in Scania (ANDIS) and Scania Diabetes Registry (SDR) cohorts, N = 11,899) and, unlike DEPICT, takes as input only genes directly containing substantial (coding) variants rather than all genes in a specified linkage disequilibrium ( Supplementary Note ). Two analyses were performed for WHRadjBMI ExomeChip: one with all variants P < 5 × 10 −4 (49 significant gene sets in 25 meta-gene sets, FDR < 0.05) and one with all variants >1 Mb from known GWAS loci 10 (26 significant gene sets in 13 meta-gene sets, FDR < 0.05). Affinity propagation clustering 123 was used to group highly correlated gene sets into ‘meta-gene sets’; for each meta-gene set, the member gene set with the best P value was used for visualization ( Supplementary Note ). EC-DEPICT was written in Python (see URLs). Pathway enrichment analyses: PASCAL We also applied PASCAL pathway analysis 22 to summary statistics from stage 1 for all coding variants. PASCAL derives gene-based scores (SUM and MAX) and tests for over-representation of high gene scores in predefined biological pathways. We performed both MAX and SUM estimations for pathway enrichment. MAX is sensitive to gene sets driven by a single signal, while SUM is better for multiple variant associations in the same gene. We used standard pathway libraries from KEGG, REACTOME and BIOCARTA, and also added dichotomized ( z > 3) reconstituted gene sets from DEPICT 20 . To accurately estimate SNP-by-SNP correlations even for rare variants, we used the UK10K data (TwinsUK 124 and ALSPAC 125 , N = 3781). To distinguish contributions of regulatory and coding variants, we also applied PASCAL to summary statistics of only regulatory variants (20 kb upstream) and regulatory+coding variants from the Shungin et al. 10 study. In this way, we could investigate what is gained by analyzing coding variants. Monogenic obesity enrichment analyses We compiled two lists consisting of 31 genes with strong evidence that disruption causes monogenic forms of insulin resistance or diabetes, and eight genes with evidence that disruption causes monogenic forms of lipodystrophy. To test for association enrichment, we conducted simulations by matching each gene with others based on gene length and number of variants tested to create 1,000 matched gene sets and assessed how often the number of variants exceeding set significance thresholds was greater than in our monogenic obesity gene set. Variance explained We estimated phenotypic variance explained by stage 1 associations in all ancestries for men, women, and combined sexes 126 . For each associated region, we pruned subsets of SNPs within 500 kb of SNPs with the lowest P value and used varying P value thresholds (ranging from 2 × 10 −7 to 0.02) from the combined sexes results. Additionally, we examined all variants and independent variants across a range of MAFs. The variance explained by each subset of SNPs in each stratum was estimated by summing the variance explained by individual top coding variants. To compare variance explained between men and women, we tested for significant differences assuming the weighted sum of χ 2 -distributed variables tend to a Gaussian distribution following Lyapunov’s central limit theorem 126 , 127 . Cross-trait lookups To evaluate relationships between WHRadjBMI and related cardiometabolic, anthropometric, and reproductive traits, association results for the 51 WHRadjBMI coding SNPs were requested from seven consortia, including ExomeChip data from GIANT (BMI, height), Global Lipids Genetics Consortium (GLGC) (total cholesterol, triglycerides, HDL-cholesterol, LDL-cholesterol), International Consortium for Blood Pressure (IBPC) 128 (systolic and diastolic blood pressure), Meta-Analyses of Glucose and Insulin-related traits Consortium (MAGIC) (glycemic traits) and DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) consortium (type 2 diabetes)) 21 , 25 , 26 , 27 , 28 , 29 . For coronary artery disease, we accessed 1000 Genomes Project-imputed GWAS data released by CARDIoGRAMplusC4D 129 and for age at menarche and menopause, we used a combination of ExomeChip and 1000 Genomes Project-imputed GWAS data from ReproGen. Heat maps were generated with gplots (R v.3.3.2) using Euclidean distance based on P value, and direction of effect and complete linkage clustering (see URLs). GWAS catalog lookups To determine whether substantial coding variants were associated with any related cardiometabolic or anthropometric traits, we also searched the NHGRI-EBI GWAS Catalog for previous associations near our lead SNPs (+/–500 kb). We used PLINK to calculate linkage disequilibrium using ARIC European participants. All GWAS Catalog SNPs in the specified regions with an r 2 > 0.7 were evaluated 37 . Consistent direction of effect was based on the WHRadjBMI-increasing allele, linkage disequilibrium and allele frequency. We do not comment on direction of effect when a GWAS Catalog variant was not identical or in high linkage disequilibrium ( r 2 > 0.9) with the WHR variant and MAF >45%. Body-fat percentage associations We performed body-fat percentage and truncal fat percentage lookups of 48 of the 56 WHRadjBMI identified variants (Tables 1 and 2 ) available in UKBB. GWAS for body-fat percentage and truncal fat percentage in UKBB excluded pregnant or possibly pregnant women, individuals with BMI <15, or those without genetically confirmed European ancestry, resulting in a sample size of 120,286. Estimated body-fat percentage and truncal fat percentage were obtained using the Tanita BC418MA body composition analyzer (Tanita). Participants were non-fasting and did not follow any specific instructions before bioimpedance measurements. SNPTEST was used to perform the analyses based on residuals adjusted for age, 15 principal components, assessment center and the genotyping chip 120 . Collider bias To evaluate SNPs for possible collider bias 17 , we used results from a recent GIANT BMI GWAS 25 . For each significant SNP from our additive models, WHRadjBMI associations were corrected for potential bias due to associations between each variant and BMI ( Supplementary Note ). Variants meeting Bonferroni-corrected significance ( P corrected < 9.09 × 10 −4 , 0.05/55 variants examined) were considered robust against collider bias. Drosophila RNAi-knockdown experiments For each gene with WHRadjBMI-associated coding variants in the final combined meta-analysis ( P < 2 × 10 −7 ), its corresponding Drosophila orthologs were identified in the Ensembl ortholog database (see URLs), when available. Drosophila triglyceride content values were mined from a publicly available genome-wide fat screen dataset 45 to identify potential genes for follow-up knockdowns. Estimated values represent fractional changes in triglyceride content in adult male flies. Data are from male progeny of crosses between male UAS-RNAi flies from the Vienna Drosophila Resource Center and Hsp70-GAL4; Tub-GAL8ts virgin females. ( Supplementary Note ). The screen comprised one to three biological replicates. We followed up each gene with a >0.2 increase or >0.4 decrease in triglyceride content. Orthologs for two genes were brought forward for follow-up, DNAH10 and PLXND1 . For both genes, we generated adipose tissue (cg-Gal4) and neuronal (elav-Gal4) specific RNAi-knockdown crosses to knockdown transcripts in a tissue-specific manner, leveraging upstream activation sequence (UAS)-inducible short-hairpin knockdown lines, available through the Vienna Drosophila Resource Center. Specifically, elav-Gal4, which drives expression of the RNAi construct in post mitotic neurons starting at embryonic stages all the way to adulthood, was used. Cg drives expression in the fat body and hemocytes starting at embryonic stage 12, all the way to adulthood. ( Supplementary Note ). Resulting triglyceride values were normalized to fly weight and larval/population density. We used the non-parametric Kruskall–Wallis test to compare wild type with knockdown lines. eQTLs analysis We queried the significant variant (Exome coding SNPs)-gene pairs associated with eGenes across five metabolically relevant tissues (skeletal muscle, subcutaneous adipose, visceral adipose, liver and pancreas) with at least 70 samples in the GTEx database 46 . For each tissue, variants were selected based on the following thresholds: the minor allele was observed in at least 10 samples and MAF ≥1%. eGenes, genes with a significant eQTL, are defined on a FDR 130 threshold of ≤0.05 of beta distribution-adjusted empirical P value from FastQTL. Nominal P values were generated for each variant-gene pair by testing the alternative hypothesis that the slope of a linear regression model between genotype and expression deviates from zero. To identify all significant variant-gene pairs associated with eGenes, a genome-wide empirical P value threshold 64 (pt) was defined as the empirical P value of the gene closest to the 0.05 FDR threshold. pt was then used to calculate a nominal P value threshold for each gene based on the beta distribution model (from FastQTL) of the minimum P value distribution f (pmin) obtained from the permutations for the gene. For each gene, variants with a nominal P value below the gene-level threshold were considered significant and included in the final list of variant-gene pairs 64 . For each eGene, we also listed the most significantly associated variants (eSNP). Only these exome SNPs with r 2 > 0.8 with eSNPs were considered for biological interpretation (supplementary eQTL GTEx). We also performed cis-eQTL analysis in 770 METSIM subcutaneous adipose tissue samples as described in Civelek et al. 131 . A FDR was calculated using all P values from the cis-eQTL detection in the q -value package in R. Variants associated with nearby genes at an FDR of less than 1% were considered to be significant (equivalent P < 2.46 × 10 −4 ). For loci with more than one microarray probeset of the same gene associated with the exome variant, we selected the probeset that provided the strongest linkage disequilibrium r 2 between the exome variant and the eSNP. In reciprocal conditional analysis, we conditioned on the lead exome variant by including it as a covariate in the cis-eQTL detection and reporting the P value of the eSNP and vice versa. Signals were considered coincident if both the lead exome variant and the eSNP were no longer significant after conditioning on the other and the variants were in high linkage disequilibrium ( r 2 > 0.80). For loci that also harbored reported GWAS variants, we performed reciprocal conditional analysis between the GWAS lead variant and the lead eSNP. For loci with more than one reported GWAS variant, the GWAS variant with the strongest linkage disequilibrium I 2 with the lead eSNP was reported. Penetrance analysis Phenotype and genotype data from UKBB were used for penetrance analysis. Three of 16 rare and low-frequency variants (MAF ≤1%) detected in the final stage 1 and 2 meta-analysis were available in UKBB and had relatively larger effect sizes (>0.90). Phenotype data for these three variants were stratified by WHR using World Health Organization (WHO) guidelines, which consider women and men with WHR greater than 0.85 and 0.90 as obese, respectively. Genotype and allele counts were used to calculate the number of carriers of the minor allele. The number of obese versus non-obese carriers for women, men and sexes combined was compared using a χ 2 test. Significance was determined using a Bonferroni correction for the number of tests performed (0.05/9 = 5.5 × 10 −3 ). Reporting Summary Further information on experimental design is available in the Nature Research Reporting Summary linked to this article. Data availability Summary statistics of all analyses are available at . Change history 07 March 2019 In the HTML version of the article originally published, the link for Supplementary Data 5 returned the file for Supplementary Data 7. The error has been corrected in the HTML version of the article.

A new breakthrough from the Genetic Investigation of Anthropometric Traits consortium, which includes many public health researchers from the University of North Carolina at Chapel Hill, identifies multiple genetic variants associated with how the body regulates and distributes body-fat tissue. The new findings broaden the understanding of how genes can predispose certain individuals to obesity. The GIANT Consortium is a major international collaboration of more than 275 scientists that seeks to identify genetic sites that affect human body size and shape, including height and measures of obesity. Kari E. North, professor of epidemiology at the University of North Carolina at Chapel Hill Gillings School of Global Public Health, is joint lead author of the new study, "Protein-Coding Variants Implicate Novel Genes Related to Lipid Homeostasis 1 Contributing to Body Fat Distribution," published February 18 in Nature Genetics. Other co-authors from the UNC Gillings School include assistant professor Kristin Young, assistant professor Misa Graff, and postdoctoral fellow Heather Highland, all in the UNC Gillings School's department of epidemiology. Identifying the genetic variants associated with obesity is central to developing targeted interventions that can reduce the risk of chronic illnesses, such as hypertension, type 2 diabetes, and heart disease, to which obesity contributes in significant ways. Genome-wide association studies previously identified 49 loci, or positions along a chromosome where the related genetic variants are located, that predispose individuals to a higher waist-to-hip ratio, which is a way to assess body-fat distribution. Lower values of WHR are associated with lower incidence of these diseases. In this study, with a specific focus on coding variation, the team found 24 coding loci—15 common and nine rare—along the chromosomes of individuals that predispose to higher WHR. Further analysis revealed pathways and gene sets that influenced not only metabolism but also the regulation of body fat tissue, bone growth and adiponectin, a hormone that controls glucose levels and breaks down fat. The team also performed functional studies across other organisms and identified two genes that were associated with a significant increase in triglyceride levels and body fat across species. "For the first time, we were able to examine, on a large scale, how low-frequency and rare variants influence body fat distribution," said Kari E. North. "These variants are rarer in the population, but the effects they have on individuals are much larger, possibly making them more clinically relevant." Another major finding from this study is the importance of lipid metabolism to body­fat distribution, which could lead to a better understanding of how obesity causes downstream diseases such as Type 2 diabetes and cardiovascular disease. "A better understanding of the genetic underpinnings of body fat distribution may lead to better treatments for obesity and the cascade of downstream diseases obesity also impacts, for example type 2 diabetes and heart disease" North said.

10.1038/s41588-018-0334-2

Chemistry

Solar cells get boost with integration of water-splitting catalyst onto semiconductor

Jinhui Yang et al. A multifunctional biphasic water splitting catalyst tailored for integration with high-performance semiconductor photoanodes, Nature Materials (2016). DOI: 10.1038/nmat4794 Journal information: Nature Materials

http://dx.doi.org/10.1038/nmat4794

https://phys.org/news/2016-11-solar-cells-boost-water-splitting-catalyst.html

Abstract Artificial photosystems are advanced by the development of conformal catalytic materials that promote desired chemical transformations, while also maintaining stability and minimizing parasitic light absorption for integration on surfaces of semiconductor light absorbers. Here, we demonstrate that multifunctional, nanoscale catalysts that enable high-performance photoelectrochemical energy conversion can be engineered by plasma-enhanced atomic layer deposition. The collective properties of tailored Co 3 O 4 /Co(OH) 2 thin films simultaneously provide high activity for water splitting, permit efficient interfacial charge transport from semiconductor substrates, and enhance durability of chemically sensitive interfaces. These films comprise compact and continuous nanocrystalline Co 3 O 4 spinel that is impervious to phase transformation and impermeable to ions, thereby providing effective protection of the underlying substrate. Moreover, a secondary phase of structurally disordered and chemically labile Co(OH) 2 is introduced to ensure a high concentration of catalytically active sites. Application of this coating to photovoltaic p + n-Si junctions yields best reported performance characteristics for crystalline Si photoanodes. Main The capture of solar energy and its conversion to storable chemical fuel by photoelectrochemical (PEC) systems provides a promising route to overcoming the current global reliance on fossil fuels 1 , 2 . In such artificial photosynthetic systems, the oxidation of water or hydroxyl ions is a requirement for providing a sufficiently abundant source of protons and electrons for sustainably driving the fuel formation reaction. However, the oxygen evolution reaction (OER) presents significant challenges in catalysis and a kinetic bottleneck for solar fuels generation. Among OER catalysts, cobalt oxides (CoO x ) have been shown to possess desirable activity over a broad range of pH values 3 . Recent breakthroughs have provided insight into the nature of catalytically active sites at the atomic scale 4 , 5 , 6 , 7 , 8 , as well as the role of dynamic structural transformations from the resting to the active states 9 , 10 , 11 . Within this context, there has been considerable interest in the development of disordered or amorphous materials that exhibit enhanced catalytic activity relative to their crystalline counterparts 12 , 13 . However, associated mechanisms of enhanced activity are only beginning to emerge. For instance, it has been shown that cobalt phosphate undergoes progressive amorphization, which results in a transition from predominantly surface-confined catalysis on crystalline surfaces to volume-active catalysis in amorphous material 10 . Likewise, Bergmann and co-workers recently demonstrated that transformation to the active phase is accompanied by reversible amorphization of a sub-nanometre-thick near-surface region of Co 3 O 4 to CoO x (OH) y (ref. 11 ). The proposed mechanisms present an intriguing picture of catalysis in dynamic materials, whereby increasing concentrations of catalytically active sites within the bulk can contribute to higher overall activity per geometric area. Although mechanistic insights are important, achieving atomic level control of structure–function relationships is essential for creating high-performance catalysts. Despite the emerging benefits of disordered or amorphous systems for OER catalysis, interfacing these materials with semiconductor light absorbers remains an outstanding challenge. For example, disordered catalyst films tend to be hydrated, and are subject to drying stress-induced cracking and delamination 14 , 15 . Such transformations introduce physical and chemical instabilities, particularly when the catalyst film is interfaced to a chemically sensitive semiconductor. In such cases, it is also necessary to form conformal coatings, while retaining optical transparency. Standard synthetic approaches for forming disordered catalysts, such as electrodeposition and other solution methods, are not well suited for achieving this combination of material properties. Here, we seek to translate catalyst design principles from solution-based synthesis approaches, in which chemically and structurally labile materials possess improved activity, to advanced vapour deposition methods capable of forming continuous and defined thin films compatible with semiconductor light absorbers. Drawing inspiration from recent studies that highlight the importance of volume transformation for increasing concentrations of OER active sites, we use plasma-enhanced atomic layer deposition (PE-ALD) to create biphasic films comprising chemically and structurally labile Co(OH) 2 on top of compact nanocrystalline Co 3 O 4 spinel layers. The presence of these two phases is important for simultaneously increasing catalytic activity and providing a stable interface to the substrate. Precise control of film thickness and conformality, afforded by PE-ALD, is crucial for creating closed films at the nanoscale, while also minimizing losses due to parasitic light absorption and eliminating traditional limitations associated with hydrated catalysts. Integration of biphasic Co 3 O 4 /Co(OH) 2 coatings with photovoltaic p + n-Si junctions yields, to the best of our knowledge, the highest photoelectrochemical activity for OER reported to date on crystalline Si. We note that sub-monolayer mixed Co 3 O 4 /Co(OH) 2 films have been previously applied to Fe 2 O 3 , resulting in considerable catalytic activity improvements 16 . Although such a coating is suitable for chemically robust light absorbers, such as haematite, it is not compatible with existing high-efficiency photoanode materials, which rapidly degrade under operational conditions. In the present work, biphasic CoO x films with defined nanocrystalline interfaces and disordered surface layers were created by PE-ALD using exposure cycles of CoCp 2 and oxygen plasma at target temperatures ranging from 40 °C to 300 °C (see Methods ) 17 , 18 . Deposition was initially performed onto the surfaces of p + -Si (0.001–0.005 Ω cm −1 ) and electrocatalytic (EC) activity for OER was evaluated under 1 M NaOH. Electrochemical properties As shown in Fig. 1a , the electrocatalytic current density is significantly increased as the deposition temperature is decreased from 300 °C to 100 °C. The dependence of activity on deposition temperature can also be seen in Fig. 1b , which shows the current density at an applied electrochemical potential of 1.8 V versus reversible hydrogen electrode (RHE), and Fig. 1c , which shows the overpotential ( η ) required to achieve current densities of 1 mA cm −2 and 10 mA cm −2 . Maximum activity is observed at a deposition temperature of approximately 100 °C. However, further decrease of the deposition temperature to 40 °C leads to poor performance ( Supplementary Fig. 1 ) due to incomplete precursor decomposition and incorporation of residual carbon into the thin film ( Supplementary Fig. 2 ). To promote nucleation, deposition was intentionally performed onto the native silicon oxide 19 , which should act as a tunnelling barrier and play an important role in defining interfacial charge transfer resistance. Nevertheless, we find improved PEC performance for PE-ALD of CoO x onto the native oxide compared to the HF-etched surface ( Supplementary Fig. 3 ). Furthermore, catalyst was also deposited onto tin-doped indium oxide (ITO), a transparent conducting oxide. As shown in Fig. 1d , we find that the catalytic activity of CoO x /ITO is also higher at lower deposition temperature, indicating that the observed differences with deposition temperature are intrinsic features of the catalysts. Figure 1: Electrocatalytic properties of CoO x films as a function of deposition temperature. a , Electrochemical current density versus applied electrochemical potential ( J–E ) curves as a function of growth temperature for PE-ALD CoO x thin films deposited on p + -Si wafers. b , Electrochemical current density from the same films at a fixed potential of 1.8 V versus RHE as a function of substrate temperature during deposition. c , Overpotential, η , required to achieve 10 mA cm −2 (solid squares) and 1 mA cm −2 (solid circles) as a function of deposition temperature. d , J–E curves from CoO x deposited onto tin-doped indium oxide (ITO) confirm that enhanced catalytic performance at reduced deposition temperature is an intrinsic feature of the films. e , Atomic force micrograph of a PE-ALD CoO x film grown at 100 °C on a (100) p + -Si substrate, revealing an r.m.s. surface roughness of 2.6 ± 0.2 Å. The corresponding measurement on material deposited at 300 °C yields an r.m.s. surface roughness of 2.8 ± 0.8 Å. These findings, together with electrochemical capacitance measurements, indicate that observed differences of catalytic performance from these highly planar films are not due to surface roughness. Full size image Capacitance measurements indicate that the electrochemically active surface areas per geometric area do not vary as a function of deposition temperature (see Supplementary Fig. 4 ). Therefore, we can conclude that changes of surface roughness are not responsible for the observed differences in catalytic activity. As shown in Fig. 1e , atomic force microscopy (AFM) of the film deposited at 100 °C reveals that the layer is planar and continuous, with a root mean square (r.m.s.) roughness of 2.6 Å. We note that solid/liquid contact area can have a considerable impact on apparent activity, with many of the catalysts with highest activity being characterized by large roughness values 20 , 21 . Given the planar morphology of the PE-ALD catalyst studied here, the overpotentials compare favourably with benchmarked CoO x catalysts possessing much higher electrochemically active surface area 20 , 21 . Atomic structure To identify the key properties that define catalytic function, we focus on comparison of films deposited at 100 °C and 300 °C. To this end, we employed advanced low-dose aberration-corrected electron microscopy, which enables characterization down to the level of single atoms, while minimizing beam–sample interactions. Cross-sectional images of CoO x films deposited onto (100) p + -Si at 100 °C and 300 °C reveal layer thicknesses of 3.2 ± 0.2 nm and 4.8 ± 0.3 nm, respectively ( Fig. 2a ). For both cases, lattice fringes in the layers originate from the nanocrystalline structure of the deposited material, and their homogeneity across the films indicates that grain diameters are comparable to the layer thicknesses. Figure 2: Structural characterization of catalysts by transmission electron microscopy. a , Cross-section micrographs of CoO x layers deposited at 100 °C (top) and 300 °C (bottom) on (100) Si, revealing thicknesses of 3.2 ± 0.2 nm and 4.8 ± 0.3 nm, respectively. A thin native silicon oxide layer is observed at the interface. b , c , Low-dose, plan-view micrographs of films grown on amorphous Si 3 N 4 membranes at 100 °C ( b ) and 300 °C ( c ). Dose rates are listed and single images were recorded with a 1 s exposure. A large underfocus ( a , f = −300 nm; b , f = −1,000 nm) forces the appearance of Fresnel fringes that encircle individual grains and are evaluated in Fourier space (insets) to measure an average grain size of 3.3 nm in a and 5.3 nm in b . The different grey values of the grains are caused by electron scattering in crystalline material along variable zone axis orientations. Bright spots in b are dominantly located in boundaries between individual grains and suggest the presence of voids. d , Analysis of grain orientations for the sample deposited at 100 °C is performed using the plan-view phase image of the electron exit wave function, reconstructed from 50 images. The grains form partly coherent interfaces, with an average particle size of 3.4 ± 0.2 nm. Approximately 30% of all grains exhibit two independent sets of reflection vectors that allow calculation of their zone axes, locally, as shown for a [211] grain and its nano-diffraction pattern (left inset). The corresponding stereographic triangle (right inset) of the cubic single crystal highlights that a range of orientations were detected. e , Samples deposited at 300 °C are characterized by voids, as shown in the reconstructed phase image, at high magnification (region bounded by dashed lines). The void shown here is bounded by crystal facets from adjacent [001], [211] and [111] oriented grains. f , The phase of the electron exit wave functions is reconstructed from a focus series of 50 images recorded on material deposited at 100 °C using a high dose rate and accumulated electron dose of 9 × 10 5 e − Å −2 . A [011] oriented grain is shown, together with the projected crystal structure of Co 3 O 4 spinel. Columns of Co 3+ , Co 2+ and O 2− , can be readily distinguished. Grain boundaries are irregular and partly coherent (arrows) but void formation at interfaces is not observed, despite considerable beam-induced diffusion and coarsening. A movie depicting dynamic processes arising from beam–sample interactions is provided in the Supplementary Methods . Full size image To image individual grains, plan-view samples were prepared by PE-ALD onto Si 3 N 4 membranes. Figure 2b, c shows low-dose-rate (<300 e − Å −2 s −1 ) images of pristine layers grown at 100 °C and 300 °C, respectively. Nanocrystalline grains are observed that form continuous layers and exhibit irregular shapes. The absence of moiré patterns suggests that a monolayer of nanocrystals is imaged. Supplementary Fig. 5 shows phase images at a higher magnification, revealing that the films consist of nanocrystals with partly coherent interfaces and grain sizes of 3.4 ± 0.2 nm and 5.2 ± 0.3 nm, respectively. Corresponding nano-diffraction patterns are assigned to electron scattering from Co 3 O 4 spinel 22 , in agreement with extended X-ray absorption fine structure (EXAFS) measurements ( Supplementary Fig. 6 ). Figure 2d (inset) shows an analysis of the film texture of the sample grown at 100 °C, which is also characteristic for films grown at 300 °C ( Supplementary Fig. 7 ), and reveals no statistically significant differences to account for the observed increase of catalytic activity at reduced deposition temperature. A striking difference observed between Fig. 2b and c is the existence of bright contrast features in the grain boundaries of the sample deposited at 300 °C. Corresponding features are not observed from the sample deposited at 100 °C, where we find that grains coalesce to form a dense film of nanocrystalline Co 3 O 4 . These features exhibit the characteristic random pattern of an amorphous material next to crystal facets of the adjacent grains ( Fig. 2d ). The bright contrast suggests a lower mass density in these regions, which favours an assignment as pinhole defects associated with facet formation when the average grain size exceeds ∼ 5 nm. These results suggest that formation of closed nanocrystalline layers is enabled by smaller and less strongly faceted grains grown at reduced temperature. At elevated deposition temperature, increased surface diffusion and more pronounced faceting may lead to the formation of nanoscopic voids at grain boundaries. Although imaging of pristine grains was accomplished by limiting total doses to <300 e − Å −2 , additional insight into the sensitivity of films to nanoscale void formation is achieved by exploiting beam–sample interactions to intentionally drive in situ atom diffusion and grain growth. As shown in Supplementary Fig. 8 , along with the associated movie, significant electron beam-induced grain coarsening of material deposited at 100 °C can be achieved by recording image series. Despite considerable atomic motion during such measurements, individual images give the appearance of a static grain structure 23 , allowing identification of the locations of all Co 3+ , Co 2+ and O 2− columns ( Fig. 2e ). Analysis of grain sizes in image series ( Supplementary Fig. 8 ) reveals surface diffusion-induced grain coarsening. Importantly, polycrystalline thin films remain void-free and compact in the presence of in situ beam-stimulated atom diffusion, grain coarsening, and grain reorientation when their average grain size is smaller than 4.5 nm. Films deposited at 300 °C exhibit a grain size of 5.2 ± 0.3 nm and a tendency to create voids between facets of adjacent crystal grains. Chemical properties Angle-resolved X-ray photoelectron spectroscopy (XPS) reveals significant differences in the chemical composition of the surface as a function of deposition temperature ( Fig. 3 ). All spectra are well described using reported parameters for Co 3 O 4 and Co(OH) 2 (ref. 24 ). The sample deposited at 300 °C is dominantly composed of Co 3 O 4 , which comprises tetrahedral Co 2+ and octahedral Co 3+ . A small contribution from Co(OH) 2 is also observed—probably due to a hydroxylated terminal surface. In contrast, the sample deposited at 100 °C is characterized by significant spectral contributions from both Co 3 O 4 and Co(OH) 2 . The components from Co(OH) 2 increase with increasing take-off angle, indicating that Co(OH) 2 is mainly present at the surface. This finding is consistent with X-ray absorption near-edge structure (XANES) measurements, which homogeneously sample the films and reveal a nearly identical overall composition of Co 3 O 4 for material deposited at 100 °C and at 300 °C ( Supplementary Fig. 9 ). Although the Co(OH) 2 phase was not directly observed by transmission electron microscopy (TEM), this is not surprising, considering that these hydroxide compounds often lack long-range crystalline order and have been described as oligomeric layered structures 25 . Together, the combination of results from TEM, XPS and XANES allows us to conclude that the bulk of both films is composed of Co 3 O 4 , whereas there is a surface layer of Co(OH) 2 atop the sample deposited at 100 °C. Such a difference in surface composition can have a dramatic effect on catalytic mechanisms, as discussed below. Figure 3: Compositions and chemical states of tailored CoO x catalyst layers. a , b , Angle-resolved Co 2 p X-ray photoelectron spectra acquired at 0° and 55° relative to the surface normal for CoO x films deposited on p + -Si at 300 °C ( a ) and 100 °C ( b ). Constrained fitting of spectra reveals the presence of two phases, Co 3 O 4 and Co(OH) 2 , with their spectral contributions shown in orange and blue, respectively. The relative intensity of the Co(OH) 2 component increases with increasing photoelectron take-off angle, indicating that this material lies at the surface, as indicated in schematic illustrations of the films deposited at 300 °C ( c ) and 100 °C ( d ). Significant differences in phase composition are observed, with the biphasic character of material deposited at 100 °C allowing for considerably improved catalytic performance, as described in the text. Full size image Chemical transformation of Co 3 O 4 surfaces to catalytically active CoO(OH) or CoO x (OH) y phases has been observed by in situ Raman spectroscopy and synchrotron X-ray spectroscopy, respectively 11 , 26 . However, electrochemical and gravimetric characterization of Co 3 O 4 spinel has demonstrated bulk phase stability under OER conditions. Transformations to CoO(OH) are restricted to the surface 27 , which has important implications for the present work. In particular, the phase stability of closed nanocrystalline films of Co 3 O 4 is desirable for ensuring a durable interface with the substrate. However, phase stability also implies an upper bound for the concentration of active sites for planar, compact thin films of Co 3 O 4 , such as those formed at 300 °C. This limitation can be overcome by the creation of biphasic Co 3 O 4 /Co(OH) 2 films at 100 °C. Unlike Co 3 O 4 , Co(OH) 2 possesses an open, layered structure that allows intercalation of water and ions from solution, as well as changes in Co oxidation state with minimal structural reorientation 6 , 27 , 28 . As such, Co(OH) 2 supports catalytic site formation not just at its surface but also within its bulk volume. Therefore, the observation of Co(OH) 2 on material deposited at low temperature suggests that the catalytic activity is enhanced by the presence of a higher concentration of active sites under OER conditions. It is known that active sites of solid-state Co-based OER catalysis proceed through redox transitions of Co 3+ → Co 4+ , with the highest activity sites provided by the generation of two adjacent and electronically coupled Co 4+ centres 5 , 9 , 29 . As shown in Fig. 4a , steady-state polarization Tafel curves yield values of ∼ 50 mV dec −1 and 70 mV dec −1 for films deposited at 100 °C and 300 °C on p + -Si, respectively. These values are in the range of those previously reported for CoO x under alkaline conditions 12 , 25 , 27 , and are consistent with fast electrochemical pre-equilibrium between Co 3+ –OH and Co 4+ –O, followed by the rate-limiting O–O bond formation step 30 , 31 , 32 . The lower Tafel slope from the 100 °C sample supports the hypothesis that the biphasic Co 3 O 4 /Co(OH) 2 promotes formation of active Co 4+ sites. The deviations from linearity at higher current densities suggest a resistance to charge transport that is higher in the thicker film deposited at 300 °C (refs 12 , 32 ). Figure 4: Electrochemical characterization of chemical transformations of CoO x catalysts. a , Steady-state Tafel plots measured in 1 M NaOH for CoO x films deposited by PE-ALD on p + -Si at 100 °C and 300 °C. The data were corrected for the IR drop in solution and yield Tafel slopes of 50 mV decade −1 and 70 mV decade −1 , respectively. b , Cyclic voltammograms (CVs) collected in the pre-OER catalytic region at 100 mV s −1 for samples deposited at 100 °C and 300 °C. Two oxidation peaks, labelled A 1 and A 2 , are assigned to Co 2+ → Co 3+ and Co 3+ → Co 4+ reactions, respectively. The enhanced redox activity of the 100 °C material is due to the biphasic Co 3 O 4 /Co(OH) 2 film. c , Comparison of first and second CV scans from material deposited at 100 °C. The transformation of Co(OH) 2 to CoO(OH) is largely irreversible, leading to a reduction in the A 1 peak after the first cycle. Activation of catalytically active Co 4+ is unaffected by cycling, as exhibited by the approximately constant area of the A 2 oxidation peak. Full size image Cyclic voltammetry (CV) in the pre-OER potential range reveals that surface redox behaviour of Co 2+ and Co 3+ is correlated with catalytic activity 6 , 12 , 25 , 27 . As shown in Fig. 4b , for the case of material deposited at 100 °C, we observe two well-defined anodic peaks. These are denoted A 1 and A 2 , and can be assigned to Co 2+ → Co 3+ and Co 3+ → Co 4+ oxidation reactions, respectively 12 , 28 . The similar areas of A 1 and A 2 are consistent with electrochemical transformation of Co 2+ in the as-deposited Co(OH) 2 phase to Co 3+ in the form of CoO(OH), followed by activation of Co 4+ (ref. 33 ). Figure 4c shows a comparison of the first and second CV sweeps. Following the first cycle, there is a considerable reduction in the magnitude of A 1 , whereas the magnitude of A 2 remains approximately constant. This is indicative of the known electrochemical behaviour of Co(OH) 2 , in which the Co 2+ → Co 3+ reaction, which does not participate directly in the catalytic cycle but is associated with phase transformation to CoO(OH) (refs 30 , 31 ), is largely irreversible 34 . Much weaker redox activity is observed in the pre-catalytic range for material deposited at 300 °C, which is consistent with surface site-confined catalysis on structurally robust Co 3 O 4 . In contrast, the larger areas of A 1 and A 2 on the material deposited at 100 °C indicate that the biphasic Co 3 O 4 /Co(OH) 2 films contain higher concentrations of active sites. Thus, the biphasic coating creates a desirable balance of properties that are well suited for integration with chemically sensitive substrates; the phase stability of Co 3 O 4 enables a durable substrate/catalyst interface, while facile chemical transformation of Co(OH) 2 promotes high catalytic activity. Integrated photoanodes We deposit the highly active Co 3 O 4 /Co(OH) 2 catalyst made by PE-ALD at 100 °C onto a p + n-Si substrate to determine if electrocatalytic performance enhancements can be translated to PEC systems. As shown in Fig. 5a , these integrated photoanodes are characterized by a photocurrent density of 30.8 mA cm −2 at 1.23 V versus RHE and a saturation photocurrent density of ∼ 37.5 mA cm −2 under 1 sun AM 1.5 simulated solar irradiation. This large saturation current is attributed to low parasitic light absorption losses enabled by the closed biphasic catalyst coating with a near-optimal thickness for maximizing optocatalytic efficiency 35 . Comparison of the EC characteristics of CoO x on p + -Si to the PEC characteristics of analogous films deposited on p + n-Si yields a photovoltage of 600 mV. Photovoltages generated from CoO x films deposited at 100 °C and 300 °C are nearly identical 17 . Therefore, we conclude that the cathodic shift of onset potential obtained with the biphasic Co 3 O 4 /Co(OH) 2 catalyst is a consequence of improved catalytic activity. To the best of our knowledge, these J–E performance characteristics are the best reported to date for catalyst-coated crystalline Si photoanodes 17 , 36 , 37 , 38 , 39 , 40 , 41 , 42 , 43 , 44 . Importantly, catalysts that undergo chemical transformation from the resting to the active state are often characterized by high ion permeability, and are susceptible to drying-induced cracking and delamination 14 , 15 . As shown in Supplementary Fig. 10 , no change of performance is observed upon repeated cycling between PEC operation and drying. Chronoamperometric tests revealed the films to be functionally stable under illuminated operation for at least 72 h ( Fig. 5b ), with no evidence of Co loss to solution, as measured by inductively coupled plasma mass spectroscopy ( Supplementary Fig. 11 ). The multifunctional properties of our biphasic coating, in which the interfacial Co 3 O 4 is chemically stable and a thin and intimately contacted surface hydroxide promotes high catalytic activity, represents an important advance for forming assemblies incorporating semiconductor light absorbers. Figure 5: Photoelectrochemical and stability characteristics of high-performance CoO x /Si photoanodes. a , Current density versus applied electrochemical potential ( J–E ) curves for biphasic Co 3 O 4 /Co(OH) 2 films deposited at 100 °C onto p + -Si dark anodes and p + n-Si photoanodes. A photovoltage of ∼ 600 mV is obtained from the difference between these curves. The photoelectrodes exhibit the highest performance characteristics reported for crystalline Si to date, with an onset potential for water oxidation of <1 V versus RHE and a saturation current density of ∼ 37.5 mA cm −2 . b , Normalized photocurrent as a function of time ( J–t ) reveals stable performance of p + n-Si/Co 3 O 4 /Co(OH) 2 photoelectrodes for at least 72 h of continuous operation in 1 M NaOH under simulated 1 sun illumination. By comparison, bare p + n-Si photoelectrodes degrade almost immediately (inset). Elemental analysis of the electrolyte during stability testing indicates no detectable transfer of Co from the film into solution ( Supplementary Fig. 11 ), even after 72 h of operation. Full size image In summary, we have fabricated conformal, biphasic CoO x catalysts that are engineered to provide high activity and compatibility with semiconductor photoelectrodes. The resulting films consist of conformal layers of spinel Co 3 O 4 nanocrystals that coalesce to form partly coherent grain boundaries. The resistance of these Co 3 O 4 spinel crystals to chemical transformation provides durable interfaces. At reduced deposition temperature, disordered surface layers of Co(OH) 2 are introduced that significantly improve electrocatalytic OER activity. This surface layer promotes chemical and structural transformation, thereby providing high concentrations of catalytically active sites. In contrast to traditional layered hydroxides, which are susceptible to drying-induced cracking and delamination, these biphasic films are simultaneously robust and active. Improved electrocatalytic activity translates to higher photoelectrochemical activity, as demonstrated by integration of the biphasic material onto the surface of p + n-Si photoelectrodes, which represent the best reported for crystalline Si to date. These results demonstrate that PE-ALD is a powerful method for synthesizing multifunctional catalysts that support desired chemical transformations, permit efficient interfacial charge transport, and minimize parasitic light absorption due to their conformal nature even at very low thicknesses. Thus, partial decoupling of ALD surface reaction kinetics from thermal activation, enabled by plasma enhancement, provides opportunities to tailor novel materials and interfaces for achieving desired functionality. Methods Deposition of CoO x by atomic layer deposition (ALD). Plasma-enhanced ALD (PE-ALD) allows partial decoupling of surface oxidation kinetics from substrate temperature by introducing a source of highly reactive oxygen radicals to the system 19 , 45 . Therefore, substrate temperature can be used to affect thin film properties with reduced influence on ALD surface reaction kinetics. This is a powerful feature of PE-ALD and offers significant, yet underexplored, potential for tailoring of catalytic materials. Here, CoO x catalyst films were deposited using a remote PE-ALD (Oxford FLexAl) process at substrate temperatures ranging from 100 °C to 300 °C (refs 17 , 18 ). The cobalt precursor was CoCp 2 (98% Strem Chemicals) and oxygen plasma was the oxidant. The precursor lines, carrier gas lines, and the reactor walls were kept at 120 °C. The CoCp 2 precursor bubbler was heated up to 80 °C and bubbled with 200 sccm of high-purity Ar gas during the precursor exposure half-cycles. Oxygen gas flow was held constant at 60 sccm throughout the deposition process. The cobalt precursor exposure half-cycle consisted of 5 s CoCp 2 dosing and 5 s purging. The remote oxygen plasma half-cycle consisted of 1 s pre-plasma treatment, 5 s plasma exposure, and 15 s purging. The applied plasma power was 300 W and was applied for 5 s during the oxygen plasma half-cycle. The deposition chamber was maintained at 15 mTorr at all times. Unless otherwise noted, 100 PE-ALD cycles were used for film formation. This was selected based on evaluation of the electrochemical potential required to reach a photocurrent density of 10 mA cm −2 from p + n-Si/CoO x electrodes as a function of the number of deposition cycles ( Supplementary Fig. 12 ), but future opportunity exists for optimization. For characterization of CoO x films as electrocatalysts, deposition was performed onto highly doped p-Si (p + -Si, B-doped, 0.001–0.005 Ω cm −1 ) and indium-doped tin oxide (ITO, 8–12 Ω cm −1 ) following ultrasonic solvent cleaning, followed by drying under flowing nitrogen. As described in the main text and shown in Supplementary Fig. 3 , we find that the electrocatalytic activities of samples made from CoO x deposited onto the native oxide of Si exhibit superior performance relative to HF-treated Si. Therefore, unless otherwise noted, all CoO x /Si samples were deposited after cleaning, but without etching the substrate in HF. Formation of p + n-Si junctions by ion implantation. Starting substrates were single-side-polished Czochralski-grown n-type (P-doped) prime grade (100) Si wafers with resistivity in the range 3.0–10 Ω cm. Ion implantation was employed for junction formation. Room temperature implantation was performed at a 7° incident angle using 11 B accelerated to 45 keV with a dose of 1 × 10 14 cm −2 and 32 keV with a dose of 5 × 10 14 cm −2 . To reduce contact resistance, the backsides of the wafers were implanted with 31 P at 140 keV with a dose of 1 × 10 14 cm −2 and 75 keV with a dose of 5 × 10 14 cm −2 . Dopant activation, both for the junction p + emitter layer and the n + back contact layer, was achieved via rapid thermal annealing at 1,000 °C for 15 s under flowing nitrogen. Photoelectrochemical testing. For both electrochemical and photoelectrochemical (PEC) characterization, cyclic voltammetry (CV) was performed using a Biologic potentiostat and a three-electrode cell using a platinum wire as counter electrode, a Hg/HgO (1 M NaOH) as reference electrode, and the CoO x -coated sample as the working electrode. Unless otherwise noted, CV data were collected at a scan rate of 100 mV s −1 . The Si working electrodes were fabricated by scratching an indium–gallium eutectic (Aldrich) into the backside of the wafer and affixing a copper wire using conductive silver epoxy (Circuit Works). Direct electrical contact to ITO working electrodes was achieved using conductive silver epoxy (Circuit Works). The copper wire was passed through a glass tube and the wafer was insulated and attached to the glass tube with Loctite 615 Hysol Epoxi-patch or 7460 adhesive. Electrodes were dried overnight before testing. The supporting electrolyte was aqueous 1 M NaOH prepared using Milli-Q water (18.2 MΩ cm). All measurements were performed using an air saturated solution. PEC CV tests were performed at 100 mW cm −2 using a Solar Light 16S-300-005 solar simulator equipped with an AM1.5 filter set, and the sample was illuminated through a quartz window of the cell. Steady-state polarization Tafel analysis was performed after IR drop correction. To determine if the surface roughness, and thus solid/liquid interfacial contact area, varies as a function of deposition temperature, the double-layer capacitance was measured by collecting CV scans as a function of scan rate. The relative electrochemically active surface areas were calculated from the linear slope of the scan-rate-dependent current density at the centre point potential of the sweep. To eliminate pseudocapacitance contributions, measurements were performed in an acetonitrile solution with 0.1 M tetrabutylammoniumhexafluorophosphate as the electrolyte instead of the aqueous NaOH solution used for the electrochemical activity measurements. Chronoamperometric stability tests were performed under simulated AM1.5G radiation at 1 sun using an Oriel Sol3A. Samples were mounted into acrylic cells using compression-fit gaskets, and electrolyte flow was established using a peristaltic pump to prevent bubble accumulation on photoelectrode surfaces. Measurements were performed at an applied electrochemical potential of 1.8 V versus RHE in 1 M NaOH. Aliquots of electrolyte were removed every ∼ 24 h for ICP-MS analysis (see below). Inductively coupled plasma mass spectrometry (ICP-MS). ICP-MS was performed using an Agilent 7900 system run using the He mode. The internal standard was Ge, selected based on its first ionization potential and M / Z as compared to Co. The standard curve was prepared from a stock solution of 10 ppm (Part no. 8500-6940, Agilent Technologies) at concentrations of 0.01, 0.1, 1, 10 and 100 ppb diluted in 1% HNO 3 prepared from 70% HNO 3 (>99.999% trace metals basis, 225711, Aldrich) and Milli-Q water. Linear fitting of the calibration curve resulted in an R 2 = 1.000. Sample solutions, collected at various times during the stability tests, were run as is. Atomic force microscopy (AFM). AFM measurements were performed using a Bruker Dimension Icon system operated in Scanasyst mode with Si tips (Bruker, Scanasyst-Air). X-ray photoelectron spectroscopy (XPS). XPS was performed using a monochromatized Al Kα source ( hν = 1,486.6 eV), operated at 150 W, on a Kratos Axis Ultra DLD system at take-off angles of 0° and 55° relative to the surface normal, and a pass energy for narrow scan spectra of 20 eV, corresponding to an instrument resolution of approximately 600 meV. Spectral fitting was done using Casa XPS analysis software. Spectral positions were corrected using adventitious carbon, by shifting the C 1 s core level position to 284.8 eV, and curves were fitted with quasi-Voigt lines following Shirley background subtraction. We note that the spectral components from cobalt oxides overlap, making it difficult to distinguish between different oxides, though differentiation of Co 2+ from Co 3+ is possible since the unpaired electrons from Co 2+ result in a distinct satellite structure at higher binding energies. In the present work, we constrained fitting to three parameters by using differential peak positions, relative amplitudes, and approximate widths reported by Biesinger and colleagues 24 . X-ray absorption spectroscopy (XAS). X-ray absorption spectra (XAS) were collected on beamline 7–3 at the Stanford Synchrotron Radiation Lightsource (SSRL) with an average current of 500 mA and an electron energy of 3.0 GeV. The radiation was monochromatized with a Si (220) double-crystal monochromator which was detuned to 50% of flux maximum at the Co K-edge. A N 2 -filled ion chamber (I 0 ) was used to monitor the intensity of the incident X-rays in front of the sample. XAS data were recorded as fluorescence excitation spectra using a 30-element Ge solid-state detector (Canberra). The monochromator energy calibration was done with the first peak maximum of the first derivative of the Co foil spectrum (7,709.5 eV). Powder reference samples were diluted with boron nitride ( ∼ 1% w/w) and then packed into 0.5-mm-thick aluminium sample holders using kapton film windows on both sides. Data reduction of the XAS spectra was performed using SamView (SixPack software, Samuel M. Webb, SSRL). Athena software (IFEFFIT package) 46 was used to subtract the pre-edge and post-edge contributions, and the results were normalized with respect to the edge jump. A five-domain cubic spline was used for the background removal in k-space. The extracted k-space data, k 3 χ ( k ), was then Fourier transformed into real space (r-space) using a k-space window of 3.0–11.30 Å −1 . Transmission electron microscopy (TEM). High-resolution electron microscopy is known to be challenging since the probing electron beam can alter the genuine structure before atomic resolution is reached. Minimization of beam–sample interactions is particularly important for understanding the structures of low-temperature and potentially disordered thin films, such as those investigated here. Recently, remarkable progress has been made in overcoming these limitations by applying low dose rate in-line electron holography 47 , 48 . In this work, electron microscopy was performed at the Molecular Foundry using microscopes operated at 300 kV. Images in Fig. 2a were recorded with the One Angstrom Microscope 49 . All other experiments were performed with TEAM 0.5 (ref. 50 ). The instrument allows control of beam–sample interactions by producing in-line holograms from image series with variable dose rates and by solving the phase problem 48 . The method builds on best practices that were developed for the imaging of biological objects, but uses large image series to obtain the needed contrast in atomic-resolution images. A resolution well below 1 Å is shown in Supplementary Fig. 5 that allows for a reliable identification of crystal structures from real-space images by indexing diffraction spots using established crystal structures 22 . All reconstructions of exit wave functions and other calculations were done with the MacTempas software package (Kilaas). In the low-magnification mode, we recorded single images at a large underfocus, f , ( Fig. 2b, c ) where Fresnel fringes become prominent and differentiate the individual grains. As a result, specific spatial frequencies occur in the low-spatial-frequency spectrum of the Fourier transform that characterize the average grain size, as shown in the insets of Fig. 2b, c . In high-resolution images, individual grains are visible that we approximate by squares of area A 2 . An average grain size, A , can be estimated by counting the number of grains, N , in a known field of view B = N × A 2 . Supplementary Fig. 8a summarizes these measurements for a recording of the same object with increasing dose rates, which allows the study of electron beam-induced grain growth and restructuring, as shown in Supplementary Fig. 8b, c , as well as the Supplementary Movie . Grain orientations were measured by taking a local Fourier transform of individual grains from the reconstructed in-line holograms (that is, complex electron exit wave functions), which are local nano-diffraction patterns ( Fig. 2d ) that cannot be obtained by direct diffraction work in broad beam mode. Texture analysis is limited by statistics, since individual zone axis orientations can only be determined for grains that exhibit two independent sets of diffraction spots. Nevertheless, roughly 30% of the grains exhibit independent sets of diffraction spots that can be indexed to calculate zone axis orientations ( Fig. 2d inset and Supplementary Fig. 7 ). The measured grain orientations cover most regions of the stereographic triangle of the face-centred cubic (fcc) single crystal. A standard cross-section sample preparation was performed using an Ar ion mill to thin samples consisting of CoO x deposited on the native oxide of crystalline (100) Si substrates ( Fig. 2a ). For plan-view observations, however, we deposited CoO x directly onto an electron-transparent silicon nitride membrane so that no additional sample preparation was necessary. Therefore, we can exclude any preparation-induced sample alteration that can occur during an exposure of small nanocrystals to the energetic beam of argon ions that is typically used for sample thinning. Depositions at both 100 °C and 300 °C yield continuous thin films that are composed of a monolayer of nanocrystalline material, which is determined by the absence of any moiré fringes. Movie. The movie depicting crystal growth by surface diffusion and grain reorientation was generated from the focus series marked by a circle in Fig. 3 . It consists of 60 images that were split into six series with ten images, each, to reconstruct six wave functions. The phase of these six reconstructed electron exit wave functions is shown as six images of a movie lasting 0.5 s. Thus, a recording time of ∼ 100 s is compressed into 0.5 s and the physical processes are shown accelerated by a factor of ∼ 200× in time. The resolution in each image is maintained around 0.6 Å, which enables a full separation of all Co and O atom columns in the [100] and [110] oriented grains that are imaged along their zone axes.

Scientists have found a way to engineer the atomic-scale chemical properties of a water-splitting catalyst for integration with a solar cell, and the result is a big boost to the stability and efficiency of artificial photosynthesis. Led by researchers at the Department of Energy's Lawrence Berkeley National Laboratory (Berkeley Lab), the project is described in a paper published this week in the journal Nature Materials. The research comes out of the Joint Center for Artificial Photosynthesis (JCAP), a DOE Energy Innovation Hub established in 2010 to develop a cost-effective method of turning sunlight, water, and carbon dioxide into fuel. JCAP is led by the California Institute of Technology with Berkeley Lab as a major partner. The goal of this study was to strike a careful balance between the contradictory needs for efficient energy conversion and chemically sensitive electronic components to develop a viable system of artificial photosynthesis to generate clean fuel. Striking the right balance "In order for an artificial photosystem to be viable, we need to be able to make it once, deploy it, and have it last for 20 or more years without repairing it," said study principal investigator Ian Sharp, head of materials integration and interface science research at JCAP. The problem is that the active chemical environments needed for artificial photosynthesis are damaging to the semiconductors used to capture solar energy and power the device. "Good protection layers are dense and chemically inactive. That is completely at odds with the characteristics of an efficient catalyst, which helps to split water to store the energy of light in chemical bonds," said Sharp, who is also a staff scientist at Berkeley Lab's Chemical Sciences Division. "The most efficient catalysts tend to be permeable and easily transform from one phase to another. These types of materials would usually be considered poor choices for protecting electronic components." By engineering an atomically precise film so that it can support chemical reactions without damaging sensitive semiconductors, the researchers managed to satisfy contradictory needs for artificial photosystems. "This gets into the key aspects of our work," said study lead author Jinhui Yang, who conducted the work as a postdoctoral researcher at JCAP. "We set out to turn the catalyst into a protective coating that balances these competing properties." Jinhui Yang performing X-ray photoelectron spectroscopy measurements, which are used to understand the chemical properties of surfaces, at the Joint Center for Artificial Photosynthesis. Credit: Robert Paz/Caltech Doing double duty The researchers knew they needed a catalyst that could not only support active and efficient chemical reactions, but one that could also provide a stable interface with the semiconductor, allow the charge generated by the absorption of light from the semiconductor to be efficiently transferred to the sites doing catalysis, and permit as much light as possible to pass through. They turned to a manufacturing technique called plasma-enhanced atomic layer deposition, performed at the Molecular Foundry at Berkeley Lab. This type of thin-film deposition is used in the semiconductor industry to manufacture integrated circuits. "This technique gave us the level of precision we needed to create the composite film," said Yang. "We were able to engineer a very thin layer to protect the sensitive semiconductor, then atomically join another active layer to carry out the catalytic reactions, all in a single process." The first layer of the film consisted of a nanocrystalline form of cobalt oxide that provided a stable, physically robust interface with the light-absorbing semiconductor. The other layer was a chemically reactive material made of cobalt dihydroxide. "The design of this composite coating was inspired by recent advances in the field that have revealed how water-splitting reactions occur, at the atomic scale, on materials. In this way, mechanistic insights guide how to make systems that have the functional properties we need," said Sharp. Using this configuration, the researchers could run photosystems continuously for three days—potentially longer—when such systems would normally fail in mere seconds. "A major impact of this work is to demonstrate the value of designing catalysts for integration with semiconductors," said Yang. "Using a combination of spectroscopic and electrochemical methods, we showed that these films can be made compact and continuous at the nanometer scale, thus minimizing parasitic light absorption when integrated on top of photoactive semiconductors." The study authors noted that while this is an important milestone, there are many more steps needed before a commercially viable artificial photosystem is ready for deployment. "In general, we need to know more about how these systems fail so we can identify areas to target for future improvement," said Sharp. "Understanding degradation is an important avenue to making something that is stable for decades."

10.1038/nmat4794

Earth

The global 'plastic flood' reaches the Arctic

Melanie Bergmann et al, Plastic pollution in the Arctic, Nature Reviews Earth & Environment (2022). DOI: 10.1038/s43017-022-00279-8

https://dx.doi.org/10.1038/s43017-022-00279-8

https://phys.org/news/2022-04-global-plastic-arctic.html

Abstract Plastic pollution is now pervasive in the Arctic, even in areas with no apparent human activity, such as the deep seafloor. In this Review, we describe the sources and impacts of Arctic plastic pollution, including plastic debris and microplastics, which have infiltrated terrestrial and aquatic systems, the cryosphere and the atmosphere. Although some pollution is from local sources — fisheries, landfills, wastewater and offshore industrial activity — distant regions are a substantial source, as plastic is carried from lower latitudes to the Arctic by ocean currents, atmospheric transport and rivers. Once in the Arctic, plastic pollution accumulates in certain areas and affects local ecosystems. Population-level information is sparse, but interactions such as entanglements and ingestion of marine debris have been recorded for mammals, seabirds, fish and invertebrates. Early evidence also suggests interactions between climate change and plastic pollution. Even if plastic emissions are halted today, fragmentation of legacy plastic will lead to an increasing microplastic burden in Arctic ecosystems, which are already under pressure from anthropogenic warming. Mitigation is urgently needed at both regional and international levels to decrease plastic production and utilization, achieve circularity and optimize solid waste management and wastewater treatment. Key points The widespread plastic pollution in the Arctic originates from both local and distant sources. Concentrations of plastic in the Arctic vary widely, with greater accumulation in certain hotspots, but are generally similar to those of more densely populated regions. Plastic has infiltrated all levels of the Arctic food web, including many endemic species, with largely unknown organismal impacts. In the fast-changing Arctic, plastic pollution adds to the effects of climate change in terms of growing sources, transport processes, potential feedback loops and ecological consequences. Mitigation of both local and distal plastic pollution is needed to prevent further ecosystem degradation. Introduction Industrial plastic production has grown rapidly since the 1950s, reaching 368 million tonnes globally per year by 2019 (ref. 1 ). Because of its low price, plastic has become one of the most widely used materials, especially in the packaging industry, and now forms an integral part of municipal waste. Every year, 19–23 million metric tonnes of mismanaged plastic waste are transferred from land-based sources to water globally 2 . As plastics are designed to be durable, they persist in the environment for long periods of time and become widely dispersed 3 , 4 . Therefore, plastic is a planetary boundary threat 5 , 6 , especially once it fragments into microplastic and nanoplastic (size ≤5 mm and ≤1 μm (ref. 7 ), respectively) due to sunlight, mechanical abrasion, biotic interaction, wave action 8 and temperature fluctuations. Plastic is also extensively used in maritime operations such as fishing, aquaculture, shipping and offshore operations, leading to substantial additional leakage into marine environments. Although millions of tonnes of plastics enter the oceans annually, it is currently unknown where in the ocean 99% of the small plastic debris ends up 9 , pointing to yet unaccounted for accumulation areas 10 . Polar regions are still perceived as pristine. However, in the past five years, high levels of plastic pollution have been found in the Arctic (Fig. 1 ). The formation of a sixth accumulation area in the Nordic Seas has been suggested by model projections 11 and is corroborated by an increase in marine debris over time 12 , 13 and comparatively high microplastic concentrations in the Arctic 14 , 15 . This evidence has prompted a Protection of the Arctic Marine Environment (PAME, Arctic Council) working group desktop study on marine litter and microplastics in the Arctic 16 to gauge the need for a Regional Action Plan, which, in turn, led to a mandate to assess the status and trends by the Arctic Monitoring and Assessment Programme 17 . Fig. 1: Overview of the pathways of plastic pollutants into the Arctic Ocean from local and distant sources. Plastic pollution can be generated by households, traffic, agriculture, wastewater treatment, landfills, illegal dumping, industry, shipyards, tourism, ships, fisheries and offshore industry, and be transported to and/or within the Arctic via the atmosphere, rivers, ocean currents, sea ice and eroding permafrost. The seafloor and sea ice are areas of plastic accumulation. The numbers in boxes refer to the abundance of plastic debris (green) or microplastics (MP, purple) in different ecosystem compartments. The ranges are based on data from 36 peer-reviewed studies reporting from 727 locations that were compiled in the database Litterbase 84 (more details on the data extraction process are provided in the Supplementary Information ). The data in each compartment were converted to common units here, but the sampling and analytical methods used in different studies varied widely, as there are currently few standardized or harmonized procedures. For example, varying size detection limits in different studies likely introduced considerable variability in the ranges shown. Figure is adapted from AWI-Infographic, CC BY 4.0 ( ). Full size image In this Review, we describe the sources of Arctic plastic debris, its distribution and its effects on Arctic biota, as well as knowledge gaps and mitigation with a broad, pan-Arctic view, complementing previous reviews focused on plastic pollution effects on Arctic biota 18 , such as seabirds 19 , or with differential geographic focuses 20 , 21 , 22 . We also discuss interactions between climate change and plastic pollution, as plastic pollution likely adds to the impacts of climate change, which has caused a three times faster increase in Arctic temperatures compared with the global average 23 . Sources of Arctic plastic debris As much of the Arctic is sparsely inhabited, relatively low local plastic pollution inputs would be expected. Yet, there are widespread observations of plastics in the region. Most model simulations and data suggest that a substantial proportion originates from the North Atlantic 9 , 24 and the North Pacific 25 , 26 (Fig. 1 ). Rivers were also suggested as a pathway of microplastic pollution in the Arctic 27 . Although the Arctic Ocean contains only ~1% of the global ocean volume, it receives >10% of the global river discharge 28 . Transport of plastic pollutants to and in the Arctic is governed by processes from large-scale ocean currents to small-scale phenomena, such as windrows and sea ice drift 29 , 30 (Fig. 2 ). Model simulations and data from global studies on microfibres suggest that some regions of the Arctic are accumulation areas for plastic pollutants 11 , 14 , 15 . In order to support the design of efficient regulatory schemes to mitigate plastic pollution, it is common to distinguish between land-based and sea-based sources from both local and distant origins, as discussed here. Fig. 2: The main pathways of pollution transport to the Arctic. Plastic pollution is transported to the Arctic via atmospheric and aquatic circulation systems, which could promote their accumulation in certain areas. The main ocean currents that move pollution to and within the Arctic are shown as thin red, blue and green arrows, and the ten largest rivers that release 10% of the global river discharge into the Arctic Ocean are illustrated by thick blue arrows. Numbers in parentheses refer to average annual discharge in km 3 (ref. 198 ). The prevailing atmospheric circulation pattern is shown as translucent arrows. The solid and dashed blue lines indicate the main Arctic river basin and watershed, respectively. Figure adapted with permission from ref. 198 , Elsevier. Full size image Local sources of plastic include the key sectors of maritime activity in the Arctic, such as hydrocarbon exploration, aquaculture and ship traffic, including cruise tourism and fisheries (Fig. 1 ). For example, abandoned, lost or otherwise discarded fishing gear is a major source of plastic debris, especially in the Greenland, Norwegian and Barents Seas 31 , 32 , Kara Sea 33 and subarctic North Atlantic 34 , 35 and North Pacific oceans 36 , 37 . On the beaches of Svalbard, plastic debris from fisheries accounted for 27–100% of beach litter 38 , 39 , 40 , 41 . Fisheries are also an important source at Novaya Zemlya, especially in terms of strapping bands 42 , and at Franz Josef Land, Barents Sea, where they accounted for 51% of the debris 31 , although they do not appear to be major sources in the Canadian Arctic 43 . Recognizable items from the Eurasian Arctic originated mostly from Russian and Scandinavian trawlers but also from the UK, Iceland, Faroe Islands, the Netherlands, Germany, Italy, Spain, Canada, Argentina, Brazil and the USA 16 , 32 , 39 , 44 . Fibres or threads from fishing nets were the most important source of microplastics in the Barents Sea 45 and the second most abundant type of microplastics in southwest Greenland 46 . Notably, 80–90% of the fishing nets found on Svalbard had been discarded deliberately by fishers after mending nets 32 . Much of the material used is positively buoyant, such that it drifts and washes ashore. Some of the items could also come from the intensifying aquaculture, but it is difficult to differentiate between fishing and aquaculture sources. Fisheries regulations such as conservation zones and fishing permits reduce the number of fishers operating in an area and can help to reduce fisheries-related debris, as shown in the 1980s in Alaska 47 , 48 . Another source is plastic debris from domestic sources, as evidenced by reports of bottles, containers, plastic bags and fabrics 31 . However, because such items are also used on ships, it is difficult to attribute such plastics to land-based versus sea-based sources, and input from sea-based sources was rated more important than land-based sources in the Arctic 24 . For example, large food containers amongst the household plastics found on northwest Svalbard point to the disposal of galley waste, which is of sea-based origin 32 . Litter quantities on the seafloor of the Fram Strait have been correlated with increasing activities in both the fisheries and the tourism sectors west of Svalbard 49 . The prevalence of fast-sinking glass debris on the deep Arctic seafloor also corroborates the importance of local sea-based sources 12 . Arctic ship traffic is due to increase as new and faster trans‐Arctic routes open, and the shipping season extends as sea ice declines 50 , potentially leading to increased local plastic inputs. A major challenge to minimizing the input of waste from land into the ocean globally is the lack of adequate waste management facilities in coastal regions 51 . As Arctic population densities are low, waste collection and disposal is very basic. Recycling and baling facilities are rare and limited to large Arctic communities. Waste collection in larger communities often relies on community haul systems, whereas in small communities, it is typically by self-haul 52 , which can be less efficient in preventing waste leakage into the environment. In some communities, traditional waste management solutions are landfills and uncontrolled dumpsites, sometimes next to the sea, and simple incinerators with no or limited flue gas treatment 53 , as seen in Greenland 54 and Iqaluit, Canada 46 , 55 . Beach litter assessments 56 report input from inadequate waste facilities on the western shores of Greenland, where 90% of the Greenlandic population lives. In the Canadian Arctic, plastic litter densities were seven times higher near communities compared with more remote locations 43 . Open dumpsites and winter travel activities were identified as potential sources 43 . Numerous open waste disposal sites and abandoned landfills were also identified as an important source of plastic pollution distributed over the flat tundra by high winds of the Archangelsk region of Russia 57 . Microplastics are also widely distributed in the Arctic, transported by ocean and atmospheric currents and biota from both distant and local sources. Microplastics are either manufactured directly, for example, as pre-production pellets and microbeads, or formed through weathering and breakdown of larger plastic items. Data from the east Canadian Arctic suggest primarily distal sources of microplastic 46 , 58 or a combination of distant and local sources 46 . Substantial quantities of microfibres are found in sediments from the Canadian Arctic (1,930 fibres per kg dry weight), 51% and 20% of which were acetate cellulose and indigo denim, respectively, indicating long-range transport from southern wastewater source regions 59 . In other regions, local sources play a prominent role. High concentrations of microplastic in surface waters off west Greenland likely originate from the capital Nuuk, which harbours 18,000 inhabitants 60 . One local source could be effluent from sewage and wastewater treatment, which is often only mechanically treated or not treated at all in Nuuk or Svalbard 60 , 61 , 62 . Indeed, large quantities of microplastic fibres are shed during washing of synthetic textiles 63 , which are disproportionately much worn in cold polar regions and can leak into the ocean through inadequately treated wastewater. Local wastewater could also be one of the sources of microplastic in the White Sea basin 57 . Six million microlitter particles per hour were emitted into the ocean (≥100 µm, ~1,500 particles m −3 ) by a wastewater treatment plant in Reykjavik, Iceland, that only used mechanical treatment 64 . Exceptionally high levels of microplastic were also recorded from a sandy beach near Reykjavik 65 , Iceland, which is located near a harbour and waste management facility. Therefore, even adequate waste management systems can act as sources if located close to the shore. Still, the introduction of mechanical and biological treatment at a wastewater treatment plant in Ny-Ålesund, Svalbard, has cut anthropogenic microparticle emissions by 99%, highlighting that systems are available to reduce further emissions from Arctic communities 66 . Other potential but poorly constrained local sources of microplastic include particles shed from ship paint, skidoos and other vehicles used on ice, as well as grey water released by rising numbers of ships operating in the area 67 . Paint-derived fragments were found in southwest Greenland 46 and dominated microplastic in water samples from a National Wildlife Area on Baffin Island, hundreds of kilometres away from any major settlement, highlighting both local and distant sources in these coastal areas 68 . The expanding hydrocarbon industry could be another source of litter and microplastic — tube-dwelling worms and sediments taken near oil and gas platforms in the North Sea bore significantly higher microplastic burdens than those collected further away, especially the viscosity-enhancer polyacrylamide 69 . However, quantitative information on microplastic inputs from shipping and the hydrocarbon industry is lacking for the Arctic region. Distribution and transport Buoyant plastic can float with ocean surface currents to higher latitudes 24 , 70 , 71 , 72 , with most plastic transport into the Arctic from the Atlantic 24 and modest transport of microplastic through the Bering Strait 73 (Fig. 2 ). Surface transport is accelerated by storms through wave-driven Stokes drift 74 or direct windage 75 . Mesoscale eddies also affect the transport of debris or other materials 76 , 77 , 78 , 79 , 80 , as can subsurface transport of less buoyant plastic at depths below 50 m (refs 81 , 82 ). Biota can disperse plastic debris through ingestion, migration and egestion 83 . Some of the floating macroplastic becomes intercepted by uninhabited Arctic beaches of Svalbard 39 , 40 , 41 , the Novaya Zemlya archipelago 79 , the Russian Far East 44 , Alaska 37 , Arctic Canada and west Greenland 43 at quantities ranging from 200 to 498,000 items km −2 or from 8,830 to 523,680 kg km −2 in terms of mass 84 . Much less is known about the transport processes of plastics within the Arctic because of scarce measurements. The available data show that plastic debris (0–7.97 items km −2 ) 13 , 72 and microplastics (0–1,287 particles m −3 ) 27 , 81 , 85 are widely distributed in Arctic surface waters (Fig. 3 ). Because of pollution transport from both the south (North Atlantic Current) and the north (Transpolar Drift), plastic quantities are likely higher in the Eurasian basin 73 , which is corroborated by less weathered plastic microfibres and three times higher microfibre concentrations in the western Arctic 86 . However, more field data are needed to verify the lower concentrations on the Amerasian side. There, Pacific water does not spread over the whole Arctic Basin, as it circulates primarily around the Beaufort Gyre before leaving with the Atlantic water via the Canadian Arctic and past west Greenland 77 (Fig. 2 ). Still, it has been suggested that, during this transport, microplastic from the North Pacific enters the western Arctic, concentrates in the Beaufort Gyre and is carried to the central Arctic and Eurasian basins 26 . Fig. 3: Plastic pollution recorded in different Arctic ecosystem compartments. Plastic pollution is widely spread in different ecosystem compartments of the Arctic Ocean. All yellow symbols refer to locations at the sea surface, water column, sea ice, seafloor and beaches where plastic pollution was recorded in 62 peer-reviewed studies compiled in the database Litterbase 84 (more details on the data extraction process are provided in the Supplementary Information ). White symbols refer to locations where no litter was observed. Grey symbols refer to locations where other types of litter (but no plastic items) were observed. Full size image High microplastic loads in Arctic sea ice (31.75–12,000,000 particles m −3 ) 29 , 58 and models both suggest that sea ice drift supports basin-scale transport of ice-rafted plastic 26 , 73 , 87 , 88 . For example, during the formation of sea ice in the Kara and Laptev Seas, microplastic from the sea surface becomes entrained in the ice matrix. In spring and summer, the sea ice breaks up and microplastics travel with ice floes to the Fram Strait via the Transpolar Drift 26 , 29 (Fig. 2 ), where the ice melts and releases its legacy to the water. The presence of ice algae and sticky extracellular polymeric substances in sea ice 89 could enable heteroaggregation of particles and, thus, promote their sinking to the seafloor 90 , as could ballasting via sea-ice-derived cryogenic gypsum from under-ice Phaeocystis blooms 91 . These mechanisms could be one reason for the high quantities of microplastics (6,595 and 13,331 particles kg −1 sediment) observed in the Fram Strait near the marginal ice zone 81 , 92 . Backward drift trajectories of ice cores taken in the central Arctic indicate that they originated from the Siberian shelves, western and central Arctic 29 , 88 or circulate in the Beaufort Gyre 26 . Much of the sea ice is formed in regions 29 that receive water from Siberian rivers (Fig. 2 ). Siberian rivers have huge catchment areas and cross big cities, industrial and agricultural areas, and receive wastewater effluents of unknown treatment level. Even further upstream, the Ob’ and Tom rivers already contain high microplastic concentrations (44.2–51.2 particles m −3 ) 93 . The Severnaya Dvina river plays a major role in the transfer of microplastics to the White Sea 45 and river discharge was identified as the second largest source of the microplastic pollution in the Eurasian basin 27 . Yet, low levels were reported from three rivers feeding into the White Sea basin (0–6 particles m −3 ) 57 . Furthermore, litter quantities from the Russian Arctic indicate low riverine contributions in autumn 72 . Still, during late spring, when river ice is melting and the greatest discharge into the ocean occurs, pollution levels could be higher. Half of the plastic from municipal waste is denser than seawater 94 and sinks directly to the seafloor. However, even positively buoyant plastic is recorded in the water column and on the seafloor 81 , 95 . Mean quantities of 0.011 mg plastic debris m −3 prevailed in the upper 60 m of the Barents Sea 95 . Waters above the deep Arctic seafloor harboured microplastic concentrations of 0–375 particles m −3 (refs 27 , 81 , 85 ). Although no vertical trend was found in the Arctic Central Basin (0–375 particles m −3 ) 85 , in the Fram Strait, the mean MP concentration decreased sixfold towards 1,000 m depth with profiles similar to those of particulate organic carbon 81 . Hence, biological processes such as incorporation in marine snow, fast-sinking aggregates of ice algae or phytoplankton and faecal pellets likely enhance the vertical flux of microplastic 81 , 96 , along with vertical advection and diffusion in the water column 97 . Three-dimensional modelling of particles from the deep Fram Strait emphasized the importance of lateral advection and settling velocities in the vertical dispersal, with trajectories as long as 653 km (ref. 81 ). Most of the modelled particles likely come from the North Atlantic, but sea ice appears to be a source of microplastic tracked back from the east Greenland slope 81 . Deep-water cascading events such as the Storfjorden overflow in Svalbard could also enhance downward particle flux 98 . Plastic pollution has been recorded from various regions of the Arctic seafloor, including the Norwegian Sea 99 , 100 , 101 , Fram Strait 12 , 49 , 92 , 102 , east Greenland slope 103 , Barents Sea 95 , central Arctic Basin 104 , Bering and Chukchi seas 105 and east Canadian Arctic 58 (Fig. 3 ). However, unlike bottom trawls from the Chukchi and Kara Seas, trawls from the East Siberian and Laptev Seas returned no litter 33 . The former was attributed to fishing activities in the Barents Sea 33 . Quantities of plastic debris on the seafloor range between 0 and 24,500 items km −2 (refs 12 , 49 , 100 , 101 ) and have increased from 813 to 5,970 items km −2 between 2004 and 2017 in the Fram Strait 12 . The absence of light, low temperatures and stable conditions lead to degradation rates that are particularly low in the deep sea, as indicated by 30-year-old plastic recovered from the Sea of Japan without any signs of deterioration 106 . Bottom currents can carry microplastics on the seafloor to accumulation areas that also happen to be biodiversity hotspots 107 . In the Arctic deep sea, microplastic concentrations range between 0 and 16,041 particles kg −1 sediment 58 , 108 and rank amongst the highest measured concentrations globally. Atmospheric transport is also an important transport pathway, as indicated by the presence of microplastic in snow samples from ice floes in the east Canadian Arctic 58 , western Arctic 26 , Svalbard, Fram Strait 109 and Icelandic ice cap 110 ranging from 0 to 14,400,000 particles m −3 (refs 58 , 109 ). Atmospheric transport could also be a pathway to lakes, although early evidence from four lakes in the Archangelsk region of Russia suggests low pollution levels (0–2 particles m −3 ) 20 . As with mercury pollution, atmospheric circulation patterns including the Icelandic Low, North American High, Aleutian Low and Siberian High could carry air masses with microplastic and nanoplastic from urban eastern and western Europe, North America, East Asia and Siberia to the Arctic, where they can fall out by wet and dry deposition and accumulate in the ocean, cryosphere and permafrost 111 (Fig. 2 ). Airborne microplastic emissions from car tyres and brakes could be as high as riverine or direct inputs of these sources to the ocean 112 . Models suggest that tyre-wear particle concentrations in Arctic snow range between 6 and 150 ng kg −1 for particles ≤10 µm and that Greenland and the Arctic Ocean are important receptor regions 112 . The ocean itself also appears to be a conduit of atmospheric transport, as indicated by microplastic in sea spray mist and onshore winds from the open Atlantic 113 . Interactions with Arctic wildlife Pervasive contamination of plastic pollution in the Arctic has led to wildlife exposure to both macroplastic and microplastic pollution (Fig. 4 ). Wildfire and plastics interact through colonization or rafting on marine debris, ingestion, entanglement and smothering, affecting a total of currently 131 species in the Arctic (based on available information as of November 2021) 84 . Interactions can occur both at sea and on land, either with beached debris or with waste from open dumpsites 31 , 39 , 114 . Fig. 4: Interactions between marine debris and Arctic biota. Most interaction records refer to the ingestion of plastic and come from studies in the European and Canadian Arctic, with much of the Arctic region underrepresented in sampling efforts. All symbols refer to locations where interactions such as entanglement (nine species), ingestion (31 species), coverage of biota (five species), rafting on marine debris (72 species) and colonization (96 species) affecting a total of currently 131 species in the Arctic (as of November 2021) were recorded in 46 peer-reviewed studies compiled in the database Litterbase 84 (more details on the data extraction process are provided in the Supplementary Information ). Full size image Ingestion of plastics among Arctic species Ingestion of plastic debris by organisms does not always lead to direct harm, but it creates the potential for malnutrition, internal injury, obstruction of the intestinal tract causing starvation or rupture, and potentially death 83 . Plastic ingestion has been reported across various regions of the Arctic (Fig. 4 ) across several levels of the food web, including in zooplankton from the east Canadian Arctic 58 and the Fram Strait 115 . A range of other marine invertebrates also ingest microplastic such as sea anemones, starfish, brittlestars, shrimps, crabs, whelks, bivalves 116 , 117 , 118 , amphipods 119 and tube worms 69 . Plastic has been found in Arctic fish such as sculpin ( Triglops nybelini ) 120 , saithe ( Pollachius virens ) 121 , polar cod ( Boreogadus saida ) 120 , 122 , Atlantic cod ( Gadus morhua ) 121 , 123 , 124 and Greenland shark ( Somniosus microcephalus ) 125 , 126 . Because fish are indicators of ecosystem health, important links in Arctic food webs and part of the human diet, further research on plastic contamination in Arctic fish is warranted. Seabirds are amongst the most studied biota in terms of plastic pollution, both globally and in the Arctic. Early reports of plastic ingestion by herring gulls ( Larus smithsonianus ) and parakeet auklet ( Aethia psittacula ) date back to the 1970s 19 . A total of 51 species of seabirds breed in the Arctic region and the ingestion of plastic is widespread among them 19 . It was common among 12 seabird species from the Russian Arctic, for instance, ~60% of Chaun Bay gull nests containing boluses with plastic likely from a nearby dumpsite 31 . The northern fulmar ( Fulmarus glacialis ) is the most widely studied species for plastic ingestion in the Arctic and globally, and has been sampled in a handful of Arctic regions repeatedly since 2001. Plastic ingestion levels vary with latitude, with fulmars sampled closer to the pole having lower levels (87% of the birds examined) 127 then their counterparts from other regions 128 , which could reflect lower pollution levels in their feeding grounds. There are only a few records of plastic ingestion by Arctic mammals 18 , most of which are from whales, including sperm whales ( Physeter macrocephalus ) 129 , belugas ( Delphinapterus leucas ) 130 , fin whales ( Balaenoptera physalus ), bowhead whales ( Balaena mysticetus ) 131 and Stejneger’s beaked whales ( Mesoplodon stejnegeri ) 132 . Only a handful of pinnipeds (seals, sea lions, walrus) have been examined in the Arctic region. No plastic pieces above 425 µm were detected in the stomachs of ringed seals ( Phoca hispida ), bearded seals ( Erignathus barbatus ) and harbour seals ( Phoca vitulina ) 133 . Similarly, no plastic pieces larger than 5 mm were found in harp seals ( Pagophilus groenlandicus ) in Greenland, but two plastic sheets were reported in a 20-day-old hooded seal pup ( Cystophora cristata ) from the Greenland Sea 134 . Seventy percent of walrus faeces in Svalbard contained microfibres larger than 1 μm (ref. 135 ). Although current knowledge suggests relatively low plastic ingestion levels of mammals overall, no firm conclusion can yet be drawn from the current data. Plastics as a vector of chemicals Plastic ingestion can expose organisms to harmful legacy pollutants from the environment or chemicals added during manufacturing 136 (Fig. 5 ). Consequently, there is a large body of work on plastics as a vector for chemicals to wildlife. In the Arctic, biota have been monitored for decades for environmental contaminants, including metals such as mercury and persistent organic pollutants (POPs). Although there are some indications that metals and POPs typically found in the environment are positively correlated with plastic ingestion in seabirds in non-Arctic regions 137 , POP levels have, so far, not been linked with plastic levels in Arctic species 138 , 139 . However, research on northern fulmars suggests that ingested plastic can be a route for a congener of polybrominated diphenyl ethers 140 . More work is needed on the transport and fate of these contaminants to determine whether plastics are an important vector. Fig. 5: Arctic food web and biotic interactions with plastic pollution. Invertebrates, fish, birds and mammals in the Arctic have been examined for plastic ingestion (indicated by coloured symbols) and have been reported to become entangled in plastic litter. Although ingested microplastics have been found across several taxa, seabird species that feed at the sea surface are potentially the most vulnerable to accumulating plastic pollution. Adapted from an image courtesy of Julia Baak. Full size image An area of emerging concern in the Arctic is the effect of plastic additives, chemicals directly linked to plastic pollution 141 . For example, ultraviolet (UV) stabilizers and substituted diphenylamine antioxidants — both plastic additives — were detected in ringed seals, northern fulmars and black-legged kittiwakes ( Rissa tridactyla ) from the Canadian Arctic 142 . These additives were also detected in seabird eggs from Alaska and northern Canada, indicating transfer to the next generation 142 , 143 . Although the effect of the small concentrations on seabirds is unknown, given that phthalates, UV stabilizers and substituted diphenylamine antioxidants are endocrine disruptors, more work is needed to understand how even small amounts affect the developmental stages of Arctic biota. Important to consider in the Arctic region is that many wild species are harvested. Future work should focus on examining plastic additives in consumed species to inform links to human health as well as the health of Arctic biota. Effects of plastic debris on Arctic wildlife Entanglement in plastic debris can have deleterious effects, such as injury, restrained movement, starvation, strangulation and suffocation if air-breathing animals cannot return to the sea surface 83 . Entanglement has been reported for Arctic terns on Svalbard ( Sterna paradisaea ) and seven other seabird species in the Russian Arctic 31 , 39 . Thirteen seabird species have also been found to incorporate plastic debris in their nests 31 , which can cause entanglement. Notably, almost all nests of two of the existing northern gannet ( Morus bassanus ) colonies at the Murman coast and 10% of an ivory gull ( Pagophila eburnea ) colony from the Kara Sea contained plastic 31 . Polar bears ( Ursus maritimus ), Arctic foxes ( Alopex lagopus ), bowhead whales, reindeer ( Rangifer tarandus ), bearded seals, harbour seals, Greenland halibut ( Reinhardtius hippoglossoides ), Atlantic cod and snow crabs ( Chionoecetes opilio ) 31 , 39 , 144 , 145 , 146 , 147 also experience entanglement. Plastic debris can also act as a raft to transport animals from one location to another. Six percent of the plastic items stranded on Svalbard were colonized by bryozoans ( Membranipora membranacea ) and barnacles ( Semibalanus balanoides ) 148 . Macroalgae, bryozoans, barnacles ( Semibalanus sp., Lepas anatifera ) and blue mussels ( Mytilus sp.) also inhabited beach debris on Svalbard 41 . Rafting of adult groups could favour dispersal over larval transport and be one of the drivers behind the reappearance of Mytilus after 1,000 years of absence 149 . The invasion of xenobionts via rafting can have population-level or community-level effects 150 , posing a potential threat to Arctic ecosystems. Ecological effects of plastic Because of the widespread contamination of plastic pollution in wildlife, there is urgency to answer questions related to ecological impact 150 , 151 . In non-Arctic systems, there is overwhelming evidence of detrimental effects from macroplastics to individuals and compelling evidence for effects to populations, communities and ecosystems 150 . For microplastics, impacts have been demonstrated across several levels of biological organization 150 , 152 , including oxidative stress 153 , changes in gene expression 136 , 154 , inflammation 155 and reduced growth 156 and reproduction 157 rates. Although these effects could apply to closely related Arctic species, too, there has been little research on the ecological effects of plastic debris in Arctic ecosystems, which are already under stress due to climate change 158 . One of the few studies available on the effects of plastic on benthic species is in the deep Fram Strait, where 45% of the plastic debris observed showed interactions with epibenthic megafauna, such as entanglement in up to 31% of the sponge colonies 12 . Although data on effects are lacking in this case, entangled fishing gear caused tissue abrasion and (partial) mortality in sponges from Florida, rendering the organisms more susceptible to pathogens, predation and overgrowth 159 . As with cold-water coral 160 , coverage of the sponge’s feeding apparatus could impair water-exchange processes, prey capture and growth. Another frequent observation was the colonization of plastic debris by sessile biota such as sea anemones 12 , 103 , which affects diversity. In general, the presence of plastic debris in benthic sediments can alter community structure 161 . Plastic items covering sediments can also affect biogeochemical processes, which could alter bottom-dwelling communities, as shown in an intertidal zone in Ireland with anoxic conditions, reduced organic matter and lower densities of sediment-inhabiting invertebrates nine weeks after coverage with plastic bags 162 . Although sediments from the Fram Strait and Canada contain up to 13,000 and 16,000 small-sized microplastics kg −1 sediment 58 , 81 and are, thus, amongst the most polluted in the world, the effects on deposit-feeding organisms such as sea cucumbers, nematodes or other worms are currently largely unknown. Sea ice also harbours high concentrations of microplastics 29 , which likely affect this ecosystem. Experimental evidence suggests that the presence of microplastic reduces the colonization of already formed sea ice by ice algae, a process that is important to transfer sea ice species from multi-year to first-year ice 90 . If added during the process of ice formation, however, microplastic did not affect algal concentrations in sea ice. Data on contamination are often collected before digging deeper into effects. Here, we suggest that it is time for a new research priority: understanding the effects of plastics in the Arctic across organismal and ecosystem scales. These efforts are especially important, as the Arctic is vulnerable to a combination of many stressors (for instance, fast warming and a sink for organic pollutants), and the addition of microplastics raises concern about multi-stressor effects to wildlife. Plastic pollution and climate change Although they are often thought of separately, climate change and plastic pollution are directly and indirectly linked, and both are amongst the biggest ecological challenges faced today globally and in the Arctic (Fig. 6 ), not least they share the same fossil origin, oil and gas. Global heating is three times faster in the Arctic compared with the rest of the planet 23 , such that Arctic ecosystems are already under severe stress 158 . One of the most prominent effects of climate change is the melting of the cryosphere. Sea ice entrains microplastic during its formation 90 , 163 and releases it during melting 26 , 29 , 61 . Changes in ice properties and its distribution will, therefore, affect the levels and spatial distribution of microplastics in the environment. Increasing quantities of released plastic particles in the water column, along with extracellular polymeric substances from ice algae 90 , could promote the formation of heteroaggregates, affecting the nutrient availability and turbidity in habitats of cyanobacteria and phytoplankton communities 6 . A decline in their populations could reduce the sequestration of carbon from the atmosphere and, thereby, fuel climate change instead 6 , 164 . On a smaller scale, a positive correlation has been found between salinity and microplastic concentrations in sea ice brine 90 , 163 . The microplastic levels reported in Arctic sea ice could increase the albedo effect by 11% and alter both the permeability of sea ice and the absorption of solar radiation, with a feedback on sea ice melting 29 , 163 . However, it is also conceivable that high concentrations of particles darker than the cryosphere promote solar absorption and, thus, melting. Fig. 6: The interaction between climate change and plastic pollution. Climate change and plastic pollution are interconnected. Several meteorological or physical impacts of climate change are known to influence the concentrations and distribution of plastic in the world, at different scales. All of these lead to an increase in plastic concentration, at least locally. Blue boxes refer to processes specific to polar regions. This figure highlights the complexity of those interconnections and how two major anthropogenic challenges are influencing each other. Figure adapted with permission from ref. 167 , Society of Environmental Toxicology and Chemistry (Wiley). © 2017 SETAC. Full size image In the atmosphere, airborne microplastic and nanoplastic can also enhance ice nucleation and, thereby, cloud formation and climate change 165 if they contribute to atmospheric trapping of infrared radiation from the Earth surface, instead of enhancing the reflection of sunlight. This process is important for the hydrological cycle, as more than 50% of the Earth’s precipitation is induced in the ice phase 165 . Through atmospheric fallout and glacial meltwater, microplastics could also penetrate and affect permafrost, and be released to rivers and the Arctic Ocean with accelerating permafrost thaw 166 . Airborne microplastics have also infiltrated snow on glaciers, potentially affecting their light absorbance, structural and general rheological properties, and could, thereby, promote the ongoing fast melting of glaciers, the greatest cause of rising sea levels 110 . Growing inputs of freshwater to the Arctic Ocean lead to a decrease in the relative buoyancy of plastics debris 167 and a weakening thermohaline circulation 168 , which could eventually slow down the poleward transport of plastic pollution (Fig. 6 ). Global warming also amplifies poleward winds 169 , which define convergence zones and surface currents and, thus, influence plastic transport, as convergence zones are accumulation areas for plastic debris 10 . Furthermore, higher wind speeds promote the vertical mixing of small plastics into deeper waters 170 . In addition, warming surface waters result in a higher frequency of storms 171 , which break up the sea ice and enhance melting 172 . Sea level rise and storm events bring about higher inputs of plastic debris from land to the ocean via water runoff 173 and wind transport. Over time, these processes could also lead to higher pollution levels in the Arctic Ocean 11 , 15 , 24 . In addition to direct effects, there are many indirect links between plastic pollution and climate change. For example, climate change causes a decrease in the sea ice thickness and extent 174 . As a result, maritime traffic in the Arctic is on the rise 175 , leading to higher levels of plastic pollution, for example, from fishing vessels, merchant shipping or tourist activities 49 . Plastic production also fuels climate change, as it accounts for 6% of the global oil consumption and could reach 20% by 2050 (ref. 176 ). Fossil-based plastics produced in 2015 emitted 1.8 gigatons of equivalent CO 2 over their life cycle 177 . Under the current trajectory, plastic-related CO 2 emissions could rise to 6.5 gigatons by 2050, which will accelerate climate change and could use up 10–13% of the remaining SR15 carbon budget of 570 gigatons to limit warming to a 66% chance of staying below 1.5 °C (ref. 178 ). Furthermore, greenhouse gases such as methane, ethylene, ethane and propylene are released during degradation of some common plastic polymers throughout their lifetime 179 . Polyethylene, the most produced plastic polymer 1 , releases the highest levels of methane and ethylene. Once initiated by solar radiation, such as in the surface ocean, this process continues in the dark 179 . The scale of greenhouse gas emissions from these processes are currently unknown. Mitigation Plastic pollution is a transboundary problem, especially in the Arctic, where it stems from both distant and local sources. The problem, thus, needs to be tackled both regionally and internationally. Plastic pollution is a function of increasing plastic production coupled with inadequate waste management. Therefore, an effective upstream reduction in the global production of plastic waste via binding targets set in international treaties similar to the Paris Agreement or Montreal Protocol 2 , 180 is warranted. In addition, a circular use of plastic and of sustainable and truly biodegradable alternatives are needed alongside improved municipal waste collection and management to help reduce leakage to the environment 2 , 180 . Manual clean-ups on shorelines, harbours and riverbanks can help to mitigate pollution if impact assessments show that benefits outweigh environmental cost 181 , such as disturbance and increased mortality of biota due to incidental by-catch caused by non-selective removal technologies or operational greenhouse gas emissions. Emissions from sea-based sources lead directly to marine pollution because of the direct input pathways. As much of the plastic debris in the Arctic region stems from local and distant commercial fisheries, mitigation in this sector would reduce plastic pollution particularly efficiently. Gear-marking schemes can prevent fishing gear loss and discarding 182 , along with incentives for adequate waste disposal 183 . Programmes for reporting and recovery of lost fishing gear are already in place in Norway and should be extended to other regions 184 , as should be schemes to recycle fishing gear, which are currently practiced in Iceland. In the long run, the use of fully biodegradable material for nets 185 , 186 , along with bans on particularly short-lived components, such as dolly ropes, that become abraded during a trawl’s passage on the seafloor, could help to reduce leakage to the environment. Education awareness campaigns designed for fishers, for example, during mandatory sea survival courses, help to shift perception in the industry but must be accompanied by institutionalized and well-organized waste facilities at fish landings and harbours to foster behavioural change 184 . The disposal of plastic in the Arctic Ocean and adjacent areas could be reduced through improved port reception facilities following a regional reception facilities plan, as is currently underway under the International Maritime Organization in the Pacific region. Lower harbour fees for ships with better waste facilities on board, a ‘No Special Fee’ system similar to HELCOM 187 and on-port recycling hubs could help to alleviate illegal dumping of waste at sea. Given that ship traffic has already increased and will further increase in the Arctic due to vanishing sea ice, this sector deserves particular attention, including improved surveillance schemes. In many locations throughout the Arctic, open landfills are still in use 56 , and it is clear that investments in local waste management solutions will reduce the leakage of plastic pollution to the environment. Rural Arctic communities that desire efficient waste collection and management schemes need financial and logistical support, for instance, through extended producer responsibility schemes or governments to establish or improve waste management and treatment. Importantly, coupled with community-based monitoring programmes 57 , sources and effectiveness of policy changes can be detected at the local scale relatively quickly 16 . Waste management studies and investments must be a priority to stem the tide of plastics from sources within the Arctic. Reducing emissions from diffuse sources is necessary but challenging. Improved material design could reduce emissions from automotive vehicle tyres and brakes, which is one of the most important sources of microplastics globally 112 , as well as from ship paint from (ice-breaking) vessels. Collection schemes of road runoff could mitigate some of the pollution as well. New regulation aimed at improvements of wastewater treatment on land, offshore and on ships could help reduce inputs of plastic microfibres. Finally, communication and community action are needed. Global audiences must be taught about plastic pollution in the Arctic, as distant sources contribute to the plastics burden in the Arctic. It is important to include local voices in both research 46 and actions aimed at reducing plastic pollution. Listening to indigenous voices has been recognized as a critical part of communication strategies under the Arctic Council 188 . For many, plastic pollution is affecting their way of life. In northern Canada, the community focus on understanding plastic pollution in the Arctic is illustrated by the variety of community-based research programmes on litter and microplastics funded under the Northern Contaminants Program. For this reason, a course including plastic pollution as a contaminant in the Arctic has been taught at Nunavut Arctic College in Iqaluit, Canada, each year since 2009. The students learn, share stories and knowledge, and participate in local research on plastic pollution. As stated by Aggeuq Ashoona, a college student who participated in this course from Kinngait, Nunavut, “This is affecting Inuit very much […]. To find plastic in their [wildlife] stomach is heart-breaking, because these are our food”. Summary and future perspectives Regardless of its remoteness, plastic pollution has infiltrated the Arctic from the atmosphere to the deep ocean floor, with pollution levels sufficiently high for some regions to be considered accumulation areas 15 , 24 , 81 . Despite recent advances in research, there is still a lack of understanding of the importance of different transport processes within the Arctic and the role of local sources, rivers and the atmosphere. It is clear, however, that plastic pollution exacerbates the impacts of climate change. These effects seem particularly clear in the Arctic, where not only are climate change effects occurring faster than elsewhere 23 but where these changes likely strongly influence the sources and transport of plastic debris, perhaps more so than in other regions. Still, we have barely scratched the surface when it comes to impacts on Arctic life, including human communities in the Arctic, requiring further and urgent research. Plastic pollution research is particularly challenging in the Arctic because of its remoteness, lack of infrastructure and harsh environmental conditions. Conventional scientific sampling is often restricted to summer months and requires the use of aircraft, research bases and/or ice-class ships. Even then, fieldwork can be jeopardized by low visibility, polar bears, ice and low temperatures defying technology. Arctic landscapes are often characterized by coarse sediments, permafrost, snow and/or ice, which lack coherent survey guidelines, and, overall, these environments are currently undersampled 189 . Another common approach to quantify plastic pollution, which is to count litter floating at the sea surface by ship-based observers, is often difficult or impossible due to fog or sea ice, which can also impede sampling by surface trawls. These examples highlight that we currently lack the basic methodology to determine pollution levels in certain areas of the Arctic and during significant periods of time. In some areas, these challenges can be overcome by the use of year-round moored sampling devices 190 , drones or collaborative research with citizen scientists 39 , 57 , 109 , 191 or local communities 192 . For example, many scientists work directly with local Inuit communities in the Canadian Arctic to design sampling schemes, sample and interpret results 46 , 133 , 193 . During the COVID-19 pandemic, many researchers in Canada could not access field sites in the Arctic, and, in some cases, local communities were compensated to undertake annual sampling. In Russia, a programme was developed to enable monitoring by local school children and students 57 . Such schemes complement professional science and should be expanded to fill knowledge gaps. In addition to difficulties that arise while conducting fieldwork in the Arctic, there is currently a lack of standardized sampling and analytical methodologies or even harmonized procedures, especially in terms of microplastics. This lack of standardization is concerning, as different analytical approaches can cause several orders of magnitude differences in the results obtained 60 , 194 . Therefore, despite a surge in plastic research in the Arctic, the results are often not comparable between studies, hampering efforts to describe the sources, sinks and large-scale distribution patterns of Arctic plastic pollution. However, the research and monitoring recommendations recently set out by the Arctic Monitoring and Assessment Programme (AMAP) 17 could inform a more harmonized research approach, which would also benefit from a common database for the upload of recorded pollution data. Nanoplastics in the Arctic have largely not been investigated, including their distribution amongst different ecosystem compartments and how they interact with microplastics as the sea ice forms and melts. It is conceivable, for example, that nanoplastic interacts with sea ice in a similar way as, for example, salt and is rejected from the ice matrix as sea ice forms. Data on nanoplastic are particularly important, as particles of this size fraction can pass biological membranes and, thus, translocate to organs, where they could elicit a strong biologic response 195 . Progress in the development of sampling and analytical methods have not only demonstrated the presence of nanoplastic in glacial ice from Greenland but will also help us to fill this knowledge gap 196 . Currently, there are no plastic budget data on relative contributions of various sources of plastic to the Arctic, such as local versus long-distance sources. Current understanding suggests that, along with local emissions, inputs of Atlantic origin could be most important, but data from the Amerasian Arctic have only begun to emerge, so no firm conclusions can yet be drawn. Information on the sources of pollution is needed to assess pan-Arctic exchange — how much plastic debris leaks from North America to Europe and vice versa. As outlined in this Review, such assessments are currently hampered by the lack of harmonized data. Another major knowledge gap pertains to atmospheric transport, which allows microplastic and nanoplastic to infiltrate even the most remote ecosystems on our planet via precipitation. Although this pathway is important for other pollutants such as mercury 111 , its contribution to the Arctic’s overall plastic burden is unknown. Integrating microplastic sampling into research cruises and ongoing air pollution observation programmes could improve our understanding of the role of airborne microplastics 197 . The amount of plastic debris entering the Arctic Ocean through rivers is unclear, but could be important, owing to their enormous catchment areas that lie beyond the Arctic borders, some of which pass through big cities. Arctic rivers are a conduit of land-based plastic pollution into the ocean, and their massive discharge every spring or summer makes the impact potentially substantial. With over 37 million people living along these waterways 16 , understanding plastic pollution in rivers that drain into the Arctic Ocean is crucial. It also increases our knowledge of terrestrial sources, which can help mitigate its input in the long run. Especially as local people depend on freshwater and land for subsistence and culture, understanding the effects of plastic pollution in these systems is a priority. Given the interest in litter and microplastics in northern and indigenous communities, and the breadth of community-based research and monitoring projects across the Arctic, locally designed and implemented projects should be prioritized within research planning strategies 46 , 57 . This strategy will ensure that local and regional research needs are included, and local communities are engaged in result discussions throughout the process and can relay this information as directly into policy solutions as needed. The propagation and impact of microplastic within the Arctic food web (Fig. 5 ), which is already under pressure from fast climate forcing, is another source of major uncertainty. Targeted work that examines plastic pollution throughout the food web is needed in order to understand where plastic pollution accumulates and the actual effects on biota. Although studies have focused hitherto on single species, future studies should take an ecosystem approach, with sampling of biota across trophic levels 153 , and in relation to environmental compartments where they feed 17 . This knowledge will help tease apart questions relating to bioaccumulation, biomagnification, excretion and, thus, cycling of both plastic pollution and contaminants that are both sorbed and derived from plastic pollution. We are also only beginning to investigate the effects of microplastic and nanoplastic on important physical processes, such as soil functions, biogeochemistry, ice properties (melting, UV reflectance and attenuation), weather (condensation, precipitation) and particle flux through the water column (biological pump), all of which have repercussions for the functioning of our Earth system, especially in a changing Arctic. However, it is already clear that effective mitigation is urgently needed to prevent further deterioration of Arctic ecosystems and communities. Change history 06 May 2022 A Correction to this paper has been published:

Even the High North can't escape the global threat of plastic pollution. An international review study just released by the Alfred Wegener Institute shows the flood of plastic has reached all spheres of the Arctic. Large quantities of plastic—transported by rivers, the air and shipping–can now be found in the Arctic Ocean. High concentrations of microplastic can be found in the water, on the seafloor, remote beaches, in rivers, and even in ice and snow. The plastic is not only a burden for ecosystems; it could also worsen climate change. The study was just released in the journal Nature Reviews Earth & Environment. The numbers speak for themselves. Today, between 19 and 23 million metric tons of plastic litter per year end up in the waters of the world—that's two truckloads per minute. Since plastic is also very stable, it accumulates in the oceans, where it gradually breaks down into ever smaller pieces—from macro- to micro- and nanoplastic and can even enter the human bloodstream. And the flood of debris is bound to get worse: global plastic production is expected to double by 2045. The consequences are serious. Today, virtually all marine organisms investigated—from plankton to sperm whales—come into contact with plastic debris and microplastic. And this applies to all areas of the world's oceans—from tropical beaches to the deepest oceanic trenches. As the study published by the Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research (AWI) now shows, the High North is no exception. "The Arctic is still assumed to be a largely untouched wilderness," says AWI expert Dr. Melanie Bergmann. "In our review, which we jointly conducted with colleagues from Norway, Canada and the Netherlands, we show that this perception no longer reflects the reality. Our northernmost ecosystems are already particularly hard hit by climate change. This is now exacerbated by plastic pollution. And our own research has shown that the pollution continues to worsen." The article paints a grim picture. Although the Arctic is sparsely populated, in virtually all habitats—from beaches and the water column, to the seafloor—it shows a similar level of plastic pollution as densely populated regions around the globe. The pollution stems from both local and distant sources. Ocean currents from the Atlantic and the North Sea, and from the North Pacific over the Bering Strait, especially contribute to this. Tiny microplastic particles are also carried northward by wind. Then there are the rivers: though the Arctic Ocean makes up only one percent of the total volume of the world's oceans, it receives more than 10 percent of the global water discharge from rivers, which carry plastic into the ocean, for example, from Siberia. When seawater off the coast of Siberia freezes in the autumn, suspended microplastic becomes trapped in the ice. The Transpolar Drift transports the ice floes to Fram Strait between Greenland and Svalbard, where it melts in the summer, releasing its plastic cargo. Some of the most important local sources of pollution are municipal waste and wastewater from Arctic communities and plastic debris from ships—especially fishing vessels, whose nets and ropes pose a serious problem. Either intentionally dumped in the ocean or unintentionally lost, they account for a large share of the plastic debris in the European sector of the Arctic: On one beach on Svalbard, almost 100 percent of the plastic mass washed ashore came from fisheries according to an AWI study. "Unfortunately, there are very few studies on the effects of the plastic on marine organisms in the Arctic," Bergmann explains. "But there is evidence that the consequences there are similar to those in better-studied regions: in the Arctic, too, many animals—polar bears, seals, reindeer and seabirds—become entangled in plastic and die. In the Arctic, too, unintentionally ingested microplastic likely leads to reduced growth and reproduction, to physiological stress and inflammations in the tissues of marine animals, and even runs in the blood of humans." The available data on potential feedback effects between plastic debris and climate change is particularly thin. "Here, there is an urgent need for further research," says the AWI expert. "Initial studies indicate that trapped microplastic changes the characteristics of sea ice and snow." For example, dark particles could mean the ice absorbs more sunlight and therefore melts more rapidly. In turn, due to what is known as ice-albedo feedback this can intensify global warming. Moreover, plastic particles in the atmosphere provide condensation nuclei for clouds and rain, which means they could influence the weather and, in the long term, the climate. And last but not least, throughout their lifecycle, plastics are currently responsible for 4.5 percent of global greenhouse-gas emissions. "Our review shows that the levels of plastic pollution in the Arctic match those of other regions around the world. This concurs with model simulations that predict an additional accumulation zone in the Arctic," says Bergmann. "But the consequences might be even more serious. As climate change progresses, the Arctic is warming three times faster than the rest of the world. Consequently, the plastic flood is hitting ecosystems that are already seriously strained. "The resolution for a global plastic treaty, passed at the UN Environment Assembly this February, is an important first step. In the course of the negotiations over the next two years, effective, legally binding measures must be adopted including reduction targets in plastic production. "In this regard, the European countries including Germany must cut their plastic output, just as the rich Arctic States have to reduce pollution from local sources and improve the often virtually non-existent waste and wastewater management in their communities. In addition, more regulation and controls are called for—with regard to plastic debris from international shipping, and fisheries."

10.1038/s43017-022-00279-8

Chemistry

A step ahead in pharmaceutical research

Hannes Schihada et al, A universal bioluminescence resonance energy transfer sensor design enables high-sensitivity screening of GPCR activation dynamics, Communications Biology (2018). DOI: 10.1038/s42003-018-0072-0

http://dx.doi.org/10.1038/s42003-018-0072-0

https://phys.org/news/2018-09-pharmaceutical.html

Abstract G-protein-coupled receptors (GPCRs) represent one of the most important classes of drug targets. The discovery of new GCPR therapeutics would greatly benefit from the development of a generalizable high-throughput assay to directly monitor their activation or de-activation. Here we screened a variety of labels inserted into the third intracellular loop and the C-terminus of the α 2A -adrenergic receptor and used fluorescence (FRET) and bioluminescence resonance energy transfer (BRET) to monitor ligand-binding and activation dynamics. We then developed a universal intramolecular BRET receptor sensor design to quantify efficacy and potency of GPCR ligands in intact cells and real time. We demonstrate the transferability of the sensor design by cloning β 2 -adrenergic and PTH1-receptor BRET sensors and monitored their efficacy and potency. For all biosensors, the Z factors were well above 0.5 showing the suitability of such design for microtiter plate assays. This technology will aid the identification of novel types of GPCR ligands. Introduction G-protein-coupled receptors (GPCRs) constitute the largest and the most diverse group of membrane receptors in eukaryotes. They play a role in a plethora of cellular processes. Despite their functional diversity, they share a similar molecular architecture of seven transmembrane helices and a conserved mechanism of activation. Binding of agonist ligands to their cognate receptor leads to a change in the arrangement of distinct transmembrane helices, and a pronounced outward movement of helix 6 (up to 14 Å) 1 , which is then propagated to the third intracellular loop that connects helix 5 with 6, thus enabling the engagement of several downstream signaling pathways, most importantly the activation of G-proteins. We previously demonstrated that such ligand-induced conformational changes can be visualized in living cells by Förster Resonance Energy Transfer (FRET). The average distance between the third intracellular loop and the C-terminus of receptors (e.g., β 2 AR 6.2 nm) 2 is within the range addressable by FRET (2.4–7.2 nm) 3 . Therefore, tagging these conformationally sensitive sites with fluorescent donors and acceptors permits the recording of receptor activation as a change in energy transfer between these chromophores. Following this principle, a series of intramolecular FRET-based GPCR-biosensors have been generated, notably for the α 2A -adrenergic receptor (α 2A AR), employing the fluorescent proteins CFP and YFP (α 2A AR CFP/YFP ) 4 , 5 or CFP with the small Fluorescein Arsenical Hairpin Binder (FlAsH) as fluorophores 6 . These FRET-based GPCR biosensors have become widely used tools 7 and represent the most direct unbiased way to determine the effects of ligands on a given receptor. Their employment has helped to elucidate several aspects of receptor pharmacology and kinetics 8 . However, the combination of donor and acceptor fluorophores used so far suffers from a low signal-to-noise ratio and high fluorescence background which limits their use to single-cell experiments, slowing down the characterization of new pharmacological compounds as GPCR-directed therapeutics. More recently, Bioluminescence Resonance Energy Transfer (BRET) has been tested as an alternative approach to monitor the conformational change of receptors. BRET occurs between proximally situated donor–acceptor pairs (1.6–8.5 nm) 9 , but here a light emitting enzyme luciferase is used as a donor, sidestepping many of the deficiencies associated with direct illumination of the sample. Earlier attempts to generate intramolecular GPCR BRET-based biosensors employing Renilla Luciferase as a donor in combination with GFP or FlAsH as acceptor showed poor agonist-induced BRET changes 10 , 11 , 12 , 13 . Despite the many efforts in optimizing the way in which we study GPCRs and a plethora of methods to assess ligand binding or downstream signaling, the lack of a generalizable assay to monitor directly in living cells receptor activation or de-activation in a format suitable for high-throughput screening is slowing down the progress in discovering new GPCR therapeutics. In the present study, we therefore set out to permutate the well-characterized FRET-based biosensor α 2A AR CFP/YFP in order to develop a universal and versatile sensor design to reach the stringent requirements of monitoring receptor conformational dynamics in intact cells in microtiter plates. Preserving the fluorophore insertion sites in the third intracellular loop and C- terminus, we investigated whether combining different donors and acceptors would improve the transfer of energy to enable the visualization of the receptor’s conformational changes in microtiter plates. Among the 10 FRET-based and the 11 BRET-based α 2A AR biosensors generated in this study, the highest amplitude in the signals upon norepinephrine stimulation was recorded with the BRET-based α 2A AR biosensors combining NanoLuc 14 luciferase as the donor and the self-labeling protein tag Halo labeled with the HaloTag dye NanoBRET 618 as the acceptor. The EC 50 -values calculated for these BRET experiments were in line with binding affinities, demonstrating that the developed BRET biosensor α 2A AR Nluc/Halo assay faithfully reports ligand affinities. This feature was conserved also for a β 2 -adrenergic and a PTH1 receptor BRET-based biosensors. Also for these two receptors, the agonist-induced BRET changes were in line with the binding data suggesting that our receptor biosensor design is a generalizable design that may be use in lieu of the endogenous receptor to determine efficacy and potency of ligands. To assess the applicability of the biosensor design to microtiter plate and its scalability, we performed for each receptor a Z-factor analysis 15 to quantify the quality and reproducibility of the assays. For all receptors analyzed, the Z-factor was well above 0.5—which characterizes an excellent assay suitable to be used in microtiter plates. We propose that this technology will speed the characterization of new pharmacological compounds acting at GPCRs. Results Dynamic range of FRET- and BRET-based α 2A AR biosensors With the goal of generating a GPCR biosensor design to monitor receptor activation in microtiter plates, we permutated the well-characterized FRET-based biosensor α 2A AR CFP/YFP . We preserved the fluorophore insertion sites in the third intracellular loop and C-terminus (Fig. 1a ) but we substituted the original YFP with its brighter variant cpVenus or with self-labeling protein tags: SNAP (α 2A AR CFP/SNAP ; 20 kDa) 16 or Halo (α 2A AR CFP/Halo ; 36 kDa) 17 . Both of these tags can be labeled covalently with exogenous cell-permeable fluorescent dyes. Their diverse emission and excitation peaks (Fig. 1b , Supplementary Fig. 1 ), degrees of spectral overlap with the donor emission (CFP) and photophysical properties impacted the FRET behavior that we observed, allowing us to optimize our approach using a modular fluorophore within the receptor chassis. In this optimization phase, HEK cells were transiently transfected with the different α 2A AR biosensors and, after specific labeling, experiments were performed in 96-well plates (Fig. 1c ). Each FRET-based pair exhibited a different degree of basal transfer of energy—as demonstrated by their respective FRET emission spectra (Fig. 1d ). However, only five of the ten FRET-based pairs tested showed a detectable agonist-induced change in FRET (ΔFRET%) that did not exceed 5% when stimulated by the full agonist norepinephrine (100 µM) (Fig. 1e ). Fig. 1 Evaluation of intramolecular FRET and BRET α 2A -adrenergic receptor biosensors. a Schematic of the biosensor. b Emission peaks of chromophores. c Workflow. d FRET emission spectra of the CFP-label alone or with fluorescent acceptors (each N = 1). e FRET changes (%) induced by norepinephrine normalized for buffer (YFP: N = 7, cpVenus173: N = 4, diAcFAM: N = 6, Oregon Green: N = 5, R110 Direct: N = 3,TMR Direct: N = 6, NanoBRET 618: N = 6, 505-star: N = 5, TMR-star: N = 5, 647SiR: N = 4). f BRET emission spectra of the NanoLuc-label alone or with fluorescent acceptors (each N = 1). g BRET changes (%) induced by norepinephrine normalized for buffer (cpVenus173: N = 7, TagRFP: N = 4, mCherry: N = 4, diAcFAM: N = 3, Oregon Green: N = 3, R110 Direct: N = 3, TMR Direct: N = 3, NanoBRET 618: N = 7, 505-star: N = 3, TMR-star: N = 5, 647SiR: N = 5). Data in e and g show box and whisker plots. Difference was analyzed by two-way ANOVA followed by Bonferroni post hoc test. * p ≤ 0.05 vs. buffer Full size image With the goal of obtaining a higher dynamic range, we tested BRET as possibly more sensitive approach. Here, we used the relatively small-sized NanoLuc 14 luciferase (19 kDa; Nluc) that possesses, in presence of its substrate (furimazine), a narrow bioluminescence spectrum, high brightness, and physical stability. We generated 11 BRET-based α 2A AR biosensors combining the donor Nluc with fluorescent proteins of different color from yellow (cpVenus) to orange-red (mCherry; TagRFP), or with the self-labeling Halo or SNAP tags. Applying the workflow developed above (Fig. 1c ), we tested the ability of the α 2A AR BRET biosensors to report effects of the full agonist norepinephrine in microtiter plate format. All the tested receptor BRET-sensors exhibited a basal energy transfer (Fig. 1f ), while the activation of the receptor induced by norepinephrine at a concentration of 100 μM was detectable in seven of them. The highest amplitude was recorded for Nluc combined with the HaloTag dye NanoBRET 618 18 (α 2A -AR Nluc/Halo 618 ; ΔBRET % = 8.15 ± 0.72) (Fig. 1g )—about two-fold higher than all other 20 FRET- and BRET-biosensors tested. No further improvement in the amplitude of ΔBRET was obtained by swapping the positions of donor and acceptor (Supplementary Fig. 2 ). Pharmacology of the α 2A -AR Nluc/Halo618 BRET-based biosensor We then sought to ascertain that the α 2A AR Nluc/Halo 618 faithfully recorded efficacies and potencies of a panel of known α 2A AR ligands (Supplementary Table 1 ). We created a cell line stably expressing this biosensor, and after optimizing the assay conditions (Supplementary Fig. 3 ), we measured ligand-induced BRET-responses. Stimulation with the full agonist norepinephrine (100 μM) and the inverse agonist yohimbine (100 µM) induced changes in BRET with opposite directions (Fig. 2a ; agonist: positive ΔBRET blue line; inverse agonist: negative ΔBRET black line)—mirroring the opposite pharmacological effect exerted at the receptor level. The changes in BRET evoked by both ligands were fast, reaching a plateau 120 s after their injection, and remained stable for more than 30 min (Supplementary Fig. 4 ). The injection of the antagonist phentolamine (1 µM) reverted the effect of norepinephrine (Fig. 2a ; red line), demonstrating the applicability of the biosensor to monitor inactivation kinetics. A panel of seven additional ligands triggered responses ranging between the two extremes, norepinephrine and yohimbine, and compatible with their known efficacies as full, partial, or inverse agonists (Fig. 2b , Supplementary Table 1 ). Full concentration–response curves were performed for all ten compounds (Fig. 2c ). Radioligand binding experiments demonstrated that the BRET biosensor α 2A AR Nluc/Halo had binding affinities (pK i -values) similar to the wild-type receptor ( Supplementary Table 2 ) , and they correlate with the measured EC 50 -values ( Supplementary Table 3 ) , indicating that our assay faithfully reports ligand affinities. Interestingly, for agonists these EC 50 -values were similar to the high-affinity component of the competition curves (pK H ) generally assumed to represent the receptor/G-protein complex 19 . This indicates some constitutive activity of the α 2A AR Nluc/Halo biosensor and its interaction with endogenous G-proteins. The relative BRET-change induced by neutral ligands and also the characterization of G-protein activation (Supplementary Fig. 5 ) demonstrated some constitutive activity of the biosensor. Such constitutive activity, not seen with analogous GFP-based sensors 20 , might be due to the larger size of the tags employed that might separate helices 5 and 6, which would shift the receptor to a more active state. However, this effect did not change the expected order of efficacies of the various ligands and even facilitated the characterization of both, inverse agonists and antagonists. Overall, these data indicate that the intramolecular α 2A AR Nluc/Halo 618 BRET-biosensor faithfully reports the activation state of the α 2A AR in 96-well plates. Fig. 2 Pharmacological characterization of the α 2A AR Nluc/Halo618 BRET-based biosensor, and evaluation of its applicability for microtiter plate screening. a Time-course of the normalized BRET ratio upon ligand stimulation (each N = 3). b Ligand-induced maximal BRET changes (Yohimbine: N = 7, Tyramine: N = 10, Phentolamine: N = 11, Clonidine: N = 10, Octopmaine: N = 5, Oxymetazoline: N = 6, Dopamine: N = 6, UK 14,304: N = 10, Norepinephrine: N = 25, Epinephrine: N = 23). c Concentration–response curves of different α 2A AR ligands (Yohimbine: N = 7, Tyramine: N = 4, Phentolamine: N = 5, Clonidine: N = 4, Oxymetazoline: N = 4, Octopamine: N = 4, Dopamine: N = 4, UK 14,304: N = 3, Norepinephrine: N = 6,Epinephrine: N = 4). Comparison of the Z-factors for cells expressing the FRET CFP/YFP or BRET Nluc/Halo sensor of α 2A AR. d FRET or e BRET changes 2 min after 100 μM epinephrine or buffer stimulation are plotted for each well of representative plates and f the average values (each N = 4). g Z-factor at different time points after addition of 100 μM epinephrine or buffer ( N = 4). Data show box and whisker plots ( b ) or mean ± s.e.m ( a , c , f , g ). Difference was analyzed by two-way ANOVA followed by Bonferroni post hoc test. * p ≤ 0.05 vs. buffer Full size image Microtiter plate suitability of the α 2A AR Nluc/Halo618 To validate the suitability of the novel BRET-based biosensor α 2A AR Nluc/Halo 618 to monitor receptor conformation in microtiter plate format, we performed the Z-factor analysis 15 to quantify the quality and reproducibility of the assay. In this analysis, values Z < 0 defines unusable assays, Z = 1 approximates a “perfect” assay, and Z ≥ 0.5 characterizes an excellent assay. Compared to the original α 2A AR CFP/YFP , classified as unusable ( Z = –1.60 ± 0.97; Fig. 2d, f ), α 2A AR Nluc/Halo 618 yielded an excellent assay ( Z = 0.65 ± 0.01; Fig. 2e, f ) that was stable for more than 20 min (Fig. 2g ). β 2 -adrenergic receptor BRET-biosensor (β 2 AR Nluc/Halo618 ) To demonstrate the transferability of the BRET Nluc/Halo 618 as a general design to pharmacologically profile any GPCR, we generated an analogous β 2 -adrenergic receptor biosensor (β 2 AR Nluc/Halo 618 ) and measured efficacy and potency of different ligands in microtiter plate format. The β 2 AR Nluc/Halo 618 biosensor showed signaling activity similar to the wild-type receptor (Supplementary Fig. 6a–c ). Full agonist stimulation induced a change in BRET opposite to that of an inverse agonist, while the amplitudes for antagonists and partial agonists were intermediate (Fig. 3a ), demonstrating that the ability of reporting the efficacy of ligands was preserved. The EC 50 -values obtained from concentration–response curves for the full agonist epinephrine, the inverse agonist ICI 118,551 and the neutral antagonist carvedilol (Fig. 3b ), were similar to data obtained by radioligand binding assay (Supplementary Table 4 ), demonstrating the β 2 AR BRET-based biosensor was reporting wild-type affinity and efficacy. Interestingly, we found that norepinephrine, usually considered a full agonist when monitored at the second messenger level 21 (cAMP), induced only a partial conformational change of β 2 AR Nluc/Halo 618 —in line with other conformational studies 22 , 23 . Our data would also support the new evidence suggesting that UK 14,304 behaves as a partial 24 more than a full agonist. This demonstrates that BRET-based biosensors are a powerful unbiased approach, since they bypass the effect of signal amplification and receptor reserve. Again, the Z -factor of ≈ 0.8 was greatly improved compared to the earlier FRET β 2 AR-biosensor and indicated the suitability for high-throughput screening (Fig. 3c ; Supplementary Fig. 8 ) Fig. 3 Transferability and microtiter plate applicability of the new intramolecular receptor-biosensor design. BRET changes reported by β 2 AR Nluc/Halo618 expressing cells upon ( a ) saturating ligand stimulation (ICI 118.551: N = 6, Propranolol: N = 6, Metoprolol: N = 6, Carvedilol: N = 12, Labetalol: N = 16, Salbutamol: N = 8, Terbutaline: N = 8, Salmeterol: N = 11, Norepinephrine: N = 8, Formoterol: N = 11, Isoprenaline: N = 21, Epinephrine: N = 21) and b serial ligand dilutions to obtain concentration–response curves (Epnephrine: N = 4, Carvedilol: N = 4, ICI 118.551: N = 6). c Mean of the Z -factor of β 2 AR CFP/YFP and β 2 AR Nluc/Halo618 (each N = 4). BRET changes reported by PTHR1 Nluc/Halo618 expressing cells upon ( d ) saturating ligand stimulation (PTH(7-34): N = 12, (dW)-PTH(7-34): N = 12, PTH(3-34): N = 14, PTH(1-31): N = 6, PTHrP(1-34): N = 14,PTH(1-34): N = 22) and e serial ligand dilutions to obtain concentration–response curves (PTH(1-34): N = 3, PTHrP(1-34): N = 3, PTH(3-34): N = 6). f Mean of the Z -factor of PTHR1 CFP/YFP and PTHR1 Nluc/Halo618 (each N = 4). Data are expressed as box and whisker plots ( a , d ) or mean ± s.e.m ( b , c , e , f ). Difference was analyzed by two-way ANOVA followed by Bonferroni post hoc test. * p ≤ 0.05 vs. buffer Full size image Parathyroid hormone receptor 1 BRET-biosensor (PTHR1 Nluc/Halo618 ) To further substantiate the transferability of the BRET Nluc/Halo 618 sensor design, we devised an analogous class B parathyroid hormone receptor 1 (PTHR1) BRET-biosensor (PTHR1 Nluc/Halo 618 ). Again, this biosensor retained signaling activity (Supplementary Fig. 6d–f ), efficacy (Fig. 3d ) and affinity (Fig. 3e ) similar to the wild-type receptors (Supplementary Table 5 ) for the analyzed peptides. Application of the full agonist PTH(1-34) evoked a ≈10% change in BRET. Antagonists had no detectable effects (Fig. 3d ), but retained their ability of inhibiting receptor activation in competition experiments (Supplementary Fig. 7 ). In contrast to the α 2A AR and β 2 AR, no inverse agonists are available for the PTHR1 that might cause effects opposite to those of PTH(1-34). The Z -factor for PTHR1 Nluc/Halo 618 (0.52 ± 0.02) demonstrated that also this BRET-biosensor was suitable for high-throughput screening (Fig. 3f ; Supplementary Fig. 8 ). Discussion Taken together, we have developed a universal sensor design for GPCRs that retains the signaling capacity of the receptors, resolves both efficacy and potency of ligands, and works in intact cells and in real time in a microtiter plate format. In a single assay, this approach offers the analysis of activation/de-activation kinetics, potency and direct efficacy–which makes it a rapid and high content screening approach. The throughput of this assay essentially depends on the equipment available for pipetting and plate reading. Under optimal conditions, it depends only on the time required to read the plate before and after addition of compounds, while under the most basic conditions, the time required for a single plate would encompass the times required for the basal read, addition of compounds, and a second read. If the labeling procedure is carried out in serum- and phenol red-free medium, then no washing step is required. As any optical detection method, the readout of our BRET-based receptor biosensor may be impaired by molecules that interfere with the absorption properties; bind to the energy donor or acceptor interfering with their emission properties; and compete with the substrate, causing inhibition of the luciferase 25 . The chemical evaluation of compounds of a screening library can help discriminating false positives. Appropriate controls for these issues therefore need to be included in such a screening approach. An advantage of our receptor biosensor is that the tags are localized intracellularly, which limits their susceptibility to cell permeable compounds. To the best of our knowledge, this generalizable design offers the first possibility to upscale the study of receptor activation and deactivation - which represents the most direct and unbiased way to estimate the effect of any chemical entity on receptors of interest, facilitating the discovery of new therapeutic compounds. These GPCR biosensors should prove useful in determining ligand properties at known GPCRs and in elucidating the binding properties of orphan receptors. Furthermore, their high sensitivity may allow their use to monitor GPCR activation in situ using new knock-in technologies 26 . Methods cDNA constructs The FRET sensor α 2A AR CFP/YFP 4 was used as starting construct to generate the various α 2A -adrenergic receptor FRET and BRET sensors (α 2A AR donor/acceptor ) described in this study. The donors (CFP or Nluc) were fused to Val461 at the C-terminus while all acceptors tested were placed in the third intracellular loop between Ala250 and Ser371. The BRET-based β 2 -adrenergic receptor biosensor β 2 AR Nluc/Halo was cloned starting from the previously described FRET version 22 introducing HaloTag in the third intracellular loop between Asp251 and Gly252 and Nluc in the C-terminus at Glu369. The BRET sensor for the parathyroid hormone receptor (PTHR1) was cloned starting from the previously described FRET version 4 . HaloTag was inserted in the third intracellular loop between Gly395 and Arg396 and Nluc was fused to Gly497 of the C terminus. All tag exchanges were performed employing established PCR strategies and restriction and ligation enzymes. Constructs were cloned into a pcDNA3 vector and verified by sequencing. Plasmids cDNA encoding the fluorescent protein cpVenus 173 was amplified from the Gα i2 -sensor v2.0 27 . The SNAPtag sequence was amplified using a SNAPtag-GABA B1 template kindly provided by J.P. Pin (Institut de Génomique Fonctionnelle, Montpellier, France) 28 . Sequences encoding TagRFP (pTagRFP-C vector) and mCherry were purchased from Evrogen and Addgene, respectively. cDNA encoding HaloTag (pFC14K HaloTag® CMV Flexi® Vector) and NanoLuciferase (pFC32K Nluc CMV-neo Flexi® Vector) were purchased from Promega. Gα i2 -FRET 27 sensor was kindly provided by J. Goedhart (Section Molecular Cytology, University of Amsterdam, Amsterdam, The Netherlands), the H187-EPAC-FRET sensor 29 was kindly provided by K. Jalink (The Netherlands Cancer Institute, Amsterdam, The Netherlands). Reagents (−)-Epinephrine, l -(−)-norepinephrine (+)-bitartrate salt monohydrate, UK 14,304, dopamine hydrochloride, (±)-octopamine hydrochloride, clonidine hydrochloride, tyramine hydrochloride, phentolamine hydrochloride, yohimbine hydrochloride, isoprenaline hydrochloride, formoterol fumarate dihydrate, terbutaline hemisulfate salt, salmeterol xinofoate, salbutamol hemisulfate salt, metoprolol tartrate, (±)-propranolol hydrochloride, ICI 118,551 hydrochloride, labetalol hydrochloride, carvedilol, GTP, poly-D-lysine, G-418 and the fluorescent monoclonal antibody Anti-Flag® M2-Cy3 and Millipore glass-fiber filters for radioligand saturation binding were from Sigma-Aldrich. [ 3 H]RX821002 was purchased from Hartmann Analytic. The HA-tag monoclonal antibody (16B12) Alexa Fluor 488 was from ThermoFisher Scientific. Oxymetazoline hydrochloride was purchased from Tocris. The peptide ligands PTH(1-34) (catalog number: H-4835), PTH(7-34) (catalog number: N-1110), (dw)-PTH(7-34) (catalog number: H-9115), PTHrP(1-34) (H-6630), PTH(1-31) (catalog number: H-3408) and PTH(3-34) (catalog number: H-3088) were from BACHEM. All HaloTag fluorescent dyes were purchased from Promega. SNAP fluorescent dyes were from NEB. White-wall, white bottomed and black-wall, black-bottomed 96-well plates were purchased from Brand. MultiScreen® Filter plates for radioligand competition binding were from Millipore. Cell culture and transfection HEK-TSA cells used for transient expression of constructs, were grown in Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 2 mM glutamine, 10% fetal calf serum, 0.1 mg mL −1 streptomycin, and 100 units per mL penicillin at 37 °C with 5% CO 2 . HEK293 cells were used for the development of stable BRET sensor cell lines. Cells grown in 100 mm dishes were transfected at a confluence of 50–70% with 5 µg of DNA using Effectene Transfection Reagent Kit (Qiagen) according to the manufacturer’s instructions. Transfected clones were selected with 600 μg mL −1 of G-418 and clonal lines were maintained in DMEM supplemented with 200 μg mL −1 G-418, 2 mM glutamine, 10% fetal calf serum, 0.1 mg mL −1 streptomycin, and 100 units per mL penicillin at 37 °C with 5% CO 2 . Transient transfection and plating For transient expression of the sensors, 1.5 × 10 6 HEK-TSA cells were seeded onto a 55 mm dish and transfected the day after with 2 µg of plasmids encoding the biosensors using Effectene Transfection Reagent (Qiagen) according to the manufacturer’s protocol. In case two different plasmids were co-transfected, 4 µg was used as total amount of DNA in a 1:1 ratio of the two plasmids. Twenty-four hours after transfection, cells were transferred to poly- d -lysine pre-coated black-wall, black-bottomed (FRET experiments) or white-wall, white-bottomed (BRET experiments) 96-well plates at a density of 50,000 (FRET) or 20,000 (BRET) cells per well. Fluorescence labeling of FRET and BRET acceptors Labeling with all dyes was performed at 37 °C and 5% CO 2 in 96-well plates. All dyes were dissolved in DMEM. HaloTag® diAcFAM (1 µM), HaloTag® Oregon Green® (1 µM), SNAP-cell 505-Star (10 µM), SNAP-cell TMR-Star (3 µM) and SNAP-cell 647SiR (3 µM) were incubated for 30 min 48 h after transfections. Excessive dye was washed out three times followed by incubation with fresh DMEM for additional 30 min (37 °C and 5% CO 2 ). HaloTag® R110Direct, HaloTag® TMRDirect and HaloTag® NanoBRET 618 required overnight labeling at a concentration of 100 nM. A minimum of 4 wells remained unlabeled to serve as correction for donor bleedthrough (unlabeled control). Measurement of fluorescence excitation and emission spectra HEK-TSA cells were transfected with different BRET based α 2A AR biosensors and labeled as described above, without substrate, in order to read only the emission and excitation spectra of the different acceptors. YFP and CFP spectra were collected using the FRET α 2A AR CFP/Halo and α 2A AR YFP/Halo biosensors without HaloTag labeling. All spectra were measured in buffer (2 mM HEPES, 28 mM NaCl, 1.08 mM KCl, 0.2 mM MgCl 2 , 0.4 mM CaCl 2 , pH 7.3) with 2 nm resolution from 400 to 700 nm using a CLARIOstar plate reader (BMG). Spectra are expressed as a percentage of the respective maximal excitation or emission peak. Measurement of FRET and BRET emission spectra HEK-TSA cells were transfected and labeled as described above. Emission spectra were recorded in buffer with 2 nm resolution from 400 to 700 nm upon donor excitation at 420 nm (FRET sensors) or addition of 1:1000 furimazine dilution (BRET) using a CLARIOstar plate reader (BMG). Spectra are expressed as a percentage of the maximal donor emission peak. FRET measurements Cells expressing the FRET sensors were washed to substitute the DMEM with the experimental buffer. Basal FRET ratio was measured in 90 µL buffer. Subsequently, 10 µL of 10-fold ligand solution or buffer (negative control) was applied to each well and the stimulated FRET ratio was recorded. All FRET experiments were conducted at 37 °C with a Synergy Neo2 plate reader (BioTEK) equipped with 420/50 nm excitation and 485/20 nm emission filters for CFP. Acceptor emission of YFP, HaloTag® R110, HaloTag® diAcFAM, HaloTag® Oregon Green® and SNAP-cell 505-Star were detected with a 540/25 nm (FRET) filter. To measure the emission of HaloTag® TMR-Direct and SNAP-cell TMR-Star a 590/35 nm filter was used. Emission of HaloTag® NanoBRET 618 and SNAP-cell 647SiR were detected with a 620/15 nm and 680/20 nm filter, respectively. Fifty excitation flashes were applied per data point. BRET measurements Cells transiently or stably expressing the BRET-biosensors were washed to substitute DMEM with the experimental buffer and incubated with substrate (90 µL of 1:1000 for β 2 AR Nluc/Halo 618 and PTHR1 Nluc/Halo 618 ; 1:4000 for α 2A AR Nluc/Halo 618 ) for 2–5 min at 37 °C to allow for substrate diffusion and the basal BRET ratio was measured. Following this, 10 µL of 10-fold ligand solution or buffer was applied to each well and the stimulated BRET ratio was recorded. To reduce the fluctuation of the BRET ratio in Z -factor experiments, seven individual BRET ratios within 5 min were measured and averaged before and after ligand addition. BRET experiments were performed at 37 °C with a GloMAX Discover (Promega) or Synergy Neo2 (BioTEK) plate reader equipped with a 460/40 nm filter to select the NanoLuc emission. For cpVenus173, HaloTag® R110, HaloTag® diAcFAM, HaloTag® Oregon Green® and SNAP-cell 505-Star a 520/20 nm (BRET) filter was used to select the acceptor emission peaks. TagRFP, HaloTag® TMR-Direct and SNAP-cell TMR-Star emissions were detected with a 530 nm long pass filter. For HaloTag® NanoBRET 618 a 620/20 nm filter was used and a 600 nm long pass filter was applied for the BRET acceptors mCherry and SNAP-cell 647SiR. The integration time per data point was set to 0.3 s. Experiments with higher temporal resolution were performed employing the Synergy Neo2 (BioTEK) plate reader, which is equipped with injectors and has a faster acquisition time. Data were acquired in well-mode, the acquisition interval was set to 1 s and the integration time to 0.3 s. After acquisition of baseline for 180 s, 10 µL of solution with or without ligand (buffer control) were injected with a speed of 225 µL s −1 (delivery time = 44 ms) and the signal was recorded for 180–360 s. Receptor staining Cells were co-transfected with FRET-based sensors to monitor downstream signaling and the wild type or the BRET-biosensor receptor as described above. Staining of the plasma membrane portion of the receptors was evaluated using a cell-impermeable anti HA-tag conjugated with AlexaFluor594 (Anti-HA-AlexaFluor594 ThermoFischer) or Anti-Flag® M2 conjugated with Cy3 (Anti-Flag® M2 Cy3, Sigma). The fluorescent antibodies were diluted in DMEM to a concentration of 10 µg mL −1 and incubated for 1 h at 37 °C in the 96-well plates. Subsequently, cells were rinsed three times and incubated additional 30 min with fresh DMEM. Subsequently, the emission intensity of HEK-TSA cells were measured using the Synergy Neo2 plate reader. Therefore, cells stained with Anti-Flag® M2 Cy3 (β 2 AR NLuc/Halo or β 2 AR) were excited using a 540/20 nm excitation and the emission intensity was recorded using 590/35 nm emission filter. Fluorescence intensities of HEK-TSA cells stained with Anti-HA-AlexaFluor594 were measured using a 590/20 nm (excitation)–620/15 nm (emission) filter combination. Expression levels of Gα i2 -FRET and H187-EPAC-FRET sensor The Synergy Neo2 plate reader was employed to assess the expression levels of the downstream sensors (Gα i2 -FRET and H187-EPAC-FRET sensor). Therefore, the FRET acceptors (cpVenus173 and tandem cpVenus173) were directly excited using a 500/20 nm excitation filter. Emission intensities were detected with a 540/20 nm filter. Membrane preparations Membranes expressing wild-type α 2A AR (α 2A AR-wt) were harvested from HEK-TSA cells grown in 15 cm dishes, 48 h after transfection. Membranes expressing the BRET-based α 2A AR sensor (α 2A AR Nluc/Halo ) were obtained from HEK293 stably expressing the sensor. Cells were detached from the dishes with a cell scraper and suspended in Tris buffer (5 mM Tris, 2 mM EDTA, pH 7.4). After centrifugation for 10 min at 1000 × g , cells were re-suspended in buffer 1 (20 mM HEPES, 10 mM EDTA, PBS, pH 7.4) and homogenized using twice Ultraturax for 15 s. The suspension was centrifuged for 10 min at 3200 × g . The resulting supernatant was further centrifuged for 45 min at 37,000 × g and 4 °C. The pellet was resuspended and the last two centrifugation steps were repeated. The pellet was then suspended in binding buffer (50 mM Tris, 100 mM NaCl, 3 mM MgCl 2 , pH 7.4) and the amount of total membrane protein was measured using the Pierce BCA Protein Assay Kit according to the manufacturer’s instructions. Radioligand binding Total radioligand binding was assessed by incubating 5 µg of membrane protein with different concentrations (0.04–12 nM) of the antagonist α 2A AR radioligand [ 3 H]RX821002. To define unspecific binding, 20 µM phentolamine was added. Competition binding was performed by incubating 2 µg membrane protein with 0.3–2.0 nM [ 3 H]RX821002 and increasing concentrations of the different α 2A AR ligands in the presence (=low affinity state for agonists) and absence (=high-affinity state for agonists) of 10 µM GTP. Following incubation for 1 h at room temperature, membranes were transferred to Millipore glass-fiber filters via vacuum filtration. These filters were incubated with scintillation cocktail and membrane-bound radioactivity was measured with a scintillation counter. Data analysis and statistics FRET and BRET ratios before (Ratio basal ) and after ligand or buffer application (Ratio stim ) were defined as acceptor emission/donor emission and corrected for donor bleedthrough into the acceptor channel by subtracting the averaged ratio of the unlabeled control (UC). For cells expressing biosensors with a fluorescent protein as acceptor, the averaged UC ratio of the analogous HaloTag construct was considered for bleedthrough correction. To quantify the ligand induced conformational change, ∆FRET or ∆BRET was calculated for each well as a percent over basal (((Ratio stim − Ratio basal )/Ratio basal ) × 100) and subtracted by the averaged ∆FRET or ∆BRET of buffer. Z -factors expressing the high-throughput suitability were calculated with the following equation: $$Z = 1 - \frac{{\left( {3\sigma _S + 3\sigma _C} \right)}}{{\left( {{\mathrm{\mu }}_S - {\mathrm{\mu }}_C} \right)}}$$ where σ s and σ c are the standard deviations of ∆FRET or ∆BRET. μ S and μ C express the mean of ∆FRET or ∆BRET values of positive and negative control, respectively. If the positive control induced a decrease in the energy transfer (negative ∆RET as for α 2A AR CFP/YFP , β 2 AR NanoLuc/Halo 618 , PTHR1 NanoLuc/Halo 618 , PTHR1 CFP/YFP ) the denominator in equation is inverted ( μ C − μ S ). As a positive control, we defined epinephrine for the α 2A AR- and β 2 AR-sensors and PTH(1-34) for PTHR1 sensors. Buffer was used as a negative control in all Z factor experiments. For simplicity, all agonist-induced RET changes were consistently plotted as ascending curves or bars. Therefore, y -axes in all figures were inverted if agonists for the respective biosensor induced a reduction of the ratio. Data were analyzed using Prism 5.0 software (GraphPad) and expressed as mean ± s.e.m. Data from concentration–response experiments were fitted using a mono-exponential curve four-parameter fit. Radioactivity values from binding experiments were analyzed using a one-site fitting model if GTP was added prior the experiment. Data from competition-binding experiments without exogenously added GTP were first analyzed for the statistically preferred fitting model applying extra-sum-of squares F-test comparing a one-component vs. two-component fit. Superiority of the two-component model was confirmed for all agonists (partial or full) tested. The two-component fit was then conducted with the fraction of the high-affinity component ( R H ) fixed to 0.58 which is the mean R H of all data where this model was applied. Statistical differences were evaluated using one-way ANOVA test followed by Bonferroni multiple comparison, Student’s t -test or extra-sum-of squares F-test. Differences were considered significant for values of p < 0.05. Data availability The datasets generated during and/or analysed during the current study are available at the homepage of the Institute of Pharmacology and Toxicology

Hormones and other neurotransmitters, but also drugs, act upon receptors. "Their active substances bind to the receptors and modify the three-dimensional receptor arrangement regulating the downstream signal pathways," says Hannes Schihada from the Institute for Pharmacology and Toxicology at the University of Würzburg (JMU). A special case are G protein-coupled receptors (GPCR). "About 30 percent of all authorized drugs worldwide act upon these receptors," explains Hannes Schihada, "but their potential is not yet fully utilized." To date, it has not been possible to test the effect of millions of potential drugs on the GPCR arrangement within a very short time. "This has been a stumbling block to the discovery of novel pharmaceutical substances and the research of still unknown GPCRs," says Dr. Isabella Maiellaro, who is in charge of the project together with Professor Martin Lohse. The JMU team has now developed a method that allows the determination of both activity and potency of GPCR ligands in living cells using high throughput technology. The scientists have published their results in the journal Communications Biology. What the new method can do The name of the method is BRET (bioluminescence resonance energy transfer-based sensor design). "It can be used not only for GPCRs but for a lot of different biomolecules," explains Schihada. The universal sensor design now allows the exploration of receptor conformational changes in living cells with the high-throughput method. This enables a much faster pharmacological characterization by a high number of test combounds that directly act on the receptor, independently of their downstream receptor signal pathways. "This technology can contribute to a faster and better understanding of the different levels of effectiveness of drugs and thus drive the development of novel therapeutic concepts," says Schihada. The study of novel receptor targets will yield a higher array to develop drugs that have less side-effects and are more efficient. Furthermore, the sensors could help to better understand what is called orphan GPCRs—GPCRs whose function and ligands are still largely unknown. "With these findings we can lay the foundation for the treatment of severe diseases that were hitherto difficult to treat, such as Alzheimer's or multiple sclerosis," says the scientist. The scientists now want to expand their range of sensors suitable for high throughput.

10.1038/s42003-018-0072-0

Space

Scientists unveil a unified theory for rocky planet formation

Konstantin Batygin et al, Formation of rocky super-earths from a narrow ring of planetesimals, Nature Astronomy (2023). DOI: 10.1038/s41550-022-01850-5 Journal information: Astrophysical Journal , Nature Astronomy

https://dx.doi.org/10.1038/s41550-022-01850-5

https://phys.org/news/2023-01-scientists-unveil-theory-rocky-planet.html

Abstract The formation of super-Earths, the most abundant planets in the Galaxy, remains elusive. These planets have masses that typically exceed that of the Earth by a factor of a few, appear to be predominantly rocky, although often surrounded by H/He atmospheres, and frequently occur in multiples. Moreover, planets that encircle the same star tend to have similar masses and radii, whereas those belonging to different systems exhibit remarkable overall diversity. Here we advance a theoretical picture for rocky planet formation that satisfies the aforementioned constraints: building upon recent work, which has demonstrated that planetesimals can form rapidly at discrete locations in the disk, we propose that super-Earths originate inside rings of silicate-rich planetesimals at approximately ~1 au. Within the context of this picture, we show that planets grow primarily through pairwise collisions among rocky planetesimals until they achieve terminal masses that are regulated by isolation and orbital migration. We quantify our model with numerical simulations and demonstrate that our synthetic planetary systems bear a close resemblance to compact, multi-resonant progenitors of the observed population of short-period extrasolar planets. Main It has long been known that the genesis of planets begins through the coalescence of solids within protoplanetary nebulae, and models of planet formation have traditionally assumed that dust within circumstellar disks is smoothly distributed. Despite being common, this simplifying assumption may be unfounded. Several lines of evidence have recently been marshalled in support of the notion that rather than arising from a smooth gradient of solids, planetesimal formation unfolds in a small number of discrete rings 1 , 2 , 3 , 4 , 5 , 6 . In this vein, the work reported in ref. 3 has proposed that protoplanetary nebulae generally originate as decretion disks that spread radially from tenths of an au, facilitating the condensation of outward-diffusing silicate vapour into rocky dust grains at the disk’s primordial silicate sublimation line. Importantly, this process naturally leads to the formation of rocky planetesimals at a stellocentric distance comparable to the Earth’s orbital radius (as well as the generation of more distant icy bodies close to Jupiter’s present-day orbit) through gravito-hydrodynamic instabilities 7 , 8 . Such a model further yields a self-consistent explanation for the isotopic dichotomy of carbonaceous and non-carbonaceous iron meteorites, as well as the physical origins of the Solar System’s broader architecture 3 , 6 . Within the framework of the aforementioned disk model, the mass budget of silicate material that forms at the rock line is distinctively variable (Extended Data Fig. 1 ). That is, depending on the specific combination of disk viscosity and metallicity, the cumulative mass of rocky planetesimals entrained within the silicate annulus residing at an orbital radius of r ≈ 1 au can readily reach tens of Earth masses (although we note that it can also be null if the threshold for planetesimal formation through gravitational collapse is not met). Moreover, numerical modelling indicates that planetesimal formation is expected to occur over a relatively short temporal burst, such that dust is incorporated into planetary building blocks over a timescale of ~10 5 years. Adopting the ringed planetesimal formation paradigm as a platform, a key goal of our work is to consider the possibility that a typical system of extrasolar super-Earths originates within such a radially confined annulus of rocky material. As we describe below, the process of planetary conglomeration within a narrow ring of silicate-rich planetesimals naturally yields a characteristic multi- M ⊕ mass scale of the resulting planets, and the simultaneous operation of accretion and orbital migration regulates the emergence of uniformity among the growing planetary embryos (Fig. 1 ). Fig. 1: Diagram of the planet formation scenario considered in this work, in chronological order from top to bottom. In the top image, a protoplanetary nebula, represented by a yellow cone, originates from infall of gas and dust (shown with a vertical blue arrow). Viscous heating and stellar irradiation regulate the temperature of the gas, T , which in turn determines the disk’s geometric aspect ratio, h / r . Owing to strong magnetic braking, the centrifugal radius of the infalling material remains at a few tenths of an au. Driven by turbulent viscosity (which is parameterized by the Shakura–Sunayev α parameter), the disk spreads outwards to large stellocentric distances, transporting vapour and minute dust grains outwards. Silicate vapour entrained within the gas continues to be carried outwards with a velocity v r (solid blue arrows). Beyond the r ≈ 1 au silicate condensation front, however, dust grains grow and drift inwards due to the sub-Keplerian azimuthal flow of the gas, v ϕ . In the middle image, the resulting accumulation of rocky material at the silicate condensation line facilitates the formation of ~100 km planetesimals through gravito-hydrodynamic instabilities. In the bottom image, pairwise collisions among planetesimals generate multi- M ⊕ objects (brown sphere) that experience substantial orbital decay. Depletion of the local supply of solids as well as the removal of the planets from the planetesimal ring through disk-driven migration regulate the terminal planetary mass, leading to the emergence of intra-system uniformity. Full size image Results The starting point of our calculation corresponds to the epoch of large-scale planetesimal formation within a protoplanetary disk. For definitiveness, here we adopt disk conditions derived from the simulations reported in ref. 3 , although we note that for the purposes of our calculations, any ringed planetesimal formation scenario is likely to lead to similar results. Our fiducial disk model is initialized with a gas surface density of Σ 0 = 2,500 g cm −2 at 1 au, a corresponding peak dust surface density of Σ • = 500 g cm −2 and a dust-grain radius of s • = 1 mm, consistent with fragmentation-limited growth 9 . Owing primarily to viscous energy dissipation, the disk maintains an appreciable aspect ratio of scale height to orbital radius of h / r ≈ 0.05 throughout the planetesimal formation epoch. While the gas surface density is taken to dissipate exponentially with a time constant of τ disk = 1.5 Myr, the dust surface density decays much more rapidly, owing to the fact that pebbles get incorporated into a planetesimal swarm of mass M ring ≈ 20 M ⊕ over an ~10 5 year timescale. The specific functional parameterizations of these quantities are delineated in Methods . As clouds of dust within the r ≈ 1 au silicate ring consolidate into planetary building blocks by means of gravitational collapse, their continued growth can proceed through two distinct channels: pairwise mergers among planetesimals and pebble accretion. The efficiency of planetesimal accretion is controlled by the extent to which gravitational focusing can increase the collisional cross-section of protoplanetary embryos. Pebble accretion, on the other hand, depends critically on whether the capture of dust proceeds in the two- or three-dimensional (3D) regimes, a determination that is sensitive to the characteristic size of dust particles. Generically speaking, the process of collisional fragmentation inhibits the growth of silicate grains beyond the millimetre-scale within protoplanetary disks, ensuring that even in relatively quiescent nebulae, turbulent stirring can maintain the dust sub-disk’s aspect ratio at an inflated level 10 . Correspondingly, pebble accretion proceeds in the comparatively inefficient 3D regime, contributing very little to the planetary conglomeration process during the planetesimal formation epoch. We further find that leftover dust that is not incorporated into planetary building blocks through gravito-hydrodynamic instabilities rapidly flows away from the planetesimal ring as the nebula matures into an accretion disk, and our estimates (‘Planetary growth’ section in Methods ) indicate that any auxiliary exterior flux of pebbles plays a negligible role in driving the formation of rocky super-Earths (we confirm these expectations with numerical simulations below). Analytical estimates In contrast with the relative inefficiency of pebble accretion in the inner regions of a protoplanetary disk, the efficacy of planetesimal accretion within a narrow annulus of rocky planetesimals is strongly enhanced. The reasons for this are two-fold. First, by concentrating tens of Earth masses of solids into a radially confined ring of planetesimals, the rate of collisions among the constituent bodies is strongly amplified. Second, the combined action of aerodynamic drag and inelastic collisions among planetesimals constitutes a fast-acting damping mechanism for the planetesimal velocity dispersion, magnifying the effect of gravitational focusing. In this regime, the associated mass-accretion rate of a planetary embryo can be deduced from an n – σ – v relation (where n is the number density, σ is the cross-section and v is the velocity dispersion), and the result is well-known 11 , 12 : \(\dot{M} \approx {{{\varSigma }}}_{{{{\mathrm{pl}}}}}\,\uppi \,{R}^{2}\,{{\varOmega }}\,(1+{{\varTheta }})\) , where Σ pl ≈ Σ • is the planetesimal surface density, R is the physical radius, Ω is the orbital frequency and Θ = ( v esc /〈 v 〉 pl ) 2 is the Safronov number (that is, the ratio of the square of the escape velocity to the square of the planetesimal velocity dispersion 13 ). Moreover, under the simplifying assumption of strong and time-invariant gravitational focusing, it is straightforward to show that a crude estimate for the timescale for an Earth-mass body to emerge within the ring of rocky planetesimals is given by \({\mathcal{T}}_{\oplus } \approx \bar{\rho }\,{R}_{\oplus }/({{{\varSigma }}}_{{{{\mathrm{pl}}}}}\,{{\varOmega }}\,{{\varTheta }})\) (ref. 14 ), where \(\bar{\rho } \approx 3\,{\mathrm{g}}\,{\mathrm{cc}}^{-1}\) is the embryo’s density. If we adopt the fiducial parameters of our model and assume that the escape velocity of the planetary embryo exceeds the planetesimal velocity dispersion by a factor of a few (corresponding to Θ ≈ 10), \({\mathcal{T}}_{\oplus }\) can be as short as ~10 5 years. No matter the dominant growth mode, planetary accretion cannot proceed without bounds. For the problem at hand, two distinct processes constitute natural termination mechanisms for planetary conglomeration, the first being isolation. Isolation occurs due to the depletion of planetesimals from the local feeding zone of the embryo. The expression for the isolation scale is easily obtained by equating the cumulative mass of planetesimals within the feeding zone (approximately two Hill radii, R Hill , within a heavily dissipated disk) and the planetary mass itself, to yield \(M \approx 8\,{\uppi }^{3/2}\,{{{\varSigma }}}_{{{{\mathrm{pl}}}}}^{3/2}\,{r}^{3}/\sqrt{3\,{M}_{\star }}\) , where M ★ represents the host star’s mass. Given our fiducial parameters, the isolation mass within the planetesimal ring evaluates to M ≈ 3 M ⊕ . A second growth-limiting process that ensues in our model is gas-driven orbital migration. As a growing planet becomes massive enough to raise a substantial wake within the gaseous nebula, the gravitational back-reaction of the wake upon the planet drives an exchange of energy and angular momentum between the planet and the disk, which in turn expels the planet from the planetesimal ring altogether. Thus, within the framework of our theoretical picture, an approximate equivalence between the mass-doubling timescale \({\mathcal{T}}{{{{\mathrm{mass}}}}} \approx 3\,{M}^{1/3}\,{\bar{\rho }}^{2/3}/({{{\varSigma }}}_{{{{\mathrm{pl}}}}}\,{{\varOmega }}\,{{\varTheta}}\!)\) and the migration timescale 15 \({\mathcal{T}}_\mathrm{mig }\) ≈ (4/ Ω )( M ★ / M )( M ★ / Σ 0 r 2 )( h / r ) 2 yields an estimate for the mass of planets that are expected to emerge from the rocky annulus of planetesimals. Auspiciously, for the aforementioned nominal parameters, the planetary mass scale that comes out from this relation also evaluates to M ≈ 3 M ⊕ . Because the isolation and accretion–migration terminal mass scales are similar, the process of planetary conglomeration is unlikely to depend sensitively on the detailed character of type-I torques, and the envisioned fiducial picture is expected to hold even in a scenario where inward migration is initially suppressed or even directed outward. And while these mass-limiting mechanisms operate simultaneously, it is nevertheless important to note that they scale differently with the planetesimal surface density. Crucially, this scaling is superlinear ( \(M\propto {{{\varSigma }}}_{{{{\mathrm{pl}}}}}^{3/2}\) ) and sub-linear ( \(M\propto {{{\varSigma }}}_{{{{\mathrm{pl}}}}}^{3/4}\) ) for isolation and migration-regulated growth, respectively, meaning that migration is expected to act as the primary accretion-quenching mechanism for massive systems (that yield M ≳ 3 M ⊕ planets), while isolation regulates the formation of planets in lower-mass rings. Beyond yielding a mass scale that is broadly consistent with the observational sample of extrasolar super-Earths 16 , 17 , the scenario described above entails the emergence of a pattern of intra-system uniformity among the forming bodies 18 , 19 , 20 . That is, while parameters such as the surface density, disk aspect ratio and so on may differ from system to system, the masses of planets that form within a given disk are likely to be similar, since they are largely determined by the isolation and accretion–migration relations. We further remark that this paradigm is not specific to circumstellar nebulae: the standard model for the formation of giant planet satellites 14 , 21 follows an analogous narrative, suggesting that architectural similarities between extrasolar multi-planet systems and the Galilean moons are not coincidental. Numerical simulations The analytical estimates quoted above provide a useful reference point for the planetary growth sequence that is expected to ensue within a ringed disk. Nevertheless, the chaotic intricacy of planet formation cannot be captured with such considerations alone. Accordingly, we have simulated the formation of super-Earths within an annulus of rocky planetesimals employing a full-fledged numerical model. Our simulations build upon an N -body framework and augment the self-consistent treatment of gravitational dynamics of planetesimals and planetary embryos with the effects of aerodynamic drag exerted on planetesimals by the gas, collisional damping, disk-driven (type-I) migration and pebble accretion, in a parameterized manner. The details of our implementation are provided in Methods . In our numerical experiments, the ring of rocky planetesimals is modelled as a population of N pl = 1,000 super-particles that form a Gaussian distribution of solid material centred on 1 au, with an initial radial spread of Δ r = 0.1 au. Along with N emb = 10 lunar-mass planetary embryos, the planetesimals were introduced into the simulation gradually over a course of 10 5 years, while the ambient surface density of solid dust is reduced in concert. Our numerical experiments followed the conventional ‘big–small’ categorization of bodies, wherein planetesimals could accrete onto the embryos but not onto each other. We further note that although the total number of particles in our simulations is limited by computational cost, we have also found that increasing the number of planetary seeds beyond 10 does not yield materially different results, and can even diminish the realism of our simulations by over-exaggerating the effects of dynamical heating (see Methods for a discussion). Cumulatively, we carried out our calculations over a time span of 3 Myr, in agreement with typical lifetimes of protoplanetary disks. All in all, our numerical experiments follow a similar narrative to that outlined by the analytic estimates, and a typical formation sequence observed in our fiducial simulation suite is depicted in Fig. 2 . As dust is converted into planetary building blocks over the course of the first 100,000 years, planetary embryos experience rapid growth. Simultaneously, the initially narrow ring of solid material spreads radially due to gravitational stirring within the system. As such, a small number of separated massive objects emerge even before the epoch of large-scale planetesimal formation is complete. Over the following ~10 5 years, a chaotic phase of impacts ensues, generating super-Earth-class objects. As collisional accretion wanes and the embryos dynamically isolate themselves within the planetesimal swarm, the planets enter a prolonged period of inward migration. In due course, the planetary orbits lock into a multi-resonant chain and stabilize near the disk’s assumed inner edge at radii spanning tenths of an au, where they eventually become observable by photometric surveys. Fig. 2: The formation sequence of a mass-uniform exoplanetary system. a – h , Orbital eccentricities as a function of semi-major axes. Inclinations are delineated with the grey, orange, red, green and blue colour code in a . Over the course of the first 100,000 years ( t = 5 ( a ), 25 ( b ), 50 ( c ), 75 ( d ) and 100 ( e ) kyr), \({\mathcal{D}}\) = 100 km super-planetesimals (points) and lunar-mass planetary embryos (purple circles), comprising M ring ≈ 20 M ⊕ in total, are gradually introduced into the simulation domain. These objects originate with eccentricities and inclinations of 〈 e 〉 ≈ 〈 i 〉 ≈ 0.01, across a radial range spanned by the horizontal line shown in a . Growth of planetary embryos is driven primarily by accretion of planetesimals, with aerodynamic drag and collisional damping facilitating enhanced gravitational focusing ( c – e ). Injection of new material into the system terminates at the t = 10 5 year mark ( e ), and over the course of the following few hundred thousand years, multi- M ⊕ planets emerge, with the conglomeration process largely completed within the first 0.5 Myr ( f , g ). Over the course of the remaining lifetime of the disk, the formed planets migrate inwards, locking into a mass-uniform multi-resonant chain ( h ). Recent work 23 , 24 has shown that tightly packed multi-resonant planetary configurations serve as ideal initial conditions for reproducing both the period-ratio distribution of observed extrasolar planets as well as their inferred degree of mass uniformity. Full size image Discussion The degree of intra-system mass uniformity within our modelized planetary systems is keenly reminiscent of that observed in the data. As a specific example, the final results of the simulation shown in Fig. 2 are characterized by a normalized mass dispersion of \({\mathcal{D}}\) M /〈 M 〉 ≈ 0.13. We have also carried out a variant of the same numerical experiment where the phase of long-range inward migration is delayed until after the principal phase of planetary growth is complete and obtained comparable average mass of 〈 M 〉 ≈ 2 M ⊕ and a normalized dispersion of \({\mathcal{D}}\) M /〈 M 〉 ≈ 0.42. While these numbers are low, they are not anomalous: a normalized mass dispersion of \({\mathcal{D}}\) M /〈 M 〉 ≲ 0.5 is an expected outcome of our proposed formation scenario. To this end, we have run a series of 12 numerical experiments akin to that depicted in Fig. 2 , each time randomizing the initial conditions and varying the mass of the planetesimal ring from 5 to 40 M ⊕ . The census of the generated systems is depicted in Fig. 3 . Overall, these numerical experiments confirm that in M ring ⩾ 20 M ⊕ systems, the typical planetary mass scales as the ~3/4th power of the initial planetesimal surface density, as is expected from the accretion–migration timescale balance. Conversely, for planets generated within M ring ⩽ 20 M ⊕ rings, this dependence is slightly superlinear (~8/7 index), signalling the increasingly important role of isolation for lower-mass systems. Regardless of the relevant scaling, the generated sample of synthetic systems conforms to a clear pattern of mass (and presumably radius) homogeneity. In agreement with recent observational determinations 22 , however, our model also predicts that this ‘peas-in-a-pod’ pattern is limited to short-period orbits and does not extend beyond r ≈ 0.5 au. Conversely, larger stellocentric distances are primarily occupied by stranded low-mass planetary objects. Finally, beyond reproducing the rocky composition of the planets themselves, the orbital architectures of our synthetic planetary systems are markedly resonant, and require post-nebular dynamical instabilities to generate the observed period-ratio distribution of the Galactic Planetary Census 23 , 24 . Fig. 3: Architectures of exoplanetary systems at time of disk dispersal, generated within the framework of our model. Each circle represents a planet, and its size represents the planetary mass. The formation and evolution of the system highlighted with a light-blue-green rectangle is depicted in Fig. 2 . As the mass of the planetesimal ring is increased from M ring = 5 M ⊕ to 40 M ⊕ , both the number of embryos that achieve the planetary mass scale, as well as their average mass itself, 〈 M 〉, increase. Within the numerical model, radial migration is taken to smoothly terminate at r = 0.5 au across a characteristic length-scale of ±0.1 au (see the ‘Planet–disk interactions’ section in Methods ). This stellocentric distance is marked on top of the figure and emphasized by a vertical dotted line. This threshold also represents the boundary between short-period progenitors to the observed population of super-Earths and low-mass objects that remain stranded close to their formation site. By and large, these ( r < 0.5 au) synthetic planetary systems adhere to a pattern of intra-system mass uniformity with the normalized mass dispersion, D M /〈 M 〉, that systematically reaches values comparable to, or smaller than, the observed value of ( D M /〈 M 〉) data ≈ 0.48. Cumulatively, these results explain how planetary systems can emerge with a broad diversity of masses while retaining an unexpectedly high degree of self-uniformity. Full size image We have not simulated the onset of such post-nebular instabilities here because the terminal point of our model corresponds to the epoch of nebular dissipation. Nevertheless, a broad array of dynamical pathways through which short-period resonant chains of planets can become unstable, including mass-loss through photo-evaporation 25 as well as interactions with a fading quadrupolar moment of the host star 26 , has been well-documented in the literature, and large-scale operation of such instabilities has already been shown to be likely 27 . It is further worth noting that destabilization of resonant chains of planets yields a period-ratio distribution that is indistinguishable from the data 23 and recent work has demonstrated that the degradation in the planetary mass homogeneity is generally very mild during a dynamical instability, such that systems that experience transient epochs of scattering remain fully consistent with the ‘peas in a pod’ pattern of uniformity observed in the population of extrasolar super-Earths 24 . We conclude this work by remarking upon the connection between our exoplanet formation model and the formation of the Earth itself 6 , as the two are indeed related within the context of our picture. Despite the innate complexity inherent to planetary accretion, the terminal outcome of our envisioned scenario is determined chiefly by the mass of the planetesimal ring, which in turn depends on a variety of the disk’s physical properties, with turbulent viscosity playing a central role 3 . This dependence is driven by the fact that the viscosity (and overall metallicity) controls the cumulative mass entrained within the population of rocky planetesimals (Extended Data Fig. 1 ). To this end, we note that beyond the masses of the planets themselves, their terminal orbital architecture is also sensitive to this parameter since planets that do not accrete rapidly enough do not experience long-range inward migration and remain close to their original formation site. Thus, within the framework of our model, one of the key reasons that the sun is encircled by the Earth, and not a group of more-massive short-period planets, is simply that the protosolar nebula was sufficiently turbulent to inhibit the agglomeration of a more massive ring of rocky planetesimals at 1 au, which prevented the terrestrial planets from growing massive enough to migrate inwards before the nebular clock had run out. If the general predictions of our model endure, a unifying model for the origins of the Earth, the moons of Jupiter and Saturn, and extrasolar planets may finally lie within reach. Methods To quantify the concurrent processes of planetary accretion and orbital migration, we have employed the mercury6 N -body simulation code 28 and have augmented it to account for a series of effects that are expected to arise within protoplanetary disks. Some of these effects were modelled self-consistently, while others were implemented through ad hoc parameterizations for the sake of computational efficiency. Below, we describe the individual elements of our model and their physical rationale. Simulation setup, resolution and initial conditions Following conventional practice of N -body calculations of planet formation, we break up our simulations into two classes of particles: fully self-gravitating planetary ‘embryos’ and semi-active ‘planetesimals’, which interact with embryos but not one another (strictly speaking, only direct coupling between planetesimals is suppressed; indirect interactions among these particles that are transmitted through the reflex motion of the central body remain, and drive a minor but non-physical excitation of the planetesimals’ velocity dispersion 29 ). The computational cost of a given numerical experiment scales quadratically with the number of embryos ( \(\propto {N}_{{{{\rm{emb}}}}}^{2}\) ) and linearly with the number of planetesimals ( ∝ N emb N pl ). Because N pl is taken to exceed N emb by a large margin in typical planet formation calculations, the computational cost is primarily controlled by the product of N emb and N pl . For this reason, we capped the planetesimal count in our simulations at N pl = 10 3 . The initial embryo masses used in our simulations were informed by particle-in-a-box calculations of collisional growth within a 20 M ⊕ planetesimal ring. To this end, we employed the Boulder code 30 to simulate the growth of protoplanetary embryos originating from 100 km objects, accounting for self-stirring of the planetesimal velocity dispersion as well as collisional damping, gas drag and dynamical friction. This calculation showed the emergence of lunar-mass planetary embryos after ~10,000 years of evolution. It is important to keep in mind that even in the most numerically heavy calculations, N pl is much smaller than the actual number of planetesimals that exist within protoplanetary disks. This means that each of our model planetesimals represents a ‘super-particle’ of mass M pl = M ring / N pl , and its dynamical evolution should be interpreted as a tracer of a large consortium of small bodies. Further, to simulate the inherent time dependence of the planetesimal formation process, we injected the particles into the simulation domain at a constant rate over a span of 100,000 years, consistent with the duration of the planetesimal formation epoch in the model outlined in ref. 3 . Both the embryos and planetesimals were introduced following a Gaussian profile in the semi-major axis, centred at 〈 a 〉 = 1 au and a standard deviation of Δ r = 0.1 au. Choosing a smaller value of Δ r yielded similar results because the gravitational scattering within the planetesimal swarm facilitates a relatively rapid radial spreading of the system. The initial eccentricities and inclinations were drawn from the Rayleigh distribution with a scale parameter of Δ e = Δ i = 0.01, while all orbital angles were uniformly randomized. In principle, the ratio N pl / N emb is an adjustable (and somewhat arbitrary) parameter of the numerical model. Due to chaotic self-regulation that ensues within the planet-forming region, however, we found that the results of our calculations are only weakly dependent on the number of embryos that are injected into the annulus of rocky material during the planetesimal formation epoch. To quantify this relative insensitivity, we carried out a series of 12 simulations where the total mass of the super-planetesimal disk was fixed at 20 M ⊕ but the number of lunar-mass embryos was varied from N emb = 5 to 10 to 20, running 4 realizations of each case. From these numerical experiments, we found that at the 300,000 year mark, the median number of proto-super-Earths that attain a mass greater than 1 M ⊕ clocked in at 3.5, 4.5 and 5, for simulations with 5, 10 and 20 embryos, respectively. Furthermore, we found a broad consistency in the properties of the emergent planets with average masses of M = 2.5 ± 0.8 M ⊕ , M = 2.3 ± 1.3 M ⊕ and M = 2.4 ± 1.4 M ⊕ (mean ± s.d.) for the 3 simulation subsets. Cumulatively, these results indicate that the dependence on the number of accreting particles saturates around N emb ≈ 10. We did, however, find that unlike simulations with 10 embryos, those with N emb = 20 consistently demonstrated a pronounced (and almost certainly unphysical) difference in the velocity dispersion of the super-planetesimals and low-mass protoplanets, with the former having a factor of ~2 lower eccentricities on average. Consequently, we chose to adopt N emb = 10 as a fiducial parameter in our model. With these parameters, the completion of a single run of the model on a 2.3 GHz machine required 60–80 CPU hours. N -body integration scheme and time step The gravitational dynamics of our system of planetary embryos and planetesimals were solved using the hybrid Wisdom–Holman/Bulirsch–Stoer integration algorithm 31 , 32 , implemented within the mercury6 software package 28 . Because our particle swarm was initialized in the vicinity of r ≈ 1 au, we adopted an initial time step of Δ t = 10 days. However, as disk-driven orbital evolution caused planetary orbits to decay towards their host star, we found it necessary to reduce the time step to Δ t = 1 day at the t = 300,000 year mark (and in some cases to an even lower value at later stages), to ensure that the symplectic time step remained smaller than ~1/20th of the shortest orbital period of any particle within the simulations. An adaptive time step was used to resolve close encounters, with an interparticle separation of Δ r = 3 R Hill marking the change-over radius for the symplectic-to-conventional integration scheme. The Bulirsch–Stoer accuracy parameter was set to \(\hat{\epsilon }=1{0}^{-8}\) . Finally, the radii of embryos were computed assuming a bulk density of 3 g cc −1 , and all collisions were treated as perfect mergers. Notably, we found that in our simulations, the collision cross-section was entirely dominated by that of the embryos, meaning that our results are insensitive to the choice of the mean density of the super-planetesimals. To demonstrate this, we carried out a simulation where the physical collision radii of super-planetesimals were effectively suppressed (by choosing a corresponding density of 10 4 g cc −1 ) while keeping the rest of the calculation (including damping prescriptions) unchanged. With such a setup, we obtained essentially identical results to our nominal simulation with planetesimal densities of 3 g cc −1 . Consequently, we adopted equal bulk densities of planetesimals and embryos, for definitiveness. Gas and dust profiles of the protoplanetary disk The presence of the gaseous component of the protoplanetary disk plays a key role in driving the early evolution of a forming planetary system. For definitiveness, in our work, we adopted various parameters directly derived from the ringed disk model delineated in ref. 3 as a guide for our functional parameterizations. More specifically, we first assumed that the gas surface density followed a Mestel-like profile 33 that decays exponentially in time: $${{\varSigma }}={{{\varSigma }}}_{0}\,\left(\frac{{r}_{0}}{r}\right)\,\exp \left(-t/{\tau }_{{{{\mathrm{disk}}}}}\right),$$ (1) where the initial value of the surface density at r 0 = 1 au is equal to Σ 0 = 2,500 g cm −2 and the disk decay constant is set to τ disk = 1.5 Myr. Second, we adopted a constant disk aspect ratio of h / r = 0.05 throughout the simulation. To this end, we note that while h / r does in principle change in time and is in general dependent on the disk viscosity itself 34 , this variation is not central to our numerical experiments, and is expected to only influence the results on a detailed level. The dust component of the disk is assumed to be composed of s • = 1 mm particles, in agreement with experimentally derived fragmentation thresholds of silicate grains 9 as well as theoretical computation of the Stokes number within the inner nebula 10 . The dust surface density itself is envisioned to be comprised of a ‘local’ component, which stems from an aerodynamically assisted buildup of solids at the silicate sublimation front, as well as an externally supplied flux of pebbles, F • = 10 −4 M ⊕ yr −1 (refs. 35 , 36 ), which is facilitated by the radial drift of solids 37 : $$\begin{array}{ll}{{{\varSigma }}}_{\bullet }&={{{\varSigma }}}_{\bullet 0}\,\exp \left[-{\left(\frac{r-{r}_{0}}{{{{\varDelta }}}_{{\mathrm{r}}}}\right)}^{2}\,\right]\,\exp \left[-{\left(\frac{t}{{\tau }_{\bullet }}\right)}^{2}\,\right]\\ &+\frac{{F}_{{{\bullet }}}}{4\,\uppi \,r\,{v}_{{{{\mathrm{kep}}}}}\,\eta \,{{{\mathrm{St}}}}}\,\exp \left[-\frac{t}{{\tau }_{{{{\mathrm{disk}}}}}}\right].\end{array}$$ (2) In the above expression, v kep is the Keplerian speed, Σ •0 = 500 g cm −2 , \({\mathcal{T}}\) • = 10 5 yr is the dust-depletion timescale and the functional form of the leading term was chosen to adequately approximate the spatial and temporal profiles of the r ≈ 1 au dust ring modelled in ref. 3 . Strictly speaking, F • represents the maximal pebble flux of the nebula, and it would have been appropriate to reduce this value in accordance with sequestration of material at the rock and ice sublimation lines, to account for the finite supply of solids in the disk. As our calculations show, however, the process of pebble accretion is highly inefficient in the inner disk, so lowering the pebble flux would merely diminish an already negligible effect. For the adopted surface density profile, the sub-Keplerian factor η ≈ (3/2)( h / r ) 2 = 3.75 × 10 −3 . Additionally, in the Epstein drag regime (applicable for the problem at hand) the Stokes number takes the form: $${{{\rm{St}}}}=\sqrt{\frac{\uppi }{8}}\frac{{\rho }_{\bullet }}{\rho }\frac{{s}_{\bullet }}{{c}_{s}}\,{{{\varOmega }}}_{{{}}}=\frac{\uppi \,{s}_{\bullet }\,{\rho }_{\bullet }}{2\,{{\varSigma }}},$$ (3) where ρ • = 3 g/cc is the particle density, ρ = Σ/(√(2 π ) h ) is the gas density, and c s = h Ω is the speed of sound. Planetesimal evolution facilitated by aerodynamic drag and collisional damping Interactions between planetesimals and the considerably more massive gaseous component of the disk ensue primarily through aerodynamic drag 37 . In the high-Reynolds-number regime (appropriate for planetesimals with \({{{\mathcal{R}}}}\gtrsim 1\,{\mathrm{km}}\) , where ℛ is the orbital radius (i.e. distance away from the star; ref. 38 ), the relevant drag acceleration has the form 37 , 39 : $${\bf{a}}\approx (1+\xi )\,\frac{3\,\rho }{16\,\bar{\rho }\,{{{\mathcal{R}}}}}\,{v}_{{{{\mathrm{rel}}}}}\,{\bf{v}}_{{{{\mathrm{rel}}}}},$$ (4) where v rel ≈ v − v kep refers to the relative velocity between a planetesimal and gas, and ξ ⩾ 0 is a numerical factor that can be used to mimic the effects of non-aerodynamic damping (see below). This acceleration was implemented into our model assuming that rocky planetesimals born at r ≈ 1 au have diameters of \(D=2\,{{{\mathcal{R}}}}=100\,{\mathrm{km}}\) and bulk densities of \(\bar{\rho }=3\,{\mathrm{g}}\,{\mathrm{cc}}^{-1}\) . A second important dissipative effect that affects planetesimals within the context of our calculations is collisional damping. Fundamentally, this process is driven by inelastic collisions among planetesimals, which occur with a characteristic frequency 40 : $$\frac{1}{{\tau }_{{{{\mathrm{coll}}}}}} \approx \frac{3\,{{{\varSigma }}}_{{{{\mathrm{pl}}}}}\,{{\varOmega }}}{8\,\bar{\rho }\,{{{\mathcal{R}}}}}.$$ (5) Although it is impossible to resolve this effect self-consistently in our super-particle calculations, we can crudely account for it by assuming that its effective functional form is similar to that of aerodynamic drag. Under this assumption, we can envision modelling the consequences of collisions (at least during the first few hundred thousand years, when the planetesimals’ velocity dispersion matters most) through the enhancement factor ξ introduced in equation ( 4 ). To quantify the relative importance of collisional damping to aerodynamic drag, and thereby determine the value of ξ , we begin by noting that the rate of aerodynamic damping of eccentricity ( e ) and inclination ( i ) is given by 39 : $$\begin{array}{ll}\frac{1}{{\tau }_{{{{\mathrm{aero}}}}}}=-\frac{1}{e}\frac{{\mathrm{d}}e}{{\mathrm{d}}t}=-\frac{2}{i}\frac{{\mathrm{d}}i}{{\mathrm{d}}t}=\frac{3\,\rho \,{v}_{{{{\mathrm{kep}}}}}}{16\,\bar{\rho }\,{{{\mathcal{R}}}}}\\ \qquad\quad\times {\left(\frac{5}{8}{e}^{2}+\frac{1}{2}{i}^{2}+{\eta }^{2}\right)}^{1/2} \approx \frac{3\,\eta \,\rho \,{v}_{{{{\mathrm{kep}}}}}}{16\,\bar{\rho }\,{{{\mathcal{R}}}}}.\end{array}$$ (6) The value of ξ can then be approximated as: $$\xi =\frac{{\tau }_{{{{\mathrm{aero}}}}}}{{\tau }_{{{{\mathrm{coll}}}}}} \approx \frac{\sqrt{8\,\uppi }}{\eta }\frac{{{{\varSigma }}}_{{{{\mathrm{pl}}}}}}{{{\varSigma }}}\frac{h}{r} \approx \sqrt{\frac{1}{2\,\uppi }}\frac{1}{\eta }\frac{{M}_{{{{\mathrm{ring}}}}}}{{{\varSigma }}\,{r}_{0}\,{{{\varDelta }}}_{r}}\frac{h}{r}.$$ (7) For nominal parameters quoted above, this factor evaluates to ξ ≈ 10. While we adopt this value as an upper limit in our calculations, to approximately capture the dependence of the collisional damping process on the planetesimal surface density, we scale ξ by the ratio \(({N}_{{{{\mathrm{pl}}}}}(t)/{N}_{{{{\mathrm{pl}}}}}^{{{{\mathrm{tot}}}}})\) in our implementation, where N pl ( t ) is the number of super-planetesimals present in the simulation at a given time step and \({N}_{{{{\mathrm{pl}}}}}^{{{{\mathrm{tot}}}}}=1{0}^{3}\) is the total number of super-planetesimals injected into the simulation. Importantly, this prescription is only designed to represent the relevant physics during the first ~2–3 × 10 5 years of the simulation, a period when the planetesimal population remains radially concentrated and collisional damping aids the accretion process. At later times, planetesimal dynamics play a negligible role in dictating the architecture of the emergent planetary system. Planet–disk interactions Generally speaking, the interactions between the gaseous nebula and solid material are not limited to planetesimals and aerodynamic drag. As planetary embryos accrete a sufficient amount of mass to raise substantial wakes in the gaseous disk, the gravitational back-reaction of the spiral density waves upon the planet drives an exchange of energy and angular momentum 41 , 42 . In a chemically inhomogeneous disk model, both the rate and the direction of the resulting planetary migration can depend sensitively on a variety of local disk properties 42 , 43 , 44 , as well as the planetary mass itself. In other words, even at the qualitative level, the picture of planetary migration can be rather complex in a detailed model of the nebula. In a simpler, power-law, locally isothermal disk model (of the type we adopt in this work), however, the sense of migration is strictly inward, and the semi-major axis, eccentricity and inclination damping timescales are well defined 15 , 45 : $$\begin{array}{ll}& {\mathcal{T}}_{\mathrm{mig}}=\frac{\gamma }{{{{\varOmega }}}_{\mathrm{kep}}}\frac{{M}_{\star }}{M}\frac{{M}_{\star }}{{{\varSigma }}\,{r}^{2}}{\left(\frac{h}{r}\right)}^{2}\\ &{{\mathcal{T}}}_{\mathrm{damp}}=\frac{{T}_{\mathrm{mig}}}{2}{\left(\frac{h}{r}\right)}^{2},\end{array}$$ (8) where the dimensionless constant γ ≈ 4. The effects of migration were implemented into our N -body scheme through auxiliary accelerations of the form 46 : $${\bf{a}}=-\frac{\zeta }{{T}_{{\mathrm{mig}}}}\,{\bf{v}}-\frac{2}{{{T}}_{{\mathrm{damp}}}}\left(\frac{({\bf{v}}\cdot {\bf{r}})\,{\bf{r}}}{{r}^{2}}+({\bf{v}}\cdot \hat{z})\,\hat{z}\right),$$ (9) where \(\hat{z}\) is the orbit-normal of the protoplanetary disk, that were only applied to the planetary embryos. A practically important and well-known attribute of protoplanetary disks is that they do not extend all the way down to stellar surfaces, but are instead truncated by their host stars’ magnetospheres 47 . Due to a dramatic enhancement of the co-rotation torque associated with a sharp surface density gradient, the migration directions reverses at the disk’s inner edge, meaning that disk cavities act as effective planet traps 48 , 49 . The resulting stalling of inward migration at r ≈ 10–20 R ⊙ facilitates the formation of resonant chains that span the a ≈ 0.1–0.5 au range 23 , 24 , 50 . A number of distinct approaches have been employed to model this magnetospheric cavity-driven torque reversal within the framework of N -body simulations, including semi-major axis re-normalization 14 , 51 , as well as ad hoc (for example, sinusoidal) modifications of the disk-driven accelerations 35 , 52 . While our numerical experiments are not sufficiently idealized for semi-major-axis rescaling to be applicable, a drawback of the latter approach is that it requires careful tuning of the damping effects, including the introduction of nonlinear eccentricity dependence in T damp and so on, to suppress unphysical excitation of the orbits. To avoid the usual difficulties of modelling the disk’s inner edge, here we opted for a simpler approach of smoothly diminishing the migration torque interior to r mig = 0.5 au such that our resonant chains stabilized with their outermost member at a ≲ r mig . We implemented this by choosing the following functional form for the multiplicative constant ζ in equation ( 9 ): $$\zeta =\frac{1}{2}\,\left(1+{{{\mathrm{erf}}}}\,\left[\frac{a-{r}_{{{{\mathrm{mig}}}}}}{{r}_{{{{\mathrm{mig}}}}}/10}\right]\right).$$ (10) In practice, we found that lowering the threshold semi-major axis to r mig = 0.25 au did not alter the structure of the emergent resonant chains in any appreciable manner. Pebble accretion In addition to growth facilitated by pairwise collisions between planetary embryos and planetesimals, in our simulations we also modelled direct capture of solid dust by the growing protoplanets. In general, this process, routinely referred to as pebble accretion 53 , 54 , can proceed in a number of distinct physical regimes, with their relative efficiency determined by the dust particle size, disk turbulence, planetary mass and so on. Recent work 10 has argued that in typical disks, pebble accretion is expected to unfold in the comparatively inefficient 3D regime interior to the water–ice sublimation line. Our assumption of a fixed s • = 1 mm particle radius yields results that are consistent with this presumption. By and large, this is due to the fact that in our model, the Stokes number (given by equation ( 3 )) evaluates to very small values; for example, St ≈ 2 × 10 −4 at r ≈ 1 au. In turn, this implies that even for relatively low values of the turbulence parameter, α , vertical settling is prohibitively inefficient, and the dust layer’s thickness remains comparable to that of the scale height of the gaseous nebula 55 : $$\frac{{h}_{\bullet }}{h}=\frac{1}{\sqrt{1+{{{\mathrm{Sc}}}}\,{{{\mathrm{St}}}}/\alpha }} \approx 1.$$ (11) We remark that the constant particle radius assumption can be lifted in favour of a more self-consistent theory for dust-gas coupling that accounts for turbulent stirring of dust grains and collisional fragmentation. Importantly, such a model also predicts that dust should be well-mixed with the gas in the inner regions of the disk, in agreement with the above estimate 10 . Because the dust component of the nebula never forms a thin sub-disk, the two-dimensional regime of pebble accretion never ensues. However, for consistency with the model of ref. 3 , where the value of the turbulent Schmidt number is taken to be on the order of Sc ≈ 10, we take h • / h = 2/5 in our model, such that h • / r = 0.02. As we discuss below, our results are not sensitive to this choice. The rate of 3D pebble accretion is given by 56 : $$\begin{array}{ll}{\dot{M}}_{{{{\mathrm{3D}}}}}&=6\,\uppi \,{{{\mathrm{St}}}}\,{R}_{{{{\mathrm{Hill}}}}}^{3}\,{{\varOmega }}\,\left(\frac{{{{\varSigma }}}_{\bullet }}{\sqrt{2\,\uppi }\,{h}_{\bullet }}\right)\\ &=\sqrt{2\,\uppi }\,{{{\mathrm{St}}}}\,{{{\varSigma }}}_{\bullet }\,{r}^{2}\,{{\varOmega }}\,\frac{M}{{M}_{\star }}\,\frac{r}{{h}_{\bullet }}.\end{array}$$ (12) Although this growth mode was implemented in our simulations, we found pebble accretion to be largely inconsequential in our calculations. This can be understood as follows. As the most favourable scenario for dust capture, let us consider an isolated protoplanetary embryo of mass m 0 that accretes pebbles at r = 1 au, where the primordial solid surface density is maximized. Substituting the functional form (equation ( 2 )) for Σ • , it is straightforward to show that the accreted mass is bounded from above by: $$\begin{array}{ll}&{{\Delta }}M < \int\nolimits_{0}^{\infty }{\dot{M}}_{{\mathrm{3D}}}\,{\mathrm{d}}t=\frac{\uppi }{\sqrt{2}}\,{\mathrm{St}}\,{\varSigma }_{\bullet 0}\,{r}^{2}\,{{\varOmega }}\,{\tau }_{\bullet }\,\frac{{m}_{0}}{{M}_{\star }}\frac{r}{{h}_{\bullet }}\\ &+\frac{{F}_{\bullet }\,{\tau }_{{\mathrm{disk}}}}{2\,\sqrt{2\,\uppi }\,\eta }\frac{{m}_{0}}{{M}_{\star }}\,\frac{r}{{h}_{\bullet }}\approx 5.2\,{m}_{0},\end{array}$$ (13) with the dominant contribution arising from the leading term (the externally supplied pebble flux accounts for less than a quarter of the accreted mass). This simple estimate indicates that even under the most optimistic conditions, pebble accretion can only boost the mass of an embryo by a factor of a few. Moreover, it is important to keep in mind that the efficiency of this process diminishes rapidly over the course of the first ~10 5 years, as the majority of pebbles are converted into planetesimals (because of the smallness of this effect, we did not implement filtering of pebble flux in our model). Accordingly, while we include the pebble accretion process in our simulations for completeness, our numerical experiments indicate that a far more dominant role in facilitating planetary growth is played by the accretion of planetesimals and mergers among protoplanetary embryos. Data availability Ascii output files summarizing the time series of our reference simulation (with an output interval of 1,000 years, totalling 1,010 files) are provided at . Code availability This work utilizes the publicly available mercury6 code ( ). The subroutine detailing user-defined forces is available on request from the corresponding author (K.B.).

A new theory for how rocky planets form could explain the origin of so-called "super-Earths"—a class of exoplanets a few times more massive than Earth that are the most abundant type of planet in the galaxy. Further, it could explain why super-Earths within a single planetary system often wind up looking strangely similar in size, as though each system were only capable of producing a single kind of planet. "As our observations of exoplanets have grown over the past decade, it has become clear that the standard theory of planet formation needs to be revised, starting with the fundamentals. We need a theory that can simultaneously explain the formation of the terrestrial planets in our solar system as well as the origins of self-similar systems of super-Earths, many of which appear rocky in composition," says Caltech professor of planetary science Konstantin Batygin, who collaborated with Alessandro Morbidelli of the Observatoire de la Côte d'Azur in France on the new theory. A paper explaining their work was published by Nature Astronomy on Jan. 12. Planetary systems begin their lifecycles as large spinning disks of gas and dust that consolidate over the course of a few million years or so. Most of the gas accretes into the star at the center of the system, while solid material slowly coalesces into asteroids, comets, planets, and moons. In our solar system, there are two distinct types of planets: the smaller rocky inner planets closest to the sun and the outer larger water- and hydrogen-rich gas giants that are farther from the sun. In an earlier study published in Nature Astronomy at the end of 2021, this dichotomy led Morbidelli, Batygin, and colleagues to suggest that planet formation in our solar system occurred in two distinct rings in the protoplanetary disk: an inner one where the small rocky planets formed and an outer one for the more massive icy planets (two of which—Jupiter and Saturn—later grew into gas giants). Super-Earths, as the name suggests, are more massive than the Earth. Some even have hydrogen atmospheres, which makes them appear almost gas giant-like. Moreover, they are often found orbiting close to their stars, suggesting that they migrated to their current location from more distant orbits. "A few years ago we built a model where super-Earths formed in the icy part of the protoplanetary disk and migrated all the way to the inner edge of the disk, near the star," says Morbidelli. "The model could explain the masses and orbits of super-Earths but predicted that all are water-rich. Recent observations, however, have demonstrated that most super-Earths are rocky, like the Earth, even if surrounded by a hydrogen atmosphere. That was the death sentence for our old model." Over the past five years, the story has gotten even weirder as scientists—including a team led by Andrew Howard, professor of astronomy at Caltech; Lauren Weiss, assistant professor at the University of Notre Dame; and Erik Petigura, formerly a Sagan Postdoctoral Scholar in Astronomy at Caltech and now a professor at UCLA—have studied these exoplanets and made an unusual discovery: while there exists a wide variety of types of super-Earths, all of the super-Earths within a single planetary system tend to be similar in terms of orbital spacing, size, mass, and other key features. "Lauren discovered that within a single planetary system, super-Earths are like 'peas in a pod,'" says Howard, who was not directly connected with the Batygin-Morbidelli paper but has reviewed it. "You basically have a planet factory that only knows how to make planets of one mass, and it just squirts them out one after the other." So, what single process could have given rise to the rocky planets in our solar system but also to uniform systems of rocky super-Earths? "The answer turns out to be related to something we figured out in 2020 but didn't realize applied to planetary formation more broadly," Batygin says. In a 2020 paper published in The Astrophysical Journal, Batygin and Morbidelli proposed a new theory for the formation of Jupiter's four largest moons (Io, Europa, Ganymede, and Callisto). In essence, they demonstrated that for a specific size range of dust grains, the force dragging the grains toward Jupiter and the force (or entrainment) carrying those grains in an outward flow of gas cancel each other perfectly. That balance in forces created a ring of material that constituted the solid building blocks for the subsequent formation of the moons. Further, the theory suggests that bodies would grow in the ring until they become large enough to exit the ring due to gas-driven migration. After that, they stop growing, which explains why the process produces bodies of similar sizes. In their new paper, Batygin and Morbidelli suggest that the mechanism for forming planets around stars is largely the same. In the planetary case, the large-scale concentration of solid rocky material occurs at a narrow band in the disk called the silicate sublimation line—a region where silicate vapors condense to form solid, rocky pebbles. "If you're a dust grain, you feel considerable headwind in the disk because the gas is orbiting a bit more slowly, and you spiral toward the star; but if you're in vapor form, you simply spiral outward, together with the gas in the expanding disk. So that place where you turn from vapor into solids is where material accumulates," Batygin says. The new theory identifies this band as the likely site for a "planet factory," that over time, can produce several similarly sized rocky planets. Moreover, as planets grow sufficiently massive, their interactions with the disk will tend to draw these worlds inward, closer to the star. Batygin and Morbidelli's theory is backed up by extensive computer modeling but began with a simple question. "We looked at the existing model of planet formation, knowing that it does not reproduce what we see, and asked, 'What assertion are we taking for granted?'" Batygin says. "The trick is to look at something that everybody takes to be true but for no good reason." In this case, the assumption was that solid material is dispersed throughout the protoplanetary disks. By jettisoning that assumption and instead supposing that the first solid bodies form in rings, the new theory can explain different types of planetary systems with a unified framework, Batygin says. If the rocky ring contains a lot of mass, planets grow until they migrate away from the ring, resulting in a system of similar super-Earths. If the ring contains little mass, it produces a system that looks much more like our solar system's terrestrial planets. "I'm an observer and an instrument builder, but I pay extremely close attention to the literature," Howard says. "We get a regular dribble of little-but-still-important contributions. But every five years or so, someone comes out with something that creates a seismic shift in the field. This is one of those papers."

10.1038/s41550-022-01850-5

Medicine

Synaptic protein regulates anxiety behaviour

Olga Babaev et al. IgSF9b regulates anxiety behaviors through effects on centromedial amygdala inhibitory synapses, Nature Communications (2018). DOI: 10.1038/s41467-018-07762-1 Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-018-07762-1

https://medicalxpress.com/news/2018-12-synaptic-protein-anxiety-behaviour.html

Abstract Abnormalities in synaptic inhibition play a critical role in psychiatric disorders, and accordingly, it is essential to understand the molecular mechanisms linking components of the inhibitory postsynapse to psychiatrically relevant neural circuits and behaviors. Here we study the role of IgSF9b, an adhesion protein that has been associated with affective disorders, in the amygdala anxiety circuitry. We show that deletion of IgSF9b normalizes anxiety-related behaviors and neural processing in mice lacking the synapse organizer Neuroligin-2 (Nlgn2), which was proposed to complex with IgSF9b. This normalization occurs through differential effects of Nlgn2 and IgSF9b at inhibitory synapses in the basal and centromedial amygdala (CeM), respectively. Moreover, deletion of IgSF9b in the CeM of adult Nlgn2 knockout mice has a prominent anxiolytic effect. Our data place IgSF9b as a key regulator of inhibition in the amygdala and indicate that IgSF9b-expressing synapses in the CeM may represent a target for anxiolytic therapies. Introduction Understanding the molecular basis of psychiatric disorders is one of the foremost medical challenges of our times, and substantial interest has arisen in the role of synaptic dysfunction in psychiatric pathophysiology 1 , 2 . The vast majority of corresponding studies have focused on glutamatergic, excitatory synapses, but it is becoming increasingly clear that an equally important contribution stems from abnormalities in inhibitory synaptic transmission 3 , 4 , 5 , 6 , 7 , 8 . This is particularly true for anxiety disorders, which have been linked to decreased inhibition in multiple brain regions 9 , 10 , 11 . To date however, astonishingly little is known about the molecular mechanisms by which abnormalities at inhibitory synapses contribute to the pathophysiology of anxiety disorders. A primary reason for this dearth of information is that the molecular composition of inhibitory postsynapses is only just being uncovered. In the past 5 years, the list of known inhibitory synapse organizers has expanded dramatically due to major technological advances 12 . A key question now is whether these molecules function uniformly across inhibitory synapses, or whether they display the same heterogeneity of function that is observed at the cellular level in the inhibitory network 3 , 13 . Understanding this organization will not only be essential for identifying disease mechanisms in pathological anxiety and other disorders, but may also provide a unique opportunity for uncovering selective targets for circuit-specific therapeutic interventions. A particularly interesting candidate molecule is IgSF9b, a member of the immunoglobulin superfamily of cell adhesion proteins that was recently shown to localize to inhibitory synapses in dissociated neuron cultures 14 , 15 , 16 . Variants in IgSF9b are associated with major depression and the negative symptoms of schizophrenia 17 , 18 , 19 , indicating that IgSF9b may regulate affect in human patients. IgSF9b was recently reported to act in a complex with Neuroligin-2 (Nlgn2) 14 , a member of the Neuroligin family of adhesion proteins specifically found at inhibitory synapses 12 , 20 , 21 , 22 . Loss-of-function mutations in Nlgn2 have been identified in patients with schizophrenia or pathological anxiety and autism 23 , 24 , and in mice, deletion of Nlgn2 robustly enhances anxiety-related behaviors 8 , 25 . Together, these findings raise the intriguing possibility that IgSF9b, like Nlgn2, may contribute to psychiatric pathophysiology by regulating synaptic inhibition in behaviorally relevant circuitry. To date, however, the molecular, cellular and circuitry functions of IgSF9b and its potential interactions with Nlgn2 in vivo remain completely unexplored. In the present study, we investigated the role of synaptic inhibition in anxiety by addressing two key questions: First, does IgSF9b regulate anxiety-related behavior and neural processing in the anxiety circuitry of WT and/or pathologically anxious Nlgn2 KO mice? Second, do IgSF9b and/or Nlgn2 act at specific inhibitory synapses within the anxiety circuitry and thus offer synapse-specific targets for interventions? To this end, we investigated the consequences of IgSF9b deletion and IgSF9b x Nlgn2 double deletion on anxiety-related behavior and circuits, as well as on inhibitory synapses in the amygdala, a brain region that plays a central role in the processing of anxiety information 9 , 10 , 26 , 27 . We show that IgSF9b regulates anxiety-like behaviors in Nlgn2 KO mice and that IgSF9b and Nlgn2 exert differential effects on distinct components of the amygdala inhibitory circuitry. Our data provide the first characterization of IgSF9b function in vivo and uncover a prominent anxiolytic consequence of IgSF9b deletion in Nlgn2 knockout (KO) mice, implicating IgSF9b-expressing neurons in the centromedial (CeM) amygdala as potential targets for anxiolytic therapies. Results Deletion of IgSF9b normalizes anxiety in Nlgn2 KO mice To address the role of IgSF9b in anxiety processing, we first tested whether IgSF9b deletion would alter anxiety-related behaviors and/or modulate the anxiety phenotype observed in Nlgn2 KO mice. To this end, we assessed the behavior of male adult WT, Nlgn2 single KO, IgSF9b single KO, and Nlgn2 x IgSF9b double KO mice in an open field (OF) apparatus (Fig. 1a–c ). As expected 8 , 25 , Nlgn2 KO mice spent significantly less time, traveled less distance, and made fewer entries into the anxiogenic center of the OF (Fig. 1d–g , red bars; ANOVA comparisons listed in Table 1 ; error bars represent SEM), indicative of enhanced anxiety-related behavior. Based on previous findings indicating that IgSF9b and Nlgn2 are components of the same molecular complex at inhibitory synapses 14 , we expected that deletion of IgSF9b may increase anxiety and/or exacerbate the phenotype of Nlgn2 KO mice. To our surprise, however, IgSF9b KO mice demonstrated a trend towards decreased anxiety (Fig. 1e, f , Supplementary Fig. 1d, e , blue bars). Most strikingly, the prominent anxiety phenotype of the single Nlgn2 KO mice was completely abolished in the double KO mice, with all measures unchanged from WT levels (Fig. 1 d–g, Supplementary Fig. 1 a–h, purple bars). An identical pattern of behavior was observed in female mice (Supplementary Fig. 1 i–m). Moreover, Nlgn2 KO mice spent less time, traveled less distance and made fewer entries into the anxiogenic open arms of the elevated plus maze (EPM) (Fig. 1h–m , red bars), while IgSF9b deletion abolished the anxiety phenotype of Nlgn2 KO mice (Fig. 1h–m , purple bars). Fig. 1 Deletion of IgSF9b normalizes anxiety-related behavior in Nlgn2 KO mice. a , b WT, Nlgn2 KO, IgSF9b KO, and double KO mice were assessed for anxiety-related behavior in an OF test. c Representative tracks of OF exploration. d Time spent in the anxiogenic center of the OF. e Distance traveled in the center of the OF, expressed as percentage of total distance traveled. f Number of entries into the center of the OF. g Total distance traveled in the OF. h Mice were assessed for anxiety-related behavior in EPM. i Representative tracks of EPM exploration. j Time spent in the anxiogenic open arms of the EPM. k Distance traveled in the open arms of the EPM, expressed as percentage of total distance traveled. l Number of entries into the open arms of the EPM. m Total distance traveled in the EPM. Statistically significant ANOVA comparisons are marked in gray at the top of the panels and are listed in Table 1 . For all other ANOVA comparisons, F < 1. Post-hoc analysis: * p < 0.05 relative to WT, ** p < 0.01 relative to WT, *** p < 0.001 relative to WT, # p < 0.05 relative to double KO, ## p < 0.01 relative to double KO, ### p < 0.001 relative to double KO. Error bars represent SEM, n = 10–12 mice per genotype for OF, n = 9–14 mice per genotype for EPM Full size image Table 1 Two-way ANOVA comparisons Full size table Interestingly, IgSF9b KO mice showed a significant hyperactivity in both the OF and the EPM (Fig. 1g, m , blue bar) that was not a result of a basic locomotor phenotype (Supplementary Fig. 2 a–i). This increased exploratory activity may be an antagonistic behavior to freezing typically demonstrated by anxious mice in novel environment 28 , and indeed “anxious” Nlgn2 KO mice demonstrate reduction in locomotor activity specifically under anxiogenic conditions (Fig. 1 g, red bar; see also Babaev et al., 2016 8 ). Taken together, these findings indicate a robust anxiolytic effect of IgSF9b deletion in Nlgn2 KO mice. Regulation of the anxiety circuitry by IgSF9b and Nlgn2 To understand the mechanism underlying this behavioral normalization, we sought to identify the relevant neural circuits using immunostaining for the immediate early gene cFos as a marker for neuronal activity 29 . We focused on the amygdala due to its well-established central role in the processing of anxiety information 9 , 26 , 27 (Fig. 2 a). Consistent with our previous report 8 , neurons in the basal (BA) but not the LA were overactivated in Nlgn2 KO mice following exposure to the OF (Fig. 2a , left panels, red bars). However, this overactivation was not reversed in double KO mice (Fig. 2a , lower left panel, purple bar), and a trend towards an increase in activation in the IgSF9b KO was observed (Fig. 2a , lower left panel, blue bar). The observed increase in cFos was induced by exposure to the OF, as BA cFos in mice taken directly from their home cages did not differ among genotypes (Fig. 2b ). Co-immunostaining with the interneuron markers parvalbumin (PV) and somatostatin (SOM) showed that the increase in double KO mice may be partially due to an increase in the number of activated PV-positive interneurons in these mice (Supplementary Fig. 3 a–h). Fig. 2 Deletion of IgSF9b normalizes neuronal activation in the CeM of Nlgn2 KO mice. a cFos immunolabeling in the LA, BA, CeL, and CeM in all four genotypes. Representative images and quantification of cFos-positive cells in the LA (upper left panel), the BA (lower left panel), the CeL (upper right panel), and the CeM (lower right panel). Scale bar, 50 μm. b Number of cFos-positive cells in the BA and CeM in WT, Nlgn2 KO, IgSF9b KO, and double KO mice under home cage conditions. c Experimental design of retrograde labeling. Mice were injected with Retrobeads into the CeM, and 1 week later they were exposed to the OF for 10 min, perfused 90 min later and analyzed for double labeling of cFos and Retrobeads in the BA and CeL. Scale bar, 100 μm. d Representative image and quantification of cFos- and Retrobead-double positive cells in BA in all four genotypes. Scale bar, 20 μm. e Representative image and quantification of cFos- and Retrobead-double positive cells in the CeL in all four genotypes. Scale bar, 20 μm. Statistically significant ANOVA comparisons are marked in gray at the top of the panels and are listed in Table 1 . For all other ANOVA comparisons, F < 1. Post-hoc analysis: * p < 0.05 relative to WT, ** p < 0.01 relative to WT, # p < 0.05 relative to double KO, ## p < 0.01 relative to double KO, ### p < 0.001 relative to double KO. Error bars represent SEM, n = 5–8 mice per genotype. WT, white bars; Nlgn2 KO, red bars; IgSF9b KO, blue bars; double KO, purple bars Full size image BA neurons project to the centrolateral (CeL) and the CeM nucleus (Fig. 2a ). Given that the CeM represents the primary output nucleus of the amygdala, which projects directly to downstream targets to mediate anxiety-related behaviors 9 , 27 , we tested whether Nlgn2 and IgSF9b deletion may have opposite effects on CeM activation that underlie their opposing anxiety-related consequences. Analysis of CeM activation during OF exposure revealed an overactivation of the CeM in Nlgn2 KO mice (Fig. 2a , lower right panel, red bar), with no effect in IgSF9b KO mice (Fig. 2a , lower right panel, blue bar). Strikingly, the overactivation of the CeM was completely normalized in double KO mice (Fig. 2a , lower right panel, purple bar), which is highly consistent with their WT-like anxiety-related behavior. Similar analysis of the CeL revealed no effect in either Nlgn2 KO or IgSF9b KO mice (Fig. 2a , upper right panel). To confirm that suppression of BA → CeM projection neurons is indeed the mechanism behind suppression of anxiety-related activity in the CeM, we stereotaxically injected a fluorescent Retrobead retrograde tracer into the CeM to label neuronal populations that project directly to the CeM (Fig. 2c and Supplementary Fig. 3i-m ). We measured activation of Retrobead-positive neurons in the BA under anxiogenic conditions using cFos immunohistochemistry (Fig. 2d ). Importantly, the stereotaxic surgery did not alter anxiety in any of the KO mice (Supplementary Fig. 3 l–m). Analysis of Retrobead- and cFos-double positive neurons revealed that BA → CeM direct projections were strongly overactivated in Nlgn2 KO mice (Fig. 2d , red bar), and surprisingly, overactivation persisted in double KO mice (Fig. 2d , purple bar). In contrast, activation of inhibitory CeL → CeM projections did not differ among any of the genotypes (Fig. 2e ). Together, these data indicate that Nlgn2 and IgSF9b deletion affect distinct targets within the amygdala: While Nlgn2 regulates anxiety-induced activation of projection neurons in BA, IgSF9b may normalize the CeM anxiogenic output by local mechanism within the CeM. IgSF9b and Nlgn2 bidirectionally regulate CeM activity To confirm that changes in the neural activity of the CeM accompany anxiety-related behavior, and to investigate whether IgSF9b and/or Nlgn2 alter specific neural substrates in the CeM in vivo, we recorded local field potential (LFP) oscillations in the CeM in freely moving mice during exploration of the OF (Fig. 3a, b and Supplementary Fig. 4a-b ). Spectral analysis of CeM LFPs revealed that exposure to the OF increased oscillatory activity in all genotypes compared to home cage CeM activity, particularly in the theta (4–12 Hz) and beta (18–30 Hz) ranges (Supplementary Fig. 4b-c ). The increase in the beta frequency band was most prominent in Nlgn2 KO mice, indicating that Nlgn2 deletion modifies anxiety-related CeM activity in the beta frequency range. To further explore whether this increase in beta power was modulated by anxiogenic conditions, we compared CeM activity during exploration (defined as speed > 1 cm/s) of the relative safety of the periphery with the potentially anxiogenic center of the OF (Fig. 3c, d ). Beta power increased in Nlgn2 KO mice during exploration of the OF center, and this increase was completely abolished in double KO mice (Fig. 3e , Supplementary Fig. 4d ). Furthermore, we found a significant correlation between the magnitude in beta power and the distance from the center of the OF specifically in Nlgn2 KO mice (Fig. 3 f–i). The increase in beta power was not induced by changes in locomotion, since beta activity was not modulated by speed in any genotype (Supplementary Fig. 4e ). These results indicate that deletion of Nlgn2 increases beta oscillatory activity in the CeM under anxiogenic conditions, while deletion of IgSF9b normalizes anxiety-related beta activity in double KO mice (Fig. 3d, e ), a mechanism that may underlie the normalization of anxiety-like behavior. Fig. 3 Neuronal activity in the beta frequency range is normalized in the CeM of double KO mice. a Representative image of the location of electrode in the CeM and experimental design. Scale bar, 500 µm. After electrode implantation mice were exposed to the OF for 15 min. b Representative traces of CeM LFPs from all four genotypes during SAP. c , d LFP power spectrum for all four genotypes during movement in the periphery ( c ) and center ( d ) of the OF. e Average power increase in the center relative to the periphery of the OF for the beta frequency range (18–30 Hz). f – i Increase in normalized beta power as a function of distance from the center in WT ( f ), Nlgn2 KO ( g ), IgSF9b KO ( h ), and double KO ( i ) mice. Statistical analysis of correlations is indicated in the figure. j Schematic of the stretch-attend posture (SAP) scored during the OF test. k Representative wavelet transforms of CeM LFP of the four genotypes during exploration of the OF. White horizontal lines indicate the duration of a representative SAP. l Average beta power during SAP. Statistically significant ANOVA comparisons are marked in gray at the top of the panels and are listed in Table 1 . For all other ANOVA comparisons, F < 1. Post-hoc analysis: * p < 0.05 relative to WT, *** p < 0.001 relative to WT, # p < 0.05 relative to double KO, ## p < 0.01 relative to double KO, ### p < 0.001 relative to double KO. SAPs and mice: WT n = 77/ 6 mice; Nlgn2 KO n = 62/ 5 mice; IgSF9b KO n = 84/ 6 mice; double KO n = 71/ 5 mice. Error bars represent SEM. WT, white bars; Nlgn2 KO, red bars; IgSF9b KO, blue bars; double KO, purple bars Full size image To further characterize the increase in beta oscillatory activity in Nlgn2 KO mice and the normalization in double KO mice, we analyzed the temporal dynamics in beta power during transitions from the periphery to the center of the OF. In agreement with their robust avoidance of the center of the OF, the number of transitions of Nlgn2 KO mice was too low to perform a powerful statistical analysis (4.4 ± 2.6 transitions). However, we noticed that Nlgn2 KO mice displayed frequent stretch-attend postures (SAPs), a risk-assessment behavior that reflects an internal conflict between anxiety and the exploratory drive 30 , 31 (Fig. 3j ). Nlgn2 KO mice showed a significant increase in the magnitude of beta power during SAPs compared to WT mice, indicative of a correlation between beta oscillations and anxiogenic conditions (Fig. 3k–l , red bar). Strikingly, IgSF9b showed a significant reduction in beta power, while double KO mice showed similar levels as WT mice (Fig. 3l, purple bar ). Together, these data highlight that (1) CeM neural activity is modulated by the valence of the environment (“safe” vs “anxiogenic”), (2) beta oscillatory activity in the CeM represents a novel neural signature of pathological anxiety induced by deletion of Nlgn2, and (3) Nlgn2 and IgSF9b bidirectionally modulate anxiety-related neural activity in the CeM particularly during risk-assessment behavior. IgSF9b knockdown in CeM normalizes anxiety in Nlgn2 KO mice To further confirm that IgSF9b acts in the CeM to modulate anxiety, and to investigate whether targeting IgSF9b-containing synapses in the adult amygdala may recapitulate the anxiolytic effects in the constitutive global KO, we locally reduced IgSF9b levels by adeno-associated virus (AAV)-mediated expression of IgSF9b shRNA (Fig. 4 a–b). AAV particles encoding IgSF9b shRNA or control shRNA (a mutant construct that lacks knockdown activity, as reported previously 14 ) were injected into the CeM of 8–12-week-old WT and Nlgn2 KO mice using stereotaxic surgery (Fig. 4 c–d and Supplementary Fig. 5a ), generating four experimental groups (Fig. 4e ). Fig. 4 Local knockdown of IgSF9b in the adult CeM ameliorates anxiety-related behaviors in Nlgn2 KO mice. a Immunoblot for IgSF9b in HEK cell lysates transfected with a Myc-IgSF9b construct and IgSF9b shRNA or control shRNA 14 . Numbers next to the protein ladder represent molecular weight in kDa. b Representative images and quantification of perisomatic IgSF9b puncta in the CeM following injection of control shRNA or IgSF9b shRNA. n = 3 mice. Scale bar, 20 μm. c Schematic diagram showing experimental design of IgSF9b shRNA experiment. Mice were tested in the OF 24 h before (pre) and 6 weeks after (post) injection. d Representative image of GFP-positive neurons in viral injection site. e Four experimental groups were generated: WT + control shRNA (white), WT + IgSF9b shRNA (white and blue striped), Nlgn2 KO + control shRNA (red), and Nlgn2 KO + IgSF9b shRNA (red and blue striped). f Time in center post-injection expressed as % pre-injection. g Normalized distance in center post-injection expressed as % pre-injection. h Representative tracks of OF exploration of Nlgn2 KO mice pre- and post-injection. Statistically significant ANOVA comparisons are marked in gray at the top of the panels and are listed in Table 1 . For all other ANOVA comparisons F < 1. Post-hoc analysis: * p < 0.05 relative to WT, ** p < 0.01 relative to WT. Error bars represent SEM, n = 6–9 mice per group Full size image One day before and 6 weeks after surgery, anxiety-related behavior was assessed in an OF (Fig. 4c ). IgSF9b shRNA had no significant effect on either the time or the distance traveled in the center of the OF in WT mice (Fig. 4f–g , white vs. blue and white shaded bars). In contrast, Nlgn2 KO mice injected with IgSF9b shRNA showed a pronounced reduction of anxiety-related behaviors compared to Nlgn2 KO mice injected with control shRNA, as evidenced by a significant increase in both the time and the distance traveled in the center (Fig. 4f–g , red vs. red and blue shaded bars, and Fig. 4h , representative traces). Interestingly, anxiety levels appeared to be exacerbated in Nlgn2 KO but not in WT mice over the observed 6-week time period, even in mice that had not undergone surgery (Supplementary Fig. 5b-d ), and this effect was completely reversed by local reduction of IgSF9b (Fig. 4f–h , Supplementary Fig. 5e-f ). Given that IgSF9b is highly expressed in the CeM (Supplementary Fig. 6a-d ), these data confirm that IgSF9b modulates anxiety-related behaviors through a CeM-specific mechanism, and indicates that targeting inhibitory transmission in the CeM can ameliorate anxiety-related behaviors. IgSF9b deletion alters CeM inhibitory synapse function To elucidate the synaptic mechanisms that may be responsible for the anxiolytic effect of IgSF9b deletion in the CeM, we recorded miniature inhibitory postsynaptic currents (mIPSCs) from acute brain slices obtained from adult mice (Fig. 5 ). To assist in the identification of the relevant brain structures and to distinguish between excitatory and inhibitory neurons, we used mice of all four genotypes that additionally expressed the fluorescent marker Venus under a vesicular inhibitory amino acid transporter (VIAAT) promoter 32 . Given that CeM neurons consisted primarily of inhibitory (Venus-positive) neurons, we restricted our analysis to these neurons. Fig. 5 IgSF9b deletion upregulates inhibitory synapse function in the CeM. a Schematic diagram showing location of mIPSC and sEPSC recordings in the CeM. b Representative firing patterns of CeM neurons. c – e Mean number of action potentials (APs) in response to a 500 ms current step injection ( c ), mean amplitude of the first AP ( d ), and mean AP firing threshold ( e ) in CeM neurons of WT, Nlgn2 KO, IgSF9b KO, and double KO mice. f , g Mean mIPSCs and representative mIPSC traces. h Average cumulative distribution of mIPSC inter-event intervals and quantification of mean mIPSC frequency in the CeM. i Probability distribution of mean mIPSC frequencies among all analyzed cells in the CeM. Kolmogorov–Smirnov test: WT vs. IgSF9b KO, p < 0.0001; WT vs. Double KO, p < 0.001; Nlgn2 KO vs. IgSF9b, p < 0.0001; IgSF9b KO vs. Double KO, p < 0.01. j Average cumulative distribution of mIPSC amplitudes and quantification of mean mIPSC amplitude in the CeM. k Probability distribution of mean mIPSC amplitude of each analyzed cell in the CeM. Kolmogorov–Smirnov test: WT vs. Nlgn2 KO, p < 0.001. n = 17–22 cells/6–7 mice per genotype. l , m Mean sEPSCs and representative sEPSC traces. n Average cumulative distribution of sEPSC inter-event intervals and quantification of mean sEPSC frequency in the CeM. o Probability distribution of mean sEPSC frequencies among all analyzed cells in the CeM. Kolmogorov–Smirnov test: WT vs. Double KO, p < 0.001; IgSF9b KO vs. Double KO, p < 0.001. p Average cumulative distribution of sEPSC amplitudes and quantification of mean sEPSC amplitude in the CeM. q Probability distribution of mean sEPSC amplitude of each analyzed cell in the CeM. n = 13–15 cells/3–5 mice per genotype. Statistically significant ANOVA comparisons are marked in gray at the top of the panels and are listed in Table 1 . For all other ANOVA comparisons, F < 1. Post-hoc analysis: * p < 0.05 relative to WT, ** p < 0.01 relative to WT. Error bars represent SEM. Box plots represent median and 25 th and 75 th percentiles and whiskers are drawn from the minimum to the maximum value Full size image Neurons in the CeM showed typical firing patterns as previously reported (Fig. 5a, b ) 33 , and all parameters that reflect membrane excitability were similar between groups (Fig. 5c–e and Table 2 ). Surprisingly, however, IgSF9b deletion significantly increased mean mIPSC frequency while leaving mean mIPSC amplitudes unaffected (Fig. 5f–k , blue bars). This increase was due to a subset of IgSF9b KO neurons with substantially larger mean mIPSC frequencies, as revealed by a significant shift in the distribution of mean mIPSC frequencies among the neurons tested (Fig. 5i , blue vs. black line). Double KO mice showed a similar albeit less pronounced phenotype, with a trend towards an increase in mean mIPSC frequency (Fig. 5h , purple bar) and a significant shift in the distribution of mean mIPSC frequencies in the neuronal population (Fig. 5 i, purple vs. black line). Consistent with our previous findings 8 , Nlgn2 deletion did not affect mean mIPSC frequency, and only modestly reduced mean mIPSC amplitude (Fig. 5h–k , red bars). No effects of any genotype were observed on the frequency and amplitude of spontaneous postsynaptic excitatory currents (sEPSCs) in the CeM (Fig. 5 l–q), although, interestingly, comparison of the probability distribution of mean frequencies among all cells to the corresponding WT distribution (Fig. 5o ) revealed a small shift towards lower mean frequencies in double KO. Rise time and decay time of mIPSCs and sEPSCs were not significantly altered (Table 2 ). Together, our observations indicate that deletion of IgSF9b specifically alters the function of inhibitory, but not excitatory, synapses. These findings indicate that IgSF9b deletion suppresses neuronal overactivation in the CeM of double KO mice, and hence may reduce their anxiety-related behavior, through increased inhibitory input onto CeM projection neurons. Table 2 Kinetics of PSCs and passive membrane properties of inhibitory neurons in CeM Full size table IgSF9b deletion increases VIAAT puncta in the CeM We next investigated the molecular mechanisms by which IgSF9b deletion increases synaptic inhibition in the CeM. Immunohistochemistry for IgSF9b revealed punctate staining in the BA and CeM of adult WT mice, which was absent in IgSF9b KO mice (Supplementary Fig. 6 a–d). Co-staining of IgSF9b with Nlgn2 revealed that in the CeM, but not in the BA, these proteins are localized in adjacent overlapping domains (Supplementary Fig. 6 e–f), consistent with previous findings in hippocampal cultures 14 . These data indicate that IgSF9b and Nlgn2 likely function in separate subdomains of the inhibitory synapse as previously proposed 14 . To assess the molecular basis of the increase in inhibitory transmission in IgSF9b KO and double KO mice, we performed immunohistochemistry for several markers of inhibitory synapses, including VIAAT, which labels inhibitory presynaptic terminals; gephyrin, the core scaffolding protein at inhibitory postsynapses; S-SCAM, a scaffolding protein that is only known synaptic interaction partner of IgSF9b; and the GABA A R subunits γ2 and α1 (Fig. 6a–g , Supplementary Fig. 6g ) 12 , 14 . Strikingly, IgSF9b KO and double KO mice showed a significant increase in the number of VIAAT puncta (Fig. 6c , blue and purple bars, respectively), consistent with the increase in mIPSC frequency, indicating that IgSF9b deletion increases the number and/or VIAAT content of inhibitory presynaptic terminals in the CeM. Gephyrin and S-SCAM remained unchanged (Fig. 6d, e , blue bars). IgSF9b deletion also resulted in a trend towards an increase in GABA A R γ2 subunit puncta, as well as a significant reduction in GABA A R α1 subunit puncta (Fig. 6f, g , blue bars). Nlgn2 KO mice showed a slight reduction in the number of GABA A R α1 subunit puncta, with no effects on any other synaptic marker (Fig. 6c–g , red bars), confirming the observation from the mIPSC recordings that Nlgn2 deletion has only very subtle effects in the CeM. Fig. 6 IgSF9b deletion increases VIAAT puncta in the CeM. a , b Schematic diagram showing location of immunohistochemistry analysis in the CeM ( a ) and genotypes assessed ( b ). c–g Photomicrographs and quantification of the number of perisomatic puncta of c VIAAT, d gephyrin, e S-SCAM, f GABA A Rγ2 and g GABA A Rα1 in all four genotypes. Scale bar, 2 μm. Error bars represent SEM, n = 5–8 mice per genotype. h Schematic diagram showing experimental design and representative photomicrographs of VIAAT puncta surrounding GFP-positive neurons in the CeM of Nlgn2 KO mice injected with control shRNA (left image) or IgSF9b shRNA (right image). Scale bar, 2 μm. i , j Number ( i ) and size ( j ) of VIAAT puncta in WT + control shRNA (white), WT + IgSF9b shRNA (white and blue striped), Nlgn2 KO + control shRNA (red), Nlgn2 KO + IgSF9b shRNA (red and blue striped). n = 20 cells/3 mice per experimental group. Statistically significant ANOVA comparisons are marked in gray at the top of the panels and are listed in Table 1 . For all other ANOVA comparisons F < 1. Post-hoc analysis: * p < 0.05 relative to WT, ** p < 0.01 relative to WT, *** p < 0.001 relative to WT. Error bars represent SEM. WT, white bars; Nlgn2 KO, red bars; IgSF9b KO, blue bars; double KO, purple bars Full size image To test whether the increase in VIAAT puncta in the CeM reflects an acute function of IgSF9b on inhibitory synapses or whether this is a developmental consequence, we quantified VIAAT puncta following injection of IgSF9b shRNA into the CeM (Fig. 6h–j ). Deletion of IgSF9b increased the number of VIAAT puncta in Nlgn2 KO mice (Fig. 6i , red and blue striped bars), an effect that may underlie the rescue of anxiety-related behavior in Nlgn2 KO mice following local deletion of IgSF9b in the CeM (Fig. 4 ). Moreover, acute deletion of IgSF9b increased the size of VIAAT puncta in WT mice (Fig. 6j , white and white and blue bars), consistent with the upregulation of VIAAT puncta in single IgSF9b KO mice (Fig. 6c , blue bar). Together, our data indicate that deletion of IgSF9b in the CeM results in an increase in inhibitory synapse function, which may underlie its behavioral consequences in the anxiety circuitry. IgSF9b deletion does not affect BA inhibitory synapses Finally, we investigated the consequences of IgSF9b deletion on synaptic inhibition in the BA, in order to determine whether the differential effects of IgSF9b deletion on anxiety-related neuronal activation in the BA and the CeM are also reflected at the synaptic level. We first recorded mIPSCs in excitatory (in our case Venus-negative) neurons in the BA (Fig. 7a–f ). Consistent with previous reports, Nlgn2 deletion resulted in a pronounced reduction in mean frequency and mean amplitude of mIPSCs in the BA (Fig. 7c, e , red bars) 8 , 20 , 34 , 35 . In striking contrast, IgSF9b deletion affected neither mean mIPSC frequency nor mean IPSC amplitude (Fig. 7c, e , blue bars). Double KO mice showed a trend towards a reduction in mean frequency and mean amplitude of IPSCs that was similar to that of Nlgn2 KO mice (Fig. 7c, e , purple bars). Comparison of the probability distribution of mean frequencies or mean amplitudes among all cells to the corresponding WT distribution (Fig. 7d, f ) revealed a significant shift towards lower mean frequencies in double KO cells (Fig. 7d , black vs. purple line), indicating that individual subpopulations of cells may be differentially affected in double KO mice. Consistent with the decrease in mIPSC frequency and amplitude, Nlgn2 deletion significantly reduced the number and size of perisomatic gephyrin and GABA A R α1 puncta, while no changes were observed for VIAAT or, interestingly, the GABA A R γ2 subunit (Fig. 7g–j , red bars). IgSF9b deletion did not affect any inhibitory synapse markers in the BA, while double KO mice showed reductions in gephyrin and GABA A R α1 staining that were identical to those observed in Nlgn2 KO mice (Fig. 7g–j , blue and purple bars, respectively), consistent with the notion that IgSF9b does not normalize anxiety through a synaptic mechanism in the BA. Thus, combined with the functional changes observed in whole-cell recordings in amygdala slices, these morphological data indicate that IgSF9b and Nlgn2 differentially affect inhibitory synapses in the amygdala, with Nlgn2 deletion primarily impairing synaptic inhibition in the BA, and IgSF9b deletion primarily enhancing synaptic inhibition in the CeM. Altogether, our findings are consistent with a model in which IgSF9b deletion normalizes anxiety-related behaviors and neuronal activity by increasing inhibition onto CeM output neurons and hence counteracting the anxiety-related overactivation of the CeM (Fig. 8 ). Fig. 7 IgSF9b deletion does not affect inhibitory synapses in the BA. a , b Schematic diagram illustrating recording sites in the BA, mean mIPSCs and representative mIPSC traces from the BA of WT, Nlgn2 KO, IgSF9b KO, and double KO mice. c Average cumulative distribution of mIPSC inter-event intervals and quantification of mean mIPSC frequency in the BA. d Probability distribution of mean mIPSC frequency of each analyzed cell in the BA. Kolmogorov–Smirnov test: WT vs. Double KO, p < 0.0001. e Average cumulative distribution of mIPSC amplitudes and quantification of mean mIPSC amplitude in the BA. f Probability distribution of mean mIPSC amplitude among all analyzed cells in the BA. n = 13–18 cells/5–6 mice per genotype. g – j Photomicrographs and quantification of the number of perisomatic puncta of g VIAAT, h gephyrin, i GABA A Rγ2, and j GABA A Rα1 in all four genotypes. Scale bar, 2 μm. n = 3–8 per genotype. Statistically significant ANOVA comparisons are marked in gray at the top of the panels and are listed in Table 1 . For all other ANOVA comparisons F < 1. Post-hoc analysis: * p < 0.05 relative to WT, ** p < 0.01 relative to WT, *** p < 0.001 relative to WT, # p < 0.05 relative to double KO. Error bars represent SEM. Box plots represent median and 25 th and 75 th percentiles and whiskers are drawn from the minimum to the maximum value. WT, white bars; Nlgn2 KO, red bars; IgSF9b KO, blue bars; double KO, purple bars Full size image Fig. 8 Model of the circuitry-based normalization of the anxiety phenotype in Nlgn2 x IgSF9b double KO mice. Based on our findings, we propose the following model for the normalization of anxiety-related behaviors in Nlgn2 x IgSF9b double KO mice: Nlgn2 deletion enhances activity of excitatory BA → CeM projection neurons, resulting in overactivation of anxiogenic projection neurons in the CeM and hence increased anxiety-related behaviors. IgSF9b deletion enhances synaptic inhibition onto anxiogenic projection neurons in the CeM, which results in reduced activation of CeM projection neurons and hence reduced anxiety-related behaviors. In Nlgn2 x IgSF9b double KO mice, the combination of these two effects results in normalization of the activity of CeM projection neurons and hence a normalization of anxiety-related behaviors Full size image Discussion In the present study we sought to elucidate the role of the cell adhesion molecule IgSF9b and its interactions with Nlgn2 in amygdala circuits, synapses, and behaviors related to anxiety processing. We show that deletion of IgSF9b has anxiolytic consequences and normalizes the prominent anxiety phenotype observed in Nlgn2 KO mice. This normalization does not occur through a mechanistic interaction of Nlgn2 and IgSF9b at the same synapses, but likely through differential effects on different inhibitory synapses in the BA and CeM, respectively. Specifically, our data support a model in which reduced inhibition in BA of Nlgn2 KO mice results in overactivation of BA → CeM projection neurons under anxiogenic conditions, which is counteracted in the CeM by the increased inhibition resulting from additional deletion of IgSF9b (Fig. 8 ). Together, our data provide the first description of IgSF9b function in vivo and uncover a novel role for IgSF9b in anxiety-related behavior and amygdala inhibitory synapses. Moreover, our findings highlight that IgSF9b-expressing synapses and neurons in the CeM may represent an important common target for anxiolytic treatments that is independent of individual upstream mutations. Major efforts have recently been invested in determining the full complement of proteins that governs the formation and function of inhibitory synapses 12 , 15 , 36 , 37 , 38 , 39 , 40 . The cell adhesion molecule IgSF9b was identified as one of these candidate molecules in cell cultures 14 , 15 , but the mechanism by which it regulates inhibitory synapse function in vivo in intact circuits remained completely unknown. Here we show that in the CeM, deletion of IgSF9b results in an enhancement of inhibitory synaptic transmission, likely through a presynaptic mechanism. In particular, the increases in mIPSC frequency and VIAAT staining without a concomitant increase in gephyrin indicate that IgSF9b deletion may result in an increase in the number of VIAAT-positive vesicles per synaptic terminal, analogous to effects recently observed in IgSF21 KO mice (albeit with opposite polarity) 41 . Together with the observation that IgSF9b, like its mammalian paralog IgSF9 and its Drosophila ortholog Turtle, forms primarily or exclusively homophilic interactions 14 , 16 , 42 , our data indicate that IgSF9b provides a transsynaptic signaling complex that regulates the organization of the inhibitory presynaptic terminal. Moreover, the differential effects of IgSF9b deletion on GABA A R α1 and γ2 subunits in the CeM indicate that IgSF9b may also regulate the subunit composition of GABA A Rs at inhibitory postsynapses, although the mechanism or functional significance of this effect remains unknown. Intriguingly, IgSF9b appears to play distinct roles at different inhibitory synapses, as evidenced by our finding that IgSF9b deletion affects inhibitory synaptic transmission and synaptic markers only in the CeM but not the BA, despite being expressed in both (Supplementary Fig. 6 ). One possible explanation for this difference lies in the differential composition of the two structures with respect to neuron types 9 , 43 . The BA is a cortex-like structure composed primarily of excitatory projection neurons with a small number of inhibitory interneurons, while the CeA is similar to striatum in that it is composed almost exclusively of inhibitory neurons. Therefore, the vast majority of inhibitory synapses sampled in the BA and CeM are made onto excitatory and inhibitory neurons, respectively, raising the possibility that IgSF9b differentially affects these synapse subtypes. Consistent with this notion, IgSF9b was previously proposed to function primarily at inhibitory synapses onto interneurons in hippocampal cultures 14 . In contrast, there is growing evidence that other known synapse organizers such as Nlgn2, MDGA1, and GARLH/LHFPL4 function exclusively at inhibitory synapses onto excitatory neurons 8 , 44 , 45 . Together, these findings give rise to the notion that excitatory and inhibitory neurons may utilize an entirely different complement of organizer proteins at inhibitory postsynapses, and that IgSF9b may be the first example of a synapse organizer with specificity for inhibitory synapses onto inhibitory neurons. Moreover, the function of IgSF9b at inhibitory synapses may also depend on the identity of the presynaptic neuron. In hippocampal cultures, shRNA-mediated deletion of IgSF9b resulted in a decrease in mIPSC frequency and a reduction in VIAAT-positive gephyrin clusters 14 , in contrast to the increase in mIPSC frequency and VIAAT puncta observed here in the CeM (Figs. 5 , 6 ). Since the inhibitory neuron subtypes found in the CeM are entirely distinct from those observed in the hippocampus 10 , 11 , it is conceivable that these different subtypes may differentially employ IgSF9b at inhibitory postsynapses. Alternatively, IgSF9b may play distinct roles at different stages of synaptic development (i.e., at developing synapses in neuronal cultures vs. mature synapses in adult mice) or in dissociated cultures vs. intact neuronal networks. Cell type- and circuit-specific analysis of IgSF9b function will be essential for fully understanding its role in the brain. Given that IgSF9b and Nlgn2 function at distinct synapses and do not appear to interact in a cell-autonomous manner, at least in the amygdala, how does deletion of IgSF9b normalize the prominent anxiety phenotype observed in Nlgn2 KO mice? Combined evidence from our cFos analysis (Fig. 2a ), retrograde tracing experiments (Fig., 2c–e ), and in vivo electrophysiology (Fig. 3 ) indicates that IgSF9b deletion normalizes anxiety-related output specifically in the CeM of Nlgn2 x IgSF9b double KO mice. Given that the normalization of anxiety-related behaviors can be mimicked by local shRNA-mediated knockdown of IgSF9b in the CeM of Nlgn2 KO mice (Fig. 4 ), it is likely to occur through a local mechanism within the CeM. In light of our observation that IgSF9b deletion increases inhibitory synaptic transmission in the CeM (Figs. 5 , 6 ), the most parsimonious explanation is that this increased inhibition onto CeM output neurons counteracts the increased excitation originating from Nlgn2 KO BA projection neurons, thus balancing the activity of CeM neurons mediating anxiogenic projections to the brainstem 9 , 10 (Fig. 8 ). This model may also explain why the anxiolytic effect of IgSF9b in WT mice is relatively modest: in the absence of a pathological overactivation of CeM projection neurons such as that induced by Nlgn2 deletion, the increased inhibition may have relatively minor effects on WT CeM output. An interesting remaining question regards the source of the inhibitory inputs to the CeM that are upregulated by IgSF9b deletion. CeM neurons are known to receive feedforward inhibitory projections from the CeL, the intercalated nucleus and the bed nucleus of the stria terminalis 9 , 10 , and synaptic transmission at any of these inhibitory connections or at local interneurons may be upregulated in response to IgSF9b deletion. Importantly, our model does not require the neuronal activity of the upstream inhibitory neurons to be increased, and indeed we show that CeL inputs to the CeM are not differentially activated (Fig. 2e ). We cannot rule out that other CeM inputs may show altered activation in IgSF9b or double KO mice. However, the effectiveness of the local deletion of IgSF9b in CeM in mimicking the double KO phenotype indicates that any such upstream alterations cannot be required for the normalization of anxiety-related behaviors by IgSF9b deletion. Our findings, therefore, identify local information processing in the CeM as a key mediator of anxiety-related behaviors and highlight the fundamental importance of better understanding the role of the CeM in anxiety processing. The strong correlation of power of beta oscillations with distance from the center of the OF in Nlgn2 KO mice and its exacerbated increase during risk-assessment behavior implicates beta oscillations in the CeM as a key neural signature of pathological anxiety. To our knowledge this is the first time that oscillatory activity in the CeM is described in anxiety processing in general, and beta oscillations during risk-assessment behavior in particular. At present, both the function of these beta oscillations in anxiety processing and the mechanisms that underlie their modulation by Nlgn2 and IgSF9b remain unknown. Beta oscillations have previously been observed in the basal ganglia and somatosensory cortex during decision-making tasks 46 , 47 , in the hippocampus upon exposure to novelty, and in the mediodorsal thalamus during working memory-related tasks 48 , 49 , highlighting their multifaceted role in information processing. However, their precise function in mediating these behavioral processes remains controversial and an area of active investigation. Similarly, the cellular and synaptic mechanisms leading to the generation of beta rhythms are largely unknown, although changes in inhibitory synaptic activity are known to affect LFPs in general 50 and beta activity in specific 51 . Further experiments will be essential to fully understand how Nlgn2 and IgSF9b regulate beta oscillations in the CeM and how this contributes to the generation and normalization of pathological anxiety processing. The anxiolytic effects of IgSF9b deletion in Nlgn2 KO mice are intriguing in light of previous studies showing that variants in IgSF9b are associated with major depression and the negative (mainly affective) symptoms of schizophrenia 17 , 18 , 19 . It is conceivable that the IgSF9b variants in human patients are in fact gain-of-function rather than loss-of-function mutations, which based on our findings, could result in reduced synaptic inhibition in the CeM. Indeed, our data indicate that the function of IgSF9b-containing synapses in the CeM may be key in determining the vulnerability or resilience of individuals towards upstream anxiogenic factors. The precise molecular mechanisms remain to be elucidated both in mice and humans, but it is clear that IgSF9b is emerging as an important regulator of affective behaviors, and that further investigations into its function may substantially contribute to our understanding of a range of psychiatric disorders. Ultimately, perhaps the most exciting conclusion arising from our study is that IgSF9b-expressing neurons and synapses in CeM may represent a viable common target for anxiolytic therapies, independent of the upstream anxiogenic mutations. Given that both global and local deletion of IgSF9b lead to a remarkably specific anxiolytic effect, targeting IgSF9b pharmacologically may provide a promising strategy for the development of more selective anxiolytic therapies. Moreover, it is tempting to speculate that in the age of circuit psychiatry 26 , targeting IgSF9b-expressing neurons with viral vectors will become feasible, offering entirely new treatment options for patients with anxiety and co-morbid psychiatric disorders. Methods Experimental subjects Nlgn2 KO mice 21 were generated in our laboratory on an 129/Sv background and were backcrossed onto a C57BL/6 J background for at least six generations. IgSF9b KO mice were obtained from Lexicon Pharmaceuticals (The Woodlands, TX, U.S.A.; Omnibank clone 281214, generated through insertion of the Omnibank gene trap vector 48 into the IgSF9b gene in Sv129 ES cells) and were backcrossed onto a C57BL/6 J background for at least six generations. Nlgn2 KO mice and IgSF9b KO mice were crossed to generate Nlgn2 het x IgSF9b het breeding pairs, from which four experimental genotypes were obtained as littermates, i.e., WT, Nlgn2 KO, IgSF9b KO, and double KO mice. For in vitro electrophysiology experiments, these mice were additionally crossed with mice expressing a Venus transgene under the control of the VIAAT promoter to label inhibitory interneurons (generated using a Venus construct generously provided by Dr. Atsushi Miyawaki, RIKEN) 32 . All mice were 2–3 months old at the beginning of the experiment. For experiments involving OF testing (Figs. 1 – 4 and Supplementary Fig. 1 – 4 ), only male mice were used unless specified otherwise. For the EPM experiment (Fig. 1h–m ) and for molecular and slice electrophysiology experiments (Figs. 5 – 7 ), both male and female mice were used, based on the observation that the OF phenotype is identical in both sexes. Animals were maintained on a 12 h light/dark cycle (7 am/7 pm), with food and water ad libitum. All experiments were performed during the light cycle (with the exception of home cage activity monitoring as described below, Supplementary Fig. 2 ). The experimenter was blind to genotype during all stages of data acquisition and analysis. All procedures were approved by the State of Niedersachsen (Landesamt für Verbraucherschutz und Lebensmittelsicherheit) and were carried out in agreement with the guidelines for the welfare of experimental animals issued by the Federal Government of Germany and the Max Planck Society. Behavioral characterization The OF test was conducted in a square arena made of white plastic (50 cm × 50 cm). Mice were placed in the corner of the OF and were permitted to explore the arena for 10 min. The EPM was conducted in an apparatus made of gray plastic, with two open arms and two closed arms (28 cm × 5 cm each, elevated 45 cm from the ground). Mice were placed in the closed arm and were permitted to explore the apparatus for 5 min. Performance was recorded using an overhead camera system and scored automatically using the Viewer software (Biobserve, St. Augustin, Germany). Between each mouse, the arena was cleaned thoroughly with 70% ethanol followed by water to eliminate any odors left by the previous mouse. In our previous study 8 , the center zone of the OF was defined as a square of 25 × 25 cm. However, closer assessment revealed that a second, intermediate zone (12.5 cm–25 cm from the walls of the chamber) was also avoided by the Nlgn2 KO mice (Supplementary Fig. 1a-h ), and that in fact the behavior in this intermediate zone resembled the behavior in the center, rather than the periphery, in all four genotypes assessed in the present study. For this reason, an extended center zone of 37.5 × 37.5 cm was used throughout this study to more accurately reflect the anxiogenic area. Recording of home cage activity was performed using the LABORAS system and software (Metris, Hoofddorp, The Netherlands). Mice were habituated to the LABORAS cages for two days. On the third day, activity was recorded for two 6-hour periods, from 9 pm to 3 am (dark cycle) and from 9 am to 3 pm (light cycle). The following parameters were assessed: total duration of locomotor activity, immobility, grooming, and climbing, as well total distance traveled and average velocity during locomotion. cFos induction assay To assess anxiety-induced cFos activation, mice were first exposed to the OF arena for 10 min. Ninety minutes after exposure, they were anesthetized with Avertin (Tribromoethanol, Sigma) and perfused transcardially first with saline, then with 4% paraformaldehyde (PFA) in 0.1 M phosphate buffer (PB). Brains were post-fixed in PFA overnight, and cryoprotected in 30% sucrose in 0.1 M PB. Free-floating sections (40 μm thickness) were prepared using a Leica CM3050S cryostat (Leica, Wetzlar, Germany). Sections were incubated in blocking solution (3% bovine serum albumin, 10% goat serum, 3% Triton X-100 in 0.1 M phosphate-buffered saline (PBS)) for 1 h, then incubated for 12 h with Rabbit polyclonal anti-cFos antibody (catalog# sc-52, Santa Cruz Biotech, Dallas, Texas, USA) diluted 1:2000 in blocking solution; and then incubated for 2 h with Alexa Fluor 488 goat anti-rabbit antibody (Invitrogen, Eugene, OR, USA) diluted 1:600 in blocking solution. The sections were washed with (PBS) after each incubation, and were finally mounted on glass slides using Aqua-Poly/Mount (Polysciences, Eppelheim, Germany). For PV/cFos or SOM/cFos immunolabeling, amygdala sections were processed as described above for cFos immunohistochemistry. All secondary antibodies were diluted 1:600 in respective blocking solution and obtained from Invitrogen, Eugene, USA. For PV immunolabeling, sections were incubated for 12 h with mouse monoclonal anti-PV antibody (catalog# 235, SWANT, Bellinzona, Switzerland) diluted 1:2000 in blocking solution; and then incubated for two hours with Alexa Fluor 555 goat anti-mouse antibody. For SOM immunolabeling, the blocking solution contained 10% donkey serum and 0.3% Triton X-100 in PBS 0.1 M. Sections were incubated for 12 h in goat polyclonal anti-SOM antibody (catalog# sc-7819, Santa Cruz Biotech, Dallas, Texas, USA) diluted 1:1000 in blocking solution; and then with Alexa Fluor 488 donkey anti-goat antibody. Sections were washed with PBS after each incubation, and were finally mounted on glass slides using Aqua-Poly/Mount (Polysciences, Eppelheim, Germany). Images of cFos and cellular markers were obtained using a confocal laser scanning microscope (Leica SP2) with a ×40 oil immersion objective. For each set, sections were anatomically matched and the settings for laser power, gain and offset were kept constant during imaging. 2 stacks of 5 μm thickness and containing 2 optical sections each were obtained from each amygdala section. In total, 10–12 stacks were imaged from 5–6 amygdala sections per mouse in each group. cFos images were thresholded manually in ImageJ with threshold set as 3*background intensity and the same threshold value was applied for WT, Nlgn2 KO, IgSF9b KO, and double KO in each group. Single-labeled cells were quantified using the Create Spots tool in Imaris (Bitplane, Zurich, Switzerland). To obtain the number of double- labeled cells, the Colocalize Spots tool in Imaris was used. Retrograde labeling Mice received an intraperitoneal (i.p.) injection of Carprofen (5 mg/kg) to reduce post-surgery pain 30 min prior to surgery. Mice were anaesthetized with Avertin by i.p. injection (20 ml/kg body weight) and placed in digital stereotaxic frame. To bilaterally label BA neurons projecting to the CeM, 50 nl of red Retrobeads (excitation maximum at 530 nm and emission maximum at 590 nm) were injected into the CeM (0.70 mm posterior, ± 2.35 mm medial, and 5.08 mm ventral from bregma). A Hamilton syringe (1 μl) was used to manually deliver the Retrobeads at the rate of 0.5 nl/sec. After the injection was completed, the tip of the syringe was raised by 100 μm and left for 3 min to allow diffusion of the Retrobeads at the injection site; and then slowly withdrawn at the rate of 1 mm/min. Following surgery, mice received Metamizol with drinking water (200 mg/kg/day, drinking rate estimated at 3 ml/day) for 3 days to reduce pain and risk of inflammation. Mice were single housed for 7 days post-surgery to allow their full recovery and traveling of beads from the injection site up the axons of BA neurons to their somata. To induce anxiety-associated neuronal activation, mice were subjected to the OF for 10 min, and cFos was subsequently assessed as described above. Only slices from brains in which injection sites did not exceed the borders of CeM (confirmed by visualizing the sites of injection on 5 subsequent coronal sections of amygdala spanning 400 µm of tissue) were included in subsequent imaging and analysis. Data acquisition and quantification of cFos-positive cell bodies containing Retrobeads were performed similarly to quantification of double labeled neurons as described above. Sections were also used for histological verification of the Retrobead injection site, and mice were excluded from analysis if the injection site lay outside the CeM (Supplementary Fig. 3j ). In vivo electrophysiology Mice were anesthetized with Avertin (loading dose 20 ml/kg, maintenance dose 2 ml/kg i.p.) and placed in a stereotaxic frame, and their body temperature was monitored by a rectal probe and maintained at 36 °C. An incision in the midline of the scalp was made to expose the skull. Bregma and lambda were aligned to a plane level ± 50 μm. A multi-wire electrode array was unilaterally implanted targeting the left CeM (0.9 mm posterior, 2.3 mm lateral and 5.04 mm ventral to bregma). The electrodes consisted of 2 bundles (spaced 750–950 μm) of 8 individual insulated tungsten wires (13 µm inner diameter, impedance 60–100 kΩ) inserted into a polymide tube (127 µm inner diameter) and attached to an 18-pin connector. A reference screw was implanted above the cerebellum. The implant was secured with two screws implanted in the skull ~300 µm lateral and anterior to the electrode and bonded with dental cement. Immediately after the surgery, mice subcutaneously received an analgesic (Carprofen 5 mg/kg) and an antibiotic (Baytril 5 mg/kg). Twenty four hours after the surgery, mice received Carprofen (5 mg/kg) subcutaneously and Baytril in the drinking water (0.2 mg/ml). Mice were sacrificed following the OF recording for histological verification of the recording site (Supplementary Fig. 4 ). Mice in which the electrode was not implanted in the CeM were excluded from the analysis. Male mice were single housed for 7 days after the surgery and before the recording. For data acquisition, the mice were connected to the electrophysiological equipment and placed in the OF chamber, where they were allowed to explore for 15 min. The electrophysiological signal was amplified and sent to the acquisition board. The raw signal was acquired at 32 kHz sampling rate, band pass filtered (0.1–9000 Hz), and stored for offline analysis. During the experiment, simultaneous electrophysiological and video recordings were made by the Cheetah Data Acquisition System. LFPs were analyzed using custom-written MATLAB scripts. The signal was filtered between 0.7 and 400 Hz using a zero-phase distortion FIR filter and down sampled to 1 kHz. The multitaper method was used for the power analysis (Chronux Package) 52 . The following time windows were used: theta range (4–12 Hz), 1 s with 0.8 s of overlap; beta range (18–30 Hz), 1 s with 0.5 s of overlap; gamma range, 0.15 s with 0.1 s of overlap. To calculate the power spectra during the entire OF session, 5 tapers were used with a time bandwidth of 3. For each mouse, the time periods and the tracks of the movement in the periphery and center were extracted using a modified version of the autotyping toolbox 53 . To evaluate the relative increase in beta power as a function of distance from the center, the distance from the center during the entire OF session was binned for each mouse (3 cm bins). To compute the beta power, the power was first summed across the beta band, and then the summed values were averaged for each location bin. The individual bin values were normalized by the average power in the periphery ( > 18 cm from center) and the linear correlation coefficients ( fitlm function, MATLAB) per genotype were computed. To calculate the correlation between speed and beta power, the beta power was averaged in speed bins of 5 cm/s for each mice and the linear correlation coefficients computed as above. To evaluate power changes during the SAP, the time events at which the mouse showed a clear SAP from the periphery towards the center of the OF were manually extracted. The events were identified by a typical elongation of the body and a very slow forward movement that was always followed by a retreat movement 30 , 31 . For the beta band, the power across frequencies was summed to produce one value for each time point. Then, the power values at each time point during the SAP were averaged and the mean power values per event were averaged per group. The Morlet wavelet transform was used to visualize the power at different frequencies ranges as shown in Fig. 3k , with 40 wavelets at centered frequencies ranging from 1 to 120 Hz and a length of 10 cycles. Analysis of synaptic markers Immunolabeling for VIAAT was performed on perfusion-fixed brains as described for the cFos assay. Briefly, sections were incubated for 24 h with rabbit polyclonal anti-VIAAT antibody (catalog# 131002, Synaptic Systems, Goettingen, Germany) diluted 1:1000 in blocking solution (3% bovine serum albumin, 10% goat serum, 3% Triton X-100 in 0.1 M PBS), washed, and then incubated for 2 h with Alexa Fluor 488 goat anti-rabbit antibody (Invitrogen, Eugene, OR, USA) diluted 1:600 in blocking buffer. Immunolabeling for IgSF9b, Nlgn2, gephyrin, S-SCAM, GABA A Rα1, and GABA A Rγ2 was performed on methanol-fixed fresh frozen brain sections using a modified version of a published protocol 54 . Brains were frozen immediately after dissection in an isopentane bath at −35 °C to −40 °C. Coronal sections were prepared using a Leica CM3050S cryostat (Leica, Wetzlar, Germany), mounted on glass slides, and dried at room temperature. Sections were then fixed in methanol at −20 °C for 5 min, and blocked for 1 h. The duration of incubation was 12 h with primary antibody and 2 h with secondary antibody for each immunolabeling. The following primary antibodies were used, diluted in blocking solution: Rabbit polyclonal anti-IgSF9b (catalog# HPA010802, Sigma Aldrich, Darmstadt, Germany) at 1:1000 mouse monoclonal anti-Nlgn2 (catalog# 129511, Synaptic Systems, Goettingen, Germany) at 1:1000; mouse monoclonal anti-gephyrin (catalog# 147111BT, Synaptic Systems, Goettingen, Germany) at 1:1000; rabbit polyclonal anti-MAGI2 ( = S-SCAM, catalog# M2441, Sigma Aldrich, Darmstadt, Germany); rabbit polyclonal anti-GABA A Rα1 (catalog# 224203, Synaptic Systems, Goettingen, Germany) at 1:1000; guinea pig polyclonal GABA A Rγ2 (generously provided by Dr. Jean-Marc Fritschy, University of Zürich) at 1:1000. The following secondary antibodies were obtained from Invitrogen, Eugene, USA, and diluted 1:600 in the blocking solution: Alexa Fluor 555 goat anti-rabbit antibody, Alexa Fluor 488 goat anti-rabbit antibody, Alexa Fluor 555 goat anti-mouse antibody. Sections were washed with PBS after each incubation. The slides were then dried overnight at 4 °C, and covered with mounting media (Aqua-Poly/Mount; Polysciences, Eppelheim, Germany) and glass coverslips. Images of synaptic markers were obtained using a confocal laser scanning microscope (Leica SP2) with a ×63 oil immersion objective and ×8 digital zoom. For each set, sections were anatomically matched and the settings for laser power, gain and offset were kept constant during imaging. 12 stacks of 2 μm thickness and containing 2 optical sections each were obtained from each amygdala section (12 stacks from 4 sections for each mouse in total). Images were thresholded in ImageJ, with same threshold applied to all mice in each set. To quantify perisomatic synapses, the perisomatic area was identified by manually tracing the perimeter of the cell body (defined as a circular area largely devoid of immunofluorescence) 6 , 8 . The perimeter was then expanded by 1.4 μm or 2 μm in each direction for quantification of postsynaptic puncta or presynaptic puncta, respectively. The number of particles was quantified in this area using the “count particles” module in ImageJ, and the number of particles per area was divided by the length of the cell body perimeter to obtain the final result. In vitro electrophysiology Adult (8–12-week-old) WT, Nlgn2 KO, IgSF9b KO, and double KO mice additionally expressing a VIAAT-Venus transgene 32 were anesthetized with Avertin and perfused transcardially for 90 s with an ice-cold sucrose-based solution (6 mM MgCl 2, 0.1 mM CaCl 2, 50 mM sucrose, 2.5 mM glucose, and 3 mM kynurenic acid diluted in artificial cerebrospinal fluid (aCSF, 124 mM NaCl, 2.7 mM KCl, 26 mM NaHCO 3 , and 1.25 mM NaH 2 PO 4 )) as described previously for the preparation of amygdala slices from adult mice 55 . The brains were rapidly dissected and placed in the same ice-cold sucrose-based solution. The brainstem was removed and the brains were mounted on a holder and transferred to the vibratome chamber for preparation of 300 μm coronal sections. Slices containing the BA and central amygdala were transferred to a chamber filled with aCSF (see above) with additional 2 mM CaCl 2 and 1.3 mM MgCl 2 and equilibrated with 95% O 2 /5% CO 2 . Slices were allowed to recover for 20 min at 33 °C and maintained at room temperature thereafter. All chemicals were obtained from Merck Millipore (Molsheim, France), Sigma Aldrich (Darmstadt, Germany), or Tocris Bioscience (Bristol, UK). Whole-cell patch-clamp recordings were obtained at room temperature (~22 °C) with an EPC10 amplifier (HEKA Elektronik, Germany). Slices were kept in a recording chamber and perfused with aCSF with additional 1.3 mM MgCl 2 , 2 mM CaCl 2, 18.6 mM glucose, and 2.25 mM ascorbic acid (osmolarity ≈320 mOsm) at a rate of 1–2 ml/min. Neurons were visually identified with infrared video microscopy using an upright microscope equipped with a ×60 objective. VIAAT-positive neurons were identified by Venus expression. For recordings in BA, VIAAT-Venus-negative neurons were targeted, while for recordings in CeM, VIAAT-Venus-positive neurons were targeted. Patch electrodes (3–5 MΩ open tip resistance when filled with internal solution) were pulled from borosilicate glass tubes. For voltage-clamp experiments to record miniature inhibitory postsynaptic currents (mIPSCs), patch electrodes were filled with Cs-based internal solution containing (in mM) 110 CsCl, 30 K-gluconate, 1.1 EGTA, 10 HEPES, 0.1 CaCl 2 , 4 Mg-ATP, 0.3 Na-GTP, and 4 N-(2,6-Dimethylphenylcarbamoylmethyl) triethylammonium bromide (QX-314; Tocris-Cookson, Ellisville, MO); pH = 7.3 (adjusted with CsOH, 280 mOsm). To block glutamatergic EPSCs, 2 μM NBQX (6-cyano-7-nitroquinoxaline-2,3-dione) and 2 μM CPP (( RS )−3-(2-Carboxypiperazin-4-yl)-propyl-1-phosphonic acid),) were added to the bath. Action potential (AP) firing was suppressed by adding 1 µM tetrodotoxin (TTX) to the aCSF. For current-clamp experiments and for recordings of spontaneous postsynaptic excitatory currents (sEPSCs), patch electrodes were filled with K-gluconate-based internal solution containing (in mM) 125 K-gluconate, 20 KCl, 0.2 EGTA, 2 MgCl 2 , 10 HEPES, and 2 Na 2 ATP; pH = 7.3 (adjusted with KCl, 280 mOsm). To block GABAergic EPSCs, 25 μM bicuculline methiodide (Tocris-Cookson, Ellisville, MO) was added to the bath. Firing thresholds were estimated from AP phase-plane plots and corrected for a measured liquid junction potential of 7.9 mV. To monitor series resistance on-line and to allow offline estimation of whole-cell membrane resistance and membrane capacitance, a voltage step (10 mV amplitude, 20 msec duration) was delivered at the beginning of each sweep during whole-cell voltage-clamp experiments. Capacitive current transients were analyzed using a simplified two-compartment equivalent circuit model 56 . Mean current transients obtained by averaging ≥ 30 consecutive sweeps were fitted using a double-exponential function \(I(t) = A_1 \times e^{ - t/\tau _1} + A_2 \times e^{ - t/\tau _2} + A_\infty\) , where I ( t ) is the amplitude of the current at time t , A l and τ 1 denote the amplitude and time constant of the fast component of decay, A 2 and τ 2 represent the amplitude and time constant of the slower component of decay, and A ∞ , is the difference between the holding current and the final steady-state current at the end of the depolarizing pulse. The holding potential for whole-cell voltage-clamp recordings was set to −70 mV. Whole-cell voltage-clamp recordings were included in the analysis if the access resistance was initially ≤ 13 MΩ and did not change by more than 20% during the recording. Recordings with a leak current > 200 pA were rejected. Data were acquired with Patchmaster software (HEKA Elektronik, Germany), low-pass filtered at cut-off frequency of 5 kHz using a Bessel filter and digitized at 20 kHz. All offline analysis was performed with IgorPro (Wavemetrics, USA). mIPSCs were detected using a sliding template-matching algorithm implemented in IgorPro 57 . Generation and stereotaxic injection of AAV To generate shRNA-expressing AAV particles, IgSF9b and control shRNA sequences were first cloned into an AAV-shRNA-GFP vector 58 generously provided by Dr. Ralph DiLeone (Yale University). The following shRNA sequences were used (modified from ref. 14 ): IgSF9b, TCATCAAGTTTGGCTACTAT; control (point mutant lacking knockdown activity), TCAT A AAGTT C GGCTACTAT. To confirm efficacy of knockdown, the resulting plasmids were co-transfected into HEK cells with a Myc-IgSF9b construct 14 generously provided by Dr. Eunjoon Kim (Korea Advanced Institute of Science and Technology), and IgSF9b levels were quantified using standard immunoblotting procedures 6 (primary antibody: rabbit anti-IgSF9b, Sigma, diluted 1:1000). AAV particles were generated using the pDPrs1/pDPrs2 packaging system (PlasmidFactory, Bielefeld, Germany) and the AAV-shRNA-GFP vectors described above. Plasmids were transfected into HEK cells using calcium phosphate transfection, and virus particles were harvested 72 h later 58 , 59 . To this end, HEK cells were lysed for 30 min at 37 °C in 20 mM Tris, pH 8.0, containing 150 mM NaCl, 0.5% sodium deoxycholate and benzonase, followed by incubation in 1 M NaCl at 56 °C for 30 min. Lysates were stored at −80 °C overnight, thawed, and purified on a 15%/25%/40%/54% iodixanol gradient by ultracentrifugation (90 min at 370,000× g ). The 40% fraction was collected, diluted in PBS containing 1 mM MgCl 2 and 2.5 mM KCl, and concentrated using an Amicon 100 K MWCO filter. Surgical procedures were exactly as described for the injection of Retrobeads, except that 1 μl of virus was injected bilaterally into CeA using a Nanoject II Microinjector (Drummond, Broomall, PA, USA) and a Micro pump controller (WPI). Mice were alternately assigned to receive AAV-control shRNA or AAV-IgSF9b-shRNA injections based on order of birth. The following coordinates relative to Bregma were used: AP (anteroposterior) −0.58, ML (mediolateral) ±2.48, DV (dorsoventral) and −5.4. After surgery, mice were housed in pairs and were allowed to recover for 6 weeks before assessment of behavior in the OF as described above. Mice were sacrificed following OF exposure for verification of the injection site as defined by GFP expression (Supplementary Fig. 5 ). Only mice in which both bilateral injection sites were correctly positioned in CeM were included in the study. Moreover, mice with any GFP expression in BLA or CeL were excluded, although minor, low-expression leakage into other border areas was tolerated. In total, 11 mice were excluded due to mistargeting (WT + Ctrl shRNA, 4 animals; WT + IgSF9b shRNA, 1 animal; Nlgn2 KO + Ctrl shRNA, 3 animals; and Nlgn2 KO + IgSF9b shRNA, 3 animals). Statistical Analysis Sample sizes were estimated based on prior experience with the methods used in this study 6 , 8 , 60 . All data were analyzed statistically using Prism (GraphPad Software, La Jolla, CA, USA) or Matlab. Outliers were identified using the Grubb’s test and were removed prior to statistical analysis. Behavioral scores were subjected to two-way ANOVA with post-hoc Tukey’s tests for comparison between groups. Data obtained from histological experiments were analyzed using two-way ANOVA with post-hoc paired, two tailed Student’s t -tests for comparison between groups. Data obtained from in vitro electrophysiological experiments were analyzed using two-way ANOVA with post-hoc Tukey’s test for comparison between groups. Distributions of mean mIPSC frequencies and amplitudes were analyzed using the Kolmogorov–Smirnov test. Data obtained from in vivo electrophysiological experiments were analyzed using two-way ANOVA with post-hoc Tukey’s test for comparisons between groups. Code availability Custom MATLAB scripts written for the analysis of LFPs are available from the corresponding author upon request. Data availability All data produced from this study are available from the corresponding author upon request.

Anxiety disorders are severe mental disorders in which patients suffer from intense fears and anxiety or from sudden, inexplicable panic attacks. In extreme cases, the affected individuals barely leave their homes, which can have serious consequences for their relationships with family and friends as well as for their professional lives. Scientists at the Max Planck Institute for Experimental Medicine in Göttingen have now identified a synaptic protein which, when inactivated, has an anxiolytic effect in mice. Around 10 percent of the population suffer from anxiety disorders, and current treatment options only offer effective help for a proportion of those affected. One of the changes observed in the brains of patients with anxiety disorders is an increased neuronal activity in the amygdala, a brain region that plays a key role in processing emotions such as anxiety or fear. An overactivation of the amygdala is thought to be involved in causing exaggerated anxiety. Many anxiolytic medications such as benzodiazepines presumably normalize this overactivation by strengthening the function of inhibitory synapses. Synapses are connections between nerve cells in the brain, at which information is transmitted from one nerve cell to another. At inhibitory synapses, this transmission results in a reduction in the activity of the neighbouring nerve cells. In the amygdala, for instance, this inhibits the transmission of stimuli that trigger fear and anxiety. Benzodiazepines strengthen this inhibitory effect—but unfortunately they affect not only those inhibitory synapses that transmit anxiogenic stimuli but also many other inhibitory synapses in the brain. This can lead to significant side effects such as pronounced sedation and impaired concentration. Accordingly, scientists are searching for new, more specific targets for anxiolytic medications. Mice with anxiety disorder Animal research with mice played a key role in helping the researchers in Göttingen to study anxiety disorders. Whereas healthy animals curiously investigate an empty test chamber, rodents with a pathological anxiety phenotype withdraw into a corner because they are afraid. However, when the scientists blocked the production of the recently discovered protein IgSF9b in these mice, the animals moved freely around the chamber again. IgSF9b produces a protein bridge at inhibitory synapses between two neighbouring nerve cells. "Blocking IgSF9b in pathologically anxious mice has an anxiolytic effect and normalises anxiety behaviour in these animals. This protein could therefore be a target for pharmacological approaches to treating anxiety disorders," explains Olga Babaev from the Max Planck Institute for Experimental Medicine who carried out the experiments as part of her doctoral work.. An investigation of the amygdala in these animals revealed that the overactivation of the amygdala was normalized, and that this effect resulted from a strengthening of the synaptic transmission at inhibitory synapses in the amygdala. "Our research shows that protein structures at inhibitory synapses in the centromedial amygdala, and particularly the protein IgSF9b, constitute promising new targets for potential treatments. It thus provides an important contribution toward understanding the biological causes of anxiety disorders and for the development of new anxiolytic medications," study leader Dilja Krueger-Burg says.

10.1038/s41467-018-07762-1

Physics

Scientists identify liquid-like atoms in densely packed solid glasses

C. Chang et al, Liquid-like atoms in dense-packed solid glasses, Nature Materials (2022). DOI: 10.1038/s41563-022-01327-w Journal information: Nature Materials

https://dx.doi.org/10.1038/s41563-022-01327-w

https://phys.org/news/2022-08-scientists-liquid-like-atoms-densely-solid.html

Abstract Revealing the microscopic structural and dynamic pictures of glasses is a long-standing challenge for scientists 1 , 2 . Extensive studies on the structure and relaxation dynamics of glasses have constructed the current classical picture 3 , 4 , 5 : glasses consist of some ‘soft zones’ of loosely bound atoms embedded in a tightly bound atomic matrix. Recent experiments have found an additional fast process in the relaxation spectra 6 , 7 , 8 , 9 , but the underlying physics of this process remains unclear. Here, combining extensive dynamic experiments and computer simulations, we reveal that this fast relaxation is associated with string-like diffusion of liquid-like atoms, which are inherited from the high-temperature liquids. Even at room temperature, some atoms in dense-packed metallic glasses can diffuse just as easily as they would in liquid states, with an experimentally determined viscosity as low as 10 7 Pa·s. This finding extends our current microscopic picture of glass solids and might help establish the dynamics–property relationship of glasses 4 . Main Matter can generally be classified into solid, liquid and gas states. Under extreme conditions or in specific systems, special states of matter exist that simultaneously exhibit some properties of both liquid and solid. In this case, solid matter contains rapidly diffusing, liquid-like atoms that can move fast even at low temperatures. Although it seems incredible, the existence of rapidly diffusing, liquid-like atoms in solids has been reported many times. For example, it is presumed that solid 4 He is able enter a supersolid state in which 4 He atoms can move fast and freely without any resistance 10 . In addition, a superionic state in which small/light atoms can diffuse freely while large/heavy atoms are fixed in their sublattices 11 , 12 , 13 , 14 has also been observed in materials, including ices 11 , helium–water compounds 12 , copper-based thermoelectric materials 13 and lithium-based conductors 14 , attracting growing attention in both materials science and engineering. Because glassy solids are generally recognized to be frozen liquids 2 , 4 , 5 , it is reasonable to imagine that there are more possibilities to find rapidly diffusing liquid-like atoms in glasses. In fact, a previous study of amorphous SiO 2 has even imaged liquid-like fast atoms at the edge of a two-dimensional sheet 15 . However, this fast motion was not ascribed to the amorphous nature but to the chemical-bond hierarchy 15 . Liquid-like atoms have rarely been observed in simple glasses, such as dense-packed metallic glasses (MGs). Recently, an unusual fast-relaxation process was observed in the dynamics of MGs at low temperatures where α and β relaxations are too slow to be tracked 6 , 7 , 8 , 9 . This process was regarded as the precursor of β relaxation with more localized atomic motion, probably related to certain chemical bonds, and as the initiation of plasticity in MGs. However, the underlying mechanism of this fast relaxation has yet to be elucidated. It is unclear whether such a fast relaxation in glasses is related to the liquid-like atoms. Answers to these questions may change our basic understanding of glasses. In this study we found that the fast relaxation in glasses is a continuation of the dynamics of high-temperature liquids, and its carriers are rapidly diffusing, liquid-like atoms inherited from high-temperature liquids. At room temperature (r.t.), these liquid-like atoms can diffuse rapidly with a viscosity as low as 10 7 Pa · s, which is six orders of magnetite lower than that of glasses (10 13 Pa · s) at the glass transition temperature ( T g ). Molecular dynamics (MD) simulations reveal that this fast relaxation arises from string-like diffusion of the liquid-like atoms in MGs. These findings not only clarify the mechanism of fast relaxation, but also reveal deeper insights into the nature of glasses and the glass transition. The initial clue to the existence of rapidly diffusing, liquid-like atoms in MGs comes from dynamics rather than from atomic structures. In fact, although current technology can capture the three-dimensional atomic structure of MGs 16 , it is still difficult to image the atomic motions in MGs. Dynamic spectra can provide a powerful tool to probe the features and basic properties of glasses 4 , 5 , 6 , 17 , 18 . Figure 1 shows the temperature-dependent loss modulus spectra of several typical MGs, where all samples exhibit three well-separated relaxation processes. . Figure 1a shows that in Al 85 La 10 Ni 5 MG, besides the α relaxation, a β-relaxation hump at about 470 K ( ∼ 0.83 T g ) and a distinct peak at a lower temperature of 270 K ( ∼ 0.48 T g ) occur, indicating that there are some atoms moving much faster than atoms involved in β relaxation. Combining our results and previous reports 6 , 7 , 8 , 9 , it can be concluded that this fast process is not restricted to a specific composition but is a common phenomenon in MGs. Fig. 1: Emergence of the fast-relaxation process in MGs on the dynamic mechanical spectra. a , Loss modulus E ″ at a testing frequency of 1 Hz for Al 85 La 10 Ni 5 MG, showing a distinct hump of β relaxation and a pronounced fast relaxation at temperatures far below that of β relaxation. The coloured regions provide a guide to the eye. b , Loss modulus E ″ for many other MGs, including Al 90 La 10 , Er 55 Co 20 Al 25 , La 68.5 Ni 15 Co 1.5 Al 15 , La 60 Ni 15 Al 25 , Y 68.9 Co 31.1 and (La 0.5 Ce 0.5 ) 65 Co 25 Al 10 , showing a universal fast relaxation at low temperatures. Note that both Al 85 La 10 Ni 5 and Al 90 La 10 show an obvious peak even in linear coordinates. Full size image Next, to prove that the atoms which carry fast relaxation at low temperatures are intrinsically liquid-like, the activation energies were systematically studied. The fast-relaxation activation energy can be obtained from Arrhenius fitting to the frequency-dependent relaxation peak 17 , as shown in Fig. 2a–c . The fast-relaxation activation energies for Y 68.9 Co 31.1 , Ce 70 Cu 20 Al 10 and Al 90 La 10 MGs are 0.60, 0.46 and 0.47 eV, respectively. We further investigated the activation energy of atoms in high-temperature liquids. It has been well established that liquid dynamics exhibits simple Arrhenius behaviour at high temperatures above a crossover temperature 19 , 20 , 21 . That is, the activation energy of liquid dynamics has a fixed value in high-temperature liquids 1 , 19 . For the Y 68.9 Co 31.1 alloy, the viscosities η were measured at high temperatures by electrostatic levitation 21 . For the Ce 70 Cu 20 Al 10 alloy, the self-diffusion coefficients D at high temperatures were obtained by quasi-elastic neutron scattering 22 . Meanwhile, we performed MD simulations for high-temperature Al 90 La 10 liquids to calculate the self-diffusion coefficients. The Arrhenius fittings to η or D give accurate estimates of the activation energies of the atomic dynamics in liquids, which are 0.63, 0.47 and 0.41 eV for Y 68.9 Co 31.1 , Ce 70 Cu 20 Al 10 and Al 90 La 10 , respectively. It is striking that the activation energies of the atomic dynamics in high-temperature liquids and the fast relaxation in glasses are almost identical (Fig. 2a–c ). Furthermore, we summarized the values of activation energies of β relaxation, fast relaxation and high-temperature liquid dynamics collected from the literature for various typical MGs 6 , 7 , 8 , 17 , 21 , 23 . As shown in Fig. 2d , the activation energies of fast relaxation and viscosity/diffusion of high-temperature liquids are quite close for different MGs, and an approximately linear relationship of activation energy ∼ 12 k B T g (where k B is the Boltzmann constant) can be obtained, demonstrating an inherent correlation between the two processes. Fig. 2: Correlation between the activation energies of fast relaxation and high-temperature liquids. a , The activation energy of fast relaxation of Y 68.9 Co 31.1 MG determined by DMA experiments and that of Y 68.9 Co 31.1 liquid obtained by viscosity measurement 21 . b , The activation energy of fast relaxation of Ce 70 Cu 20 Al 10 MG determined by DMA experiments and that of self-diffusion of copper atoms in liquid Ce 70 Cu 20 Al 10 derived from neutron scattering data 22 . c , The activation energy of fast relaxation of Al 90 La 10 MG determined by DMA experiments and that of liquid Al 90 La 10 calculated by MD simulations. d , Statistics of the activation energy of β relaxation (full-filling), fast relaxation (half-left-filling) and high-temperature liquid (half-up-filling) for a variety of MGs (different symbols), showing a close link between fast relaxation and high-temperature liquids. The data are listed in Supplementary Tables 1 – 3 . Full size image In addition to the statistics of activation energy, we also compared the relaxation times of different dynamic processes in Y 68.9 Co 31.1 . The relaxation time τ of a high-temperature liquid can be extracted from the viscosity η via the Maxwell relationship η = G ∞ × τ , where G ∞ is the instantaneous shear modulus. Figure 3 shows that the dynamics of a high-temperature liquid and fast relaxation collapses on a single master curve (the Arrhenius relation for liquids), revealing clearly the strong correlation between them and suggesting the inheritance of rapidly diffusing liquid-like atoms from high-temperature liquids into glass solids. Fig. 3: Relaxation map from experimental results of a Y 68.9 Co 31.1 MG as a function of the inverse of temperature, illustrating how a deep glassy solid inherits liquid-like atoms from a high-temperature liquid. The red line is the Vogel–Fulcher–Tammann 1 fit to the α-relaxation data; the orange line is the Arrhenius fit to the β-relaxation data; the blue line is the Arrhenius fit to the fast-relaxation data, and the dark-green line is the Arrhenius fit to the viscosity data 21 . The half-filled blue star represents the measured viscosity at r.t. for liquid-like atoms via a nanoindentation experiment. The α-relaxation, β-relaxation and fast-relaxation data are measured by DMA. The dotted circles represent crossover points T αβ and T A . Full size image Previous studies have generally treated supercooled liquids as the parent phase of glass and ignored the possible inheritance from high-temperature liquids 1 , 2 , 8 , 17 , 18 . The finding of the inheritance of liquid-like atoms in glass from high-temperature liquids provides a very useful picture for glass structures and dynamics. Figure 3 shows that as temperature decreases, the dynamics of liquids deviates from Arrhenius behaviour, followed by subsequent decoupling of α and β relaxations at the crossover temperature T αβ in a moderately supercooled range 1 , 17 , 18 . Our results suggest an additional decoupling related to the fast relaxation at another crossover temperature. Many studies defined this crossover temperature as T A and revealed that T A is the onset temperature where flow starts to be microscopically cooperative 20 , 21 , 24 . Thus, this decoupling suggests that when the viscosity of a liquid deviates from Arrhenius behaviour, not all of the atoms take part in the cooperative flow; some atoms can maintain liquid Arrhenius behaviour even in glassy states, appearing as persistent liquid-like atoms. Similar decoupling was observed in a previous study of a polymer glass-forming system 25 , implying that the inheritance of liquid-like atoms might be a universal feature of disordered systems. Next, we evaluated the viscosity of rapidly diffusing, liquid-like atoms at r.t. The motion of liquid-like atoms will cause an obvious anelastic hysteresis under superfast loading/unloading processes, which can be directly measured by nanoindentation experiments 26 (Supplementary Fig. 6 ). The viscosity of liquid-like atoms in the Y 68.9 Co 31.1 at r.t. was determined to be 10 6.93 ± 0.02 Pa·s. According to our inheritance scenario, the viscosity of inherited liquid-like atoms at r.t. is estimated to be 10 6.44 ± 0.02 Pa·s. The experimental viscosity (Fig. 3 , half-filled blue star) is close to the predicted value (Fig. 3 , blue curve), confirming the inheritance scenario. As a comparison, a low viscosity of 10 8 Pa·s for nanosized Pd–Si MGs has been reported 27 , revealing the liquid-like nature of MG nanoparticles. The viscosity for the atoms carrying fast relaxation here (10 7 Pa·s) is even smaller than that in Pd–Si nanoparticles, implying that these atoms actually exhibit liquid behaviour. Further, we performed extensive MD simulations on Al 90 La 10 MG with pronounced fast relaxation. Figure 4a,b plot the trajectories of aluminium and lanthanum atoms in Al 90 La 10 at 300 K during 100 ns simulations. All lanthanum atoms vibrate around their equilibrium positions. In contrast, some aluminium atoms can move out of their cages. The analysis of mean square displacements (MSDs) in Fig. 4c shows clearly that these fast aluminium atoms can diffuse as in liquid states, which is further confirmed by calculating the Linderman parameter 28 , 29 (Supplementary Fig. 11 ). Fig. 4: Identification and characterization of liquid-like atoms in Al 90 La 10 MG near r.t. (300 K) by MD simulations. a , b , Atomic trajectories (yellow lines) during 100 ns in MD simulations for aluminium ( a ) and lanthanum ( b ). The red and blue balls represent the initial positions of lanthanum and aluminium atoms, respectively. The trajectories in a are only shown for fast atoms that end up more than 2.0 Å away from their initial positions. c , MSDs for different atoms. d , The spatial distribution of the liquid-like atoms (red balls) and their atomic displacements (yellow arrows). The circles indicate the clusters shown in f . The blue circles show the high-symmetric cluster around a slow atom and the orange circles show the low-symmetric cluster around a liquid-like atom in the string. e , The atomic displacement distribution p ( u , t ) and the pair correlation function g ( r ). f , The probability distribution for the bond-orientational order parameter ŵ 6 . For comparison, the ŵ 6 of aluminium atoms in high-temperature liquids (2,000 K) is also presented. Insets: an icosahedral cluster with high symmetry, and a low-symmetric cluster around fast-diffusive atoms marked in red. Both clusters are circled in d . Full size image Importantly, Fig. 4d shows that the motions of liquid-like atoms form string-like structures, suggesting that fast diffusion of liquid-like atoms in MGs is not random but cooperative 30 . The atomic displacements (Fig. 4e ) display clear multi-peaks, corresponding to the peaks in pair correlation functions. This confirms that the liquid-like atoms diffuse in a cooperative manner, that is, one atom jumps to the adjacent position that was previously occupied by another atom, generally one of its nearest neighbours. There is almost no difference between the atomic volumes of fast and slow aluminium atoms (Supplementary Fig. 12c ). In contrast, the statistics of atomic-bond-orientational order parameter ŵ 6 (ref. 31 ) in Fig. 4f shows that liquid-like atoms exist within more disordered environments and share more similarity with atoms in a liquid, implying that liquid-like atoms usually have lower-symmetric cages that are easier to break. When a neighbouring atom moves away, the centre atom simultaneously hops to this neighbouring site, and another neighbouring atom hops to the centre site at the same time, forming string-like structures (Fig. 4d ). Next, we investigated the inheritance process of liquid-like atoms. Because the liquid-like atoms diffuse in a string-like manner, we analysed the ratio of atoms participating in string-like diffusion to the total number of fast-moving atoms, from high-temperature liquids to glasses. The ratio in Fig. 5a shows clear peaks at all temperatures. As the temperature decreases, the peak shifts to longer times, indicating the slowing down of string-like diffusion. Meanwhile, the peak intensity increases strongly with decreasing temperature, implying that the string-like diffusion becomes increasingly pronounced. We hypothesize that most of the fast-moving atoms will participate in string-like diffusion at cryogenic temperatures, resulting in the fast relaxation. Furthermore, the relaxation time for string-like diffusion is calculated as the time when the ratio reaches its maximum. The relaxation time continues to follow the Arrhenius relation of liquid (Fig. 5b ), and its extension goes across the experimental dynamic mechanical analysis (DMA) data of fast relaxation. This provides strong evidence that the fast relaxation is caused by string-like diffusion inherited from the high-temperature liquid. Fig. 5: Inheritance of string-like diffusion in the Al 90 La 10 system. a , The ratio of atoms participating in string-like diffusion to the total number of fast-moving atoms versus time at different temperatures. b , The various relaxation times, including the string-like diffusion, the high-frequency DMA (1 GHz) testing in simulations, the measured fast relaxation in DMA experiments, and the superfast anelastic behaviour at r.t. measured by nanoindentation. Full size image For further evidence, high-frequency DMA measurement at 1 GHz was performed via MD simulations. The loss modulus spectrum shows clearly two peaks at 665 and 528 K, respectively (Supplementary Fig. 13 ). The peak at 528 K was plotted as a magenta sphere in Fig. 5b , in agreement with the master curve of string-like diffusion. This provides direct evidence that the fast relaxation revealed by DMA experiments is associated with string-like diffusion. Nanoindentation was conducted to detect the rapidly diffusing, liquid-like atoms in La 10 Al 90 at r.t. It gives a relaxation time of 0.70 ms, which is consistent with that of string-like diffusion. Based on these experimental and simulated results, we concluded that the fast relaxation is the inherited motion from liquids. The observation of liquid-like atoms in dense MGs deepens our understanding of the nature of glasses. Although a previous study of amorphous SiO 2 imaged liquid-like fast atoms at the edge of a two-dimensional sheet 15 , the fast motion there was due to the chemical-bond hierarchy. The case in MGs is quite different. First, unlike oxide glasses, which have complex covalent network structures, MGs have simple atomic structures similar to the random packing of spheres that are unlikely to result in chemical-bond hierarchy 18 , 32 . Second, MGs are quite densely packed compared with SiO 2 glass 33 . Therefore, the finding of rapidly diffusing, liquid-like atoms in dense-packed MGs reveals a more universal and extraordinary feature of glass solids: liquid-like atoms can be inherited from a high-temperature liquid into a glass during the glass formation, and a glassy solid is essentially part solid and part liquid. The inheritance scenario demonstrated here challenges the previous understanding of fast relaxation. Fast relaxation has long been taken to be a more localized mode and a precursor of β relaxation 6 , 7 , 8 , 34 . Under this localized scenario, fast relaxation is related closely to β relaxation and is expected to evolve into β relaxation like β relaxation evolving into α relaxation. However, this scenario is not supported by experimental results. On the one hand, according to a wide range of experimental results, the composition dependences of fast and β relaxations are quite irrelevant, that is, a system with pronounced fast relaxation does not necessarily exhibit pronounced β relaxation. On the other hand, fast relaxation will not merge into β relaxation but directly into the Arrhenius relation of liquid (Fig. 3 ). This is not observed exclusively in Y–Co but is a universal feature of various MGs (Supplementary Fig. 7 ). These facts suggest that fast relaxation is probably not a more localized motion or a precursor of β relaxation, but an inheritance of liquid dynamics demonstrated here. In contrast to the localized motion, the liquid-like atoms carrying fast relaxation may be located inside or outside the flow units of β relaxation. When these liquid-like atoms are activated, our simulation shows that they have clearly broken their low-symmetry cages and exhibit string-like cooperative diffusion. Unlike the β relaxation induced by local, loosely packed atoms, these liquid-like atoms are associated with an extremely disordered local structure. β relaxation is believed to be the inheritance from supercooled liquids, whereas fast relaxation is the inheritance of liquid-like atoms from high-temperature liquids. The finding of liquid-like atoms in MGs provides more complete microcosmic structural and dynamic pictures of glasses, which can help understand the dynamics–property relationship in MGs. In this work, we confirm that the liquid-like atoms control the superfast anelasticity of MGs at r.t., which is crucial for the application of MGs in microelectromechanical systems 35 and high-precision sensors 36 . The liquid-like atoms exist within highly disordered local environments which may act as defects to promote the formation of multiple shear bands and therefore increase the plasticity of MGs 37 . For example, recent studies reported a synchronously non-monatomic change between compression plasticity and fast relaxation in a high-entropy MG 38 . Meanwhile, the time scale of fast relaxation at r.t. is comparable with the frequency at which ultrasonic moulding can join MG ribbons together into bulk 39 , implying that liquid-like atoms play a role in the mechanism of cold-joining processes. In other systems these liquid-like atoms have also been found to influence thermal conductivity 13 , 28 , which helps design functional glasses. In summary, by combining dynamic mechanical studies and detailed MD simulations, we demonstrate that the general fast dynamics of MGs at low temperatures is related to liquid-like atoms in glass solids. The liquid-like atoms are inherited from high-temperature liquids and exhibit behaviour similar to that in liquids, including comparable activation energies and low viscosities of ∼ 10 7 Pa·s at r.t. These findings extend our current understanding of the microscopic structural and dynamic pictures of glasses and have implications for establishing the dynamics–property relationship in glasses. Methods Material preparation The MG-forming alloy ingots with various nominal compositions were prepared by arc-melting mixtures of pure metals in a titanium-gettered argon atmosphere. Each ingot was melted at least five times to ensure chemical homogeneity. Eventually, glassy ribbons with thickness of about 30 μm were fabricated by the single-roll melt-spinning in an argon atmosphere. The amorphous nature of the samples was verified by the X-ray diffraction using Cu Kα radiation (Supplementary Fig. 1 ). DMA The relaxation behaviours of selected MGs were characterized by a TA Q800 dynamic mechanical analyser. The dynamic modulus \(E^ \ast \left( \omega \right) = E^\prime \left( \omega \right) + iE^{\prime\prime} \left( \omega \right)\) , where the real part E ′ and imaginary part E ″ represent the storage and loss modulus, respectively, was measured in film tension mode with a constant heating rate of 3 K min −1 and different single testing frequencies or discrete frequencies of 1, 2, 4, 8 and 16 Hz (Supplementary Figs. 2 – 4 ). Nanoindentation Nanoindentation was performed on a TI 980 TriboIndenter system with a 5 μm spherical indenter at r.t. (293 K). The peak loads of Al 90 La 10 and Y 68.9 Co 31.1 were set to 500 and 200 μN, respectively; and the fast-loading times of Al 90 La 10 and Y 68.9 Co 31.1 were set to 0.005 and 0.01 s, respectively. In all of these tests, the slow-loading, holding and unloading time were both fixed at 0.1 s (Supplementary Figs. 5 and 6 ). MD simulation MD simulations were conducted within the LAMMPS software package to prepare and analyse Al 90 La 10 alloy using the embedded atom method potentials. In MD simulations, each sample contains 4,000 atoms in a cubic box with periodic boundary conditions applied in three directions. The samples were first equilibrated at 2,000 K for 2.0 ns, followed by hyperquenching to 100 K with cooling rate of 1.0 × 10 12 K s −1 , and further relaxed at 100 K for 2.0 ns (Supplementary Fig. 8 ). A time step of 2.0 fs was selected. The quenching process was performed in the NPT ensemble with zero pressure. The quenched sample was further relaxed at different temperatures in the NVT ensemble for statistics and analysis. More details about the simulation and calculations can be found in the Supplementary Information . Calculation of self-diffusion coefficient D The self-diffusion coefficient D in liquids was calculated via linear fitting of the long-time MSD at different temperatures. The MSD is defined as \({\Delta}r^2(t) = N^{ - 1}\mathop {\sum}\nolimits_{j = 1}^N {\left\langle {\left( {{{{\mathbf{r}}}}_{j,t} - {{{\mathbf{r}}}}_{j,0}} \right)^2} \right\rangle }\) , where N is the number of atoms, and r j ,0 and r j , t represent the coordinate vector of atom j at zero and time t , respectively. The self-diffusion coefficient D can be calculated as \(D = \mathop {{\lim }}\limits_{t \to \infty } \frac{{{\Delta}r^2(t)}}{{6t}}.\) More details can be found in Supplementary Fig. 9 . Atomic volume and bond-orientational order parameters To characterize the local atomic structures, Voronoi tessellation was used to divide space into close-packed polyhedra around atoms by constructing bisecting planes along the lines joining the central atom and all of its neighbours. The volume of the Voronoi polyhedral surrounding each atom is calculated as the atomic volume V a . To quantitatively describe the bond-orientational symmetry around the central atom, the bond-orientational order (BOO) parameters were employed via spherical harmonic expansion. Specifically, the BOO of atom i can be represented by the l -fold symmetry as a 2 l + 1 vector defined as \(q_{lm}\left( i \right) = \frac{1}{{N_i}}\mathop {\sum}\nolimits_{j = 1}^{N_i} {Y_{lm}} \left( {\theta \left( {{{{\mathbf{r}}}}_{ij}} \right),\phi \left( {{{{\mathbf{r}}}}_{ij}} \right)} \right)\) , where Y lm are spherical harmonics and N i is the number of nearest neighbours (Voronoi neighbours in our case) around atom i . The rotational invariants for atom i can be defined as \(q_l = \sqrt {\frac{{4\uppi }}{{2l + 1}}\mathop {\sum}\nolimits_{m = - l}^l {\left| {q_{lm}} \right|^2} }\) and \(w_l = {\sum} {\left( {\begin{array}{*{20}{c}} l & l & l \\ {m_1} & {m_2} & {m_3} \end{array}} \right)} \cdot q_{lm_1} \cdot q_{lm_2} \cdot q_{lm_3}.\) Here the term in brackets is the Wigner 3- j symbol. Instead of w l , the normalized parameter defined as \(\hat w_l = w_l \cdot \left( {\mathop {\sum}\nolimits_{m = - l}^l {\left| {q_{lm}} \right|^2} } \right)^{ - 3/2}\) is a more sensitive measure of the different orientational symmetries and is often used to evaluate the BOO. For example, the \(\hat w_6\) values of face-centred cubic, hexagonal close-packed and icosahedron clusters are −0.01316, −0.01244 and −0.16975, respectively. Ratio of string-like diffusion Here we define the atom j as a fast-moving atom if it moves more than 2.0 Å during a time interval t ( \(\left| {{{{\mathbf{r}}}}_{j,t} - {{{\mathbf{r}}}}_{j,0}} \right| >2.0\,{\mathrm{\AA}}\) ). And define the atom j as a string-like atom if it moves to the original position of another atom ( \(\left| {{{{\mathbf{r}}}}_{j,t} - {{{\mathbf{r}}}}_{k,0}} \right|<0.3\,{\mathrm{\AA}}\) ). We count the total number of fast-moving atoms ( N fast ) and number of string-like atoms ( N string ), and use their ratio ( N string / N fast ) to describe the intensity of string-like diffusion. High-frequency DMA testing To prove that fast relaxation does exist in the high-temperature liquid, we also performed high-frequency DMA testing via MD simulations. A sinusoidal strain \(\varepsilon \left( t \right) = \varepsilon _{{{\mathrm{A}}}}\sin \left( {2\uppi f_{{{\mathrm{A}}}}t} \right)\) was applied to the samples along the x – y direction, where f A is the frequency and selected as 1 GHz, and ε A is fixed at 1%. For statistics, 100 independent sinusoidal shear deformation simulations were performed at each temperature. For each MD–DMS loading, ten cycles were used. We fitted the stress as \(\sigma \left( t \right) = \sigma _0 + \sigma _{{{\mathrm{A}}}}\sin \left( {2\uppi f_{{{\mathrm{A}}}}t + \delta } \right)\) , where σ 0 is a constant term and δ is the phase difference between stress and strain. Loss moduli are calculated according to \(E^{\prime\prime} = \sigma _{{{\mathrm{A}}}}/\varepsilon _{{{\mathrm{A}}}}\sin \left( \delta \right)\) (Supplementary Fig. 13 ). Data availability The data supporting the findings of this work are included within the paper and its Supplementary Information files. Extra data are available from the corresponding authors upon reasonable request.

Metallic glass is an important advanced alloy, holding promise for broad engineering applications. It appears as a solid form in many aspects, with beautiful metal appearance, exceeding elasticity, high strength, and a densely packed atomic structure. However, this all-solid notion has now been challenged. Prof. Bai Haiyang from the Institute of Physics of the Chinese Academy of Sciences (CAS) has recently shown the existence of liquid-like atoms in metallic glasses. These atoms inherit the dynamics of high-temperature liquid atoms, revealing the nature of metallic glasses as part-solid and part-liquid. Results were published in Nature Materials. Condensed matter can generally be classified into solid and liquid states. Under extreme conditions or in specific systems, matter exists in special states that simultaneously exhibit some properties of both solids and liquids. In this case, solids may contain rapidly diffusing, liquid-like atoms that can move fast even at low temperatures. For example, ice enters a "superionic" state under high pressure at high temperatures. In this state, H atoms can diffuse freely while O atoms are fixed in their sublattices. Such special states are also observed in Earth's inner core and in the Li-conducting materials of advanced batteries, which are drawing growing attention in science and engineering. Visualization of Liquid-like atoms in Al90La10 MG at 300K by MD simulations. Credit: Institute of Physics In this study, the researchers revealed that liquid-like atoms exist in densely packed metallic glasses. Combining extensive dynamical experiments and computer simulations, they found that when the viscosity of a liquid deviates from Arrhenius behavior, not all atoms take part in cooperative flow and subsequent solidification. In fact, some atoms can maintain liquid Arrhenius behavior even when the system is cooled down to a glass state, thus appearing as persistent liquid-like atoms that lead to fast relaxation at rather low temperatures. "A glassy solid is essentially mostly solid and a small part liquid. Even at room temperature, liquid-like atoms in a glassy solid can diffuse just as easily as in its liquid state, with an experimentally determined viscosity as low as 107 Pa·s, while the viscosity of the solid part is larger than 1013 Pa·s," said Prof. Bai. These findings provide a clearer microscopic picture of glasses. This new picture can help scientists better understand how the properties of glass materials are related to their dynamics. For example, liquid-like atoms control the anelasticity of glasses and may affect their ductility. Moreover, the strong relationship between liquid-like atoms and disordered structure also has implications for studying the topological origin of fast diffusion in solids, such as superionic-state matters and ion conductors.

10.1038/s41563-022-01327-w

Biology

Unnoticed sex reversal in amphibians due to artificial estrogen from pills

Stephanie Tamschick et al. Sex reversal assessments reveal different vulnerability to endocrine disruption between deeply diverged anuran lineages, Scientific Reports (2016). DOI: 10.1038/srep23825 Journal information: Scientific Reports

http://dx.doi.org/10.1038/srep23825

https://phys.org/news/2016-04-unnoticed-sex-reversal-amphibians-due.html

Abstract Multiple anthropogenic stressors cause worldwide amphibian declines. Among several poorly investigated causes is global pollution of aquatic ecosystems with endocrine disrupting compounds (EDCs). These substances interfere with the endocrine system and can affect the sexual development of vertebrates including amphibians. We test the susceptibility to an environmentally relevant contraceptive, the artificial estrogen 17α-ethinylestradiol (EE2), simultaneously in three deeply divergent systematic anuran families, a model-species, Xenopus laevis (Pipidae) and two non-models, Hyla arborea (Hylidae) and Bufo viridis (Bufonidae). Our new approach combines synchronized tadpole exposure to three EE2-concentrations (50, 500, 5,000 ng/L) in a flow-through-system and pioneers genetic and histological sexing of metamorphs in non-model anurans for EDC-studies. This novel methodology reveals striking quantitative differences in genetic-male-to-phenotypic-female sex reversal in non-model vs. model species. Our findings qualify molecular sexing in EDC-analyses as requirement to identify sex reversals and state-of-the-art approaches as mandatory to detect species-specific vulnerabilities to EDCs in amphibians. Introduction Amphibians face a global ongoing decline 1 , 2 . Anthropogenic causes such as industrial agriculture 3 , habitat destruction 4 , 5 , invasive species 6 , climate change 7 , land use 8 and infectious diseases 9 , including several forms of chytridiomycosis 10 , 11 , are among the major threats. However, the sum of multiple stressors 1 , 7 , some of which poorly known, is considered to be the true reason for the massive population declines. One potential cause represents endocrine disrupting compounds (EDCs) 12 . Besides pesticides, EDCs comprise either natural products or synthetic chemicals that mimic, enhance (an agonist), or inhibit (an antagonist) the action of hormones and in this way interfere with the synthesis, secretion, transport, binding, action, or elimination of natural hormones, which are responsible for the maintenance of homeostasis, reproduction, development, and/or behavior 13 . Considerable amounts of EDCs are globally found in waste and surface waters 14 , 15 and can easily enter the body of aquatic organisms and impair their natural hormonal pathways. EDCs are well known for their negative impacts on the sexual development of aquatic organisms such as fish 16 , 17 and are suspected to cause fertility problems in humans 18 , 19 . However, their impact to non-model amphibians with aquatic larvae is not well studied, despite recent evidence for high EDC-relevance to suburban frog populations 20 . One globally relevant EDC is 17α-ethinylestradiol (EE2), a synthetically stabilized estrogen and main ingredient of many female contraceptive pills. The inert EE2 is then excreted and insufficiently eliminated by sewage plants and hence reaches aquatic ecosystems 14 . It is a main hormonal pollutant, resistant to degradation, that accumulates in sediments and biota 14 . Concentrations from 24 to 831 ng/L have been detected in European and American surface waters 21 , 22 , 23 . Such concentrations have been shown to alter behavior and somatic and sexual development in fish and amphibians 12 , 14 , 15 . Due to their semi-aquatic life cycle, often aquatic reproduction and a highly permeable skin, amphibians are especially sensitive to EDCs. Effects on development and reproduction are best examined in clawed frogs, Xenopus laevis and X. tropicalis . In these amphibian models, EE2-concentrations as low as 0.3 ng/L have been shown to affect calling behavior and mating success 24 . Higher but still environmentally relevant amounts of EE2 (29 to 840 ng/L) have been shown to affect body morphology, metamorphosis and hemoglobin catabolism 25 , 26 . Importantly, EE2 can lead to impaired sexual development as mirrored by gonad histomorphology, demonstrating that male clawed frogs ( X. laevis ) develop mixed sex (=‘intersex’, see below) gonads or even show complete phenotypic sex reversal 26 , 27 , 28 , 29 . The undifferentiated anuran gonad is bipotential and can develop into either ovary or testis 30 . Therefore, exogenous hormones can override the primary genetic sex determination signal and lead to developmental disturbances, mixed sexes or complete sex reversal. One major obstacle of studying EDC-effects in amphibians has been the mostly inaccessible information about genetic sex. In most previous EDC-studies, sex reversal had to be inferred by comparing sex ratios of control and exposed frogs, assuming a normal 1:1 proportion, which may have easily led to wrong conclusions about EDC-impacts on sex ratios. While all amphibian species investigated show genetic sex determination 31 , exhibiting either male (XX/XY) or female (ZZ/ZW) heterogamety, an extrapolated 96% of all species have microscopically indistinguishable sex chromosomes 32 , requiring molecular sexing methods. Although EDC-studies with molecular sexing were applied to the model Xenopus 26 , 33 , sex markers have become only recently available for some non-model anurans 32 , 34 , 35 , 36 , 37 , 38 and have not been used in EDC-experiments. Using a high-standard flow-through-system and the first direct experimental approach of its kind, we simultaneously exposed European tree frogs ( H. arborea ), green toads ( B. viridis ) and the well investigated but deeply diverged model-species X. laevis to EE2, applied molecular sexing followed by histological analysis and compared impacts on their sexual development. We found striking differences in the susceptibility to sex reversal between model and non-model species, showing that state-of-the-art approaches are an important prerequisite to detect species-specific vulnerabilities to EDCs in amphibians. Results Phenotypic sex reversal of genetic males Among all three anuran species, simultaneous exposure to three EE2-concentrations under flow-through-conditions resulted in different proportions of male-to-female sex reversal, ranging from 15 to 100% ( Table 1 and Fig. 1 ), which was solely revealed when comparing genetic and phenotypic sex of experimental animals. Importantly, no sex reversal occurred in control groups. While sex reversal ( Figs 1 and 2 ) was generally correlated to EE2-concentration, interspecies differences (p ≤ 0.010) between clawed frogs ( X. laevis ) and tree frogs ( H. arborea ) were found at all concentrations and between clawed frogs and green toads ( B. viridis ) at the highest concentration (5,000 ng/L; p ≤ 0.001). While EE2-treatment produced similar percentages of sex-reversed tree frogs and green toads (15 to 36%), clawed frogs appeared most susceptible (up to 100%). At the lowest concentration (50 ng/L) 31.3% of genetically male clawed frogs developed female phenotypes, i.e. ovaries, while no sex reversal occurred in the non-model species ( H. arborea, B. viridis ). As expected for a feminizing EDC, sex reversal occurred always from genetic male to phenotypic female. According to gross morphological observation and histological evidence, sex-reversed genetic male frogs and toads developed ovaries that showed no difference to those of genetic control and untreated 39 females. Table 1 Effects of three 17α-ethinylestradiol (EE2) concentrations (50, 500 and 5,000 ng/L) on the sexual development of model and non-model amphibian species. Full size table Figure 1 Quantities of sex reversal (contradiction between genetic and phenotypic sex) under the influence of 17α-ethinylestradiol (EE2) in three deeply diverged anuran amphibians. Percentages of genetic-male-to-phenotypic-female sex reversal in African clawed frogs ( Xenopus laevis , red), European tree frogs ( Hyla arborea , green) and European green toads ( Bufo viridis , blue) exposed to three concentrations of EE2 and in control animals; pooled data from two replicate experiments for each treatment or control. Susceptibility differences in genetic-male-to-phenotypic-female sex reversal occurred at all concentrations: (*) significant differences between clawed frogs and tree frogs (p ≤ 0.010); ( x ) significant differences between clawed frogs and green toads (p ≤ 0.001). Statistical analyses were conducted using cross-tabulation, Chi square and Fisher´s exact tests (α = 0.05). Full size image Figure 2 Histological sections of three anuran species under the influence of 17α-ethinylestradiol (EE2). ( a–c ) Normal male, normal female and phenotypically sex-reversed gonad of African clawed frog ( Xenopus laevis ). (d–f) Normal male, normal female and phenotypically sex-reversed gonad of European green toad ( Bufo viridis ). (g–i) Normal male, normal female and phenotypically sex-reversed gonad of European tree frog ( Hyla arborea ). Bo – Bidder’s organ, characteristic of bufonid gonads (for details: Methods); fb – fat body; o – ovary; t – testis; arrows indicate seminiferous tubules; *ovarian cavity; arrowheads – diplotene oocytes. Scale bars are 100 micrometers. Full size image Mixed sex gonads In addition to sex reversals, EE2-treatment provoked the development of various percentages of mixed sex 40 gonads (equivalent to ‘intersexes’ of some authors 41 , 42 , 43 ) that were histologically recorded in all three species ( Fig. 3 and Table 1 ). Such altered gonads are characterized by the presence of ovarian within testicular tissue in genetic males and were never found in control groups. In contrast to the sex reversal analyses, X. laevis formed fewer mixed sex gonads than B. viridis (p ≤ 0.026). No significant susceptibility differences between H. arborea and the model species were found. Both non-model species also differed in their susceptibility at the lowest concentration (50 ng/L; p ≤ 0.015). Figure 3 Histological sections of mixed sex gonads of three anuran species under the influence of 17α-ethinylestradiol (EE2). (a) African clawed frog ( Xenopus laevis ), (b) European green toad ( Bufo viridis ), (c) European tree frog ( Hyla arborea ); Fig. 2 for control and sex-reversed individuals. Bo – Bidder’s organ, specific of bufonid toads’ gonads; fb – fat body; m – meiocytes; o – ovary; st – seminiferous tubules; t – testis; *a cavity separating testicular and ovarian parts of the mixed sex gonad; white arrow indicates ovarian cavity in the ovarian portion of the mixed gonad; white arrowheads show diplotene oocytes; yellow dotted lines separate testicular and ovarian parts of the mixed sex gonads. Scale bars represent 100 micrometers. Full size image Discussion Using a new combination of experimental features, we provide evidence for different quantities of genetic-male-to-phenotypic-female sex reversal in three amphibian species, diverged between 78 million years 44 ( Hyla, Bufo ) and 206 My ( Xenopus ), under exposure to the estrogen EE2. This synthetic substance is globally of high relevance for EDC-pollution of aquatic ecosystems 14 , 15 . Our new approach combined simultaneous exposure of tadpoles to three EE2-concentrations in a flow-through-system and genetic sexing of metamorphs of model and non-model experimental anurans. We applied environmentally (pollution) and physiologically (expected effects in X. laevis ) relevant concentrations of EE2. Genetic sexing of metamorphosed tree frogs and green toads revealed these two non-model species to have similar susceptibilities to sex reversal among each other, while both significantly differed from X. laevis . This model-species, in which genetic sex is governed by a female heterogametic (ZZ/ZW) chromosome system 45 , proved to be more sensitive to EE2 with a lower dose provoking sex reversals and more affected animals ( Table 1 ). On the other hand, B. viridis and H. arborea , both diverged 206 My from X. laevis and possessing male heterogametic (XX/XY) sex chromosomes 32 , 35 , showed higher percentages of mixed sex individuals than X. laevis . All of this suggests that species-specific developmental stages, sex determination systems or endocrine pathways, shaped by long separate evolutionary histories, were differently affected by EE2 and such a wide spectrum of effects can be generally expected also for other EDCs among diverged anuran lineages. The occurrence of more than 50% of genetic females among the randomly chosen hatchlings in several of our test tanks underlines the importance of genetic sexing. Unavailability of genetic sexing, as in many previous studies, could easily lead to wrong conclusions about the strength of feminizing (or masculinizing) effects of EDCs when determining “no observed effect concentration” (NOEC) and “lowest observed effect concentration” (LOEC) for endocrine active substances. Different estrogenic compounds with concentrations reaching from the low nanogram- to the high microgram-per-liter range have been shown to provoke phenotypic male-to-female sex reversals in X. laevis 46 , 47 , 48 models. To our knowledge, only one previous study 26 has examined sex reversals after EE2-exposure using molecular sexing in X. laevis , examining a similar range of concentrations (90, 840, 8,810 ng/L). In contrast to our study, male-to-female sex reversals were not detected under the 90 ng/L treatment and at the higher concentrations with only 7 and 17%, respectively. However, these authors used a static and not a flow-through-system, which may explain the deviating results to our study, as EE2-concentrations may stronger fluctuate due to effects of metabolic activity of microorganisms in tanks 49 , 50 , due to greater biomass sorption of EE2 51 , or due to simple adsorption to surfaces of exposure tanks. Beyond the synthetic EE2, on which we focused due to its high environmental relevance, previous sex reversal estimates in X. laevis , evaluating only sex ratios, involved the natural, ephemeral 17β-estradiol (E2). Such E2-treatments provoked skewed sex ratios 40 , 48 , 52 , 53 , 54 or complete feminization 46 , 53 , 55 , 56 . In H. arborea and B. viridis effects have only been studied 56 at the very high 100,000 ng/L E2-concentration. In both species, no female-biased sex ratios but a high percentage (59.3%) of undifferentiated gonads in B. viridis were found. Since gonad differentiation in bufonid toads is slower compared to the other species at this developmental stage 39 , we assume that the time of dissection at metamorphosis may have influenced these results. Several inconsistent outcomes in the literature may be explicable by the potentially wrong assumption of initial 1:1 sex ratios of experimental amphibians. Based on our data, we strongly recommend genetic sexing, whenever available, as a hallmark of appropriate evaluation of EDC-effects in amphibians. This demand can be extended to other vertebrates and generalized to EDC-research in organisms with homomorphic sex chromosomes, including invertebrates 57 . Otherwise, as shown here, complete sex reversal as a very profound EDC-effect, occurring at low concentrations, may be completely overlooked. Furthermore, deep phylogenetic differences may result in strong susceptibility differences towards EDCs. Though we do not advocate the extensive use of endangered amphibians, we conclude that results gained from earlier studies in X. laevis in general and without genetic sex information specifically should not be uncritically extrapolated to other anuran species. Methods Animals This experiment was approved by the German State Office of Health and Social Affairs (LaGeSo, Berlin, Germany; G0359/12); all methods were carried out in accordance with approved guidelines. Xenopus laevis tadpoles were obtained from the stock at the Leibniz-Institute for Freshwater Ecology and Inland Fisheries. Induction of spawning and tadpole husbandry followed standard methods 58 . Parental animals of B. viridis and H. arborea were caught at several localities in Greece ( Supplementary Table 1 ) and non-invasively DNA-sampled 59 . Parts of their clutches were transferred to IGB (permit 115790/229) and acclimated at 22 ± 1 °C in 10 L Milli-Q grade water, supplemented with 2.5 g marine salt (Tagis, Germany). Hormone exposure and experimental conditions 17α-Ethinylestradiol (Sigma-Aldrich, Germany), dissolved in dimethyl sulfoxide (DMSO; Roth, Germany), was applied in nominal concentrations of 50, 500 and 5,000 ng/L ( Supplementary Fig. 1 and Supplementary Table 2 , for measurements during the experiment); control animals received 0.00001% DMSO. EE2-concentrations in test tanks were checked weekly by high performance liquid chromatography/mass spectrometry (HPLC-MS/MS) and adjusted if required. In order to minimize adsorption or release of EDCs, we used glass tanks and all connections of the flow through system consisted of inert materials involving mainly PTFE (Polytetrafluoroethylene, “Teflon”)-coating or Platinum-cured Silicon tubing (Cole-Parmer). Exposure of tadpoles started at Gosner 60 -stage 22–23 in B. viridis and H. arborea , equivalent to Nieuwkoop-Faber 61 stage 42–44 in X. laevis , distinctly prior to the sensitive phase of sex determination in all species 30 , 62 . Twenty randomly chosen individuals per species and treatment were transferred into each test tank in a high-standard flow-through-system (details 52 ). Two replicates per exposure group (including control) comprised in total 160 tadpoles per species. Stock solutions and water were piped via a peristaltic pump into a mixing chamber, mixed to final EE2-concentrations and supplied to a cluster of three test tanks each. Concentrations were thus identical for all three species in each treatment group. Tadpoles were reared in a 12/12 h light/dark cycle at constantly 22 ± 1 °C in sufficiently aerated and regularly cleaned tanks. Weekly monitored water parameters comprised: dissolved oxygen, nitrate, ammonium, pH, conductivity and hardness; values were adequate as in previous studies involving the same equipment 40 . Tadpoles were fed SeraMicron (Sera, Germany), H. arborea and B. viridis were additionally supplied with TetraMin (Tetra, Germany). To imitate natural conditions under which H. arborea and B. viridis leave water at metamorphosis, animals were transferred to glass terraria at Gosner stage 46. Xenopus laevis were dissected at equivalent Nieuwkoop-Faber stage 66; hylids and bufonids after sufficient post-metamorphic differentiation 39 . Phenotypic sexing based on gonad gross morphology and histology Animals were anesthetized by immersion in tricaine methanesulfonate (MS 222; Sigma-Aldrich), decapitated and dissected under a binocular microscope (Olympus SZX7). Gonadal anatomy served for preliminary phenotypic sexing and detection of underdeveloped gonads. To improve visualization, a drop of Bouin’s solution (Sigma-Aldrich) was added; and in situ anatomical photographs were taken (Olympus DT5 camera). For histology, gonads were carefully dissected, separated from adjacent tissue, fixed in Bouin (24 h) and subsequently rinsed several rounds in 70% ethanol. Histological sections were prepared for 50% of study animals (from each tank 10 randomly chosen individuals, i.e. 20 per treatment group). Analyses were performed according to established protocols 30 , 39 , 63 . Using Stemi SV11 (Zeiss) microscope and camera, separated gonads were photographed, embedded in paraplast, sectioned into 7 μm longitudinal slices, stained with Mallory’s trichrome and examined using Zeiss Axioskop 20 microscope. Images were acquired by a cooled Carl Zeiss AxioCam HRc CCD camera. Histological sections were screened slide by slide to establish phenotypic sex. Ovaries were recognized by the presence of ovarian cavities, early meiotic oocytes and/or diplotenes and testes by spermatogonia, spermatocytes and/or seminiferous cords or tubules. In the case of B. viridis , the most anterior part of both male and female gonads is Bidder’s organ, an ovary-like structure, characteristic of bufonids 64 . In B. viridis mixed sex was defined when ovarian meiocytes were found inside male testicular tissue behind the physiological transition region between Bidder’s organ and the actual gonad. All phenotypic sexing was performed without prior knowledge about genetic sex of animals. Genotypic sex determination DNA extraction involved the BioSprint robotic workstation with its 96 DNA Plant Kit (Qiagen, Germany) according to the manufacturer’s protocol. To establish genetic sex, species-specific polymerase chain reactions (PCRs) were conducted on Eppendorf Mastercyclers (Ep Gradient S). For X. laevis , two genes were amplified 45 , 65 : DMRT1 and the female-specific DM-W ( Supplementary Table 3 ). Genetic sexing of non-model species involved microsatellites WHa5–201 and Ha-H108 34 , 36 ( H. arborea ) and C201 32 , 66 , HNRNPD and CHD1 67 ( B. viridis ; Supplementary Table 3 ). Genotypes were analyzed on a sequencer (3500 × L Genetic Analyzer, Applied Biosystems) and Genemapper v. 4.0 was used for visualization of peaks. DNA quality issues (four H. arborea ) and homomorphy of microsatellites (33 B. viridis ) prevented genetic sexing in these individuals that were excluded from sex reversal analyses. Detection of complete sex reversal and mixed sex Phenotypic sexing of all animals was based on gross morphology and histology of gonads 39 , 52 . Complete sex reversal was stated if genetic males showed a phenotypically female gonad, irrespective of the degree of its differentiation and not differing from those of control females. Mixed sex gonads were detected by the presence of ovarian and testicular tissue in the same gonad. Statistics All data were analyzed with SPSS Statistics 22 (IBM, Armonk, NY). Intra- and inter-specific differences in EE2-susceptibility were examined. For evaluations of sex reversal and mixed sex, we first compared both replicates per species and parameter using Fisher’s exact test. If no differences (exact p ≥ 0.05) were found, both replicates per treatment were pooled in order to compare control and exposure groups within and between species using cross-tabulations with 2-sided Chi square tests (α = 0.05). Post-hoc Fisher’s exact tests (2-sided) were applied for pairwise comparisons including False Discovery Rate corrections 68 . Additional Information How to cite this article : Tamschick, S. et al . Sex reversal assessments reveal different vulnerability to endocrine disruption between deeply diverged anuran lineages. Sci. Rep. 6 , 23825; doi: 10.1038/srep23825 (2016).

Hormonally active substances may contribute to global amphibian decline. Some compounds, for example from pharmaceuticals, occur in biologically relevant concentrations in freshwater ecosystems, and thus can affect the hormonal system and the sexual development of animals. Researchers from the IGB and the University of Wroclaw have compared the effects of the pill estrogen ethinylestradiol (EE2) in three amphibian species. The study, published in the journal Scientific Reports, shows that EE2 can lead to a complete feminization of genetic males. Without molecular establishment of the genetic sex, this has remained partly unnoticed. 17α-Ethinylestradiol (EE2) is a synthetic estrogen, which is a frequently used active ingredient of female contraceptive pills but naturally does not occur in the environment. As it is only incompletely removed in sewage treatment plants, it can reach water bodies in biologically relevant concentrations. The evolutionary biologist, Dr. Matthias Stöck, principal investigator of the study and Heisenberg-Fellow, says: "Amphibians are almost permanently exposed to such threats. Only, if we will be able to actually access these risks, we will be able to eliminate them in the long term". The sensitivity towards hormonally active substances, like EE2, is not the same in all amphibians, hypothesize the scientists. Some species have passed through several hundred million years of independent evolutionary history and have evolved different mechanisms of genetic sex determination. Thus, for the first time within the same experiment, the research team from the IGB and the University of Wroclaw tested the influence of EE2 on the development of three different amphibian species: In addition to the model species, the African clawed frog (Xenopus laevis), tadpoles of the European tree frog (Hyla arborea) and the European green toad (Bufo viridis) have been raised in water containing different concentrations of EE2, and compared to control groups. Tadpoles and adults of the European tree frog (Hyla arborea) may be exposed to hormonally active substances, like EE2, in their natural habitats. Credit: Christophe Dufresnes Remarkable in this experiment was that the genetic sex of all species was established using latest molecular approaches. At the same time, the researchers studied the phenotypic development of the sexual organs (gonads), i.e. the anatomical appearance and that of the tissues under the microscope. Only this comparison of genetic and phenotypic sex has allowed to completely capture the effects of EE2. "In addition to other threats, the feminization of populations may contribute to the extinction of amphibian species", says Matthias Stöck. The results of the study show that after exposition to EE2, in all amphibian species, a sex reversal occurred reaching from 15 to 100 percent. However, the three species exhibit different sensitivity. IGB-researcher Prof. Werner Kloas, co-author of the paper, is an internationally renowned eco-toxicologist. He says: "EE2 is also part of our water supply and, together with other estrogen-like substances, it presents a serious risk not only for amphibians but also for humans. Our study shows that the clawed frog as model species is well-suited to study the effects of hormonally active substances in the environment. The effect established in this species, however, cannot be extrapolated to other amphibian species without caution." For the study, the collaboration with the University of Wroclaw in Poland was essential. The collaborator there, Prof. Maria Ogielska, declares: "We have contributed our profound research experience from the investigation of reproduction of amphibians, namely from developmental biology, into the project". Ph.D.-student Stephanie Tamschick says: "Together with students, we have raised tadpoles over three months, under identical experimental conditions. It was new that we have established the genetic sex in all species, i.e. whether a tadpole was genetically male or female. In my thesis, especially by this approach, I could conclude that the feminizing effects of EE2 are so diverse among the examined frog and toad species." The African clawed frog (Xenopus laevis) as a model species shows high vulnerability towards the hormonally active substance EE2, an artificial Estrogen of female contraceptive pills. Credit: David Außerhofer

10.1038/srep23825

Medicine

Shape of tumor may affect whether cells can metastasize

Junmin Lee et al, Interfacial geometry dictates cancer cell tumorigenicity, Nature Materials (2016). DOI: 10.1038/nmat4610 Journal information: Nature Materials

http://dx.doi.org/10.1038/nmat4610

https://medicalxpress.com/news/2016-04-tumor-affect-cells-metastasize.html

Abstract Within the heterogeneous architecture of tumour tissue there exists an elusive population of stem-like cells that are implicated in both recurrence and metastasis 1 , 2 . Here, by using engineered extracellular matrices, we show that geometric features at the perimeter of tumour tissue will prime a population of cells with a stem-cell-like phenotype. These cells show characteristics of cancer stem cells in vitro , as well as enhanced tumorigenicity in murine models of primary tumour growth and pulmonary metastases. We also show that interfacial geometry modulates cell shape, adhesion through integrin α 5 β 1 , MAPK and STAT activity, and initiation of pluripotency signalling. Our results for several human cancer cell lines suggest that interfacial geometry triggers a general mechanism for the regulation of cancer-cell state. Similar to how a growing tumour can co-opt normal soluble signalling pathways 3 , our findings demonstrate how cancer can also exploit geometry to orchestrate oncogenesis. Main Cancer is a leading cause of death, primarily through the process of metastasis where malignant cells spread to distant organs 4 . It is believed that ‘tumour initiating cells’ or ‘cancer stem cells’ (herein referred to as CSCs) inherently possess the characteristics necessary for establishing metastases; however, within a tumour mass comprised of billions of cells, usually only a small percentage of cells exhibit a CSC phenotype 2 . This same population of cells is believed to be the root cause of recurrence after treatment, because most therapeutic regimens have not been optimized to target CSCs, and there have been multiple examples of CSCs being resistant to therapy 5 . Current evidence suggests wide-scale dynamic variation in the presence and function of CSCs across cancers and patients 6 . Deciphering the cues in the microenvironment that promote the CSC phenotype is a pressing need for understanding disease progression and developing therapeutics that can disrupt the processes involved for induction and survival of CSCs. Emerging evidence suggests that tumour cells may show ‘plasticity’ in response to microenvironmental cues. For example, melanoma cells have been shown to adopt a tumorigenic, CSC-like state and form new tumours after transplantation 7 . An exciting recent report showed how soft fibrin gels can promote selected growth of tumorigenic melanoma cells 8 , and further investigation demonstrated how the mechanical properties of the matrix can regulate Sox2 expression 9 . However, the canonical self-renewal transcription factors Oct4 and Nanog were not activated in these cells. The influence of matrix mechanics on cancer cell tumorigenicity has been demonstrated in several other cancers 10 . Taken together, these reports show that, in addition to the classical models underlying tumour heterogeneity, tumour cells may exhibit more plasticity than originally anticipated, and may be influenced through biophysical cues in the tumour microenvironment. Substrate stiffness is known to modulate cell behaviour 11 and gene expression 12 . Furthermore, the geometric organization of cells in tissue places them into variable regions of mechanical stress 13 , which can influence proliferation 14 , migration 15 , branching 16 , stem cell characteristics 17 , and cancer cell survival and invasiveness 18 , 19 . For example, Nelson and colleagues demonstrated how geometry can guide epithelial to mesenchymal transitions (EMT) through mechanical stress in micropatterned mammary epithelial cells 18 . In this letter we use soft hydrogel microengineering to pattern populations of tumour cells on two-dimensional (2D) and within three-dimensional (3D) hydrogels of variable stiffness, with combinations of perimeter geometric cues, to explore how biophysical parameters influence CSC characteristics and tumorigenicity ( Fig. 1a ). Figure 1: Interfacial geometry at perimeter features directs expression of CSC markers. a , Schematic depicting how extracellular matrix properties may guide tumorigenicity. b , Increasing micropattern size shows optimal curvature that guides expression of CSC markers in B16F0 cells ( N = 3, ∗ P < 0.05, Fisher’s exact test compared to glass); boxes represent 25th to 75th percentile and whiskers represent minimum–maximum. Horizontal lines and points within boxes represent the median and mean respectively for three replicates. c , Curvature influences expression of CSC molecular markers. Intensity of ABCB5 shown as fold change over the non-patterned (NP) condition ( N = 3). ( P -value from ANOVA analysis). R , pattern radius. d , Shapes controlling line width, curvature and perimeter to area ratio to explore the relationship of interfacial geometry and expression of CSC markers ( N = 3). e , Immunofluorescence heatmaps of B16F0 cells cultured in a panel of shapes with variable perimeter geometric features showing semiquantitative localization characteristics for CSC surface markers (ABCB5, CD271, CD133), slow-cycling related demethylase enzyme (JARID1B), intermediate filaments (Nestin) and transcription factors (Nanog, Oct4, Sox2). Far right column shows finite element models of perimeter stress in cellular sheets. Colour bar indicates minimum (bottom) to maximum (top) intensity. Scale bars, 50 μm. Error bars represent s.d. Full size image We prepared hydrazine-modified polyacrylamide (PA) hydrogels of different elasticity ( ∼ 1, 10, and 100 kPa) and used soft lithography to conjugate matrix proteins in various patterns with different sizes and shapes ( Supplementary Fig. 1 ). As a model system we selected the murine B16 melanoma cell lines and used putative CSC molecular markers CD271, CD133 and ABCB5 (refs 20 , 21 , 22 ). We first measured these markers in B16 melanoma cells cultured for 1, 3 and 5 days on circular patterns of different matrix elasticity and pattern size ( Supplementary Fig. 2 ). Expression of tumorigenic molecular markers strongly depended on culture duration (1–5 days) and colony size, with a maximum at the perimeter of circular islands ∼ 3,000–100,000 μm 2 . The stiffness of the underlying matrix did not exert a significant influence over the expression of CSC markers, thus we fixed the stiffness of our matrices at 10 kPa ( Fig. 1b and Supplementary Fig. 3 ). Analysis of cell morphology at these perimeter features reveals that, with decreasing pattern size, individual cells occupy longer arcs along the pattern perimeter, with larger subtended arc angles on average. This correlates with higher ABCB5 expressions in these cells ( Fig. 1c ). For instance, an average cell on the perimeter of a 3,000 μm 2 pattern has an edge curvature ∼ 2.2 times longer, with an angle of curvature ∼ 12.7 times larger, and shows ∼ 2.6 times higher ABCB5 expression than a cell on a 10,000 μm 2 pattern. Analysis of cell and nuclear shapes, proliferation characteristics and integrin expression shows marked differences in these parameters, which may correlate with enhanced invasiveness 23 ( Fig. 1c and Supplementary Fig. 4 ). Because cell–cell adhesion within tissue will regulate the perimeter stress, we designed straight line and torus geometries where curvature and perimeter/area can be varied. After five days of culture we see that both perimeter curvature and perimeter–area ratio (a measure proportional to interfacial energy 24 ) exert an influence on the expression of perimeter CSC markers. In all cases, convex curvature at the exterior of the torus shows higher expression of CSC markers compared to interior concave regions ( Fig. 1d and Supplementary Fig. 5 ). We designed a range of patterns comprising edges, concave and convex regions, corners of different angles and various radii of curvature, to investigate how combinations of interfacial cues at the perimeter of a population of tumour cells guide cellular organization and the expression of CSC markers ( Supplementary Fig. 6 ). Across all shapes we see higher expression of CSC markers near the periphery, with higher expression localized to convex features and corners. We note a degree of anisotropy in some of our heatmaps, which may be due to uneven initial seeding or patterning artefacts. To further verify our trends in spatial immunofluorescence, we performed segmentation analysis of CSC markers across our pattern features ( Supplementary Figs 7 and 8 ). To evaluate whether these cells show other characteristics of stem cells, we stained for molecular markers of pluripotency and tumorigenic phenotypes, including intermediate filaments (Nestin), chromatin modifying enzymes (Jarid1b) and transcription factors (Oct4, Sox2, Nanog). Strikingly, these markers co-localized with the CSC markers. We used finite element analysis to construct a simple model of the relative mechanical stress distribution of a contractile patterned monolayer, and found good correspondence between ‘hotspots’ of high CSC marker expression and regions of enhanced mechanical stress within multicellular sheets ( Fig. 1e and Supplementary Figs 9 and 10 ). Because perimeter features in cell islands, both convex and concave, give rise to cells with higher expression of CSC markers compared to cells in the interior, we designed a spiral geometry with a high interfacial boundary (perimeter/area) that exhibits an increasing radius of curvature encompassing the range depicted in Fig. 1d . Cells cultured in the spiral shape demonstrate high expression of markers associated with a CSC state ( Fig. 1e ). We selected cells cultured on this shape for flow cytometry analysis using both B16F0 and B16F10 melanoma cells cultured for five days. Similar to the immunofluorescence results, cells cultured in the spiral patterns show higher levels of stem cell and tumorigenicity markers compared to those cultured on non-patterned surfaces and those on glass ( Fig. 2a and Supplementary Fig. 11 ). Figure 2: Geometric cues activate CSCs at the perimeter through integrin α5β1, mitogen-activated protein kinase (MAPK) signalling and regulation of signal transducer and activator of transcription (STAT) pathways. a , Flow cytometry characterization of markers associated with epithelial to mesenchymal transition, CSC state and pluripotency in B16F0 cells. b , Gene expression analysis of transcripts associated with metastatic potential and MAPK/STAT signalling for cells cultured on glass (G), non-patterned hydrogel (NP) and spiral-patterned hydrogel. c , Immunofluorescence heatmaps of α 5 β 1 , Stat1 and Stat3 for B16F0 cells patterned on our panel of geometries. Colour bar indicates minimum (bottom) to maximum (top) intensity. d , CD271 expression in B16F0 cells following treatment with pharmacological inhibitors of MAPK pathways ( N = 3). e , Flow cytometry characterization of CD271, Stat1 and Stat3 positive cells with p38 inhibition. f , CD271 expression in B16F0 cells on treatment with blocking antibodies against integrin α 5 β 1 ( N = 3). g , Proposed pathway for interfacial geometry guiding tumorigenicity. Error bars represent s.d. Bold shapes depict the cell culture geometries. Scale bar, 50 μm. Full size image To gain insight into how interfacial geometry may exert an influence on the CSC state, we performed a full genome expression analysis. B16F0 and B16F10 cells were grown on spiral-patterned PA gels, non-patterned PA gels, and glass substrates for five days, followed by RNA isolation and gene expression analysis. Hierarchical clustering demonstrates segregation of B16F0 and B16F10 cells, as well as those cultured on patterned and non-patterned gels compared to glass. A panel of metastasis-related transcripts, mitogen-activated protein kinase cascades (MAPK), and signal transducers and activators of transcription (STAT) show higher levels of expression in both cell lines (B16F0 and B16F10) cultured on the patterns relative to cells cultured on non-patterned gels and glass substrates ( Fig. 2b and Supplementary Fig. 12 ). Immunostaining for integrin α 5 β 1 , Stat1 and Stat3 in patterned B16F0 and B16F10 cells shows elevated expression at the perimeter features, similar to the signature found with CSC markers and stem cell transcription factors ( Fig. 2c and Supplementary Fig. 13 ). Cells cultured on spiral shapes exhibit elevated expression of genes involved in the MAP kinase pathways linked to mechanotransduction, particularly p38 kinases and extracellular related kinases (ERK). To determine the extent to which MAPK signalling transduces signals within cells along the perimeter curvature, we supplemented our patterned culture with pharmacological inhibitors of MAPK pathways. Addition of a p38 inhibitor and an ERK 1/2 inhibitor led to a decrease in the expression of CSC markers at the perimeter, whereas addition of a JNK inhibitor resulted in more subtle, shape-dependent changes ( Fig. 2d and Supplementary Fig. 14a ). Because STAT transcriptional activity can be elevated through p38 MAPK signalling cascades 25 and has been shown to play a role in melanoma progression 26 , we also explored the ability of p38 inhibition to modulate STAT activity. Supplementing the patterned cultures with p38 inhibitor attenuated Stat1 and Stat3 perimeter localization as determined by both immunofluorescence and flow cytometry ( Fig. 2e and Supplementary Fig. 14b ). In addition, introduction of blocking antibodies against α 5 β 1 during culture leads to a partial reduction in the expression of melanoma CSC markers ( Fig. 2f and Supplementary Fig. 7c ), suggesting integrin α 5 β 1 plays a role in CSC adhesion. Taken together, we propose that interfacial geometry will modulate cell shape, and enhance α 5 β 1 adhesion, MAPK signalling and STAT activity to promote initiation of self-renewal stem cell networks ( Fig. 2g ). Recently we demonstrated how switching the biophysical microenvironment could rewire cell state using mesenchymal stem cells as a model system 27 . Using this platform we explored whether switching the microenvironments between patterned and glass substrates could rewire the tumorigenic CSC state. Cells were cultured on both types of substrate for five days, followed by transfer of spiral-patterned cells to glass and vice versa. Transfer of B16 cells from glass to patterned substrates led to increased expression of CSC markers, whereas cells transferred from patterned to glass substrates maintained some elevation of ABCB5 and CD271 after five days, suggesting the CSC state remains initially stable after removal from the patterns ( Supplementary Fig. 15 ). In our model 2D tumour microenvironments, interfacial geometry will promote signalling that establishes a tumorigenic CSC state. During tumour growth in vivo , stiffening matrices will similarly present regions of high interfacial tension at the perimeter of the growing tumour 1 . To ask whether interfaces in higher-dimensional materials—that more closely recapitulate an in vivo environment—can similarly activate a tumorigenic state, we used a templating approach to fabricate pseudo-3D microwells of PA gels ( Fig. 3a ), a 3D microfluidics polydimethylsiloxane (PDMS) device with varying geometry ( Fig. 3b ), or encapsulating groups of cells in 3D poly(ethylene glycol) (PEG) gels, all either coated or conjugated with fibronectin ( Fig. 3c ). After five days in culture, cells at the perimeter express higher levels of CSC markers in all of these experimental architectures. The consistent enhancement of CSC markers at the perimeter of our 2D and 3D tumour models gives credence to the idea that interfacial geometry may prove a general driver in coordinating cell state during oncogenesis en route to metastasis ( Fig. 3d ). Figure 3: Cells encapsulated in model 3D microenvironments demonstrate interfacial regulation of the CSC phenotype. a , Representative immunofluorescence microscopy images and immunofluorescence heatmaps of B16F0 and B16F10 cells captured within polyacrylamide (PA) pseudo-3D (2.5D) microwells with large areas (50,000 μm 2 ). Colour bar indicates minimum (bottom) to maximum (top) intensity. b , Flow cytometry characterization of CSC markers expressed at the perimeter within a PDMS microfluidic device with 3D spiral and linear channels. c , Encapsulated B16F0 and B16F10 cells in matrix metalloprotease (MMP)-degradable PEG gels showing increased localization of CSC markers at the perimeter of aggregates. Approximately 30% of cells expressing either ABCB5 or CD271 did not express both CSC markers at the same time. d , The fraction of B16F0 cells positive for CD271 in different dimensional synthetic model systems. The glass control was used to generate thresholds of the marker intensity for different substrates ( N = 3). Scale bar, 50 μm. ( ∗ P < 0.05, Fisher’s exact test). Error bars represent s.d. Bold shapes depict the cell culture geometries. Full size image To explore the metastatic potential and tumorigenicity of our engineered cells, we performed a number of in vitro and in vivo analyses. Wound healing and Boyden chamber invasion assays demonstrate enhanced migration and invasion characteristics for cells initially cultured on spiral patterns, and p38 inhibition abrogates these trends ( Fig. 4a and Supplementary Fig. 16 ). For an in vivo test of tumorigenicity, B16F0 cells were cultured for five days on spiral-patterned gels, non-patterned gels or glass substrates, followed by subcutaneous injection into 6–8-week-old C57BL/6 mice; primary tumour establishment and growth were monitored every three days with callipers. Primary tumour growth was significantly enhanced for the B16F0 cells cultured on patterned gels compared to cells cultured on non-patterned gels or glass ( Fig. 4b ). To probe the differences in growth rates for the B16 cells, we cultured cells on spiral-patterned gels, non-patterned gels or glass substrates for five days, followed by trypsinization and re-seeding on glass. Division rates were similar for both conditions ( Fig. 4c ), suggesting enhanced tumour growth in vivo for the engineered cells is due either to enhanced proliferation in vivo or on account of better survival characteristics. We performed a limited dilution study to evaluate tumorigenicity, where mice were inoculated with suspensions of 2,500, 1,000, 500 and 250 cells from our spiral-patterned gel and non-patterned gel condition. After two months we see that 4 of the 6 mice at the lowest dilution have developed tumours from spiral-patterned cells, compared to 1 of 6 in mice injected with cells from non-patterned gels ( Fig. 4d ). This result suggests cells from our patterned hydrogels exhibit enhanced tumorigenicity. With the observed difference for primary tumour growth, we sought to confirm if engineered cells would likewise possess enhanced metastatic potency. We induced experimental metastasis by tail vein injection in C57BL/6 mice of three conditions: B16F0 cells cultured on glass, B16F0 cells cultured on non-patterned gels, and B16F0 cells cultured on spiral-patterned gels. After 5 and 10 days, a cohort of mice were euthanized and histopathology performed on the lungs, with metastatic burden calculated as a normalized percentage tumour surface area. B16F0 cells cultured on spiral patterns show higher metastatic burden compared with those cultured on glass or non-patterned gels ( Fig. 4e and Supplementary Fig. 17a and b ). Correspondingly, Kaplan–Meier analysis demonstrates that mice inoculated with 3.0 × 10 4 F0 cells cultured on non-patterned gels or glass survived the longest, whereas cells grown on spiral geometries demonstrated truncated survival times ( Fig. 4f ). We measured early stage (Day 10) B16 F0 proliferation after metastasis (2.5 × 10 5 cells) and found similar proliferation, albeit slightly higher for cells inoculated from the non-patterned gel condition ( Fig. 4e and Supplementary Fig. 17c ). Considering the significantly higher tumour burden observed in lungs from mice inoculated with patterned cells, this suggests metastatic burden does not arise from increased proliferation, but rather from improved survival characteristics consistent with increased tumorigenicity. We also performed experimental metastasis to compare B16F0 cells with the highly metastatic B16F10 cells cultured on glass. Kaplan–Meier analysis shows that mice inoculated with B16F0 cells survived the longest, with comparable truncated survival times for B16F10 cells and B16F0 cells grown on spiral patterns ( Fig. 4g and Supplementary Fig. 17d ). We have shown how interfacial geometry can activate a stem-cell state in vitro ( Supplementary Fig. 8 ); however, our in vivo experiments with the spiral-patterned gels remain inconclusive as to whether curvature or the sole presence of the interface regulates cancer-cell state. Future work exploring cells patterned in other shapes that isolate positive and negative curvature may prove useful in discerning how subtle changes in perimeter geometry may guide tumorigenicity. Furthermore, it remains to be demonstrated whether curvature at the interface in a growing tumour will prime a highly metastatic CSC state. Figure 4: Activated cells show higher tumorigenicity and metastatic potency in vivo . a , Wound healing and Boyden chamber invasion assays for B16F0 (F0) cells cultured on glass, non-patterned gels, spiral-patterned gels, and on spiral-patterned gels with p38 inhibition. Scale bar, 100 μm ( N = 3, ∗ P < 0.05, # P < 0.01, ∗ ∗ P < 0.005 based on ANOVA with Tukey HSD post hoc testing). b , Tumour growth characteristics of subcutaneous implanted cells in C57BL/6 mice. Scale bar, 5 mm. c , Proliferation characteristics of patterned (P) and non-patterned (NP) cells relative to those cultured on glass ( N = 3, ∗ P < 0.05, # P < 0.01, ∗ ∗ P < 0.005 based on ANOVA with Tukey HSD post hoc testing). d , Tumorigenicity results (number of mice that develop tumours) after limited dilutions of B16F0 cells from non-patterned gels or spiral-patterned gels after 60 days implantation in C57BL/6 mice. e , Tumour surface area (at day 10) and proliferation assessment through Ki67 staining (at day 10) of excised lung tissue after experimental metastasis. ( P -value from ANOVA analysis, N is the number of lung sections used). f , Kaplan–Meier analysis of C57BL/6 mice after experimental metastasis. ( P -value from ANOVA analysis; Med., median survival time (days)). g , Histopathology of lung sections after pulmonary metastasis and immunolabelling of excised lung tissue stained for Ki67 markers after tail vein injection of B16F0 cells cultured on glass or in the spiral geometry, compared to the highly metastatic B16F10 cells (F10; positive, brown; negative, blue). ( N is the number of lung sections used, ∗ ∗ ∗ P < 0.0005 based on ANOVA with Tukey HSD post hoc testing). Error bars represent s.d.—except for e and g , where boxes represent 25th to 75th percentile and whiskers represent minimum–maximum. Scale bar, 1 mm for histopathology images and 50 μm for Ki67 staining. Full size image To ascertain whether the influence of geometry may prove to serve as a universal tumorigenicity guidance cue, we immunostained several other cancer types for the generation of heatmaps. Similar to murine B16F0 and B16F10 melanoma cells, several human cancers; human HeLa (cervical), A549 (lung), and PC3 (prostate) cell lines, all showed similar trends in CSC marker expression on 10 kPa gels ( Supplementary Fig. 18 ). These findings suggest that interfacial geometry may prove to be a general biophysical phenomena underlying cancer cell progression within a microenvironment. Our results demonstrate how the interfacial constraints imposed by perimeter geometric features in a population of tumour cells can guide cancer cells towards a stem-cell-like state. In vivo , the balance between intercellular adhesion and cortical tension act together to specify tissue surface tension 24 , which regulates the local behaviour of groups of cells 28 , 29 . Similarly, the state of a tumour cell in a multicellular aggregate may depend on the interplay between force balance, cellular tension, intercellular adhesion and relative position with respect to other cells 30 . In line with this hypothesis, we find that perimeter curvature can coordinate the spatial arrangement of cells by modifying interfacial energy, cortical tension and intercellular adhesion. We show that this coordination can foster a unique microenvironment where integrin-mediated adhesion and mechanotransduction activate a CSC phenotype. Our findings are in contrast to previous studies where ‘stemness’ is promoted in regions of low mechanical tension 9 , 31 , which suggests interfacial geometry may play a unique role in cancer through activation of a stem-like cell with a role in metastasis. This finding may help guide clinical analysis of the tumour microenvironment during biopsy or resection, and may lead to advances in the design, development and translation of patient-specific models of oncogenesis for personalized therapeutic development. Methods 2D and 3D surface preparation. For 2D polyacrylamide (PA) hydrogels, we used the previously reported protocol 32 . Briefly, PA gels with desired stiffness (1, 10 and 100 kPa) were fabricated on a glass coverslip (18 mm, Fisher Scientific) by mixing varying amounts of acrylamide and bis-acrylamide (Sigma). 0.1% ammonium persulfate (APS, Sigma) and 0.1% tetramethylethylenediamine (TEMED, Sigma) were added to initiate the reaction. Activated coverslips were functionalized with 3-aminopropyltriethoxysilane for 3 min and glutaraldehyde for 30 min (Sigma) and then 20 μl of gel solution was sandwiched between the activated coverslips and a hydrophobically treated glass slide. The gel-coated coverslips were gently detached after an appropriate polymerization time for each stiffness condition. The gel surfaces were treated with hydrazine hydrate (55%, 2 h) to modify the surface chemistry and then rinsed with 5% glacial acetic acid (1 h) followed by distilled water (1 h; ref. 33 ). Patterned or non-patterned polydimethylsiloxane (PDMS, Polysciences) stamps of desired shapes were produced by using photolithography. To create patterned stamps, PDMS was polymerized on a patterned master of photoresist (SU-8, MicroChem) created using ultraviolet photolithography through a laser printed mask. Free aldehydes were generated on fibronectin (25 μg ml −1 , Sigma) in PBS by adding sodium periodate ( ∼ 3.5 mg ml −1 , Sigma) for at least 45 min. The protein solution was pooled onto patterned or non-patterned (flat) stamps for 30 min and dried with air. The protein residues on stamps were transferred to the gel surface through covalent conjugation of free aldehyde groups in proteins to reactive hydrazide groups on the gels. For pseudo-3D microwells, an SU-8 photolithography master exhibiting the inverse features used in fabricating the PDMS stamps was used as a template to cast PA gels with microwells on the surface. The gels were chemically modified with hydrazine hydrate and the oxidized protein solution was applied. To render the external surface non-adhesive, the top layer of protein-conjugated gels was sheared off by applying an adhesive strip to the surface. In all of our experiments we ensured cells formed only a single monolayer, to ensure uniform antibody staining. For 3D poly(ethylene glycol) (PEG) gels, 10,000 MW PEG (Sigma) was modified to poly(ethylene glycol) diacrylate (PEGDA), as previously reported, by the addition of acryloyl chloride 34 . Fibronectin was acrylated by the addition of NHS-Acrylate (Sigma) under basic pH for 4 h. Matrix metalloprotease (MMP) cleavable peptides were synthesized using solid phase peptide synthesis and reacted with PEGDA via Michael addition of terminal cysteine residues. 3-(trimethoxysilyl)propyl methacrylate 2% solution in ethanol with 0.3% glacial acetic acid was applied to coverslips for 10 min for activation, followed by baking at 95C for 1 h 35 . To encapsulate cells in the degradable hydrogels, 30% (w/v) PEGDA–MMP was mixed with ultraviolet-initiator (0.05% Irgacure 2959, Sigma) and cells were centrifuged and re-suspended in this mixture. This solution was sandwiched between the activated coverslip and a hydrophobic coverslip. Ultraviolet light (5 mW cm −2 ) was applied for 10 min and the encapsulated cells were detached and placed in cell culture media. For 3D microfluidics PDMS devices, flexible rubber-coated wire (200 mm long, 2 mm diameter) was employed to design microfluidic devices with different shapes. The designed wire (line or spiral) was placed on the first layer of PDMS (flat) and the second layer of PDMS was fabricated with the wire inside. After the wire was removed from the PDMS, 0.2 mg ml −1 sulfosuccinimidyl 6 (4′-azido-2′-nitrophenyl-amino) hexanoate (Sulfo-SANPAH; Pierce), a heterobifunctional protein crosslinker, was used to covalently bind proteins to the PDMS channel inside; exposure of the PDMS in a solution of Sulfo-SANPAH with an ultraviolet light source at 365 or 320 nm covalently linked the sulfo-SANPAH to the PDMS (1 h). Sulfo-SANPAH solution was removed from the PDMS channel and the device was washed by gently adding and aspirating PBS until the PDMS channel was transparent again. Fibronectin (25 μg ml −1 ) was conjugated on the surface-modified PDMS inside the channel for 24 h. Cell source and culture. The cancer cell lines B16F0 and B16F10 (murine melanoma), and PC3 (human prostate) were obtained from American Type Culture Collection (ATCC) and cultured according to the recommended protocols. HeLa (human cervical, ATCC) cells were a gift from Andrew M. Smith’s laboratory; A549 (human lung, ATCC) cells were a kind gift from Jianjun Cheng’s laboratory. For cell culture, media was changed every three days and cells were passaged at nearly 90% confluence using 0.25% trypsin (Gibco). B16 cells were tested for mycoplasma contamination at Charles River Laboratories for cell line testing prior to in vivo inoculation. Immunofluorescence. Cells were fixed with 4% paraformaldehyde (Sigma) for 20 min and permeabilized in 0.1% Triton X-100 (Fisher) in PBS for 30 min. 1% Bovine serum albumin (BSA) was employed to block the cells for 15 min and primary antibody labelling was performed in 1% BSA in PBS for 2 h at room temperature, followed by rinsing with PBS two times. Secondary antibody labelling was performed in 2% goat serum, 1% BSA in PBS for 20 min in a humid chamber (5% CO 2 and 37 °C). Immunofluorescence microscopy was conducted using a Zeiss Axiovert 200 M inverted research-grade microscope (Carl Zeiss) or an LSM 700 (Carl Zeiss), which is a four laser, point scanning confocal with a single pinhole. Antibodies used for immunostaining, flow cytometry, and immunohistochemistry are listed in Supplementary Table 1 . Inhibition assays. MAP kinase inhibitions (FR180204 (ERK1/2), SP600125 (JNK), and SB202190 (p38)) (Calbiochem) were performed by adding media supplemented with these inhibitors at 6 μM concentration after cell seeding and with each media change. Integrin-blocking antibodies (α 5 β 1 ) were added to cells in media before seeding at 1 μg ml −1 . Wound healing assay. B16F0 and B16F10 cells were cultured for five days on spiral patterns (with or without p38 inhibitors), non-patterned gels, or glass substrates (12 identical substrates). Cells were trypsinized and re-plated on glass substrates (10 6 cells per glass) and then cultured under permissive condition to about 90% confluence. A pipet tip was employed to create a linear scratch in the confluent monolayer. Cells were allowed to migrate and close the wound for 12 h, and were observed under phase contrast microscopy. The scratch width per field of view, between the time points 0 and 12 h following wound closure, was determined using imageJ 36 , and the average percentage of wound closure, indicated by the shifted width after 12 h, was assessed. Boyden chamber assay. Invasion of B16F0 and B16F10 cells was examined using 24-well Boyden chambers (Corning) with inserts (8 μm pores) and precoated basement membrane extract (Matrigel, growth factor reduced) (BD Biosciences). Cells were cultured for five days on spiral patterns (with or without p38 inhibitors), non-patterned gels or glass substrates (12 identical substrates) and then placed on the inserts in the upper chambers (of each well) and cultured for 12 h. Cells on the upper surface of the membrane filter were removed. Cells that crossed the inserts to the lower surface were fixed with 4% paraformaldehyde and then stained with 4′,6-diamidino-2-phenylindole (DAPI). Cells per field of view were imaged under a ×10 fluorescence microscope and counted. Cell labelling and flow cytometry. B16F0 and B16F10 cells cultured for five days on spiral patterns or glass substrates (12 identical substrates) were trypsinized and broken down into a single cell suspension. Cells were fixed in 4% paraformaldehyde for 20 min and permeabilized in 0.1% Triton X-100 in PBS for 15 min. Cells were blocked in 1% BSA for 1 h. Cells were stained with primary antibodies in 1% BSA in PBS for 2 h at room temperature. Next, secondary antibodies in 2% goat serum, 1% BSA in PBS were applied for 20 min in a humid chamber (5% CO 2 and 37 °C). Before every step, cells were washed three times with PBS. Flow cytometry analysis was performed with a BD LSR Fortessa Flow Cytometry Analyzer. Cells stained without primary antibodies were used as negative controls to set the baseline. Cell proliferation assay in vitro . BrdU staining was conducted as reported previously 37 . Briefly, BrdU labelling reagent (Invitrogen) was added (1:100 v/v) before 24 h of fixing; the reagent was added after seeding, day 2, or day 4 for fixing at days 1, 3, or 5, respectively. Cells were fixed in 70% ethanol for 30 min and then denatured with 2 M HCl for 30 min. Cells were permeabilized with 0.1% Triton X-100 in PBS for 30 min, blocked with 1% BSA in PBS for 15 min, and then incubated with mouse anti-BrdU primary antibody (3 h at room temperature) followed by Alexa Fluor 647-conjugated antimouse IgG antibody (20 min in a humid chamber). Cell nuclei were stained with DAPI. For the division rate assay, B16F0 and B16F10 cells cultured for five days on spiral patterns or glass substrates (six identical substrates) were trypsinized and placed on glass. Ethics statement. All experiments using live animals were in compliance with animal welfare ethical regulations and approved by the Institute Animal Care and Use Committee before experimentation. B16 melanoma in vivo models. Six–eight-week-old female C57BL/6 mice were purchased from Charles River Laboratories for Animal Experiment. Primary localized tumours were established by subcutaneously injecting B16F0 cells (range 250–10 6 ) grown on patterned gels, non-patterned gels (NP) or glass into the right lateral flank (the information of the number of mice in each experiment is described in each figure). Macroscopic tumour growth was serially measured (maximal length and width) with callipers three times a week. Tumour growth was checked every three days and experiments were stopped when the first mouse of the respective series had a tumour exceeding 2,000 mm 3 . The volume of tumour was calculated using V = ( L × W 2 )/2, where L is length and W is width. Experimental metastases were established by injecting 2.5 × 10 5 (B16F0 grown on pattern/glass or B16F10 grown on glass) or 3.0 × 10 4 (B16F0 grown on pattern/NP/glass) melanoma cells via lateral tail vein injection. Mice were either euthanized 5, 10, and 16 days (2.5 × 10 5 cells injected) post injection and used to quantify percentage tumour surface area within the lung parenchyma, or followed for survival analysis. Mice were used for determination of primary tumorigenesis and experimental metastases. No animals or target organ samples (lung tissue) were excluded from analysis. Criteria used for primary tumorigenesis was the formation of subcutaneous tumours which were detectable by visual examination and measurable with callipers. For comparison of primary tumour formation kinetics, mice were evaluated daily until primary tumours exceeded 20 mm in diameter, then humanely euthanized. In some experiments evaluating primary tumorigenesis, study endpoints were dichotomous in nature, being either tumour formation or no tumour formation after a cumulative lapse of time (60 days). For experimental metastases, the primary endpoint was survival time, and mice were monitored daily until reaching criteria for humane euthanasia. Mice were used for determination of primary tumorigenesis and experimental metastases. Inoculation of mice with melanoma cells grown on different conditions (glass, non-pattern, and pattern) and different cell densities was not performed in a random fashion. Rather, cohorts of mice were predetermined to receive injections of melanoma cells grown under specified conditions and cell densities before inoculation. No blinding was done for these animal studies. Ki67 immunohistochemistry. Five representative lung sections fixed in 10% formalin per mouse were immunohistochemically stained for Ki67. Within each lung section, three randomly selected parenchymal areas completely effaced by melanoma cells were microscopically quantified for Ki67 nuclear positivity and expressed as a percentage using ImageJ software. Microscopy data analysis. Immunofluorescence images were analysed using ImageJ software. Multiple cells (over 20 patterns) were imaged for each condition and fluorescence intensities of single cells in patterns (after background subtraction) were used to compare marker expression. For cell curvature analysis, the number of cells in circular patterns (over 20 patterns) with different areas (3,000–100,000 μm 2 ) were counted, and cell curvature length was calculated based on the length of pattern perimeter and the number of cells at the perimeter. Average curvature angle and intensity of cells at the perimeter of the patterns were measured and plotted. For inhibition studies, positive cells which were above the maximum intensities of the glass control (ImageJ threshold) were counted, and the numbers were divided by total cells in patterns. For generating immunofluorescence heatmaps, cells cultured on various shapes were imaged on the same day using the same microscope and camera settings. Background intensities of raw fluorescent images were subtracted, and patterns were aligned in ImageJ with the same orientation as cultured across the surface, incorporated into a Z stack and the average intensity calculated for heatmap generation. For segmentation analysis, cells cultured on each shape in a single monolayer were manually segmented for at least 100 single cells through immunostaining using ImageJ. Because cells predominantly express surface markers at the surface and not within nuclei and junctions, it is possible to segment single cells at the perimeter (line, convex, or concave): we used ×20 immunofluorescence images in ImageJ; the contrast and brightness were controlled to optimize the image for segmentation analysis; the surface region of each single cell at the perimeter was selected, excluding nuclei; the original image was re-opened, and the marker intensity of segmented single cells was measured using ImageJ; the background intensity was subtracted from the measured intensity values. RNA isolation for microarray experiments. Adherent B16F0 and B16F10 cells cultured for five days on spiral patterns, non-patterned gels, or glass were lysed directly in Trizol reagent (Invitrogen) according to the vendor’s instructions. Total RNA from each sample (12 duplicates) was extracted and quantified by photospectrometry using a NanoDrop ND-1000 (ThermoFisher). RNA quality was confirmed by an Agilent Bioanalyzer, and gene expression profiling performed using Illumina iScan Sentrix BeachChip technology at the Roy J. Carver Biotechnology Center at the University of Illinois at Urbana-Champaign using standard Illumina protocols ( ). Illumina gene array data was pre-processed using GenePattern 38 . The background values were subtracted and thresholded. The data was then normalized using the quantile method. Heatmaps of fold changes over glass in gene expression were visualized using the Gene-E ( ) software package. A panel of metastasis genes was selected from a previous report by Clark and colleagues 39 . For finding relevant pathways, genes upregulated in patterns compared to glass were tested in the Database for Annotation, Visualization and Integrated Discovery (DAVID) website ( ) and genes in each pathway were selected based on DAVID, and genes with negligible expression (below ten) were not included in the analysis. Percentage tumour surface area. Five representative lung sections fixed in 10% formalin per mouse were microscopically examined at two different tissue planes separated by 50 μm. Subgross images (×1.25), including one image containing an embedded micrometer, were captured for each lung section at both tissue planes using standard microscopy imaging equipment. Images were imported into Adobe Photoshop Creative Cloud 2014 and the embedded micrometer was used to set a measurement scale of image pixels to length in mm (1,503 pixels = 5.0 mm). The parenchymal surface area of each lung lobe was subsequently measured using the Quick Selection Tool. Regions of B16 melanoma growth were then identified visually and cross-referenced with the histologic slide if necessary, surface area measured using the Magic Wand Tool, summed, and then expressed as a percentage relative to total parenchymal surface area using ImageJ software. Modelling of cell monolayer. Abaqus software was used to construct and analyse a finite element model of contractile cell monolayers as described previously 14 . Briefly, the desired geometry was modelled in two layers: an active 20-μm-thick top layer and a passive 5-μm-thick bottom layer that is constrained at the bottom surface. The active layer is made to contract isotropically by applying a 5 K temperature drop. The von Mises stress at the bottom surface is reported. We confirmed convergence by testing multiple mesh sizes andlayer properties. Statistical analysis. Data was obtained from three independent experiments and expressed as the mean ± standard deviation (s.d.) unless otherwise specified. Statistical comparisons between two groups were based on Student’s t -test and comparisons of more than two groups were performed by analysis of variance (ANOVA) with Tukey HSD Post-hoc testing to correct for multiple comparisons. Differences were considered significant at P < 0.05.

Only a few cells in a cancerous tumor are able to break away and spread to other parts of the body, but the curve along the edge of the tumor may play a large role in activating these tumor-seeding cells, according to a new University of Illinois study. Using engineered tissue environments in various shapes and patterns, the study of skin cancer found that the more curved the cell cultures were, the more cancer cells at the edges displayed markers of stem cell characteristics - the key to spreading to other tissues. This has potential for furthering our understanding of cancer as well as developing personalized treatment plans. Led by Kristopher Kilian, a professor of materials science and engineering, and Timothy Fan, a professor of veterinary medicine, the researchers published their findings in the journal Nature Materials. "The most dangerous part of cancer is metastasis," Kilian said. "Some cells that we call cancer stem cells adopt deadly characteristics where they can travel through the bloodstream to other tissue and form new tumors. There's a need for ways to find these cells and to study them, and importantly, to develop drugs that target them, because these cancer stem cells are resistant to chemotherapy drugs that target the main tumor. This causes recurrence: The cancer comes back." Kilian's group specializes in tissue engineering to create models of tumors, in order to more accurately study cancer processes in a culture dish. In the new study, the researchers cultured mouse skin-cancer colonies on various 2-D and 3-D environments of different shapes and patterns to see if the tumor shape contributes to activation of cancer stem cells, and to see where in the tumor the stem cells appeared. They found that cancer stem cells seemed to appear in the highest numbers along the edges of the engineered tumor environments, particularly where there were corners and convex curves. "It was actually quite surprising," Kilian said. "Normal stem cells prefer a soft, squishy, internal position. So for cancer, everyone had assumed that the cancer stem cells were in the middle of the tumor. We found that geometric constraints, like you would have where a tumor touches healthy tissue, seem to activate these cancer stem cells at the perimeter." The researchers did a number of tests in their engineered environments to confirm tumor-spreading ability, such as genetic analysis. They also tested other cancer lines - human cervical, lung and prostate cancers - and found that they responded to the patterned tumor environments in the same way. Then Kilian's group teamed with Fan's group to test the skin-cancer stem cells in live mice, and found that the cells taken from the patterned environments were much more likely to cause tumors than cells taken from a conventional flat dish. "We found that many more mice developed tumors when given the cells that we had engineered to have these stem cell characteristics, and they had a much higher incidence of metastasis in the lungs," Kilian said. "In a tumor, similarly, regions that develop these kinds of shapes may activate cells that can then escape and form more tumors. This may allow surgeons to look at the perimeter of a growing tumor and use the shape to guide their assessment of which regions could be more problematic - where they need to take out more tissue around the tumor and where they may not need to take as much." Kilian hopes that the patterned, engineered tissue environments will give researchers a new way to find and culture cancer stem cells, which have been very elusive in conventional cultures - less than 1 percent of cells, he said. Beyond the fundamental science of finding and understanding these cancer-spreading cells, he also sees engineered tumor environments as having therapeutic applications in personalized medicine. "You can imagine a patient has a particular tumor. You could engineer that in a dish, and using the patient's own cells, you could develop a model of their specific tumor to test out drugs," he said. "If you could take a patient's cells and within days have microtumors that you could use to screen all the available drugs, then an oncologist would be able to prescribe a treatment that's tailor-made for the patient that targets both the tumor cells and these elusive cancer stem cells that currently we can't see. "There's a lot more work to be done, but we're very excited about how a very simple materials property of a growing tumor might be a culprit of the disease spreading. We think it opens up a new avenue of investigation for drug development, guiding surgery, and understanding progression and spreading of cancer," Kilian said. "Cancer is very complex, so putting it in context is key. If there is a microenvironment that provides the context for activating cells that can spread cancer, then that's important to know."

10.1038/nmat4610

Physics

Metal halide perovskites for next-generation optoelectronics: Progress and prospects

He Dong et al, Metal Halide Perovskite for next-generation optoelectronics: progresses and prospects, eLight (2023). DOI: 10.1186/s43593-022-00033-z

https://dx.doi.org/10.1186/s43593-022-00033-z

https://phys.org/news/2023-01-metal-halide-perovskites-next-generation-optoelectronics.html

Abstract Metal halide perovskites (MHPs), emerging as innovative and promising semiconductor materials with prominent optoelectronic properties, has been pioneering a new era of light management (ranging from emission, absorption, modulation, to transmission) for next-generation optoelectronic technology. Notably, the exploration of fundamental characteristics of MHPs and their devices is the main research theme during the past decade, while in the next decade, it will be primarily critical to promote their implantation in the next-generation optoelectronics. In this review, we first retrospect the historical research milestones of MHPs and their optoelectronic devices. Thereafter, we introduce the origin of the unique optoelectronic features of MHPs, based on which we highlight the tunability of these features via regulating the phase, dimensionality, composition, and geometry of MHPs. Then, we show that owing to the convenient property control of MHPs, various optoelectronic devices with target performance can be designed. At last, we emphasize on the revolutionary applications of MHPs-based devices on the existing optoelectronic systems. This review demonstrates the key role of MHPs played in the development of modern optoelectronics, which is expected to inspire the novel research directions of MHPs and promote the widespread applications of MHPs in the next-generation optoelectronics. Peer Review reports 1 Introduction Metal halide perovskites (MHPs) have been emerging as the rising star in the field of optoelectronics during the past decade, and the state-of-the-art optoelectronic technologies based on MHPs, such as perovskite solar cells (PSCs), light emitting diodes (LEDs), photodetectors (PDs) and lasers, have been leading the prevailing paradigm owing to the intriguing optoelectronic properties of MHPs. MHPs family, taking the general formula of ABX 3 (A = CH 3 NH 3 + (MA), HC(NH 2 ) 2 + (FA), and Cs + , B = Pb 2+ , Sn 2+ , X = I − , Cl − , Br − ), possesses the merits of facile and low-cost processing and the favorable attributes of tunable optical and electronic features, providing a rich and fertile ground for the development of high-performing multifunctional optoelectronic devices and their future industrialization. The first study of optoelectronics of MHPs could be traced back to 1950s, the intensely colored crystals of CsPbX 3 (X = Cl, Br) with perovskite structure exhibits the behavior of frequency-dependent photoconductive response, implying the special electrical properties of perovskites [ 1 ]. Shortly thereafter in early 1990s, Mitzi and co-workers focused particular interest on the layered tin halide perovskites that are naturally self-assembled into quantum confinement structures (QWS) in the presence of alkylammonium cations, exhibiting obvious excitonic features and demonstrating potential application in LEDs and transistors [ 2 ], crucially with the interesting electronic properties and ability by tuning the MHP structures flexibly [ 3 , 4 , 5 ]. In the meanwhile, the first observation of light emission and lasing in (C 6 H 13 NH 3 ) 2 PbI 4 was reported in 1998 [ 6 ]. Subsequently, as the birth of first perovskite-sensitized solar cell with a power conversion efficiency (PCE) of 3.8% in 2009 [ 7 ], “perovskite fever” with a focus on solar cells application makes several milestones in PCE improvement, which has exceeded 20% since then [ 8 ]. This breakthrough process triggers the researchers’ curiosity to explore the intrinsic optoelectronic properties of MHPs, where the adjustable band gap, long carrier diffusion length, strong light absorption, low defect density and solution processability are identified [ 9 ]. In the mild-2010s, MHPs are shown to be promising for optical sources with the high optical gain and stable amplified spontaneous emission (ASE) with a threshold of 10 ± 2 μJ cm −2 at room temperature [ 10 ], and the first laser work in the field without external cavity [ 11 ] as well as the first optically pumped perovskite microcavity laser were reported shortly after [ 12 ]. In 2015, the first ASE and lasing in MHP nanocrystals of CsPbX 3 was demonstrated with low pump thresholds down to 5 ± 2 μJ cm −2 [ 13 ]. Almost the same time, the first room temperature perovskite LEDs (external quantum efficiency (EQE) ~ 0.76%) were reported by Friend and co-workers [ 14 ]. Meanwhile, the first MAPbI 3 PDs with X-ray sensitivity of 25 μC mGy air −1 cm −3 was investigated by Heiss and co-workers [ 15 ], Park and co-workers then achieved highly sensitive X-ray PDs up to 11 μC mGy air −1 cm −2 based on printable MAPbI 3 films in 2017 [ 16 ]. The next evolution of MHPs in optoelectronics comes from the revival of layered perovskites and perovskite nanocrystals (NCs) in late 2010s, and they have shown competitive photoluminescence (PL) and electroluminescence (EL) efficiency compared with bulk perovskites [ 17 ], pushing the related studies, especially of the relationships between dimensionality, size, geometry, composition, colloidal synthesis approaches and their optoelectronic properties to a new upsurge, in which the EQE of green, red, and near-infrared LEDs can be as high as 20.3%, 21.3% and 21.6% [ 18 , 19 , 20 ], respectively. Also, new breakthrough in CsPbBr 3 NCs X-ray detectors has been made with a low dose rate of 13 nGy s −1 [ 21 ]. Propelling current interest in the 2020s is that MHPs continue to exhibit new and surprising optoelectronic properties, the certified PCE for single-junction PSCs, all-perovskite tandem solar cells and hybrid tandem PSCs has been up to 25.7%, 28% and 31.3%, respectively [ 22 ], and new applications of this materials expand not only to PSCs, LEDs and PDs, but also to facilitate the integration of optoelectronic devices. Meanwhile, the challenges such as instability bottleneck has been gradually resolved. One can say that the optoelectronic properties of these materials create a host of new avenues for the scientific community to explore. Figure 1 demonstrates the representative progresses of MHPs for optoelectronics. Inspiringly, solid researches have been carried out during the course of development to disclose the unique optoelectronic features of MHPs, and their multidiscipline applications have given new momentum to the field of optoelectronics, enabling the rapid revolutionary step towards their practical use for human civilization. Fig.1 Representative progresses of MHPs for optoelectronics Full size image In this review, we dedicate to giving a panorama picture on the optoelectronic traits of MHPs and their revolutionary impact on the next-generation optoelectronics. We first introduce the structure of MHPs and the origin of their optoelectronic properties, and briefly discuss the relationship between structure and optoelectronic properties of MHPs. Then, we summarize the merits of MHPs by focusing on the fine controlling of the optoelectronic properties by regulating the structure of MHPs, including phase, dimensionality, composition, and geometry. Thereafter, we highlight the revolutionary applications of MHPs in the technologies that are related to important area of human society, which includes functional integration system, information display system, electronic communication system, and health & medical system. This perspective aims at providing critical guidance for inspiring the novel research directions of MHPs to promote the widespread application of MHPs in optoelectronics. 2 Origin of optoelectronic properties of MHPs In a typical ABX 3 perovskite, B occupies the center of an octahedral [BX 6 ] 4− cluster, which forms corner-shared octahedral BX 6 framework, while A is 12-fold cuboctahedral coordinated with X anions and occupies the interspace of octahedral BX 6 framework, as shown in Fig. 2 . For [PbI 6 ] 4− in particular, the valence band maximum (VBM) is formed by antibonding orbitals originating from Pb(6 s ) and I(5 p ) atomic states, while antibonding Pb(6 p ) and I(5 s ) orbitals form the conduction band minimum (CBM). Of critical importance, intrinsic electronic configurations of MHPs are the direct origin of their unique optoelectronic properties, including high optical absorption, high carrier mobility, high defect tolerances, long diffusion lengths, unique ambipolar charge transport, and flexible turnability. Fig. 2 Schematic representation of the halide perovskite crystal structure (left) and molecular orbitals for MAPbI3 [ 23 ] Full size image (i) High optical absorption. The high symmetry of MAPbI 3 crystal structure leads to direct band gap while the lone pair electrons on Pb s orbital results in p-p electronic transitions from VB to CB, contributing to exceptionally high optical absorption of the material [ 24 ]. (ii) High carrier mobility. The inorganic Pb-I framework in MHPs mainly contributes to the valence and conduction band-edge state, and both VB and CB of MHPs exhibit antibonding character, and the anti-bonding interaction results in a rather small effective electron mass, enabling the high charge carrier mobility of MHP with the values of tens of cm 2 /(Vs) [ 25 , 26 ]. (iii) High defect tolerance. On the one hand, orbital expands VB bandwidth and raises VB edge, suggesting that most defect states are located within or closer to VB edge. On the other hand, coupling between Pb 6p orbital and I 5p orbital is very weak, and thus CB edge exhibits an ionic character, suggesting that the energy states are not much affected by external defect levels. Therefore, MHPs show high defect-tolerance [ 27 ]. (iv) Long diffusion lengths. Long diffusion lengths in MHPs is ascribed from the high carrier mobility and high defect tolerance, and thus the ideal perovskite crystal structure shows long carrier diffusion length greater than 10 μm [ 28 ]. (v) Ambipolar charge transport. Because the electron effective mass of MHPs (0.23 m 0 ) is considerably balance with hole effective mass (0.29 m 0 ), MHPs exhibit unique ambipolar charge transport property, which play an important role in photovoltaic application [ 29 ]. (vi) Flexible tunability. Because the band structure is determined by the B and X elements, the substitution of B and/or X gives rise to the feasible tunability of VBM and CBM position as well as the band gap of MHPs. Also, deriving from the doubly bridging halide ion (B-X-B moiety), MHPs possess the unique property of successive phased transition, which can undergo a series of structural changes via changing temperature, pressure, and/or chemistry to force the B-X-B angle deviating from ideal crystal structure [ 23 , 30 ]. In sum, those unique characteristics of MHPs are derived from their crystal structure and chemical composition, which allows the unprecedented flexibility to independently and synergistically tune the optoelectronic properties of MHPs, laying an important foundation for MHPs being applicable to various optoelectronic applications [ 31 ]. In the next section, we will introduce the current understanding on the tunable optoelectronic properties of MHPs. 3 Tunable optoelectronic properties of MHPs 3.1 Phase transition The flexible nature of the BX 6 framework in ABX 3 MHPs is mainly originating from the largest metal-halide-metal (B-X-B) bond angle, providing the perovskites with another interesting property of successive phase transition, which largely influences the electronic structure and thus the resulting optoelectronic properties of MHPs (Tables 1 , 2 ) [ 23 ]. For example, in the prototypical material APbI 3 , the cation series MA + , FA + and Cs + leads to progressively larger octahedral tilting in the perovskite phase at room temperature (Φ change in Fig. 3 a) as a result of decreased cation size [ 30 ], leading to an increase in band gap from 1.48 eV in α-FAPbI 3 to 1.67 eV in γ-CsPbI 3 . Generally, a series of phase transition of perovskites can be modulated by temperature, pressure and/or chemistry. The structural phase transition of cesium halide perovskites has proven to be controlled by temperature, often between a room temperature non-perovskite phase and a high-temperature perovskite phase (Fig. 3 b). These two phases feature distinct optoelectronic properties, such as bandgap, photoluminescence quantum efficiency, charge carrier mobility, and charge carrier lifetime [ 32 ]. In addition, the phase transition could also be triggered by high-pressure, and studies have reported that under a few GPa pressure, two-dimensional perovskites (2DPKs) show optical bandgap modulations and enhanced PL effects due to the phase transition. Interestingly, the structure of 2DPK crystals can be stabilized in a metastable state or a different phase at ambient pressure after a cycle of compression/decompress and amorphization (above 10 GPa) (Fig. 3 c), resulting in a series of the strong excitonic emissions in the PL spectra [ 33 ]. Furthermore, through chemical route of acid–base interactions, reversible structural and compositional changes of CsPbX 3 nanocrystals can be induced at room temperature between cubic and orthorhombic CsPbX 3 (Fig. 3 d), which is owing to the controlled self-assembly of small CsPbX 3 nanocrystals via the control of the ligand shell environment [ 34 ]. These results highlight the potential application of doped halide perovskites in switchable optoelectronics, such as the smart windows, building-integrated photovoltaics (BIPV) and so on. Table 1 Structure phase parameters for typical MHPs Full size table Table 2 Comparison of the key parameters for photodetectors Full size table Fig. 3 a The B-X-B angle of the APbI3 series with different cations at room temperature [ 30 ]. b Illustration of the temperature-induced phase transition of CsPbI3 crystal [ 32 ]. c Illustration of the pressure-induced phase transition of (C4H9NH3)2PbI4 single crystal [ 33 ]. d Illustration of the chemistry-induced phase transition of CsPbBr3 crystal [ 34 ] Full size image 3.2 Dimensionality Relying on the variety structures and compositions of MHPs, especially in terms of the choice of R cation, the dimensionality of MHPs can be finely tailored from 3D to low-dimensional ones extending from quasi-two- (quasi-2D), two- (2D), one- (1D) to zero- (0D) dimension. Normally, the term “low-dimension” of MHPs is often classified into two types: “structure-level” and “material-level”. The “structure-level” low-dimension underlines the morphologies of MHPs, and commonly refers to those nanostructures, such as nanosheet, nanowire, and nanocrystal (Fig. 3 a(i)). By contrast, the “material-level” low-dimension emphasizes the intrinsic crystal structure of MHPs, in which the [BX 6 ] 4− octahedra are separated by the bulky dielectric spacer molecules to form bulk assembly of atom-level 0D clusters, 1D quantum wires, or 2D quantum-wells (QWs) (Fig. 4 a(ii)). Compared to their 3D counterparts, low-dimensional MHPs possess unique band structures, because the dimensionality and size of the nanostructure have direct consequences on the electronic band gap and joint density of states (DOS) close to the band gap, which is of crucial importance to determine their optoelectronic properties (Fig. 4 a(iii)) [ 64 ]. Generally, lower dimension MHPs have narrower absorption range owing to their larger bandgaps, which vary with the number of layers (Fig. 4 b) [ 65 ]. When considering a family of low dimensional MHPs structures composed of the same metal halide octahedral, the lowest energy electronic transition follows the monotonic trend of E g, 3D < E g, 2D < E g, 1D < HOMO–LUMO , 0D . Most importantly, the inclusion of dielectric and quantum confinement effects in low-dimensional MHPs can further increase the discrepancy between the optoelectronic characteristics of 3D and LD structures (Fig. 4 c) [ 66 ]. The exciton binding energy progressively decreases with the increased n , and when n ≤ 2, the perovskites exhibit strong excitonic behavior with high photoluminescence yield (PLQY) at room temperature, and this intensified oscillator strength and optical nonlinearities of low-dimensional MHPs make them suitable for light-emitting applications. While when n ≥ 3, the perovskites have a low exciton binding energy and smaller bandgap, which extends the light absorption range in the visible region, suggesting their suitability of sunlight-to-energy conversion (Fig. 4 d) [ 67 , 68 ]. Fig. 4 a Schematic representation of (i) “structure-level” and (ii) “material-level” 3D bulk, 2D, 1D, and 0D perovskite materials, and (iii) is the corresponding density of states versus the energy Eg [ 64 ]. b Electric band structure of low-dimensional MHPs with different n values [ 65 ]. c Illustration of quantum- and dielectric-confinement effects in low-dimensional MHPs [ 66 ]. d Exciton binding energy and PL emission wavelength as a function of number of n layers in low-dimensional MHPs [ 67 , 68 ] Full size image 3.3 Composition Benefiting from the ionic nature, MHPs family perform flexible chemical management via substitution of A, B and X site, which enables tunable optoelectronic properties, such as band gap, PL diversity, carrier transport, and so on. For example, it is obvious that FAPbI 3 shows the tunable band gap when using substitution of the A-site cation with Cs and B-site with Sn (Fig. 5 a–c). The tunable optical band gaps achieved by compositional change provides a general framework of design rules for desired band gap and band positions of particular MHP [ 69 ]. Particularly, Originating from the highly soluble precursors could lead to anion exchange in a few seconds without deteriorating the initial structure or forming any remarkable lattice/surface defects, anion doping is widely used to tune the chemical composition and optical properties of more stable CsPbX 3 nanostructures [ 70 ], which has enabled MHPs to be suitable for multiple applications, such as light-emitting diodes, wavelength-tunable lasers, and photodetectors that operation of wide spectral range. As shown in Fig. 5 d and e, PL emission of CsPbX 3 QDs could be varied from UV to NIR spectral regions owing to the anion-based spectral tunability, leading to the QLED devices with the emission of uniform and large-area blue, green, and orange light [ 71 ]. Fig. 5 a Optical band gap of Pb and Sn-based perovskites as a function of Cs content. b Optical band gap of mixed perovskites as a function of Sn content, corresponding to different fractions of Cs. c Two-dimensional band gap map of MHPs with the substitution of A-site and B-site [ 69 ]. d Composition-dependent PL spectra of different MHPs with X-site anion variation. e Photographs of CsPbX3 QDs showing anion-based PL tunability [ 71 ] Full size image 3.4 Geometry The optoelectronics of MHPs can also be readily modulated by changing the crystal geometry, especially in the MHP laser field, where lasing actions from different crystal geometry demonstrate different kinds of distinct microcavity effects. The key role of optical cavity is to provide optical feedback for laser oscillation, which defines the allowed cavity lasing modes within the gain spectrum as well as the spatial characteristics of the output beam [ 72 ]. There are mainly two kinds of optical microcavities for MHP lasers, including intrinsic and extrinsic cavities. The intrinsic cavity of MHPs is owing to the gain medium itself ( i.e. , MHPs) with a particular geometric structure capable of providing the optical feedback. As shown in Fig. 6 a, various geometries of MHPs micro/nanocrystals, including 1D nanowires, 2D triangular, tetragonal, hexagonal, and octagonal microdisks, and 3D microspheres, are obtained by different synthesis approaches [ 72 ], indicating a geometry-dependent microcavity effects. For example, the MAPbBr 3 perovskite 1D microwires with rectangular cross section shows Fabry-Pérot (FP) longitudinal cavity effects, while for the 2D microplates, exhibits a four-edge reflected whispering-gallery-mode (WGM) lasing action (Fig. 6 b, c) [ 73 ]. While the extrinsic cavity of MHPs is realized by introducing an additional geometry that induces the optical feedback to MHPs, including microsphere and capillary cavities [ 74 , 75 ], DFB cavity [ 76 ], and distributed Bragg reflector (DBR) cavity [ 77 ], as shown in Fig. 6 d–g. MHPs as gain media can be integrated with the extrinsic cavities through various techniques, such as spin-coating, inkjet-print, vacuum thermal evaporation, and so on. Such merits of MHPs would help us to fabricate mini-sized lasers with specific functionalities through better understanding the structure–property relations, and to promote the advancement of metal hybrid flexible optical elements into mini-sized photonic circuits with higher performance [ 73 ]. Fig. 6 a – c Intrinsic cavities of MHPs. a Perovskite crystals with different morphologies, including 1D (i) nanowires, 2D (ii) triangular, (iii) tetragonal, (iv) hexagonal, and (v) octagonal microdisks, and 3D (vi) microspheres. b Optical trajectories and c The corresponding simulated field distribution of perovskite cavities with different morphologies [ 72 ]. d – g Extrinsic cavities of MHPs, d side-view SEM image of a microsphere cavity [ 74 ]. e Schematic diagrams of capillary cavity with circular WGM laser resonances [ 75 ]. f SEM image of DFB laser on a quartz substrate [ 76 ]. g Cross-sectional SEM image of a perovskite DBR laser device [ 77 ] Full size image 4 Optoelectronic devices based on MHPs Owing to the unique optoelectronic features of MHPs, they have been extensively investigated in various optoelectronic devices, where MHPs mainly serve as light harvest or a light emitter. The light harvesting devices, such as solar cells and photodetectors, convert incident photons into free charge carries in MHPs (red boxes in Fig. 7 ). Different from the typical sandwich architecture of solar cells, PDs can be classified into vertical and lateral structures. The vertical-structure PDs ( i. e., photodiodes) have the similar configuration to the photovoltaic device, while the lateral-structure PDs ( i. e ., phototransistor and photoconductor), being known as photoconductive type, includes the metal–semiconductor-metal and the field-effect transistor configuration, showing the functions of light detection and signal amplification simultaneously. As for the light-emitting devices, a typical perovskite LED usually turns the injected charge carriers into luminescent layer under forward bias, where they recombine radiatively and emit light in all directions (green box in Fig. 7 ). While lasers essentially work based on a light-amplification process by stimulated emission of radiation (purple box in Fig. 7 ) [ 78 , 79 ]. Fig. 7 The nature of photogenerated exciton/carriers within perovskites and the photophysical process in various optoelectronic devices [ 88 ] Full size image Management of light, such as emission, absorption, modulation and transmission, is the core of optoelectronic device and is principally governed by the photo-generated carrier behavior, and the optoelectronic devices correspond to various photophysical processes, which have their own suitable working conditions, including excitation density and carrier density [ 80 ]. As shown in Fig. 7 , under low excitation fluence with carrier densities region from ~ 10 4 to ~ 10 17 cm −3 , the typical photovoltaic devices ( i.e. PDs, PSCs and LEDs) are usually evaluated, where the PDs with detectable light intensities rang from 1pW cm 2 to ~ 100 mW/cm 2 [ 81 , 82 ], PSCs with light intensities are over a range of ~ 1 μW/cm 2 to ~ 1000 mW/cm 2 [ 83 , 84 ], and LEDs can further extend the working range of the device from ~ 1 μW/cm 2 to ~ 10 mW/cm 2 [ 85 , 86 ]. As the excitation intensity reaches beyond 100 mW/cm 2 , the effect of ASE behavior can even be realized along with the carrier density reaches over ~ 10 18 cm −3 [ 10 , 87 ], leading to the realization of optically pumped perovskite lasers. 5 Revolutionary applications of MHPs-based devices Owing to the unique optoelectronic properties of MHPs, the application of MHPs-based optoelectronic devices in various fields has been widely investigated, including energy harvest/conversion, imaging/sensing, display/communication, and manufacture/medicine as shown in Fig. 8 . In these applications, MHPs exhibit superior performance compared to the traditional materials, such as high efficiency, flexibility, tunability, and versatility. Taking advantages of these merits, MHPs are expected to make indispensable contribution to the advanced revolutionary technologies that could benefit the mankind, such as functional integration system, information display system, electronic communication system, and health & medical system. In this section, the advanced development of MHPs-based optoelectronic devices used in those systems will be briefly introduced. Fig. 8 Revolutionary applications enabled by MHPs optoelectronic components [ 16 , 87 , 89 , 90 , 91 , 92 , 93 , 94 , 95 , 96 , 97 , 98 , 99 ] Full size image 5.1 Functional integration system Inspired by the unique merits of MHPs, such as low cost, semi-transparent, high energy/mass ratio, and flexibility, MHPs as photo-electric unit have been explored to develop potential applications in the next-generation functional integration system, including smart photovoltaic windows energy conversion & storage system, and self-powered system. Integrated smart photovoltaic windows based on MHPs are regarded as highly an efficient and promising green technology, which have been praised for the broaden application, including BIPVs, self-powered electronic device displays, and solar-powered automobile roofs [ 91 , 100 ]. Compared to traditional opaque solar cells, semitransparent perovskite solar cells (ST-PSCs), using phase transformative perovskites can switch between semitransparent and transparent states, demonstrating the potential in realizing the integration both functions composed of color-adjustment and power-generation, and solving the fundamental trade-off between transparency and solar energy harvesting of conventional semiconductor materials [ 101 , 102 , 103 ]. Next-generation internet of things (IoT), smart cities, and wearable electronics, are expected to be self-powered by conformable energy storage devices that can provide energy output whenever needed. Therefore, the integrated photoelectric conversion-storage systems, which can transfer the photo-generated electrons and holes from the excited perovskite and electrochemically store in the storage devices, has offered a promising solution. Also, the system gained great scientific and technological attention due to the increasing demand for green energy and the tendency for miniaturization and multi-functionalization in photo-electronic industry. Since the concept of “photo-capacitor” was first proposed by Tsutomu Miyasaka in 2004 [ 104 ], efforts for integrated photoelectric conversion-storage systems have been made, ranging from PSCs-lithium ion batteries (LIBs), PSCs-supercapacitors (SCs), and PSCs with other types of energy storage devices [ 105 , 106 ]. These integrated systems enable the development of flexible and self-powered electronics, which have attracted tremendous research interests due to their applications in wearable and portable devices. One of the critical issues needs to be addressed for integrated system is the inherent mismatch between low voltage, high current solar cells and high voltage, low current batteries. In this regard, it is challenging but of great significance to develop integrated systems with both high-power density and large energy density, leading to the high overall photoelectric conversion and storage efficiency of the integrated system. In sum, the functional integration system based on MHPs photo-electric devices may become one of the most disruptive technologies for the smart cities, IoT, and wearable electronics. 5.2 Information display system Next-generation display system require efficient light sources that combine high brightness, high color purity, high stability, compatibility with flexible substrates, and transparency. In order to meet these requirements, researchers have never stopped exploring and pursuing better display materials and technologies. Compared with the existing commercial phosphor, MHPs are shown to fulfill all the requirements above, which has been widely regarded as a promising candidate of next generation display system. LEDs, which convert electricity to light, are widely used in modern society-for example, in lighting, flat-panel displays, medical devices and many other situations [ 107 , 108 ]. Inspiringly, perovskite-based light-emitting diodes (PeLEDs) have become an ideal candidate for next-generation solid-state lightings and high-definition displays due to their high PLQY, tunable emission wavelength over the visible spectrum, and narrow emission. In recent years, prominent progress has been achieved for PeLEDs, whoes EQE has reached over 20% for green and red PeLEDs, while high-performing blue PeLEDs remains a big challenge [ 109 , 110 ]. In addition to PeLEDs, the unique properties of MHPs enable them to potentially go beyond many other information display systems. For example, the excellent performance of PeLEDs and hybrid LEDs based on inorganic QDs and perovskites in the NIR region may bring about new applications in biomedical diagnosis and a variety of wearable electronic devices [ 111 ]. Moreover, the good mechanical property of MHPs promotes their application in stretchable electronics that are soft and elastic, which has been considered as the next generation of smart electronics with enhanced functionality, usability, and aesthetics. Notably, stretchable displays and solid-state lighting systems are emerging technologies factors that can conform to diverse applications, such as expandable and foldable screens for information technologies, bio-integrated devices, and wearable electronic clothing. 5.3 Electronic communication system The continuing development of consumer electronics, mobile communications and advanced computing technologies has led to a rapid growth in data traffic, creating challenges for the communications industry. PeLEDs, which can offer tunable optoelectronics properties and solution-processable manufacturing, are of particular interest in the development of next-generation data communications [ 99 ]. In addition, photodetectors are key components of many electronic products, such as digital camera, smart phone, and medical diagnostics instrument, which are important for imaging and optical communication applications [ 112 , 113 , 114 ]. In the last decade, with the continuous rise of autonomous driving, environmental monitoring, optical communication, or biosensors, there is a more prominent demand for excellent photodetectors. Moreover, laser technology has become ubiquitous in every section of modern society, such as scientific research, manufacturing, communication and so on. The possibility of developing efficient on-chip coherent light sources from electrically pumped semiconductor laser diodes (LDs) has been the center of attention for researchers in the integrated optoelectronic field [ 115 , 116 , 117 ]. Owing to the advantages of MHPs material, it could play a key role in the next-generation electronic communication system, and MHPs-based LEDs, PDs, and LDs can be applied in widespread communication solutions in future scenarios. For instance, integrated components for short-range communications with rigorous latency requirements (for example, self-guiding vehicles), flexible biosensors (for example, real-time health monitoring and disposable lab-on-a-chip for point-of-care diagnostics), light communication, underwater communications, low-cost on-chip IoT sensors for accurate tracking and positioning, indoor data services (for example, Li-Fi and short-distance on-board communications). 5.4 Health & medical system Among the various optoelectronic applications, MHPs have also attracted significant interest in health & medical systems such as health monitoring, biomedicine, medical imaging techniques. The use of digital flat detectors in many diagnostic and interventional medical applications, including in X-ray imaging procedures, has increased rapidly over the past decades [ 112 ]. For example, medical X-ray imaging procedures require digital flat detectors operating at low doses to reduce radiation health risks, while solution processed organic–inorganic hybrid perovskites have characteristics that make them good candidates for the photoconductive layer of such sensitive detectors [ 16 , 21 ]. Recent researches on the X-ray scintillator ( i. e ., all-inorganic perovskite nanocrystals CsPbBr 3 QDs) have shown that the color-tunable perovskite nanocrystal scintillators can provide a facile visualization way for X-ray radiography [ 21 ]. In addition, perovskites integrated with chiral ligands exhibits superior chiroptoelectronic property in recent years, and circularly polarized photodetectors and light sources based on these materials are promising for flexible integrated devices [ 118 ]. However, to promote the commercialization, MHPs-based devices will face many of the same challenges that PSCs face, specifically, stability, toxicity issues, and competition from more mature competing technologies. 6 Challenges and perspectives of MHPs 6.1 Instability Despite impressive advances of MHPs, their typical challenges, such as long-term instability and toxicity, greatly impede their rapid commercialization (Fig. 9 ). The instability problem is origin from the structural degradation under certain conditions, which includes intrinsic and extrinsic sources. The initial intrinsic stability of MHP can be primarily evaluated by the Goldschmidt tolerance factor (τ), which is an empirical index widely applied to the evaluation of the geometry stability of perovskite crystal, and the value of τ in the range of 0.8–1 favored an ideal cubic perovskite structure [ 119 ]. However, for MHPs with τ value out of the 0.8–1 range ( e.g. , FAPbI 3 > 1), formation of secondary non-perovskite phases (δ-FAPbI 3 ) can also influence the device performance and stability [ 120 ]. Introduction of MA cation or Cs cation has been the most common strategy to enable the stabilization of black-phase FAPbI 3 perovskite, but the resulted mixed cation perovskite may broaden the optical bandgap and possibly inducing the inferior thermal stability. Moreover, intrinsic ion migration in MHPs has also been shown to occur under the drive of electric filed, resulting in the abnormal J - V hysteresis, phase segregation, slow photoconductivity response, and device performance degradation [ 24 ]. In addition, characteristics of polycrystalline films can also be intrinsic sources of instability too, for instance, high density of grain boundaries and thermal strain generated during the annealing process of perovskite films accelerates degradation of the perovskite. Therefore, how to stabilize black-phase FAPbI 3 to prepare high-stability and high-purity film is still the primary problem that needs to overcome to achieve efficient and stable PSCs [ 121 ]. Some volatile additives, such as FAAc, have been recently introduced to form the intermediate phase, facilitating the formation of stable α- FAPbI 3 [ 122 ]. In addition, a mechanochemical strategy to prepare high quality α-FAPbI 3 was demonstrated by Zhu and co-workers, where they prepare the δ-FAPbI 3 perovskite powder first, and redissolving it in precursor solution, leading to a high concentration of large-sized polyiodide colloids and further enabling the formation of black-phase FAPbI 3. As a result, the devices deliver a champion PCE of 24.22% [ 123 ]. Fig. 9 Instability problem of MHPs and their advanced resolution strategies Full size image In addition to the intrinsic stability of perovskites, external environments, such as moisture, oxygen, heat, light, external bias, and contact layers, are also shown to be the origin of the instability problem of MHP-based devices. Instability against moisture and O 2 can be completely solved by standard encapsulation of the device, while the instability induced by light, external bias and contact layers are unavoidable in a working device and should be seriously handled [ 124 ]. Ion migration, trap state generation, etc. have been found to be reasons for performance deterioration under these external factors, and some of the most impressive strategies have been reported by 2D/3D heterojunction, defect passivation, interfacial stabilization, structural healing, strain release and device encapsulation [ 125 , 126 , 127 , 128 ], pushing the rapid development of the long-term stability of MHPs. 6.2 Toxicity Apart from long term stability, toxicity of perovskite materials for both humans and ecosystem is an overwhelming challenge, which may hinder the commercialization pace of MHPs-based optoelectronic devices. As shown in Fig. 10 , the life cycle of MHPs typically undergoes from raw material extraction, synthesis, film fabrication to the usage and finally decommissioning stage [ 129 ]. Strategies of preventing the lead leakage, such as physical encapsulation ( e. g. , inserting the epoxy resin into the device module and the top glass cover), chemical absorption (e. g., introducing metal–organic framework as scaffolds) [ 130 ], eco-friendly perovskite materials ( e. g. , Sn 2+ , Sb 2+ , Ge 2+ , Cu 2+ and mixed monovalent and trivalent elements for double-perovskite structures) [ 131 ], and recycling strategy have been developed. Future research on the safe deployment of MHPs optoelectronic technology relies entirely on adopting precautionary measures against contamination at each stage of the device’s life, from fabrication to disposal/recycling, and test standards should be also established. Fig. 10 Toxicity problem of MHPs and their advanced resolution strategies [ 129 ] Full size image 7 Conclusions As the rising star in the field of optoelectronics, MHPs with extraordinary properties have experienced unprecedented rapid development during the past decades. Here, we give an overview on the prominent optoelectronic properties of MHPs and their revolutionary applications in advanced technologies, which might greatly change the development course of human society. Nevertheless, to promote the large-scale utilization of MHPs, there are still many technique problems to be overcome, and vast investments are required to establish the building of MHPs-based optoelectronics. Overall, one can expect that in the next decade, MHPs will be under the spotlight in the era of “light”. Availability of data and materials Not applicable.

Metal halide perovskites (MHPs) have been emerging as the rising star in the field of optoelectronics during the past decade. The state-of-the-art optoelectronic technologies based on MHPs, such as perovskite solar cells (PSCs), light emitting diodes (LEDs), photodetectors (PDs) and lasers have been leading the prevailing paradigm, owing to the intriguing optoelectronic properties of MHPs. Additionally, MHPs possesses the merits of facile and low-cost processing and favorable tunable optical and electronic features, providing a rich and fertile ground for the development of high-performing multifunctional optoelectronic devices and their future industrialization. In a new paper published in eLight, a team of scientists led by Professor Wei Huang from Frontiers Science Center for Flexible Electronics, Xi'an Institute of Flexible Electronics, Northwestern Polytechnical University, China, Key Laboratory of Flexible Electronics & Institution of Advanced Materials, Nanjing Tech University, China, Key Laboratory for Organic Electronics & Information Displays, and Institute of Advanced Materials, Nanjing University of Posts and Telecommunications, and colleagues has given a panorama picture on the optoelectronic traits of MHPs and their revolutionary impact on the next-generation optoelectronics. They first review the historical research milestones of MHPs and their optoelectronic devices. Thereafter, they introduce the origin of the unique optoelectronic features of MHPs, based on which the tunability of these features via regulating the phase, dimensionality, composition, and geometry of MHPs are highlighted. Then, they show that owing to the convenient property control of MHPs, various optoelectronic devices with target performance can be designed. At last, they emphasize on the revolutionary applications of MHPs-based devices on the existing optoelectronic systems. This perspective aims at providing critical guidance for inspiring the novel research directions of MHPs to promote the widespread application of MHPs in optoelectronics. Revolutionary applications enabled by MHPs-based optoelectronic devices. Credit: He Dong, Chenxin Ran, Weiyin Gao, Mingjie Li, Yingdong Xia, Wei Huang The unique characteristics of MHPs are derived from their crystal structure and chemical composition, which allows the unprecedented flexibility to independently and synergistically tune the optoelectronic properties of MHPs, laying an important foundation for MHPs being applicable to various optoelectronic applications, including solar cells, LEDs, PDs and lasers. The reasearchers highlight the unique optoelectronic properties of MHPs: "Of critical importance, intrinsic electronic configurations of MHPs are the direct origin of their unique optoelectronic properties, including high optical absorption, high carrier mobility, high defect tolerances, long diffusion lengths, unique ambipolar charge transport, and flexible turnability. We summarize the merits of MHPs by focusing on the fine controlling of the optoelectronic properties by regulating the structure of MHPs, including phase, dimensionality, composition, and geometry." "MHPs are expected to make indispensable contribution to the advanced revolutionary technologies that could benefit [...] mankind, such as functional integration system[s], information display system[s], electronic communication system[s], and health and medical system[s]," they add. "Metal halide perovskites (MHPs), emerging as innovative and promising semiconductor materials with prominent optoelectronic properties, have been pioneering a new era of light management (ranging from emission, absorption, modulation, to transmission) for next-generation optoelectronic technology. Nevertheless, to promote the large-scale utilization of MHPs, there are still many technique problems to be overcome, such as stability and toxicity issues, and vast investments are required to establish the building of MHPs-based optoelectronics. Overall, one can expect that in the next decade, MHPs will be under the spotlight in the era of light,'" the researchers forecast.

10.1186/s43593-022-00033-z

Medicine

Scientists pinpoint sensory links between autism and synesthesia

Jamie Ward et al, Atypical sensory sensitivity as a shared feature between synaesthesia and autism, Scientific Reports (2017). DOI: 10.1038/srep41155 Journal information: Scientific Reports

http://dx.doi.org/10.1038/srep41155

https://medicalxpress.com/news/2017-03-scientists-sensory-links-autism-synesthesia.html

Abstract Several studies have suggested that there is a link between synaesthesia and autism but the nature of that link remains poorly characterised. The present study considers whether atypical sensory sensitivity may be a common link between the conditions. Sensory hypersensitivity (aversion to certain sounds, touch, etc., or increased ability to make sensory discriminations) and/or hyposensitivity (desire to stimulate the senses , or a reduced response to sensory stimuli are a recently introduced diagnostic feature of autism spectrum conditions (ASC). Synaesthesia is defined by unusual sensory experiences and has also been linked to a typical cortical hyper-excitability. The Glasgow Sensory Questionnaire (GSQ) was administered to synaesthetes and people with ASC. Both groups reported increased sensory sensitivity relative to controls with a large effect size. Both groups also reported a similar pattern of both increased hyper- and hypo-sensitivities across multiple senses. The AQ (Autism-Spectrum Quotient) scores were elevated in the synaesthetes, and one subscale of this measure (attention to detail) placed synaesthetes within the autistic range. A standard laboratory test of visual stress (the Pattern Glare Test), administered online, corroborated the findings of increased sensitivity to aversive visual stimuli in synaesthetes. We conclude that atypical sensory sensitivity is an important shared feature between autism and synaesthesia. Introduction People with synaesthesia have unusual experiences of the world: for example, words may evoke tastes, sequences such as months and numbers may be visualised as spatial landscapes (sequence-space synaesthesia), and graphemes (i.e., letters/numbers) may evoke colours (grapheme-colour synaesthesia). The present study focuses on grapheme-colour synaesthesia. Synaesthetic experiences tend to be percept-like in nature, occur automatically, and are triggered by inducing stimuli 1 , 2 . Synaesthesia emerges during childhood, if not before 3 , has a hereditary component 4 , and is linked to structural and functional differences in the brain 5 . It is also linked to wider cognitive differences - for example, in memory 6 and mental imagery 7 - but is not linked to global difficulties in intellectual functioning 8 . However, recent research has suggested that synaesthesia and Autism Spectrum Condition (ASC) co-occur together more than would be expected by chance. Neufeld et al . 9 and Baron-Cohen et al . 10 screened samples of patients diagnosed with autism for grapheme-colour synaesthesia (primarily) and reported prevalence rates of 17.2% and 18.9% respectively. The current prevalence estimate for grapheme-colour synaesthesia is 1–2% 11 . Hence these studies suggest a link between synaesthesia and autism. But what is the nature of that link? The present research considers this from the perspective of sensory sensitivity. Autism is a heterogeneous condition that impacts on several cognitive domains. It is unclear whether all, or only some, of these domains are related to synaesthesia. Autism entails impairments in social communication alongside unusually narrow interests and repetitive behavior (DSM-5, 2013). Contemporary models of autism also attempt to account for relative strengths as well as impairments. The empathizing-systemizing model of autism emphasizes not only the socio-cognitive difficulties in understanding others (linked to empathizing) but also an interest in and aptitude for rule-based systems (hyper-systemizing) 12 . Related, others have characterized the cognitive style of autism in terms of more local information processing that may, for instance, give them an advantage on certain perceptual tasks such as finding hidden (embedded) figures 13 . What traits relating to autism are found in studies of synaesthesia? Banissy et al . 14 administered personality questionnaires to a large group of grapheme-colour synaesthetes. Two of these assessed empathy and social functioning (the Inter-Personal Reactivity Index, IRI, and Empathy Quotient, EQ) that are known to differentiate people with autism and controls 15 , 16 . However, the synaesthetes did not score differently on any subscale in these measures. In terms of autism-related abilities and differences in cognitive style, there is more positive evidence. Mealor et al . 17 developed the Sussex Cognitive Styles Questionnaire that links together mental imagery 18 , local/global bias and systemizing. This questionnaire was given to people with grapheme-colour synaesthesia, sequence-space synaesthesia, synaesthetes with both, and controls with neither (but not to an autistic group). The presence of sequence-space synaesthesia was linked to increased systemizing, increased technical/spatial processing which also contains items from the Systemising Quotient 19 , and increased local bias which contained items from the Autism-Spectrum Quotient, AQ 20 (grapheme-colour synaesthetes tended to have scores intermediate between controls and sequence-space synaesthetes). Sequence-space synaesthesia has been linked to certain forms of savant abilities, prevalent in the autistic population, such as prodigious memorisation of dates 21 . Of course, other forms of synaesthesia may be relevant to savantism too (e.g. music visualisation). The present study considers, whether sensory sensitivity is a common facet of both synaesthesia and autism. The 2013 Diagnostic Statistical Manual (DSM-5) added the following to their list of criteria: “Hyper- or hypo-reactivity to sensory input or unusual interests in sensory aspects of the environment (e.g., apparent indifference to pain/temperature, adverse response to specific sounds or textures, excessive smelling or touching of objects, visual fascination with lights or movement).” In order to explore this feature of autism experimentally, several measures have been developed 22 , 23 , 24 . The Glasgow Sensory Questionnaire (GSQ) – used in the present study – contains items relating to seven sensory modalities (vision, hearing, taste, touch, smell, vestibular, proprioception) that tap both hypesensitivity and hyposensitivity 22 . People with high autistic traits score significantly higher than controls and, interestingly, all questions loaded on a single factor (i.e. there was no evidence of fractionation according to sensory modalities or whether hypo- or hyper-). Questionnaire measures of sensory sensitivity show no evidence of a neurotypical gender difference 22 , 24 unlike other traits related to autism that do show gender differences (with males on average tending to score lower on empathizing and higher at systemizing 12 ). Grapheme-colour synaesthetes show evidence of a hyper-excitable visual cortex as shown by reduced phosphene thresholds when the brain is stimulated using TMS (Transcranial Magnetic Stimulation) 25 . They show increased amplitude visual-evoked potentials, in EEG, for certain simple visual stimuli that do not elicit synaesthesia 26 and have enhanced perceptual discrimination for colour 27 . As such, we hypothesise that grapheme-colour synaesthetes will report unusual sensitivity in the visual domain but it is less clear whether this pattern will extend to other modalities too (as found in autism). In the study below, we administer the GSQ to a group of participants with a confirmed diagnosis of autism, a group of verified grapheme-colour synaesthetes (with most reporting sequence-space synaesthesia too), and a group of controls. The AQ (Autism-Spectrum Quotient) was given to the synaesthetes and controls to assess for the broader range of traits linked autism. Finally, a common measure of ‘visual stress’ termed the Pattern Glare Test 28 is given to synaesthetes and controls. It was not given to the autism sample as they took part in other research (not reported here). The test was given to corroborate the GSQ measure to see whether questionnaire differences (group differences and/or individual differences) can be related to differences on a more psychophysical measure. In the Pattern Glare Test, participants are shown black-and-white gratings of different spatial frequencies. People with high visual stress report being highly sensitive to mid-frequency gratings (2–5 cycles per degree at which contrast sensitivity is optimal) 29 . These stimuli are reported as not only being aversive but also as inducing visual experiences such as shimmer, shapes, and colour. The fact that an achromatic stimulus (with similar physical properties to printed text 30 ) can induce colour in some people has interesting parallels with grapheme-colour synaesthesia. Those who are susceptible to this test are assumed to have a highly-excitable visual cortex 31 . Importantly, this test also contains a control stimulus (of low spatial frequency) that is not linked to visual stress and acts as a baseline to correct for any overall tendency for people to report unusual visual experiences per se. We also include a high spatial frequency grating that also tends to be linked to visual stress (see Table 1 ) although this may reflect opthalmological effects as well as cortical ones 32 . Table 1 Expected pattern of visual disturbances/discomfort according to the spatial frequency used (LSF, MSF, HSF = low, mid, and high spatial frequencies respectively) and degree of daily visual stress reported. Full size table In summary, we have three main hypotheses. First, both synaesthetes and the ASC group will have higher sensory sensitivity on the GSQ (although we are agnostic as to whether all aspects will be affected). Secondly, on the basis of previous research showing a higher prevalence of synaesthesia in autism we hypothesise that synaesthetes will have an elevated AQ score (again, we are agnostic as to which aspects of the AQ might be affected). Finally, we hypothesise that that high sensory sensitivity, on the GSQ, will be related to a psychophysical measure of visual stress (the Pattern Glare Test). Results Questionnaire Measures For the GSQ, the overall scores across the three groups were compared using a one-way ANCOVA with age and gender entered as covariates. The three groups differed in their scores (F(2,193) = 35.776, p < 0.001; η p 2 = 0.270) and post-hoc t-tests revealed that all three groups differed from each other even after a Bonferroni correction (ASC v. controls: applying Levene’s correction t(141.5) = 8.265, p < 0.001, d = 1.32; synaesthetes v. controls: t(52.3) = 4.044, p < 0.001, d = 0.81; ASC v. synaesthetes: t(111) = 2.150, p = 0.034, d = 0.41). This is shown in Fig. 1 . Neither gender (F(1,193) = 0.005, p = 0.941; η p 2 = 0.000) nor age (F(1,193) = 3.056, p = 0.082; η p 2 = 0.016) were significant covariates. Figure 1 Group differences in the Glasgow Sensory Questionnaire (GSQ) showing mean and SEM. Full size image More detailed explorations show that synaesthetes show a similar pattern of responding to the ASC sample insofar as they show both increased hypersensitivities and increased hyposensitivities across multiple modalities. The results are shown in Fig. 2 . The basic pattern of ASC > Synaesthetes > Controls holds in virtually every case. That is, synaesthetes present with a pattern of atypical sensory sensitivity that qualitatively resembles that found in autism but is not as extreme as that found in autism. A 3 × 7 ANOVA was conducted contrasting group (synaesthetes, ASC, controls) and modality on the hypersensitivity scores and hyposensitivity scores separately. Considering hypersensitivity, there was a main effect of group (F(2,196) = 36.480, p < 0.001, η p 2 = 0.271). There was also a main effect of modality (F(6,1176) = 118.151, p < 0.001, η p 2 = 0.376) and an interaction between group and modality (F(12,1176) = 2.114, p = 0.014, η p 2 = 0.021) such that the differences between groups were more pronounced for some modalities than others. However, we did not pursue multiple pairwise comparisons as these were not theoretically motivated at this finer-grained level and these modality differences are already noted in prior research on autism 33 . The pattern was similar for the hyposensitivity questions. There was a main effect of group (F(2,196) = 25.151, p < 0.001, η p 2 = 0.204). There was also a main effect of modality (F(6,1176) = 90.655, p < 0.001, η p 2 = 0.316) and an interaction between group and modality (F(12,1176) = 6.471, p < 0.001, η p 2 = 0.062) such that the differences between groups were more pronounced for some modalities (e.g. auditory) than others (e.g. olfactory). Robertson 33 noted that scores for olfactory were the lowest of all modalities, and it is to be noted that the hypo-olfactory questions had some of the lowest factor loadings in their Principal Component Analysis (i.e. are less representative of the scale as a whole). All three groups in our study show a tendency for ‘hypo’ and ‘hyper’ items to correlate (synaesthetes: r = 0.77; ASC r = 0.74; controls r = 0.64; all p < 0.001) consistent with previous findings 33 . Figure 2 The profile of hypersensitivies (top) and hyposensitivities (bottom) by modality and group (error bars show 1 SEM). Full size image The mean AQ scores of synaesthetes was 23.0 (SD = 9.3) and the mean of controls was 18.6 (SD = 8.8), with this difference being statistically significant (F(1,113) = 4.332, p = 0.040, η p 2 = 0.037) but neither gender (F(1,113) = 1.069, p = 0.303; η p 2 = 0.009) nor age (F(1,113) = 0.736, p = 0.393; η p 2 = 0.006) were significant covariates. Fig. 3 shows these scores, broken down according to the five AQ subscales (social skills, attention switching, attention to detail, communication, and imagination) and contrasted against our ASC sample (for whom this data was already available). A 5 × 3 ANCOVA contrasting subscale score and group, with age and gender as covariates, revealed main effects of group (F(2,178) = 147.764, p < 0.001, η p 2 = 0.624) and subscale (F(4,712) = 6.469, p < 0.001, η p 2 = 0.035) and an interaction between them (F(4,712) = 16.395, p < 0.001, η p 2 = 0.156) with neither age nor gender acting as significant covariates. The interaction is due to different patterns across the subscales (confirmed by post-hoc t-tests shown in Fig. 3 ): the synaesthetes statistically resembled the ASC sample on attention to detail (i.e. ASC = Syn > controls), but statistically resembled controls on the other subscales (i.e. ASC > Syn = controls). The effect size contrasting synaesthetes and controls on attention to detail was large (d = 1.0; t(90.6) = 4.809, p < 0.001 with Levene’s correction) and the difference survives Bonferroni correction for multiple comparisons (i.e. < 0.05/15). The other significant differences (between ASC and the other samples) also survived multiple comparisons. Figure 3: Subscale scores of the AQ (out of 10) showing the mean and SEM for ASC group, synaesthetes, and controls. The lines show the pattern of significant differences (all p < 0.001). Full size image Mealor et al . 17 also report that synaesthetes score higher on attention to detail (referred to in that study as local processing bias) on a questionnaire that contains items from the AQ. The effect was greater in sequence-space synaesthesia than grapheme-colour synaesthesia. Within our sample of synaesthetes, there is some evidence that sequence-space synaesthesia may be particularly relevant to the AQ attention-to-detail score (SSS mean = 7.32, SD = 1.67; non-SSS mean = 5.71, SD = 1.60; t(33) = 2.284, p = 0.029), suggesting that these subtypes need more careful contrasting in future research. By comparison, splitting our synaesthetes in this way had no significant impact on the overall GSQ scores (SSS mean = 67.3, SD = 24.2; non-SSS mean = 56.3, SD = 18.4; t(33) = 1.123, p = 0.270) suggesting that sensory sensitivity is not strongly tied to the presence of sequence-space synaesthesia. In summary, synaesthetes differ from controls on both a measure of sensory sensitivity (linked to autism) and on a broader measure of autistic tendencies (AQ) and, again, this was particularly apparent with regards to the perceptual features of autism (attention to detail/local processing). Pattern Glare Test of Visual Stress Given that this test is sensitive to the presence of migraine, it is important to note that the participants of this study, synaesthetes and controls, reported a low incidence of migraine with aura (syns = 4.3%, controls = 9.4%) and migraine without aura (syns = 8.7%, controls = 13.2%) and did not differ from each other in this regard (Fisher’s test p = 0.634; and p-.688 respectively), as noted by a previous study 34 . The results of the Pattern Glare Test are summarised in Fig. 4 for the two standard measures on this test (total number of experiences reported and level of visual comfort) and for the novel measure that we introduced (number of colours reported). Figure 4: Top: the level of visual discomfort induced by these gratings (0 = neutral, +ve = comfortable, −ve = uncomfortable). Middle: The number of visual experiences (e.g. shimmer, shapes, color, etc.) induced by different grating frequencies (LSF, MSF, HSF for low medium and high). Bottom: the number of colors induced (blue, red, green, etc.). Full size image The comfort rating scale is an 11-point measure that was treated as a parametric variable. A 2 × 3 ANOVA contrasting group (synaesthetes v. controls) and spatial frequency (high, medium, low) revealed a main effect of group (F(1,58) = 3.996, p = 0.050, η p 2 = 0.064), a main effect of spatial frequency (F(2,116) = 10.673, p < 0.001, η p 2 = 0.155), but no interaction (F(2,116) = 0.767, p = 0.467, η p 2 = 0.013). The effect of spatial frequency replicates many previous results showing that mid- and high spatial frequency gratings are linked to more visual discomfort than low frequency 29 , 35 . The main effect of synaesthesia reflects quantitatively higher level of visual discomfort for this group. In terms of the qualitative pattern, the synaesthetes reverse their ratings from ‘comfortable’ for low spatial frequency to ‘uncomfortable’ for the mid and high frequency patterns. The number of experiences reported for each stimulus was low and treated non-parametrically. The (baseline) low spatial frequency condition was subtracted from the number of experiences reported to each of the mid- and high spatial frequencies. Synaesthetes reported significantly more visual experiences than controls to both the mid- (independent samples median test, p = 0.023) and high (p = 0.025) spatial frequencies even after this baseline correction (i.e. they were not globally higher across all stimuli). The comparable analysis on total number of colours reported did not yield a difference between synaesthetes and controls for the mid (p = 0.458) or high (p = 0.681) spatial frequency stimuli relative to baseline of low spatial frequency. This reflects a generally higher tendency for the synaesthetes to report colours being induced across all stimuli. How is performance on the Pattern Glare Test related to the AQ and GSQ? This was examined by correlating the measures from the Pattern Glare Test with the overall AQ and GSQ questionnaire scores (in the combined synaesthete and control sample). The predictions were that the GSQ should be related to visual disturbances/discomfort for the mid-frequency stimulus and possibly the high-frequency stimulus, but not the low-frequency stimulus. These analyses should be considered as exploratory given the large number of correlations that do not survive correction for multiple comparisons. The AQ did not correlate with any other measures (all p’s > 0.10) but the GSQ correlated with two measures and showed one non-significant trend (all other p’s > 0.10; see Supplementary Material ). There were significant correlations between GSQ score and number of experiences reported to the mid-frequency stimulus (r = 0.329, p = 0.007) and the GSQ score and the number of colours to the mid-frequency stimulus (r = 0.273, p = 0.028), with a trend between the GSQ score and number of colours reported to the high frequency stimulus (r = 0.207, p = 0.098). Recall that it is the mid-frequency stimulus that is generally considered to be the most reliable inducer of the highest levels of visual stress 36 . Discussion Two previous studies have found an increased prevalence of synaesthesia in autism 9 , 10 . The aim of this study was to take the novel and complementary approach of looking for autistic-related traits in synaesthetes. We focussed on sensory sensitivity given: the sensory-like experiences of synaesthetes; the evidence for hyper-excitability of the visual cortex in synaesthetes 25 ; and behavioral enhancement of sensory discrimination in synaesthetes 27 . We found that grapheme-colour synaesthetes, like people with autism, scored higher on a questionnaire measure of sensory sensitivity (the GSQ) and showed a qualitatively similar pattern to the autism group (i.e. both hyper- and hypo- sensitivities across multiple senses) albeit quantitatively intermediate between the control and autistic group. We suggest that sensory sensitivity is an important link between these two conditions. Future research will need to establish whether this is the main (or indeed only) link between them. Although we show that synaesthetes do show elevated autistic traits in other domains (using the AQ) this was most strongly driven by the subscale relating to perception (‘attention to detail’ which concerns local perceptual processing). Finally, we corroborate these self-report measures using a psychophysical test that is sensitive to various clinical conditions linked to visual discomfort 29 . Previous research shows that those with very high visual discomfort in everyday life show a peak number of experiences to the mid-frequency stimulus, but those with moderate degrees of visual discomfort find the higher spatial frequency to be more potent 36 . (In both cases, the low spatial frequency stimulus acts as a control). The synaesthetes showed evidence of visual stress to both high and mid spatial frequency stimuli. This links to previous research showing enhanced visual-evoked potentials (in EEG) to mid (5 cycles per degree), but not low, spatial frequency gratings 26 . The participants, as a group, showed a significant correlation between GSQ scores and the number of induced experiences to the mid-frequency (but not high or low) stimulus. The results cannot be explained as a simple response bias because all three tests show a condition X group interaction. On the Pattern Glare Test we found effects for mid/high spatial frequency but not low spatial frequency. On the AQ and GSQ we found stronger effects on some subscales than others. The mechanism that underpins changes in subjective sensory sensitivity is not well understood. Some models have attempted to link sensory sensitivity to enhanced performance on tests of perception. Baron-Cohen et al . 37 suggested that perceptual hyper-sensitivity drives attention-to-detail that, in turn, drives systemizing (or the drive for pattern recognition and pattern creation) and can lead to savant talents. A similar account, termed Enhanced Perceptual Function (EPF), has been proposed by Mottron et al . 38 who explicitly draw a parallel between the perceptual abilities of both synaesthetes and people with autism. However, the relationship between subjective sensitivity and objective perceptual ability is unlikely to be a simple one 39 , and it is harder to explain the presence of increased hyposensitivity within these frameworks that link hypersensitivity with ability 39 . It is possible that what is termed ‘hyposensitivity’ results from an interest in the sensations produced by sensory-motor repetition, such as repeated finger flicking. Turning to synaesthesia, it is known that synaesthetes show better performance on certain perceptual tests, and this appears to be related to the type of synaesthesia they experience. Thus, synaesthetic experiences of touch are linked to increased tactile spatial acuity and synaesthetic experiences of color are linked to better color discrimination 27 . Although the present study was not able to consider different subtypes of synaesthesia in detail, it is noteworthy that our grapheme-color synaesthetes still report differences in sensory modalities that are rarely (e.g. gustatory, olfactory, vestibular) seen in synaesthesia. It may be that altered sensory sensitivity across all the senses is a general feature of the synaesthetic brain. If so, we would predict the same pattern of subjective sensitivity reported here will be found for very different types of synaesthesia (e.g. lexical-gustatory). Further research is needed to explore the underpinning mechanisms behind ‘attention to detail’ by comparing people with autism to different forms of synaesthesia (and pulling apart the contribution of grapheme-color and sequence-space synaesthesia given that both were present in our sample). Grapheme-color synaesthetes have been shown to be good at detecting a hidden figure made up of local elements (e.g. a triangle made up of 2 s) in a larger array of different elements (5 s) 40 , 41 . The standard explanation for this is that their synaesthetic colors enable them to ‘see’ the hidden shape. Ward et al . 40 provided some direct evidence for this: those synaesthetes who reported seeing many synaesthetic colors did better. However, even those synaesthetes who reported no synaesthetic colors during the trials outperformed controls. A novel suggestion is that this reflects the ability to group local elements that resembles the perceptual abilities found in autism 42 . Superficially, autism and synaesthesia seem like very different conditions and the differences between the conditions are as important to explain as the similarities. Learning disability, defined as an IQ less than 70, is present in around a third of cases of autism and, moreover, autism is around three times as common in males 43 . Synaesthesia has traditionally been considered more as a ‘gift’ (rather than linked to intellectual dysfunction) and with no male bias 44 . Our focus on the similarity in sensory sensitivity/functioning between autism and synaesthesia may offer an account of these superficial differences insofar as sensory sensitivity is not linked to gender differences. There are various limitations of the present research that require further research to address. It is not yet known how people with autism would respond to the Pattern Glare Test and, more generally, it is important to use a wider range of visual and non-visual stimuli and determine how this is linked to sensory sensitivity (both subjectively, behaviorally and neurophysiologically). The use of internet-based studies has the disadvantage of potentially introducing sources of noise into the data (which in themselves are unlikely to lead to between-group differences), but have the advantage of enabling us to recruit larger samples of rare populations and is becoming more common in perception research 45 . Sensory sensitivity has, in various accounts, been linked to cognitive abilities found in autism such as savant skills 37 , 38 , 46 . Whilst these theories emphasise a link between sensory sensitivity and some of the positive features of autism (e.g. savantism), other theories (such as Intense World Theory 47 ) suggest a causal link between increased sensory sensitivity and impaired social functioning (e.g. social withdrawal). Our results are more consistent with the former rather than the latter theoretical approach. More generally, our study attests to the value of considering synaesthesia alongside autism to explore the relationship between perceptual, cognitive and social symptoms and abilities. Method This research was given ethical approval through the Cross-Schools Science and Technology Research Ethics Committee at the University of Sussex. The research was carried out in accordance with the Code of Ethics of the World Medical Association (Declaration of Helsinki) and informed consent was obtained for all participants. Participants The synaesthetes were recruited via a database of volunteers held by the University of Sussex. This consisted of 35 participants (5 males, 29 females, 1 undisclosed; mean age = 28.9, SD = 10.5). All participants had grapheme-color synaesthesia (verified using the method described in Rothen et al . 48 ), and most (N = 28) also reported sequence-space synaesthesia. Seventy-eight participants with ASC were recruited from the Cambridge Autism Research Database (CARD) (33 males, 45 females; mean age = 36.2 years, SD = 9.2). They had all received a formal diagnosis of autism according to DSM-IV, DSM-5 or ICD-10 criteria and, where tested, scored high on a screening measure of autism, the AQ, conducted prior to this study (mean 40.2, SD = 5.3, range = 22–48; AQ scores unavailable for 12 participants in the autism sample). There were 86 control participants (28 males, 58 females; mean age = 32.8, SD = 14.1) who were recruited from a mixture of sources including non-ASC volunteers from CARD ( (N = 28) and via opportunistic sampling from the researchers (N = 58). All participants took part voluntarily without financial compensation. Procedure All participants were tested online and remotely. For the participants recruited from Cambridge (ASC and controls), they were given a link to a Qualtrics survey, an online data collection platform. The survey included the GSQ followed by a set of other questions not part of the present study (concerning savant abilities). For the participants recruited from Sussex (synaesthetes and controls) they were given a link to a Qualtrics survey that included the GSQ, the AQ, and the Pattern Glare Test. They were additionally asked basic questions about their medical history including the presence of migraine, with and without aura. The details of each measure are described in turn below. The Glasgow Sensory Questionnaire (GSQ) consists of 42 items, and investigates sensitivities across seven sensory modalities: visual, olfactory, auditory, gustatory, tactile, vestibular, and proprioceptive 22 . Each modality is assessed by six items, in which sensory hyposensitivity and hypersensitivity were determined by three questions each. All questions asked participants how frequently they experienced a sensory event and/or performed a particular behavior, and items were answered using a five-point scale (Never, Rarely, Sometimes, Often, Always). Example items included “Do you find certain noises or pitches of sound annoying?”, “Do bright lights ever hurt your eyes or cause a headache?” and “Do you ever feel ill just from smelling a certain odour?” Responses were coded on a scale from 0 (Never) to 4 (Always), with possible scores ranging from 0 to 168. The Autism-Spectrum Quotient (AQ) is a 50-item questionnaire used to measure traits associated with the autistic spectrum in adults of normal intelligence 20 . The AQ contained 10 statements tapping five different subscales: social skill, attention switching, attention to detail, imagination, and communication. Participants demonstrated their level of agreement with each statement using a four-point scale (Definitely Agree, Slightly Agree, Slightly Disagree, Definitely Disagree). Example items included “I find it hard to make new friends”, “It does not upset me if my daily routine is disturbed” and “I find it difficult to imagine what it would be like to be someone else”. Approximately half the questions are reverse coded. Each item scored one point if the respondent recorded an autistic-like behavior (poor social skill, poor attention switching/strong focus of attention, exceptional attention to detail, poor imagination, poor communication skill), either slightly or definitely. Therefore, responses were coded as either 0 or 1, and total scores ranged from 0 to 50. The Pattern Glare Test used the same achromatic stimuli and same basic procedure as Braithwaite et al . 35 . In order to present the stimuli under controlled conditions, the Qualtrics survey took participants to a link containing the test hosted by Inquisit ( ). Participants were informed that they should not participate if they had a history of epilepsy. The stimuli consisted of black and white alternating horizontal stripes presented in an oval window around a small fixation point and against a mid-grey (RGB 128,128,128) background. The low-, mid-, and high- spatial frequency stimuli comprised of 4.5, 31.5 and 130 cycles (i.e. stripes) which, when viewed at the appropriate distance, corresponded to 0.4, 3.0, and 12.4 cycles per degree. The stimuli were always presented centrally at their actual resolution of 652 × 500 pixels. Inquisit can determine the resolution of the monitor used by participants but not the physical size. To standardize viewing distances, participants were asked to input the physical size of their monitor (it was explained that this is measured diagonally from corner to corner). They were then instructed how far to sit from the monitor (in centimetres and inches) such that each stimulus subtends approximately 10.5 degrees in height. Although we are unable to assess compliance with viewing instructions, our interest lies in the relative differences across both stimuli and groups. Moreover, it would take a very large error in viewing distance to move the 3 cycle per degree stimulus outside of the desired range of 2–5 cycles per degree. Participants were informed that they should concentrate on the central fixation dot and they would see stripy patterns that, for some people, may elicit experiences such as colors, shimmering or shapes. They were reassured that there is no right or wrong answer. The stimuli were presented for 5 seconds in a random order. Following each stimulus they were asked three questions. First, they were asked how many experiences they had to the stimulus by checking as many or as few options as they liked (colors, bending of lines, blurring of lines, shimmer/flicker, fading, shadowy shapes, other/specify). As in previous research, the total number of experiences reported is summed (i.e. a score of 0 to 7). The second question was specifically introduced because of our interest in color experiences and asked participants to report the colors that they experienced selecting as many or few as they liked (yellow, red, green, blue, purple, pink, brown, orange). The final question asked how uncomfortable they found it on an 11-point visual analogue scale with endpoints marked as “extremely uncomfortable” (left) and “extremely comfortable” (right). The pointer was initially set at the centre (labelled as “neither comfortable nor uncomfortable”) and participants dragged the pointer with the mouse to indicate their response. Analysis The analyses consisted of a between-groups comparison conducted as a 3 X N ANCOVA with gender and age as covariates. For main group results on a single measure N = 1 and, when considering specific subscales (AQ, GSQ) or stimulus parameters (PGT) then N > 1. Additional Information How to cite this article : Ward, J. et al . Atypical sensory sensitivity as a shared feature between synaesthesia and autism. Sci. Rep. 7 , 41155; doi: 10.1038/srep41155 (2017). Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Concrete links between the symptoms of autism and synaesthesia have been discovered and clarified for the first time, according to new research by psychologists at the University of Sussex. The study, conducted by world-leading experts in both conditions at Sussex and the University of Cambridge and published in the journal Scientific Reports, found that both groups experience remarkably similar heightened sensory sensitivity, despite clear differences in communicative ability and social skills. Two previous studies had found an increased prevalence of synaesthesia in autistic subjects, suggesting that although they are not always found in conjunction, the two conditions occur together more often than would be expected by chance alone. However, this is the first study that has attempted to draw a definitive symptomatic link between the two. Synaesthesia and autism seem on the surface to be rather different things, with synaesthesia defined as a 'joining of the senses' in which music may trigger colours or words may trigger tastes, and autism defined by impaired social understanding and communication. The new research shows that both groups report heightened sensory sensitivity, such as an aversion to certain sounds and lights, as well as reporting differences in their tendency to attend to detail. However, the synaesthetes tended not to report difficulties on the traditional communicative symptoms that usually define autism. While the research shows that there are certainly links between the two conditions, these appear to be sensory rather than social. The study was led by Professor Jamie Ward, Professor of Cognitive Neuroscience and Co-Director Sussex Neuroscience group, alongside Sussex Psychology colleague, Professor Julia Simner; and Professor Simon Baron-Cohen, Professor of Developmental Psychopathology at the University of Cambridge and Director of the Autism Research Centre. Commenting on the research, Prof Ward said: "Synaesthesia has traditionally been considered more of a gift than an impairment, whereas the opposite could often be said of autism. Our research suggests that the two have much more in common than was previously thought, and that many of the sensory traits that autistic people possess are also found in those who experience synaesthesia. "Though further research is required, our understanding of autism in the context of synaesthetic abilities may help us unlock the secrets of some of the more positive aspects of autism, such as savantism, while also uncovering further neurological links between the two conditions." Another research paper by the group of researchers, looking more closely at the question of savantism in people with autism, is also due to be published soon. Reinforcing their initial research, it shows that synaesthesia tends to be particularly prevalent in people with autism who also have unexpected 'savant' abilities, such as superior abilities in arithmetic, memory and art. Prof Ward added: "Though some theories propose a causal link between increased sensory sensitivity and impaired social functioning in people with autism, our research so far demonstrates the value of considering synaesthesia on the same spectrum as autism from a sensory point of view. We hope in future to be able to continue to explore the relationship between perceptual, cognitive and social symptoms and abilities in autistic and synaesthetic people."

10.1038/srep41155

Biology

Mechanical weeding promotes ecosystem functions and profit in industrial oil palm, finds study

Najeeb Al-Amin Iddris et al, Mechanical weeding enhances ecosystem multifunctionality and profit in industrial oil palm, Nature Sustainability (2023). DOI: 10.1038/s41893-023-01076-x Journal information: Nature Sustainability

https://dx.doi.org/10.1038/s41893-023-01076-x

https://phys.org/news/2023-03-mechanical-weeding-ecosystem-functions-profit.html

Abstract Oil palm is the most productive oil crop, but its high productivity is associated with conventional management (that is, high fertilization rates and herbicide application), causing deleterious environmental impacts. Using a 2 2 factorial experiment, we assessed the effects of conventional vs reduced (equal to nutrients removed by fruit harvest) fertilization rates and herbicide vs mechanical weeding on ecosystem functions, biodiversity and profitability. Analysing across multiple ecosystem functions, mechanical weeding exhibited higher multifunctionality than herbicide treatment, although this effect was concealed when evaluating only for individual functions. Biodiversity was also enhanced, driven by 33% more plant species under mechanical weeding. Compared with conventional management, reduced fertilization and mechanical weeding increased profit by 12% and relative gross margin by 11% due to reductions in material costs, while attaining similar yields. Mechanical weeding with reduced, compensatory fertilization in mature oil palm plantations is a tenable management option for enhancing ecosystem multifunctionality and biodiversity and increasing profit, providing win–win situations. Main Oil palm production has increasingly expanded in large areas of Southeast Asia, with Indonesia currently the world’s largest producer of palm oil 1 , which also coincides with increased rates of deforestation in the country 2 . It is estimated that between 2001 and 2019, oil palm plantations were responsible for 32% of the total forest area lost in Indonesia 3 . Despite the high environmental costs, oil palm production is highly attractive because of its short-term economic returns and increasing global demand for food, fuel and cosmetics 4 . Industrial oil palm plantations (>50 ha planted area and owned by corporations 5 ) are currently estimated to account for about 60% of the total cultivated oil palm area in Indonesia 6 . Compared with smallholder farms (typically <50 ha of land and owned by individuals 5 ), the productivity of industrial oil palm plantations is ~50% higher 7 but is largely driven by high fertilizer and herbicide applications to optimize productivity 8 at the expense of other ecosystem processes. In contrast to the forests, industrial oil palm plantations tend to be structurally simplified and highly disturbed 9 , 10 , 11 and have reduced capacity to provide several ecosystem functions simultaneously (so-called ‘multifunctionality’ of ecosystems) 4 , 12 . Although various studies have addressed the deleterious effects of oil palm expansion on forest and biodiversity losses 4 , 12 , 13 , to date, there has been no holistic multiyear and spatially replicated assessment of the effect of different oil palm management strategies on ecosystem functions, biodiversity and economic productivity. Primary, experimental research studies that assess the effect of management on multiple outcomes, although presently rare, are pertinent to providing clearer oil palm management recommendations 14 because oil palm plantations are inherently complex agricultural systems. Application of fertilizer and herbicide at high frequencies are arguably the most important management activities in industrial oil palm plantations due to their effect on oil palm yield. Thus, a major leverage for a more environmentally sustainable production of palm oil is identifying optimal fertilizer application rates and weeding method to maintain sufficiently high productivity and economic profitability while minimizing associated losses of ecosystem functions and biodiversity. Fertilization contributes to maintaining high productivity levels but also represents a substantial share of management costs in industrial plantations. High fertilization rates are associated with high nutrient-leaching losses 15 and greenhouse gas (GHG) emissions 11 , 16 . Additionally, over-fertilization can have multiple deleterious effects on the soil, including reductions in soil microbial biomass 17 , pH, base cation availability 18 and notable changes in the composition of bacterial and fungal communities 19 . Economically, excess fertilization of oil palm may reduce profitability due to saturation of the yield curve despite increasing fertilizer application so that the linear increase in fertilizer costs could substantially diminish net returns 20 . This is vital to consider given global reduced availability and consequent price increase of fertilizers 21 . The effects of over-fertilization could be reduced by adjusting fertilization rates to levels that compensate only for the quantity of nutrients exported through harvest and through organic fertilization by putting back palm litter 22 and other processing by-products (for example, palm oil mill effluents and empty fruit bunches) 23 . Such a reduction in fertilization intensity may consequently promote efficient retention and recycling of soil nutrients, decrease soil acidification and the associated need for lime application and generally increase profitability due to fertilizer cost savings and hence present a more ecologically and economically attractive option. Another important management practice in industrial oil palm plantations is weed control, needed to balance the positive and potential negative effects of the understory vegetation and to facilitate access for management operations. Industrial plantations conventionally employ the use of herbicides for easy and quicker removal of the understory vegetation, but this is often complemented by mechanical weeding, for example, during rainy periods and for removing woody plants. Herbicide weed control results in lower understory vegetation diversity 24 while promoting the invasion of herbicide-resistant weeds 25 . It also reduces the habitat complexity, strongly impacting vegetation-dwelling biodiversity 26 , 27 , 28 . A more sustainable alternative could be mechanical weeding, which allows for fast regeneration of the understory vegetation due to preservation of their roots. This consequently leads to increased understory plants cover and diversity which can promote nutrient cycling 29 , 30 and enhance habitat complexity, supporting animal abundance and diversity 26 , 27 , 28 . Nevertheless, these ecologically desirable effects may have adverse economic effects as mechanical weeding requires higher labour input compared with herbicide application, which may reduce profits. Additionally, competition between oil palms and understory vegetation may put yields at risk, particularly during drought periods, given the sensitivity of fructification and fruit bunch weight to water stress 31 , 32 . In this study, we report primary data from a multidisciplinary management experiment in a state-owned industrial oil palm plantation in Jambi, Indonesia, that evaluates whether reduced management (that is, reduced fertilization rates and mechanical weeding) as an alternative to conventional management (that is, high fertilization rates and herbicide use) can reduce the negative impacts on ecosystem functions and biodiversity while maintaining high and stable production levels 8 . A full factorial field experiment of two fertilization rates × two weeding practices, with conventional vs reduced (that is, equal to nutrients exported with harvest) fertilization rates and herbicide vs mechanical weeding (that is, brush cutter), was established in 2016 in a mature (≥16 years old) industrial oil palm plantation. Treatments had four replicate plots of 50 m × 50 m each. During the first four years of the experiment, we measured indicators of eight ecosystem functions, seven indicators of biodiversity and six economic indicators linked to the level and stability of yield and profit (Supplementary Tables 1 and 2 ). We hypothesized that compared with conventional high fertilization rate and herbicide treatment, reduced fertilization rate and mechanical weeding will enhance ecosystem functions and biodiversity while maintaining high productivity and increased profit. Results Management effects on ecosystem functions and biodiversity Ecosystem multifunctionality, an aggregated measure of multiple ecosystem functions, was, on average, higher in mechanical weeding than herbicide treatment (Fig. 1 and Table 1 ). Similarly, threshold multifunctionality, calculated as the number of ecosystem functions above a specific threshold (Methods and Supplementary Fig. 3 ), showed a higher number of functions in mechanical weeding than herbicide treatment that were above 70% and 90% thresholds and marginally significant at thresholds of 50% and 80% (Table 1 and Supplementary Fig. 3 ). Although neither fertilization nor weeding effect was detected when analysing only for each individual ecosystem function (Table 1 and Supplementary Fig. 1 ), the higher multifunctionality of mechanical weeding over herbicide treatment was brought about by the higher mean z values of measured indicators of litter decomposition, soil fertility, water filtration and plant refugium (Supplementary Fig. 2 ). Fig. 1: Multifunctionality in different fertilization and weeding treatments in an industrial oil palm plantation in Jambi, Indonesia. Box plots (25th percentile, median and 75th percentile) and whiskers (1.5 × interquartile range) are based on the z-standardized values ((actual value − mean value across replicate plots) / standard deviation) of indicators for a specified ecosystem function. N = 4 plots. The black horizontal lines indicate the mean multifunctionality of eight ecosystem functions. 2 2 factorial treatments: + indicates conventional fertilization and herbicide treatment; − denotes reduced fertilization and mechanical weeding. Different letters denote that ecosystem multifunctionality was higher in mechanical weeding than herbicide treatment (linear mixed-effects model at P = 0.03; Table 1 ). Full size image Table 1 Statistical results using linear mixed-effects model (LME) for ecosystem functions and biodiversity and linear model (LM) for multifunctionality and economic indicators on the effects of different fertilization and weeding treatments ( N = 4 plots) in an industrial oil palm plantation in Jambi, Indonesia Full size table Biodiversity, estimated as the average of indicators of taxonomic richness across seven trophic groups, was also higher in the mechanical weeding than herbicide treatment (Fig. 2 and Table 1 ). This effect was mainly driven by the strong increase in understory vegetation species richness in the mechanical weeding during four-year measurements (2017–2020; Fig. 2 and Supplementary Table 2 ). Nevertheless, when removing understory plant species richness from the biodiversity index, there was still a positive marginal effect of mechanical weeding on biodiversity ( P = 0.09). Of the 126 understory plant species in the plantation, 33% more plant species occurred in the reduced management than in conventional management. Across the years, the most abundant plant species were the herbicide-resistant invasive shrub Clidemia hirta (L.) D.Don, the invasive herb Asystasia gangetica subsp. micrantha (Nees) Ensermu, the native grass Centotheca lappacea (L.) Desv. and the native fern Christella dentata (Forssk.) Brownsey & Jermy. Compared with herbicide treatment, mechanical weeding increased the ground cover of all these plant species ( P < 0.04) except for C. hirta for which ground cover was reduced by 55% ( P = 0.01). Fig. 2: Biodiversity taxonomic richness in different fertilization and weeding treatments in an industrial oil palm plantation in Jambi, Indonesia. Box plots (25th percentile, median and 75th percentile) and whiskers (1.5 × interquartile range) are based on the z-standardized values ((actual value − mean value across replicate plots) / standard deviation) of indicators for a specified trophic group. N = 4 plots. The black horizontal lines indicate the mean biodiversity richness of seven multitrophic groups. 2 2 factorial treatments: + indicates conventional fertilization and herbicide treatment; − denotes reduced fertilization and mechanical weeding. Different letters denote that biodiversity richness was higher in mechanical weeding than herbicide treatment (linear mixed-effects model at P = 0.01; Table 1 ). Full size image Management effects on yield and profit The four-year cumulative yield did not differ among treatments (Fig. 3a and Table 1 ). To assess the notion that reduced fertilization (with or without mechanical weeding) will have less yield stability than conventional fertilization (with or without herbicide treatment), we looked at the left end of the yield distribution. We found that cumulative yield of palms with the lowest performance (that is, the lower fifth quantile of yield per palm per plot) did not differ among treatments (Fig. 3b ). Additionally, the probability of the palm yield to fall below 75% of the average yield (that is, yield shortfall probability; Fig. 3c ) was similar among treatments, indicating that replacing herbicide with mechanical weeding did not have a negative effect on worst-case yield. When analysing separately for 2017–2018 and 2019–2020, we did not find any effect of treatments on yield and profit indicators (Supplementary Fig. 4 ), signifying that there was no evidence for possible delayed yield reductions in response to reduced management treatment during the four-year measurements. Fig. 3: Yield, costs and profit in different fertilization and weeding treatments in an industrial oil palm plantation in Jambi, Indonesia. Oil palm yield ( a ), yield per palm with the lowest performance ( b ), share of palms per plot which fell below 75% of the average yield ( c ), costs ( d ), profit ( e ) and relative gross margin ( f ) (cumulative values for 2017–2020) in different fertilization and weeding treatments ( N = 4 plots) in an industrial oil palm plantation in Jambi, Indonesia. Box plots indicate the 25th percentile, median and 75th percentile, and whiskers are 1.5 × the interquartile range. 2 2 factorial treatments: + Fert, + Herb are conventional fertilization with herbicide treatment; − Fert, + Herb are reduced fertilization with herbicide treatment; + Fert, − Herb are conventional fertilization with mechanical weeding; − Fert, − Herb are reduced fertilization with mechanical weeding. Full size image The main management costs considered in the plantation were material (chemical products) costs and labour costs for harvest and weeding operations. Material costs in the reduced fertilization and mechanical weeding treatment were 41% lower (Fig. 3d ) than in the conventional fertilization and herbicide treatment (Table 1 and Supplementary Table 2 ). Conversely, the labour costs were 10% higher in the mechanical weeding than in the herbicide treatment (Fig. 3d and Supplementary Table 2 ). The most intensive labour activity was harvesting (accounting for 39–45% of total labour costs) and cleaning of the palm circle (12–14% of total labour costs), which were the same in both weeding treatments. Consequently, profit was 12% higher (Fig. 3e ) in mechanical weeding than herbicide treatment (Table 1 and Supplementary Table 2 ). There was an increase in relative gross margin in the reduced fertilization and mechanical weeding by 11% (Fig. 3f ) compared with the conventional fertilization and herbicide treatment (Table 1 and Supplementary Table 2 ). Discussion Reduced management (reduced fertilization rates and mechanical weeding) promoted ecosystem functioning and maintained high levels of yield and profit, providing a win–win situation (Fig. 4 ). In particular, the mechanical removal of weeds instead of using herbicide was found as a tenable management practice that enhanced ecosystem multifunctionality and understory plant diversity, which together with reduced fertilization, increased profit. Fig. 4: Conventional vs reduced management and their associated ecosystem and economic functions in an industrial oil palm plantation in Jambi, Indonesia. For each petal, the centre (fifth quantile) and the outer edge (95th quantile) are based on the z-standardized values of eight ecosystem functions (purple; Fig. 1 ), seven multitrophic richness for biodiversity (green; Fig. 2 ) and six indicators for yield and profit (orange; Fig. 3 ). 2 2 factorial treatments: + indicates conventional fertilization and herbicide treatment; − denotes reduced fertilization and mechanical weeding. Full size image Improved multifunctionality and biodiversity Several ecosystem functions in oil palm plantations tend to be interrelated 23 , which makes it difficult to evaluate the effect of management practices on single functions. Analysing across multiple ecosystem functions, mechanical weeding clearly exhibited higher multifunctionality than herbicide treatment (Fig. 1 and Supplementary Fig. 3 ), which was concealed when analysing only for a single ecosystem function (Supplementary Fig. 1 ). To note, the indicators of GHG regulation function revealed that this mature plantation located on mineral soils was a slight C source 16 (that is, as indicated by the net ecosystem productivity (NEP) of which fruit harvest was subtracted; Supplementary Table 2 ). Additionally, it had also large soil N 2 O emissions due to the high N fertilization rate, which is ~50–75% higher than the N fertilizers applied in smallholder oil palm plantations 11 . Nevertheless, the lack of management effects on GHG regulation function or on any single ecosystem function may be due to the relatively short-term span of this management experiment, which was 2016–2020 against at least 16 years of conventional management when this plantation was established (1998–2000). The legacy effect of the long-term conventional management practices (Methods) may have dampened the effects of the reduced management practices. Additionally, some of the ecosystem indicators were measured as early as one year after the start of the management experiment (Supplementary Table 1 ). Indeed, studies that investigated the effects of reduced fertilization and mechanical weeding on root-associated soil biota one year after the start of the experiment found no significant treatment effects, which they attributed to legacy effect of conventional fertilization and herbicide use 33 . Similarly, different weeding treatments did not affect vegetation cover (our indicator for erosion prevention), litter decomposition rates (one of our indicators for organic matter decomposition) and soil physical and biochemical characteristics during the first one to two years after treatment 15 , 26 . When considering across the eight ecosystem functions, within four years of a shift from herbicide to mechanical weeding, an increased ecosystem multifunctionality was achieved (Fig. 1 and Supplementary Fig. 3 ). Mechanical removal of weeds promoted fast vegetation regrowth due to the preservation of the root biomass, which may have resulted in positive feedback effects on plant diversity 24 and ultimately on multifunctionality (Figs. 1 and 2 ). Experimental studies in mature oil palm plantations that compared herbicide use with manual removal of weeds have reported substantially higher levels of plant biomass and cover in manual weeding 24 , with greater potential to support ecosystem functions 26 . We expect the fast recovery of the understory vegetation under mechanical weeding to provide high organic matter input 34 and a more suitable microclimate for soil microbial and faunal activity 26 , which together promote decomposition 26 and soil nutrient retention 35 , 36 . Indeed, the positive effect of mechanical weeding on multifunctionality in our study was largely related to its effect on soil functions, such as decomposition, soil fertility and water filtration and on plant refugium functions (Supplementary Figs. 2 and 3 ). The positive effect of mechanical weeding on biodiversity was mainly driven by plant diversity effects with a small positive effect on other trophic groups (soil microorganisms, soil invertebrates and aboveground insects; Fig. 2 ). We found a strong increase in understory vegetation species richness in the mechanical weeding (Fig. 2 ), which was similar to other findings from oil palm 24 and Eucalyptus 29 plantations when changing from herbicides to manual weeding. An increase in plant diversity enhances nutrient use efficiency and availability 37 through increased decomposition of diverse plant litter and has been linked to enhanced functional diversity, promoting multifunctionality 38 . High plant diversity is likely to create several complex food webs capable of supporting high macrofaunal abundance, which can, in part, stimulate decomposer and pollinator communities 26 , 39 . Thus, an increase in plant diversity may promote biodiversity at higher trophic levels due to bottom-up effects, with direct benefits for primary consumers and soil microbes due to increased diversity of substrate 40 . The method of weeding can also have considerable effects on plant species composition due to the encroachment of non-native, invasive plant species 8 , 24 . Particularly, the suppressing effect of herbicides on vegetation diversity is well documented as it causes selective pressure on weeds and promotes herbicide-resistant species 41 . Accordingly, the plant community in the herbicide treatment plots was dominated by the herbicide-resistant weed C. hirta , but mechanical weeding was effective in slowing down the spread of this species. This has important implications for sustainable management of oil palm plantations as this highly invasive alien species, in addition to being problematic for weed management in the plantation, can also invade tropical forests and threaten their biodiversity 42 . The reduction of C. hirta in the mechanical weeding was also concomitant with a higher cover of the invasive plant A. gangetica subsp. micrantha , an attractive plant for pollinators 43 , and of some native species (such as the grass C. lappacea and the fern C. dentata ), which resulted in a higher plant refugium function (Supplementary Fig. 2 ). Given the four-year span of this management experiment, the positive effects of mechanical weeding on biodiversity and multifunctionality signified that such easily adoptable field practices can reap substantial benefits within a short period. Significant plant diversity effects take some time to propagate through trophic levels, which tend to increase considerably with experimental duration 44 . Therefore, it is possible that the positive effects of mechanical weeding on biodiversity and ecosystem functions could get stronger over time, especially on soil invertebrates and aboveground insects 45 . Employing mechanical weeding during the early stage of oil palm establishment may generate timely benefits on ecosystem multifunctionality and biodiversity, although there is no study so far investigating this in young plantations. Management effect on yield and profit Reduced management (reduced fertilization rates and mechanical weeding) was as effective in attaining similar yield as the conventional management but at lower costs (Fig. 3 , Table 1 and Supplementary Fig. 5 ). The average annual yield of 29.6 Mg ha −1 yr −1 under the reduced management system was very close to the recently published ‘attainable yield’ of 30.6 Mg ha −1 yr −1 for large-scale oil palm plantations, defined as the yield attained with the adoption of economically optimal inputs and management practices 1 . The similar yields among treatments suggest a more efficient use of applied fertilizers in the reduced management system, particularly on N, as shown by a 56% decrease in dissolved N losses in the reduced fertilization and mechanical weeding compared with conventional management (Supplementary Table 2 ) 15 . Given the time for sex ratio differentiation and fruit development of oil palm 20 , it is estimated that the effects of management practices on the yield may have a delayed response of approximately two years 46 . Our yield measurements covered four years of this management experiment, and the lack of treatment effect on four-year cumulative yield and similar yields between two separate periods (that is, 2017–2018 and 2019–2020; Supplementary Fig. 4a ) suggest that reduced fertilization was a more sustainable alternative to existing conventional fertilization regimes that can lead to higher profitability. This was further reinforced by our findings that the reduced management did not lead to an increased risk of only obtaining very low yields, for example, during a dry period in late 2018 to early 2019 caused by El Niño/Southern Oscillation, as shown by our indicators of yield stability (Fig. 3b,c ). The reduced management even tended to show a higher worst-case yield and lower probability to fall under the threshold of 75% of average yield compared with conventional management. It is important to mention that our notion of production risks refers only to yields whereas we disregarded the fluctuations of producer prices for oil palm fruit bunches on the one hand and trends in input prices on the other hand. While declines in producer prices would not lead to any changes in the ranking of the management alternatives, increasing scarcity and consequent price increase of mineral fertilizers might even increase the economic superiority of the reduced fertilization treatments. We found that the reduced fertilization rates would outperform the conventional fertilization rates in terms of profit if fertilizer costs increased by 100% (Supplementary Fig. 6 ). Global fertilizer prices are currently at near-record levels 21 , exceeding those seen during the food and energy crisis in 2008. The high fertilizer prices may last long, especially given the scarcity in P 47 , vastly over-priced K in agriculture 48 and rising consensus on the social costs of excessive N fertilization 49 , which were not monetized here. The high yield under reduced fertilization suggests that this plantation may be over fertilized. However, this statement holds for mature oil palm plantations that have been heavily fertilized since planting, such as the plantation of this study, but may not be true for younger plantations in which the soil has not accumulated nutrients from fertilization and from the decomposition of senesced fronds 22 . Therefore, fertilization rates may need to be adjusted during the life cycle of the plantation, considering the different nutrient accumulations and nutrient requirements over the course of the oil palm life cycle 50 . Lastly, the higher profit and relative gross margin in reduced compared with conventional management treatments were mainly due to a reduction in material cost and the maintenance of high yield (Fig. 3 and Supplementary Fig. 5 ). Due to the low minimum wage in Indonesia (US$184 per month per person), labour cost constituted only a small share of the total management cost (Fig. 3d ) and did not translate into low profit in the reduced management treatment. Yet, an increase in minimum wage in Indonesia or higher wages in other production areas would most likely not translate into a sharp difference in labour costs among treatments. This is because the relative change in the labour cost difference among treatments would still be small because the main time-consuming activities, such as harvesting and cleaning of the palm circle, required the same time in all treatments. Our sensitivity analysis demonstrates that the relative differences in profitability appear robust against changes in assumptions on per unit labour costs. Only at an increase of labour costs by 1,250% would the herbicide treatment (with reduced fertilization rates) become more profitable than the respective mechanical weeding (Supplementary Fig. 6 ). Nevertheless, under labour shortage, it needs to be considered that even if the difference in labour hours was small (approximately 12 extra man-hours ha −1 yr −1 ), mechanical weeding increases total labour hours of plantation management. This management experiment was carried out in an industrial plantation with all works being done by staff members of the plantation within the course of the common plantation management. While we cannot fully exclude any bias, for example, labour time as recorded by researchers observing the operations, we are still convinced that this inherent variation will be smaller than the differences in total management cost, profit and relative gross margin we found between reduced and conventional management treatments (Table 1 ). Overall, the results of our four-year management experiment provided early indications that mechanical weeding, together with reduced, compensatory fertilization rates in mature, industrial oil palm plantations, can help in minimizing soil nutrient leaching, decrease water pollution risk, eliminate the effect of herbicide on native vegetation and other non-target soil trophic groups 33 , reduce risks to health of plantation workers and thus contribute to sustainability guidelines of certification bodies such as Roundtable of Sustainable Oil Palm 51 . We acknowledge that when analysing across multiple ecosystem functions, the positive effects of the reduced management on multifunctionality may be still small during the first four years of this experiment as there was no significant effect when analysing for individual ecosystem functions. However, the improved ecosystem multifunctionality under mechanical weeding exemplifies a win–win situation, given its high yield and profit on one hand and the environmental costs associated with herbicide use on the other. Reduced management is therefore a viable management option to maintain optimal yield, lower material cost and improve biodiversity and ecosystem multifunctionality in industrial oil palm plantations that are located on mineral soils. Methods Ethics No ethics approval was required for this study. Our study was conducted in a state-owned industrial oil palm plantation where we established a cooperation with the estate owner to access the site and collect data. No endangered or protected species were sampled. Research permits were obtained from the Ministry of Research, Technology and Higher Education, and sample collection and sample export permits were obtained from the Ministry of Environment and Forestry of the Republic of Indonesia. Study area and experimental design Our study was conducted in a state-owned industrial oil palm plantation (PTPN VI) located in Jambi, Indonesia (1.719° S, 103.398° E, 73 m above sea level). Initial planting of oil palms within the 2,025 ha plantation area started in 1998 and ended in 2002; planting density was 142 palms ha −1 , spaced 8 m apart in each row and between rows, and palms were ≥16 years old during our study period of 2016–2020. The study sites have a mean annual temperature of 27.0 ± 0.2 °C and a mean annual precipitation of 2,103 ± 445 mm (2008–2017, Sultan Thaha Airport, Jambi). The management practices in large-scale oil palm plantations typically result in three contrasting management zones: (1) a 2 m radius around the base of the palm that was weeded (four times a year) and raked before fertilizer application, hereafter called the ‘palm circle’; (2) an area occurring every second inter-row, where pruned senesced palm fronds were piled up, hereafter called ‘frond piles’; and (3) the remaining area of the plantation where less weeding (two times a year) and no fertilizer were applied, hereafter called ‘inter-rows’. Within this oil palm plantation, we established a management experiment in November 2016 with full factorial treatments of two fertilization rates × two weeding practices: conventional fertilization rates at PTPN VI and other large-scale plantations (260 kg N–50 kg P–220 kg K ha −1 yr −1 ), reduced fertilization rates based on quantified nutrient export by harvest (136 kg N–17 kg P ha −1 yr −1 –187 kg K ha −1 yr −1 ), herbicide and mechanical weeding 15 . The reduced fertilization treatment was based on quantified nutrient export from fruit harvest, calculated by multiplying the nutrient content of fruit bunches with the long-term yield data of the plantation. Fertilizers were applied yearly in April and October following weeding and raking of the palm circle. The common practice at PTPN VI and other large-scale plantations on acidic Acrisol soils is to apply lime and micronutrients, and these were unchanged in our management experiment. Before each N–P–K fertilizer application, dolomite and micronutrients were applied to the palm circle in all treatment plots using the common rates) 52 : 426 kg ha −1 yr −1 dolomite and 142 kg Micro-Mag ha −1 yr −1 (containing 0.5% B 2 O 3 , 0.5% CuO, 0.25% Fe 2 O 3 , 0.15% ZnO, 0.1% MnO and 18% MgO). Herbicide treatment was carried out using glyphosate in the palm circle (1.50 l ha −1 yr −1 , split into four applications per year) and in the inter-rows (0.75 l ha −1 yr −1 , split into two applications per year). Mechanical weeding was done using a brush cutter in the same management zones and at the same frequency as the herbicide treatment. The 2 2 factorial design resulted in four treatment combinations: conventional fertilization with herbicide treatment, reduced fertilization with herbicide treatment, conventional fertilization with mechanical weeding and reduced fertilization with mechanical weeding. The four treatments were randomly assigned on 50 m ×\ 50 m plots replicated in four blocks, totalling 16 plots. The effective measurement area was the inner 30 m × 30 m area within each replicate plot to avoid any possible edge effects. For indicators (below) that were measured within subplots, these subplots were distributed randomly within the inner 30 m × 30 m of a plot. All replicate plots were located on flat terrain and on an Acrisol soil with a sandy clay loam texture. Ecosystem functions and multifunctionality Our study included multiple indicators for each of the eight ecosystem functions 23 , described in details below (Supplementary Tables 1 and 2 ). All the parameters were expressed at the plot level by taking the means of the subplots (that is, biological parameters) or the area-weighted average of the three management zones per plot (that is, soil parameters). (1) Greenhouse gas (GHG) regulation was indicated by NEP, soil organic C (SOC) and soil GHG fluxes. (2) Erosion prevention was signified by the understory vegetation cover during the four-year measurements. (3) Organic matter decomposition was indicated by leaf litter decomposition and soil animal decomposer activity. (4) Soil fertility was signified by gross N mineralization rate, effective cation exchange capacity (ECEC), base saturation and microbial biomass N. (5) Pollination potential was designated by pan-trapped arthropod abundance and nectar-feeding bird activity. As such, it does not quantify the pollination potential for oil palm, which is mainly pollinated by a single weevil species, but rather as a proxy for a general pollination potential for other co-occurring plants. (6) Water filtration (the capacity to provide clean water) was indicated by leaching losses of the major elements. (7) Plant refugium (the capacity to provide a suitable habitat for plants) as signified by the percentage ground cover of invasive plants to the total ground cover of understory vegetation during the four-year measurements. (8) Biological control (the regulation of herbivores via predation) was indicated by insectivorous bird and bat activities and the soil arthropod predator activity. All the ecosystem functions were merged into a multifunctionality index using the established average and threshold approaches 12 . For average multifunctionality, we first averaged the z-standardized values (Statistics) of indicators for each ecosystem function and calculated the mean of the eight ecosystem functions for each plot. For threshold multifunctionality, this was calculated from the number of functions that exceeds a set threshold, which is a percentage of the maximum performance level of each function 12 ; we investigated the range of thresholds from 10% to 90% to have a complete overview. The maximum performance was taken as the average of the three highest values for each indicator per ecosystem function across all plots to reduce effect of potential outliers. For each plot, we counted the number of indicators that exceeded a given threshold for each function and divided by the number of indicators for each function 12 . Indicators of GHG regulation We calculated annual NEP for each plot as: net ecosystem C exchange – harvested fruit biomass C (ref. 16 ), whereby net ecosystem C exchange = C out (or heterotrophic respiration) – C in (or net primary productivity) 53 . The net primary productivity of oil palms in each plot was the sum of aboveground biomass production (aboveground biomass C + frond litter biomass C input + fruit biomass C) and belowground biomass production. Aboveground biomass production was estimated using allometric equations developed for oil palm plantations in Indonesia 54 , using the height of palms measured yearly from 2019 to 2020. Annual frond litter biomass input was calculated from the number and dry mass of fronds pruned during harvesting events of an entire year in each plot and was averaged for 2019 and 2020. Aboveground biomass production was converted to C based on C concentrations in wood and leaf litter 55 . Annual fruit biomass C production (which is also the harvest export) was calculated from the average annual yield in 2019 and 2020 and the measured C concentrations of fruit bunches. Belowground root biomass and litter C production were taken from previous work in our study area 55 , and it was assumed constant for each plot. Heterotrophic respiration was estimated for each plot as: annual soil CO 2 C emission (below) × 0.7 (based on 30% root respiration contribution to soil respiration from a tropical forest in Sulawesi, Indonesia 56 ) + annual frond litter biomass C input × 0.8 (~80% of frond litter is decomposed within a year in this oil palm plantation 8 ). SOC was measured in March 2018 from composite samples collected from two subplots in each of the three management zones per plot down to 50 cm depth. Soil samples were air dried, finely ground and analysed for SOC using a CN analyser (Vario EL Cube, Elementar Analysis Systems). SOC stocks were calculated using the measured bulk density in each management zone, and values for each plot were the area-weighted average of the three management zones (18% for palm circle, 15% for frond piles and 67% for inter-rows) 15 , 22 . From July 2019 to June 2020, we conducted monthly measurements of soil CO 2 , CH 4 and N 2 O fluxes using vented, static chambers permanently installed in the three management zones within two subplots per plot 11 , 57 . Annual soil CO 2 , CH 4 and N 2 O fluxes were trapezoidal interpolations between measurement periods for the whole year, and values for each plot were the area-weighted average of the three management zones (above). Indicators of erosion prevention Diversity and abundance of vascular plants were assessed once a year from 2016 to 2020 before weeding in September–November. In five subplots per plot, we recorded the occurrence of all vascular plant species and estimated the percent cover of the understory vegetation. The percentage cover and plant species richness of each measurement year were expressed in ratio to that of 2016 to account for initial differences among the plots before the start of the experiment. For example, percentage cover in 2017 was: $$\mathrm{Cover}_{2017} = \frac{{\left( {\mathrm{Cover}_{2017} - \mathrm{Cover}_{2016}} \right)}}{{\mathrm{Cover}_{2016}}}$$ The values from five subplots were averaged to represent each plot. Indicators of organic matter decomposition Leaf litter decomposition was determined using litter bags (20 cm × 20 cm with 4 mm mesh size) containing 10 g of dry oil palm leaf litter 8 . Three litter bags per plot were placed on the edge of the frond piles in December 2016. After eight months of incubation in the field, we calculated leaf litter decomposition as the difference between initial litter dry mass and litter dry mass following incubation. Soil animal decomposer activity is described below (Soil arthropods). Indicators of soil fertility All these indicators were measured in February–March 2018 in the three management zones within two subplots per plot 22 . Gross N mineralization rate in the soil was measured in the top 5 cm depth on intact soil cores incubated in situ using the 15 N pool dilution technique 58 . ECEC and base saturation were measured in the top 5 cm depth as this is the depth that reacts fast to changes in management 22 . The exchangeable cation concentrations (Ca, Mg, K, Na, Al, Fe, Mn) were determined by percolating the soil with 1 mol l −1 of unbuffered NH 4 Cl, followed by analysis of the percolates using an inductively coupled plasma-atomic emission spectrometer (ICP-AES; iCAP 6300 Duo view ICP Spectrometer, Thermo Fisher Scientific). Base saturation was calculated as the percentage exchangeable bases (Mg, Ca, K and Na) on ECEC. Microbial biomass N was measured from fresh soil samples using the fumigation-extraction method 59 . The values for each plot were the mean of the two subplots that were the area-weighted average of the three management zones (above) 15 , 22 . Indicators of general pollination potential Fluorescent yellow pan traps were used to sample aboveground arthropods (to determine pollinator communities 60 ) in November 2016, September 2017 and June 2018. The traps were attached to a platform at the height of the surrounding vegetation within a 2 × 3 grid centred in the inter-rows of each plot in six clusters of three traps, totalling 18 traps per plot. Traps were exposed in the field for 48 h. We stored all trapped arthropods in 70% ethanol and later counted and identified to order and species level. The abundance of trapped arthropods in 2017 and 2018 were calculated as the ratio to the abundance in 2016 to account for initial differences among the plots before the start of the experiment. The activity of nectar-feeding birds is described below (Birds and bats). Indicators of water filtration Element leaching losses were determined from analyses of soil-pore water sampled monthly at 1.5 m depth using suction cup lysimeters (P80 ceramic, maximum pore size 1 μm; CeramTec) over the course of one year (2017–2018) 15 . Lysimeters were installed in the three management zones within two subplots per plot. Dissolved N was analysed using continuous flow injection colorimetry (SEAL Analytical AA3, SEAL Analytical), whereas these other elements were determined using ICP-AES. The values for each plot were the mean of the two subplots that were the area-weighted average of the three management zones 15 , 22 . Indicators of plant refugium In five subplots per plot, the percentage cover and species richness of invasive understory plant species were assessed once a year from 2016 to 2020 before weeding in September–November. We defined invasive species as those plants non-native to Sumatra 61 and among the ten dominant species (excluding oil palm) in the plantation for each year. The percentage cover of invasive understory plant species of each measurement year was expressed in a ratio to that of 2016 to account for initial differences among the plots before the start of the experiment. The values for each plot were represented by the average of five subplots. Indicators of biological control The activities of insectivorous birds and bats are described below (Birds and bats). In five subplots per plot, soil invertebrates were collected (Soil arthropods), counted, identified to taxonomic order level and subsequently classified according to their trophic groups that include predators 60 . The values from five subplots were average to represent each plot. Biodiversity Biodiversity was measured by the taxonomic richness of seven multitrophic groups, described in details below (Supplementary Tables 1 and 2 ). Understory plant species richness The method is described above (Indicators of erosion prevention), using the number of species as an indicator (Supplementary Table 2 ). Soil microorganism richness This was determined in May 2017 by co-extracting RNA and DNA from three soil cores (5 cm diameter, 7 cm depth) in five subplots per plot 62 . While DNA extraction describes the entire microbial community, RNA represents the active community. The v3–v4 region of the 16S rRNA gene was amplified and sequenced with a MiSeq sequencer (Illumina). Taxonomic classification was done by mapping curated sequences against the SILVA small subunit (SSU) 138 non-redundant (NR) database 63 with the Basic Local Alignment Search Tool (BLASTN) 64 . Soil arthropod order richness For determination of soil arthropods, we collected soil samples (16 cm × 16 cm, 5 cm depth) in five subplots per plot in October–November 2017. We extracted the animals from the soil using a heat-gradient extractor 65 , collected them in dimethyleneglycol-water solution (1:1) and stored in 80% ethanol. The extracted animals were counted and identified to taxonomic order level 61 . They were also assigned to the trophic groups decomposers, herbivores and predators based on the predominant food resources recorded in previous reviews and a local study 66 , 67 . Orders with diverse feeding habits were divided into several feeding groups, for example, Coleoptera were divided into mostly predatory families (Staphylinidae, Carabidae), herbivorous families (for example, Curculionidae) and decomposer families (for example, Tenebrionidae). The total number of individuals per taxonomic group in each subplot was multiplied by the group-specific metabolic rate, which were summed to calculate soil animal decomposer activity. The values from five subplots were average to represent each plot. Aboveground arthropod order and insect family richness In addition to the fluorescent yellow pan traps described above (Indicators of general pollination potential), sweep net and Malaise trap samplings were conducted in June 2018, which targeted the general flying and understory dwelling arthropod communities. Sweep net sampling was conducted within the understory vegetation along two 10 m long transects per plot, with ten sweeping strokes performed per transect. In each plot, we installed a single Malaise trap between two randomly chosen palms and exposed it for 24 h. Arthropods were counted, identified to taxonomic order level and the insects to taxonomic family level and values from the three methods were summed to represent each plot. Birds and bats Birds and bats passing at each replicate plot were sampled in September 2017 using SM2Bat + sound recorders (Wildlife Acoustics) with two microphones (SMX-II and SMX-US) placed at a height of 1.5 m in the middle of each plot 68 . We assigned the bird vocalization to species with Xeno-Canto 69 and the Macaulay library 70 . Insectivorous bat species richness was computed by dividing them into morphospecies based on the characteristics of their call (call frequency, duration, shape). In addition, we gathered information on proportional diet preferences of the bird species using the EltonTrait database 71 . We defined birds feeding on invertebrates (potential biocontrol agents) as the species with a diet of at least 80% invertebrates and feeding on nectar (potential pollinators), if the diet included at least 20% of nectar. Economic indicators We used six indicators linked to the level and stability of yield and profit: yield, lower fifth quantile of the yield per palm per plot, shortfall probability, management costs, profit and relative gross margin. We assessed fruit yield by weighing the harvested fruit bunches from each palm within the inner 30 m × 30 m area of each plot. The harvest followed the schedule and standard practices of the plantation company: each palm was harvested approximately every ten days and the lower fronds were pruned. For each plot, we calculated the average fruit yield per palm and scaled up to a hectare, considering the planting density of 142 palms per ha. Because the palms in each plot have different fruiting cycles and were harvested continuously, the calculation of an annual yield may lead to misleading differences between treatments. Therefore, we calculated the cumulative yield from the beginning of the experiment to four years (2017–2020), which should account for the inter- and intra-annual variations in fruit production of the palms in the plots and thus allowing for comparison among treatments. As effects of management practices on yield may be delayed 46 , we also calculated the cumulative yield during two consecutive years (2017–2018 and 2019–2020) and checked for treatment effects on yield and profit indicators separately for these two periods. We computed risk indicators on the cumulative yield and on the yield between the two periods. We used the lowest fifth quantile of the yield per palm per plot (left side of the distribution) to indicate the production of the palms with lowest performance. Also, we determined the yield shortfall probability (lower partial moment 0th order), defined as the share of palms that fell below a predefined threshold of yield; the thresholds chosen were 630 kg −1 per palm for cumulative yield and 300 kg −1 per palm per year for the two-year yield, which corresponded to 75% of the average yield. Revenues and costs were calculated as cumulative values during four years of the experiment (2017–2020) using the same prices and costs for all the years. This was because we were interested in assessing the economic consequences of different management treatments, and they might be difficult to interpret when changes in prices and costs between calendar years are included, which are driven by external market powers rather than the field-management practices. For the same reason, we abstained from discounting profits. Given the usually high discount rates applied to the study area, slight differences in harvesting activities between calendar years or months might lead to high systematic differences between the management treatments, which are associated with the variation in work schedule within the plantation rather than the actual difference among management treatments. Revenues were calculated from the yield and the average price of the fruit bunches in 2016 and 2017 61 . Material costs were the sum of the costs of fertilizers, herbicide and gasoline for the brush cutter. Labour costs were calculated from the minimum wage in Jambi and the time (in labour hours) needed for the harvesting, fertilizing and weeding operations, which were recorded in 2017 for each plot. The weeding labour included the labour for raking the palm circle before fertilization, which was equal in all treatments, and the weeding in the palm circle and inter-rows either with herbicide or brush cutter. In addition, we included the time to remove C. hirta , which must be removed mechanically from all plots once a year, calculated from the average weed-removal time in the palm circle and the percentage cover of C. hirta in each plot for each year. We then calculated the profit as the difference between revenues and the total management costs and the relative gross margin as the gross profit proportion of the revenues. Statistics To test for differences among management treatments for each ecosystem function and across indicators of biodiversity, the plot-level value of each indicator was first z standardized (z = (actual value − mean value across plots) / standard deviation) 4 . This prevents the dominance of one or few indicators over the others, and z standardization allows several distinct indicators to best characterize an ecosystem function or biodiversity 4 . Standardized values were inverted (multiplied by −1) for indicators of which high values signify undesirable effect (that is, NEP, soil N 2 O and CH 4 fluxes, element leaching losses, invasive plant cover, yield shortfall, management costs) for intuitive interpretations. For a specific ecosystem function (Supplementary Figs. 1 and 2 ) and across indicators of biodiversity (Fig. 2 ), linear mixed-effects (LME) models were used to assess differences among management treatments (fertilization, weeding and their interaction) as fixed effects with replicate plots and indicators (Supplementary Tables 2 and 3 ) as random effects. The significance of the fixed effects was evaluated using ANOVA 72 . The LME model performance was assessed using diagnostic residual plots 73 . As indicator variables may systematically differ in their responses to management treatments, we also tested the interaction between indicator and treatment (Table 1 ). For testing the differences among management treatments across ecosystem functions (that is, multifunctionality; Fig. 1 ), we used for each replicate plot the average of z-standardized indicators of each ecosystem function and ranges of thresholds (that is, number of functions that exceeds a set percentage of the maximum performance of each function 12 ; Supplementary Fig. 3 ). The LME models had management treatments (fertilization, weeding and their interaction) as fixed effects and replicate plots and ecosystem functions as random effects; the interaction between ecosystem function and treatment were also tested to assess if there were systematic differences in their responses to management treatments (Table 1 ). As we expected that the type of weeding will influence ground vegetation, we tested for differences in ground cover of understory vegetation, measured from 2016 to 2020, using LME with management treatments as fixed effect and replicate plots and year as random effects. Differences among management treatments (fertilization, weeding and their interaction) in yield and profit indicators, which were cumulative values over four years (Fig. 3 ) or for two separate periods (2017–2018 and 2019–2020; Supplementary Fig. 4 ), were assessed using linear model ANOVA (Table 1 ). For clear visual comparison among management treatments across ecosystem functions, multitrophic groups for biodiversity, and yield and profit indicators, the fifth and 95th percentiles of their z-standardized values were presented in a petal diagram (Fig. 4 and Supplementary Fig. 5 ). Data were analysed using R (version 4.0.4), using the R packages ‘nlme’ and ‘influence.ME’ 73 . Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Data availability Data on the indicators of ecosystem functions, biodiversity and economic productivity are publicly available from the Göttingen Research Online repository: (ref. 74 ). Taxonomic classification of ribosomal RNA was done by mapping curated sequences against the SILVA SSU 138 NR database ( ). Data on proportional diet preferences of the bird species were gathered from the EltonTrait database ( ). Bird vocalizations were assigned to species level using Xeno-Canto (Xeno-Canto Foundation, 2012) and the Macaulay library ( ).

Oil palm trees are the most productive oil crop and global demand is increasing. However, their productivity is due to conventional management practices including high fertilizer usage and herbicide application, resulting in severe environmental damage. A new study by an international, multidisciplinary research team led by the University of Göttingen, shows that shifting to mechanical weeding and reducing fertilizer usage lead to significant increases in both ecosystem multifunctionality and profit. The scientists compared different environmental measures and economic indicators in mechanical weeding, herbicide application, and combinations of these with high and reduced fertilizer usage. Their study was published in the journal Nature Sustainability. Oil palm production has increased in Indonesia, currently the world's largest producer of palm oil, which coincides with the country's increased rate of deforestation. Although oil palm production has brought socio-economic benefits, it also causes environmental problems such as biodiversity loss, nitrate leaching and greenhouse gas emissions. The team carried out their research in plantations that were at least 16 years old, starting in 2016 in Jambi, Indonesia, with the aim of testing reduced management against conventional practices. They looked at the effects of reducing fertilizer usage to compensate for the amount of nutrients taken out with harvests of oil palm fruits, and mechanical weeding using a brush cutter. Over four years, the researchers collected data on oil palm yield, material and labor costs, animals both in the soil and above ground, diversity of ground vegetation, greenhouse gas emissions, soil fertility, and nutrient leaching. Oil palm tree plantation in Jambi province, Sumatra (Indonesia) . Credit: Oliver van Straaten "Despite reduced fertilizer usage, the oil palm yields are similar to conventional management but profit significantly increases due to reductions in fertilizer costs. Biodiversity also significantly improves, driven by increased ground vegetation species with mechanical weeding," says first author Dr. Najeeb Al-Amin Iddris from the University of Göttingen's Soil Science of Tropical and Subtropical Ecosystems. As ecosystem functions tend to be interrelated, analysis was conducted across multiple ecosystem functions, known as "multifunctionality." "Mechanical weeding shows significantly higher ecosystem multifunctionality than herbicide application. It promotes fast recovery of ground vegetation and increases its species diversity, which may enhance recycling of nutrients via root uptake, and in combination with reduced fertilizer-usage decreases leaching and increases nutrient retention in the soil," explains Iddris. "The study found no reduction in greenhouse gas emissions with reduced fertilizer usage and mechanical weeding during four years of this experiment. The consequences of more than 15 years of conventional management prior to the start of the experiment may have dampened the effects of reduced management," explains lead author Dr. Marife Corre, University of Göttingen. "The positive effects of mechanical weeding on ecosystem multifunctionality and profit show that such smart practices can generate benefits even over a short period. Employing mechanical weeding during the early stage of establishing an oil palm plantation may reap even greater benefits," says Corre. "These management practices can be easily adopted in the field and should be included as one of the criteria for production of sustainable palm oil, as set out by the Roundtable for Sustainable Palm Oil, an organization of oil palm producers, processors, manufacturers, investors, environmental groups, and social development stakeholders."

10.1038/s41893-023-01076-x

Medicine

Scientists identify how some angiogenic drugs used to treat cancer and heart disease cause vascular disease

Macarena Fernández-Chacón et al, Incongruence between transcriptional and vascular pathophysiological cell states, Nature Cardiovascular Research (2023). DOI: 10.1038/s44161-023-00272-4

https://dx.doi.org/10.1038/s44161-023-00272-4

https://medicalxpress.com/news/2023-05-scientists-angiogenic-drugs-cancer-heart.html

Abstract The Notch pathway is a major regulator of endothelial transcriptional specification. Targeting the Notch receptors or Delta-like ligand 4 (Dll4) dysregulates angiogenesis. Here, by analyzing single and compound genetic mutants for all Notch signaling members, we find significant differences in the way ligands and receptors regulate liver vascular homeostasis. Loss of Notch receptors caused endothelial hypermitogenic cell-cycle arrest and senescence. Conversely, Dll4 loss triggered a strong Myc-driven transcriptional switch inducing endothelial proliferation and the tip-cell state. Myc loss suppressed the induction of angiogenesis in the absence of Dll4, without preventing the vascular enlargement and organ pathology. Similarly, inhibition of other pro-angiogenic pathways, including MAPK/ERK and mTOR, had no effect on the vascular expansion induced by Dll4 loss; however, anti-VEGFA treatment prevented it without fully suppressing the transcriptional and metabolic programs. This study shows incongruence between single-cell transcriptional states, vascular phenotypes and related pathophysiology. Our findings also suggest that the vascular structure abnormalization, rather than neoplasms, causes the reported anti-Dll4 antibody toxicity. Main Notch is a cell-to-cell ligand–receptor signaling pathway that has a major influence on cell transcription and biology 1 , playing important roles in several diseases 2 . General Notch signaling or γ-secretase inhibitors have been used in clinics with undesired side effects, including disruption of the normal intestinal stem-cell differentiation 2 , 3 . Specific blocking antibodies are now available that target the various ligands and receptors of the Notch pathway 4 , 5 , 6 , 7 , 8 . Given the specificity of Dll4 expression in endothelial cells (ECs), targeting this ligand was initially thought to be an effective and safe strategy for specifically modulating Notch signaling and angiogenesis in disease, such as during tumor growth 6 , 7 . However, anti-Dll4 treatment was later shown to induce a loss of endothelial quiescence and vascular neoplasms, which were proposed to be the main cause of pathology in several organs 5 , 8 , 9 . This toxicity diminished the clinical appeal of Dll4/Notch blockers in cancer or cardiovascular disease settings. Here, we characterized the effect of single or compound targeting of all Notch signaling members on adult mice vascular homeostasis. High-resolution single-cell RNA sequencing (scRNA-seq) and three-dimensional (3D) confocal microscopy of adult liver vessels revealed very significant differences in the way each Notch member regulates vascular signaling, structure and single-cell states. γ-Secretase inhibitors or removal of Notch receptors did not cause substantial vascular or organ disease. Abnormal proliferating and sprouting single-cell states were generated only after Dll4 targeting. Surprisingly, suppression of these angiogenic cell states by additional genetic or pharmacological targeting was insufficient to prevent vascular and organ disease. Conceptually, our data show that the major transcriptional changes and angiogenic cell states elicited by targeting Dll4 correlate with, but do not cause, the observed vascular pathophysiology. Instead, we propose that it is the unrelated vascular structure abnormalization and malfunction that leads to organ pathology and the reported toxicity of anti-Dll4 treatment 5 , 8 . Results Notch pathway expression and signaling in adult organ ECs To elucidate the role of Notch signaling in global vascular homeostasis, we first assessed its activity in different organ vascular beds by immunodetection of the activated form of the Notch1 intracellular domain (N1ICD Val1744 ). This epitope was detected in ∼ 50% of all organ ECs (Fig. 1a,b ). Bulk RNA-seq analysis revealed that Dll4 and Notch1 are the most expressed ligand–receptor pair in quiescent vessels of most organs (Fig. 1c,d and Extended Data Fig. 1a ), and that Mfng is the most strongly expressed Notch glycosyltransferase. These enzymes are known to significantly enhance Delta ligand signaling and decrease Jagged ligand signaling 1 . Adult mice with induced deletion of Dll4 in ECs ( Dll4 iDEC - Dll4 flox/flox Cdh5-CreERT2 ) led to a significant reduction in N1ICD Val1744 and Hey1 signals in most organs’ quiescent ECs (Fig. 1e–i ). This indicates that Dll4 is the main functional ligand responsible for triggering Notch activity in most quiescent vessels. We observed compensatory upregulation of Dll1 only in lungs (Fig. 1i ). Dll4 deletion elicited remarkably different gene expression signatures among different organ vascular beds, with the adult liver endothelium presenting the most pronounced changes in gene expression (Fig. 1j,k and Extended Data Fig. 1 ). Despite significant transcriptional changes in most organs’ ECs, only the endothelium of the heart, muscle and liver showed an increase in the frequency of cycling or activated Ki67 + cells upon Dll4 deletion (Fig. 1l–n ), and these were the only organs with clear alterations in the 3D vascular architecture after the loss of Dll4–Notch signaling (Fig. 1o ). The brain underwent significant changes in gene expression (Fig. 1j,k and Extended Data Fig. 1 ), but these were not accompanied by endothelial proliferation or vascular morphological changes. Fig. 1: Dll4 deletion leads to EC activation and proliferation only in some vascular beds. a , b , Notch1 signaling activity (cleaved Val1744 N1ICD) in quiescent endothelium (DAPI + Endomucin + , abbreviated as EMCN). c , Schematic representation to illustrate the bulk RNA-seq experiment performed with adult ECs isolated by FACS. d , Heatmap with RNA-seq reads per kilobase per million mapped reads (RPKM). e , Experimental layout for the inducible deletion of Dll4 in Cdh5 + ECs ( Dll4 iDEC ) with Cdh5(PAC)-CreERT2 . f , Expression of Dll4 protein in CD31 + EMCN + vessels. g , h , Dll4 deletion significantly reduces Notch signaling activity (cleaved Val1744 N1ICD) in all quiescent vascular beds. In brain micrographs, white arrowheads indicate ECs and yellow arrowheads indicate non-ECs. Note that whereas N1ICD is maintained in non-ECs, most N1ICD signal disappears from the ECs in Dll4 iDEC brains. i , Schematic representation to illustrate the bulk RNA-seq experiment performed with adult ECs. Below, a heatmap showing the relative expression of all Notch pathway components and canonical target genes in control and Dll4 iDEC mutant ECs. j , Unsupervised hierarchical clustering showing stronger gene expression changes in Dll4 iDEC liver ECs compared with the other organs. Z -score lcpm, Z -score of the logarithmic counts per million. k , Unsupervised hierarchical clustering showing strong upregulation of Myc target genes in Dll4 iDEC liver ECs compared with the other organs. l – n , Dll4 deletion results in increased EC proliferation (Ki67 + ERG + cells) in some organs but not others. o , 3D reconstruction images from thick vibratome sections show vessel (CD31 + EMCN + ) enlargement in Dll4 iDEC heart and liver but not in brain. Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bars, 100 μm. Source data Full size image Targeting Dll4 induces heterozonal responses in liver vessels The previous RNA-seq and histological data revealed the adult liver endothelium as the most reactive vascular bed to the targeting of Dll4–Notch signaling. Rats and chimpanzees treated with anti-Dll4 antibodies also developed significant liver vascular neoplasms and disease 5 , 8 ; therefore, we focused our analysis on this organ. To gain deeper insight, we performed a high-resolution spatiotemporal phenotypic and transcriptomic analysis after targeting Dll4 for 2 days to 3 weeks. In contrast to targeting Dll4 during angiogenesis, targeting Dll4 in liver sinusoidal ECs (LSECs) for 48 h, which abolishes the generation of cleaved N1ICD, did not induce major transcriptomic changes (only 11 differentially expressed genes) or vascular phenotypic changes (Extended Data Fig. 2a–e ). Gene set enrichment analysis (GSEA) revealed upregulation of only a few E2F and Myc target genes at this time point (Extended Data Fig. 2f–h ). The increase in vascular density after targeting Dll4 was relatively slow and progressive, only becoming noticeable 1 week after genetic deletion (Fig. 2a–c ). Endothelial proliferation peaked at day 4 and was sustained after, leading to a progressive increase in vascular density and the total number of ECs (Fig. 2b–d ). Proliferation of neighboring hepatocytes was also increased, peaking after the peak in endothelial proliferation (Fig. 2e ), suggesting that Dll4 KO ECs secrete angiocrine factors inducing hepatocyte proliferation, as shown previously during liver regeneration 10 . Fig. 2: Targeting Dll4 induces heterozonal responses in liver vessels. a , Experimental layout for the inducible deletion of Dll4 in Cdh5 + ECs ( Dll4 iDEC ) with Cdh5(PAC)-CreERT2 . 3D projection of confocal images from thick vibratome sections. b – e , Analysis of EC (ERG + cells) and hepatocyte (ERG − DAPI + ) proliferation (Ki67 + ) and cell number. f , Representative confocal micrographs showing that the abnormal vascular pattern observed in Dll4 iDEC livers is located in the central vein (CV)-connecting sinusoids, but not in ECs surrounding portal veins (PV). Yellow dashed lines highlight the CV affected area. g , EC density in Dll4 iDEC liver is higher in sinusoids connecting the CVs rather than those around PVs (CD34 + ). White dashed lines highlight the denser area. h , Dll4 iDEC liver section showing the increase in nuclei size mainly in CV-connecting sinusoids. White dashed lines highlight the area with higher EC density and with larger EC nuclei. Higher magnification pictures of insets a and b together with pseudocoloring of nuclear sizes (lower panels) show differences in nuclei size between CV and PV areas, respectively. Violin plots reflecting changes in cell nuclei sizes. i , Increased EC proliferation (Ki67 + ERG + ) in Dll4 iDEC liver, particularly in the sinusoids connecting the CVs. j , Myc protein is upregulated mainly in ECs (ERG + cells) around the CVs after Dll4 deletion. k , Increased apoptosis (cleaved caspase-3 (C3)) in CV areas upon Dll4 deletion. l , m , Dll4 and activated N1ICD (V1744) protein are mostly present in arterial PV areas, while being mostly undetectable in venous CV areas. n , Dll4 deletion leads to loss of N1ICD (Val1744) activation in liver ECs. o , Msr1 immunostaining showing loss of arterial identity in Dll4 iDEC vessels. p , p21 expression in Dll4 iDEC liver ECs (ERG + ) is also higher in the sinusoids around the CVs. q , Dll4 iDEC Ki67 + liver ECs are actively dividing in S phase (EdU + Ki67 + ERG + , yellow arrowheads in upper panel), and a small fraction of proliferating ECs (Ki67 + ERG + ) also expresses p21 protein (p21 + Ki67 + ERG + , yellow arrowheads in lower panel). r , Dual ifgMosaic single-cell clonal tracking after Dll4 deletion. Images showing representative dual-labeled EC clones (yellow and white arrowheads in i and asterisks in iii). Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bars, 100 μm. Source data Full size image The effect of Dll4 targeting was, however, notably heterogeneous and zonal. Only vessels around the central veins and with a known venous identity 11 had a higher number of ECs (Fig. 2f,g ), larger nuclei (Fig. 2h ), and expression of cell-cycle (Fig. 2i,j ) and apoptosis (Fig. 2k ) markers. Therefore, the previously reported anti-Dll4-driven liver histopathology and increase in cell proliferation 8 is now found to be mainly associated to the central-vein sinusoids, which become enlarged and full of blood cells (Extended Data Fig. 2i–m ). Paradoxically, the portal-vein sinusoids, which have arterial identity and the highest Dll4 expression and Notch activity (Fig. 2l–n and Extended Data Fig. 2n,o ), showed a minor increase in EC proliferation (Fig. 2i ) despite a significant loss in the expression of arterial genes (Fig. 2o and Extended Data Fig. 2p ). Besides the cell-cycle marker Ki67, we also analyzed more specific S-phase (EdU) and cell-cycle arrest/senescence (p21) markers. This analysis revealed expression of p21 in 30% of Dll4 iDEC ECs in the venous vessels around the central veins (Fig. 2p ). Among Ki67 + ECs, 40% were positive for EdU and 25% were positive for p21 (Fig. 2q ). This shows that there is a mix of productive cell division (EdU + ) and arrest (p21 + ) after Dll4 loss in liver ECs. Pulse–chase single-cell ifgMosaic tracking revealed that relatively few of the Ki67 + ECs had the ability to divide and clonally expand after Dll4 targeting, with some cells dividing 6 to 50 times more than their neighbors (Fig. 2r ). All of these progenitor cells were located in the sinusoids around central veins (Fig. 2r,iii ). Loss of Notch1 or Rbpj in LSECs induces hypermitogenic arrest Notch ligands and receptors can be targeted with a range of pharmacological compounds and antibodies 4 , 5 , 6 , 7 , and so far only Dll4-targeting antibodies have been reported to cause major vascular disease 5 , 8 . In contrast, genetic deletion of Notch1 or Rbpj in mice has been suggested to cause vascular phenotypes very similar to the genetic deletion of Dll4 , during angiogenesis and in adult vessels 9 , 12 , 13 , 14 , 15 . Therefore, we investigated if deleting Notch1 or Rbpj , the master regulator of all Notch receptor signaling, induced vascular pathology similar to that induced by the loss of Dll4 (Fig. 3a ). Surprisingly, Notch1 and Rbpj deletion for 2 weeks or 4 weeks did not significantly increase EC proliferation and related vascular pathophysiology (Fig. 3b–e and Extended Data Fig. 3a–g ), despite these mutant cells having even higher activity of phosphorylated extracellular signal-related kinase (p-ERK) than ECs lacking Dll4 (Fig. 3f–h ). Livers treated with anti-Notch1 blocking antibody 4 also lacked the major hallmarks of pathology observed in anti-Dll4-treated livers (Extended Data Fig. 3h ). Next, we compared the transcriptome of Dll4 iDEC and Rbpj iDEC vessels. ECs from both mutant lines showed a similar upregulation of genes related to cell-cycle activation and metabolism (Fig. 3i ) and had enlarged nuclei (Fig. 3j ). However, compared with Dll4 iDEC livers, Rbpj iDEC livers had significantly less vascular expansion and organ abnormalities (Fig. 3k and Extended Data Fig. 3c–g ) and stronger upregulation of p21 (Fig. 3l ), a cell-cycle inhibitor frequently upregulated in senescent or hypermitogenically arrested cells 16 . We also identified a significant increase in the number of binucleated p21 + ECs, suggestive of replicative stress and G2 arrest of the mutant cells (Fig. 3m and Extended Data Fig. 3i ). RNA-seq analysis revealed signatures of genetic pathways linked to G2/M checkpoints, chromosome segregation, and general replicative stress and senescence in Rbpj iDEC ECs (Fig. 3n and Extended Data Fig. 3j ). To determine the functional effect of p21 upregulation, we analyzed compound Rbpj iDEC p21 KO mice (Fig. 3o ). p21 loss did not affect the minor vascular sinusoid dilation seen in Rbpj iDEC livers, but did increase the frequency of cycling (Ki67 + ) and apoptotic (cleaved caspase-3 + ) cells (Fig. 3p–r ), in line with the role of p21 as a cell-cycle and apoptosis inhibitor 17 , particularly in hypermitogenically activated Rbpj KO cells. This dual and paradoxical effect of p21 loss on both cell proliferation and apoptosis may explain the relatively mild increase in EC numbers in Rbpj iDEC p21 KO livers compared with the fully arrested Rbpj iDEC liver vessels. These results suggest that loss of Dll4 induces a reduction in Notch signaling that results in a mixed population of proliferative and arrested ECs, whereas the complete loss of Notch signaling induces mostly hypermitogenic arrest, without productive cell division. Fig. 3: Deletion of Rbpj or Notch1 in liver quiescent blood vessels does not phenocopy Dll4 deletion. a , Experimental layout for the inducible deletion of Rbpj ( Rbpj iDEC ), Notch1 ( Notch1 iDEC ) and Dll4 ( Dll4 iDEC ) in Cdh5 + ECs. All mice contained the Cdh5(PAC)-CreERT2 and iSuR e-Cre (expressing MbTomato-2A-Cre) alleles to ensure genetic deletion of the floxed alleles. b – d , Increased EC density (ERG + per field) and proliferation (Ki67 + ERG + /ERG + ) were observed only in Dll4 iDEC liver ECs. e , Gross liver pathology is observed exclusively in Dll4 iDEC livers. f – h , p-ERK immunostaining and whole-liver western blot showing that the frequency of p-ERK-expressing ECs and intensity levels increase in the mutants, particularly the Notch1 and Rbpj mutants. i , Heatmap with the normalized enrichment score (NES) from significant hallmark analysis (FDR q value < 0.05) by GSEA from bulk RNA-seq data. FC, fold change. j , Mutant liver ECs have a larger nuclei size than control liver ECs. k , Vascular (CD31 + ) dilation or expansion is more pronounced in Dll4 iDEC mutants. l , p21 expression in ECs (p21 + ERG + ) is more increased in Rbpj iDEC mutants. m , Binucleated cells (white arrowheads) identified in Dll4 iDEC and Rbpj iDEC mutants. High magnification of insets a and b are shown at the bottom. n , GSEA analysis shows a positive and significant enrichment in Chromosome Segregation-related and Cellular Senescence-related genes in Rbpj iDEC mutant liver ECs as shown by the NES. o , Experimental layout for the inducible deletion of Rbpj in a p21 KO background. p , 3D projection of thick vibratome sections showing the endothelial surface marker CD31 and EMCN, and proliferation (Ki67) analysis in ECs (ERG + ). q , Analysis of the apoptosis marker cleaved caspase-3. r , The absence of p21 in a Rbpj iDEC background results in a modest increase in EC density (ERG + ), but both EC proliferation (Ki67 + ERG + ) and apoptosis (cleaved caspase-3, CC3) are significantly increased. Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bars, 100 μm, except e , 1 mm. Source data Full size image Targeting Dll4 and Notch induces incongruent cell states Next, we performed scRNA-seq to identify possible differences in vascular single-cell states induced by targeting Dll4, Notch1 or Rbpj. This analysis was performed on cells expressing the Cdh5-CreERT2 and iSuRe-Cre alleles 18 to guarantee endothelium-specific recombination, labeling and full genetic deletion of all of the floxed genes used in this study (Fig. 4a,b and Extended Data Fig. 4a ). To reduce batch effects, Tomato + CD31 + ECs were isolated on the same day from multiple control and mutant animals, tagged with different oligonucleotide-conjugated antibodies, and loaded in the same chip. The few mutant cells with mRNA expression of Dll4 and Notch1 were likely contaminants. For Rbpj , only exons 6–7 are deleted, leading to a less stable, but still detectable, 3′ mRNA. Altogether, the scRNA-seq data analysis showed the existence of ten clearly defined cell clusters (Fig. 4c–e and Extended Data Fig. 4b ). The deletion of Rbpj, Notch1 and Dll4 resulted in a significant decrease in Notch signaling and Hes1 expression (Fig. 4b ) and the loss of the arterial sinusoidal capillary transcriptional C1a cluster. In agreement with this, all of these mutants had a reduction in distal portal-vein (arterial) caliber and branching complexity (Extended Data Fig. 3c–e ). However, only the loss of Dll4 was able to induce a very pronounced loss of liver sinusoidal genes and capillarization 19 , 20 and a tip-cell transcriptional program (C4). This program was characterized by the downregulation of Gata4 19 , Maf 21 and the venous Wnt2 gene expression (Fig. 4f–h and Extended Data Fig. 4d ) and very high expression of the tip-cell markers Kcne3, Esm1, Angpt2 and Apln , as well as Myc and its canonical target Odc1 (Fig. 4i and Extended Data Fig. 4b–d ). Most of the upregulated genes in the tip-cell cluster were associated with Myc metabolism, increased ribosome biosynthesis, glycolysis, mTORC1 signaling, and fatty acid and oxidative phosphorylation (Extended Data Fig. 4e ). Paradoxically, Notch1 iDEC and Rbpj iDEC liver ECs, in which the decrease in Notch signaling was more pronounced ( Hes1 expression in Fig. 4b ), showed a more moderate metabolic activation, and most of these mutants ECs clustered in either the venous C1v cluster or the activated C3 cluster and did not reach the extreme C4 tip-cell state (Fig. 4c,d ). Fig. 4: scRNA-seq analysis reveals significant differences between targeting Dll4 and Notch signaling. a , Experimental layout for the inducible deletion of the indicated genes in Cdh5-CreERT2 + ECs and collection of the iSuRe-Cre + (Tomato-2A-Cre + ) cells to ensure genetic deletion. b , Violin plots showing Dll4 , Notch1 and Rbpj mRNA expression in single cells and the subsequent downregulation of the Notch target gene Hes1 in all mutants. c , d , UMAPs showing the ten identified clusters, and barplot showing the percentage of cells in each cluster in all samples. e , Dot plot showing the frequency (size) and intensity (color) of expression for the top cluster marker genes. f , Heatmap showing the indicated LSEC and continuous/capillary endothelial cell (CEC) gene expression signatures. g , Enrichment score analysis of LSEC and CEC signatures in Dll4 iDEC ECs. h , Violin plots showing decreased Gata4 and Wnt2 expression only in Dll4 iDEC mutants. i , Violin plots for some cluster marker genes. j , k , In Dll4 iDEC mutants, tip cells (Esm1 + ERG + ) are localized in the sinusoids around CVs, but not in PV sinusoids. l , The global cell-cycle marker Stmn1 is highly upregulated exclusively in Dll4 iDEC liver ECs. m , Most of the Esm1 + ECs are not Ki67 + , but have Esm1 − Ki67 + ECs as neighbors in the CV sinusoids. n , Violin plots for the indicated genes and conditions. o , Experimental layout for the inducible deletion of the indicated genes, their violin plots, UMAPs and barplots. p , Expression of the tip-cell marker Esm1 in ERG + ECs located in CV sinusoids. q , Violin plots showing that deletion of Notch1/2/4 results in less Notch signaling ( Hes1 ) and less arterial marker expression ( Msr1 ), but no induction of the tip-cell program ( Kcne3/Esm1/Myc/Odc1 ) or the proliferation marker Stmn1 . The cell-cycle arrest marker ( Cdkn1a ) is increased. r , Experimental layout for the inducible heterozygous deletion of Dll4 ( Dll4Het iDEC ) for 2 weeks or DBZ treatment for 4 days in Cdh5 + ECs used for scRNA-seq. s , UMAPs and barplots obtained. t , Violin plots showing expression of the canonical Notch signaling target Hes1 . Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bars, 100 μm. Source data Full size image Histology confirmed that indeed only the Dll4 iDEC mutants had a significant population of Esm1 + tip cells (Extended Data Fig. 4f ) and that these were mostly present in the venous sinusoidal capillaries interconnecting the liver central veins (Fig. 4j,k ), where EC proliferation and density are the highest (Fig. 2f–i ). The upregulation of the global cell-cycle marker Stmn1 in Dll4 iDEC livers (Fig. 4l ) correlated with the sixfold higher frequency of Ki67-protein + cells in these mutants compared with the Notch1 and Rbpj mutants (Fig. 3d ). Most Esm1 + tip cells were not Ki67 + , in accordance with their higher sprouting activity and arrested nature, but had proliferating Ki67 + cells as close neighbors (Fig. 4j,m ). Notch1 iDEC and Rbpj iDEC ECs showed significant upregulation of the replication-stress/senescence markers p21 ( cdkn1a ), p53 ( trp53 ) and p16 ( cdkn2a ) (Fig. 4n ). These cells undergo hypermitogenic S/G2/M arrest (Fig. 3m,n ) without becoming Kcne3 + /Esm1 + sprouting tip cells (Fig. 4c,d ), which is in contrast to the current understanding of sprouting angiogenesis 16 , 22 . Notch1 iDEC livers upregulated the expression of Notch4 (Extended Data Fig. 5a ), a receptor known to partially compensate for Notch1 deletion 23 . Deletion of Notch1 /2/ 4 in ECs, similarly to Rbpj loss, results in even lower Hes1 expression and higher p21 expression (arrest); however, this does not result in the induction of tip cells (Esm1 + /Kcne3 + ) or proliferating Stmn1 + cells (Fig. 4o–q and Extended Data Fig. 5b–f ). We also tested a general γ-secretase inhibitor, DBZ, which is known to block Notch signaling and elicit strong effects on tumor and retina angiogenesis 6 , similarly to anti-Dll4 treatment (Extended Data Fig. 5g ). However, this compound had a very weak effect on quiescent vessels, similar to the changes seen in Dll4 heterozygous livers (Fig. 4r–t ). We also observed by scRNA-seq that ECs with full loss of Dll4 signaling for only 4 days had already lost the arterial capillary program (C1a cluster) and become activated (C3 cluster), but had not yet had time to fully differentiate to tip cells (C4 cluster in Extended Data Fig. 5h–j ). This suggests that in order to fully activate quiescent ECs and induce significant numbers of tip cells and vascular abnormalization, pronounced and continuous loss of Dll4 signaling must be sustained for about 1 week, which can be achieved with genetic deletion or blocking antibodies 5 but not with small-molecule inhibitors targeting Notch. The difference between the liver vascular phenotypes of Dll4 and Notch receptor mutants could be also due to a role of the ligand, and not the receptors, on signaling to adjacent liver cells. scRNA-seq analysis of all other liver cell types revealed that hepatocytes did not express significant amounts of Notch receptors (Extended Data Fig. 6a–e ). Hepatic stellate cells, Kupffer cells (stellate macrophages) and some other blood cell types expressed Notch receptors, but their target genes were not significantly downregulated by endothelial Dll4 deletion, suggesting that this ligand mainly signals within ECs (Extended Data Fig. 6d–f ). Single-cell data analysis revealed a significant increase in leukocytes in Dll4 iDEC livers, particularly monocytes, neutrophils and macrophages (Extended Data Fig. 6c ), presumably due to the vascular pathology and the subsequent abnormal blood flow that leads to the accumulation of these cells and an organ pathology signature (Extended Data Figs. 2j,m and 6c,g–j ). Remarkably, EC-specific expression of N1ICD rescues the major hallmarks of the Dll4 iDEC vascular pathology at the organ and single-cell levels (Extended Data Fig. 7 ). These data suggest that it is not the loss of Dll4 signaling to non-ECs that causes the difference between Dll4 iDEC and Notch1/2/4 iDEC or Rbpj iDEC mutants. It also confirms that it is the partial downregulation of the Dll4–Notch transcriptional program in ECs, which is not matched by the complete loss of Notch receptors or Rbpj, that causes the liver vasculature abnormalization and subsequent pathology. Deletion of all other Notch ligands does not elicit pathology Besides Dll4, other Notch ligands are also expressed in liver ECs (Fig. 5a ). The Notch signaling target Hes1 is more expressed in Dll4 iDEC than in Notch1 iDEC , Rbpj iDEC or Notch1/2/4 iDEC mutants (Fig. 4b,q ), suggesting that the other weakly expressed Notch ligands (Jagged1, Jagged2 and Dll1) may partially compensate the loss of Dll4 and induce residual Notch signaling essential for the induction of the tip-cell state. Notably, Jagged1 mRNA was barely detectable in bulk or scRNA-seq data of quiescent liver ECs (Figs. 1d and 5a ), but its protein was clearly expressed in liver vessels (Fig. 5b ). Deletion of all three ligands ( Jag1, Jag2 and Dll1 ) did not alter vascular morphology, induce pathology, or increase the frequency of Ki67 + cells, confirming that Dll4 is the main Notch ligand in quiescent vessels (Fig. 5c–g ). Liver blood profiling revealed an increase in the percentage of neutrophils, but this was also seen in circulating blood, suggesting a systemic rather than organ-specific role of these ligands (Fig. 5h,i ). In agreement with this, scRNA-seq data analysis confirmed that most mutant ECs remained quiescent and did not become activated or form tip cells (Fig. 5j–l ). Moreover, deletion of Jag1 , Jag2 and Dll1 in ECs did not compromise the portal sinusoid arterial identity (Fig. 5k,m ), instead revealing a slight increase in the Notch signaling target Hes1 and the arterial gene CD34 , together with a very pronounced decrease in the expression of the venous-enriched Wnt2 gene (Fig. 5n ). This counterintuitive increase in Notch signaling was also observed previously after the loss of Jagged1 during angiogenesis 24 . Fig. 5: Deletion of Jag1 , Jag2 and Dll1 in liver ECs does not cause pathology. a , Heatmap of bulk RNA-seq reads and violin plot of single-cell data showing expression of all Notch ligands in liver ECs. b , Despite its low mRNA expression, Jag1 protein is clearly detected in the adult liver quiescent endothelium (EMCN + ) and absent in Jag1/Jag2/Dll1 iDEC mutants. c , Experimental layout for the inducible deletion of Jag1 , Jag2 and Dll1 in Cdh5 + ECs. d , CD31 and EMCN + immunostaining shows no vascular architecture changes. e , f , Macroscopic pictures and hematoxylin and eosin (H&E) staining show absence of liver pathology. g , Deletion of the three ligands does not lead to an increase in endothelial proliferation (Ki67 + /ERG + ECs) nor an increase in EC number (ERG + cells per field). h , Analysis by FACS of the percentage of different blood cells in livers. NS, not significant. i , Hematological analysis of circulating (systemic) blood cells. j , Violin plots showing expression of the four ligands in scRNA-seq data. k , UMAPs and barplot showing the ten identified clusters and the percentage of cells in each cluster in the two samples. l , Jag1/Jag2/Dll1 mutant ECs do not upregulate the tip-cell ( Esm1/Kcne3/Angpt2 ), nor metabolic ( Myc/Odc1 ), nor proliferation ( Stmn1 ) transcriptional program observed in Dll4 mutants. m , Immunostaining and scRNA-seq data showing that Jag1/Jag2/Dll1 mutant ECs do not downregulate the expression of the arterial markers Msr1 and Efnb2 . n , Violin plot showing an increase in the Notch target gene Hes1 and the arterial gene CD34 , together with a decrease in the expression of the venous Wnt2 gene in Jag1/Jag2/Dll1 mutant ECs. Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bars, 100 μm, except e , 1 mm. Source data Full size image Myc loss prevents Dll4 iDEC transcriptional states but not pathology Next, we aimed to determine the molecular mechanisms responsible for the unique EC activation, tip-cell signature, and vascular pathology induced by targeting Dll4. As mentioned above, Myc and its target Odc1 were among the most strongly upregulated genes in Dll4 mutant ECs, compared with Notch1 and Rbpj mutants. Myc is known to activate important ribosome biogenesis and protein translation pathways, favoring cell growth 25 . Dll4 iDEC livers showed upregulation of a large range of canonical E2F, Myc, mTORC1 and ribosomal ( Rpl ) genes, particularly in the activated, proliferating and endothelial tip-cell clusters (Fig. 6a and Extended Data Fig. 4 ). This hypermetabolic transcriptional status was confirmed by mass spectrometry (MS) analysis of protein lysates obtained from freshly isolated liver ECs (Fig. 6b–f ), providing a high-depth proteomic analysis of the endothelial tip-cell state induced by targeting Dll4. We also independently confirmed Myc mRNA and protein upregulation in Dll4 KO vessels (Fig. 6g,h ). Fig. 6: Myc loss prevents the Dll4 KO endothelial activation and single-cell states but not vascular pathology. a , GSEA hallmark analysis for each single-cell cluster. b , GSEA hallmark analysis performed with the Dll4 iDEC bulk proteome and transcriptome. c , Heatmaps showing log(fold change) of genes and proteins belonging to different sets. d , Barplot showing the NES in each single-cell cluster for the indicated gene sets. e , Barplot with the top differentially expressed (DE) proteins in Dll4 iDEC livers. f , Enrichment analysis showing a significant positive enrichment in translational initiation-related genes and proteins encoded by genes that are regulated by the Myc-Max transcription factors. g , Micrographs showing immunostainings for the Myc protein, which is upregulated in liver ECs (ERG + cells) after Dll4 deletion. h , Myc mRNA expression (normalized counts from bulk RNA-seq). i , Experimental layout for the inducible deletion of Dll4 and Myc in Cdh5 + and iSuRe-Cre + ECs and scRNA-seq analysis. j , UMAPs and barplot showing the ten identified clusters and the percentage of cells for each cluster in the different samples. k , Dot plot of the top upregulated genes in Dll4 iDEC liver ECs belonging to the indicated gene marker groups. l , GSEA hallmark analysis showing the decreased expression of most gene sets in Dll4/Myc iDEC . m , Double deletion of Dll4 and Myc in ECs results in a significant reversion of proliferation (Ki67 + ERG + cells) and Esm1 + expression (Esm1 + ERG + ) to control levels. n , 3D confocal micrographs from thick vibratome sections (top) or thin sections (bottom), and liver macroscopic pictures showing vessel enlargement and liver pathology in Dll4/Myc iDEC mutants similarly to Dll4 iDEC mutants. Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bar, 100 μm, except n lower panel, 1 mm. Source data Full size image Next, we investigated the implication of Myc in the Dll4 iDEC transcriptional program and subsequent vascular-related pathology. Myc loss (in Dll4/Myc iDEC animals) almost entirely blocked the EC activation induced by Dll4 loss, and very few ECs were in the activated (C3) and tip-cell (C4) clusters (Fig. 6i–l and Extended Data Fig. 8a ). Consistent with the scRNA-seq data, frequencies of proliferating (Ki67 + ) and tip (Esm1 + ) cells in Dll4/Myc iDEC mutants were similar to those in wild-type animals (Fig. 6m and Extended Data Fig. 8b, c ). Myc activity is thus essential for the strong metabolic and biosynthetic phenotype of Dll4 KO liver ECs and the appearance of the abnormal cell states. Surprisingly, despite this strong transcriptional and cell-state reversion to a quiescent state, Dll4/Myc iDEC mutant vessels were still highly abnormal and dilated (Fig. 6n and Extended Data Fig. 8d ). The vascular abnormalities in Dll4/Myc iDEC mutant livers were not in accordance with their more quiescent scRNA-seq profile (Fig. 6j–l ), nor with the significantly lower frequencies of Ki67 + and Esm1 + cells (Fig. 6m ). Interestingly, Dll4/Myc iDEC livers retained hallmarks of tissue hypoxia and inflammation (Fig. 6l and Extended Data Fig. 8e ) and had strong activation of surrounding hepatocytes already 5 days after deletion (Extended Data Fig. 8f ), despite having a quiescent endothelium. Altogether, these data indicate that the vascular structure abnormalization observed in Dll4 mutant livers is not driven by the detectable changes in endothelial transcriptional programs or the proliferative and tip EC states. Anti-VEGFA treatment prevents the Dll4 iDEC pathology with less effect on transcription Among the few GSEA hallmark pathways whose upregulation in Dll4 mutants was not altered in Dll4/Myc iDEC vessels was the hypoxia pathway and inflammatory response (Fig. 6l and Extended Data Fig. 8e ). Hypoxia is known to induce expression of vascular endothelial growth factor A (VEGFA), which can induce vascular expansion without the need for proliferation 26 . The expression of VEGFA was significantly upregulated in the Dll4 KO venous tip-cell cluster (Extended Data Fig. 9a ). Therefore, we explored if anti-VEGFA treatment could prevent the appearance of the activated vascular cell states, vascular enlargement and liver pathology induced by Dll4 deletion. Unlike Myc loss, anti-VEGFA treatment reduced both the vascular expansion and the liver pathology induced by Dll4 deletion (Fig. 7a–d and Extended Data Fig. 9b ). scRNA-seq analysis confirmed the almost-complete loss of the tip-cell (C4) and proliferating (C5) single-cell states, as well as a significant reduction in the activated cell states (C3), with a general return to the quiescent cell states, with exception of the arterial state (Fig. 7e–i and Extended Data Fig. 9c–f ). scRNA-seq and histology data also revealed a depletion of VEGFR2/Kdr + sinusoidal capillaries by anti-VEGFA treatment (Fig. 7b–e, i and Extended Data Fig. 9a ). Anti-VEGFA treatment rescued the expression of the blood flow and shear stress responsive genes Klf2 and Klf4 (Fig. 7j and Extended Data Fig. 9g ), suggesting a normalization of vessels and blood flow. Fig. 7: Vascular abnormalities and liver pathology are prevented by VEGFA antibody administration in Dll4 iDEC mutants by ERK-independent mechanisms. a , Experimental layout for the inducible deletion of Dll4 in Cdh5 + ECs and VEGFA antibody administration. b , Confocal micrographs showing reduced CD31 or EMCN vascular immunostaining after anti-VEGFA treatment. c , Stereomicroscope liver pictures. d , Vessel density is reduced in Dll4 iDEC mutants after anti-VEGFA treatment. e , UMAPs and barplot showing the identified clusters and the percentage of cells for each cluster in indicated samples. f , Unsupervised hierarchical clustering showing gene expression changes. g , Dot plot of the top upregulated genes for each indicated gene set. h , Violin plots of scRNA-seq data showing that anti-VEGFA treatment prevents the strong upregulation of Myc and its target Odc1 . i , The total number of ERG + ECs, proliferation (Ki67 + ERG + ) and Esm1 expression (Esm1 + ERG + ) return to control conditions after VEGFA antibody administration. j , Dot plot showing expression of flow/shear stress genes. k , Number of upregulated genes for each contrast and Venn diagrams showing that when compared with Myc loss, anti-VEGFA treatment has less effect on the Dll4 iDEC upregulated genetic program. DEGs, differentially expressed genes. l , GSEA hallmark analysis confirms the more moderate effect of anti-VEGFA treatment on the Dll4 iDEC genetic program when compared with Myc loss. m , Experimental layout for the inducible deletion of Dll4 and SL327 administration. n , UMAPs and barplot showing the identified clusters and the percentage of cells for each cluster in indicated samples. o , p , The administration of an ERK/MEK signaling inhibitor (SL327) results in reduced ERK phosphorylation. q , Violin plot showing that SL327 treatment partially inhibits the generation of tip cells ( Kcne3 + ). r , The administration of SL327 does not change the frequency of proliferating Ki67 + ECs (Ki67 + ERG + ). s , Abnormal vasculature (CD31 + EMCN + ) associated with liver pathology still occurs after SL327. Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bars, 100 μm, except in c and s upper panel, 1 mm. Source data Full size image These results show that anti-VEGFA treatment prevents not only the appearance of the abnormal single-cell states induced by Dll4 targeting, as Myc loss also does, but also the vascular expansion and blood flow abnormalities associated with organ pathology. However, blocking VEGF had a much lesser effect than Myc loss on the Dll4 KO transcriptional signature (Fig. 7k ). Anti-VEGFA treatment of Dll4 iDEC livers attenuated, but did not completely downregulate, many of the genes associated with metabolic and biosynthetic activities (Fig. 7l and Extended Data Fig. 9h, i ). This suggests that even though Dll4 iDEC + anti-VEGFA-treated ECs are transcriptionally and metabolically more active than Dll4/Myc iDEC ECs, only the latter form abnormal and enlarged vessels that result in organ pathology. Inhibition of major signaling pathways did not prevent Dll4 iDEC pathology VEGFA induces many important endothelial functions that are often difficult to distinguish, such as proliferation, sprouting, cell size, survival and permeability 27 , 28 , 29 . VEGF is thought to execute its effects on sprouting and angiogenesis mainly through ERK signaling 30 , 31 . However, administration of a highly effective ERK/MEK signaling inhibitor (SL327) had a much more modest effect than anti-VEGFA treatment, and only partially reduced the number of activated and tip ECs (Fig. 7m–s and Extended Data Fig. 9j ). The VEGF-dependent vascular enlargement or expansion could be alternatively mediated by increased Rac1(ref. 32 ), Pi3K/mTOR (refs. 33 , 34 ) or nitric oxide (NO) 35 , 36 signaling. However, the inhibition of these pathways also did not prevent the vascular pathophysiology induced by targeting Dll4 (Fig. 8 and Extended Data Fig. 9k,l ). Rapamycin effectively prevented the increase in the number of ECs, but not vascular dilation and pathology. Thus, the vascular pathophysiology effects of anti-VEGFA treatment, and anti-Dll4, are broader and independent of the activity of these signaling pathways. Fig. 8: Inhibition of Rac1, mTOR and NO signaling does not prevent the vascular pathophysiology induced by Dll4 targeting. a , Stereomicroscope images showing adult liver vascular defects and blood accumulation after Dll4 deletion and treatment with different inhibitors for 2 weeks. b , Confocal micrographs showing that the expansion and abnormalization of the liver sinusoids (CD31 + EMCN + ), particularly around CVs, observed after Dll4 deletion, are not prevented by the administration of the indicated compounds. On the right, images show EC (ERG + nuclei) proliferation (Ki67 + ). c , Charts showing quantification of EC density/numbers and proliferation. Note that mTOR inhibitor-treated liver ECs do not proliferate significantly (same ERG + content), despite a fraction being Ki67 + . d , Deletion of Rac1 with Cdh5-CreERT2 in adult liver endothelium (Cdh5 + ) does not prevent the vascular pathology induced by blocking Dll4 with REGN1035. e , The use of the indicated inhibitors in postnatal mouse retina angiogenesis assays for 48 h confirms that they do not prevent the increase in vascular expansion/density (isolectin B4 labeling) induced by anti-Dll4 antibody treatment (7.5 mg/kg). Note that 2 days of angiogenesis growth correspond to the vasculature formed above the red dashed line. Data are presented as mean values ± s.d. For statistics, see Source Data File 1 . Scale bars, 200 μm, except in a and d , 1 mm. Source data Full size image Overall, these results show that the genetic and pharmacological modulation of single-cell states related to endothelial dedifferentiation, activation, proliferation and sprouting often do not correlate with adult vascular phenotypes, function and ultimately organ pathology. Discussion Notch is one of the most important pathways for vascular development because it enables the necessary differentiation of ECs during angiogenesis 28 , 37 , 38 . Here, we expand on previous observations that Notch also plays an important role in the homeostasis of several organ vascular beds 8 , 9 , 12 . Dll4 is active in all organ vascular beds, and its loss affects the transcriptome of most quiescent ECs; however, Dll4 targeting effectively activates vascular growth in only the heart, muscle and liver. Even though the existence of four Notch receptors and five ligands allows for the possibility of multiple quantitative and qualitative signaling combinations and redundancy, our results confirm that Dll4 and Notch1 are clearly the most important Notch ligand–receptor pair for maintaining the global homeostasis of ECs. Previous work suggested that Dll4 and Notch1/Rbpj have similar functions in vascular development and homeostasis 6 , 7 , 8 , 15 , 23 , 24 , 39 , with only Jagged ligands shown to have opposite functions in Notch signaling and angiogenesis 24 . In this study, we show that Dll4 can have distinct functions from its receptors in vascular biology. It was possible to identify this difference only because of the use of scRNA-seq and high-resolution confocal analysis of liver vessel morphology; bulk RNA-seq analysis did not reveal significant differences between the transcriptomes of Dll4 and Rbpj mutants. The loss of Dll4, unlike the loss of Notch receptors or Rbpj, elicits a unique cascade of changes that culminates in the loss of sinusoidal marker genes and upregulation of Myc, similar to the loss of Gata4 (ref. 19 ). Dll4 iDEC liver vessels lose all quiescent arterial and venous cell states. The arterial cells become highly activated, and the venous cells show either tip-cell or proliferating cell signatures. Paradoxically, although Dll4 loss induces a weaker loss of Notch signaling than is induced by the loss of Notch receptors or Rbpj, it elicits a much stronger metabolic activation and expansion of the liver endothelium. This may be in part related to the bell-shaped response of ECs to mitogenic stimuli, as we previously showed during retina angiogenesis 16 . Our data indicate that full loss of Notch, or Rbpj, induces stronger ERK signaling and hypermitogenic arrest associated with hallmarks of cellular senescence, whereas Dll4 iDEC vessels retain a residual level of Notch signaling that instead effectively induces strong Myc-driven ribosome biogenesis and a metabolic switch toward active protein synthesis and cell growth that drives both EC proliferation and the generation of tip cells. The pro-proliferative effect of targeting Dll4 in quiescent vessels is in contrast to the hypermitogenic cell-cycle arrest that occurs after targeting Dll4 during embryonic and retina angiogenesis 16 , 40 , presumably a reflection of the significantly lower levels of growth factors, including VEGF, in adult organs. Previously, a noncanonical and N1ICD transcription-independent role for Dll4/Notch in inducing Rac1 and maintaining vascular barrier function was proposed 41 . Dll4 deletion could also affect signaling to other cell types, unlike deletion of the Notch receptors in ECs. However, our data show that the liver vascular abnormalization after targeting Dll4 can be rescued by the expression of N1ICD in ECs. This suggests that the vascular pathology is caused by the absence of Dll4 canonical signaling and transcription within the endothelium, and not due to noncanonical effects on vascular barrier function, or the loss of Notch signaling in other adjacent cell types. The observed lack of pathology in anti-Notch1-treated livers also corroborates this. High-resolution confocal microscopy revealed the heterozonal effect of Dll4 targeting. The induction of EC proliferation and tip cells was restricted to the most hypoxic liver venous sinusoids, precisely the ones with lower expression of Dll4 and Notch. Previous research showed that liver venous sinusoids have higher baseline activity of several tyrosine kinase signaling pathways 42 , which may explain the observed zonal effect of Dll4 targeting. The temporal analysis of the effects of Dll4 targeting on the adult liver vasculature also revealed that it takes at least 1 week for the full transcriptional reprogramming of quiescent ECs and the vascular expansion and organ pathology to become noticeable. During angiogenesis, this transcriptional and vascular morphology switch is already evident after 24 h of anti-Dll4 treatment 16 . This slow transcriptional reprogramming of quiescent ECs by Dll4 targeting may be related to the much lower levels of growth factors and nutrient availability in adult organs. The slow nature of this reprogramming may also explain the lack of effect of the small-molecule inhibitor DBZ on quiescent ECs. Unlike anti-Dll4 treatment or genetic deletion, which result in continuous loss of signaling, the less stable small-molecule inhibitor DBZ elicited no significant change in the quiescent vascular cell transcriptional states and phenotypes, whereas it is very effective during retina angiogenesis 15 , 16 . Anti-Notch1 4 also did not cause liver vascular pathology, despite its strong effect on angiogenesis. These findings have implications for selecting the most effective and safest way to target Notch in clinics, including blocking antibodies that target Dll4 versus antibodies that target Notch receptors, or the use of small-molecule inhibitors. Our data indicate that Notch receptor-targeting antibodies or small-molecule γ-secretase inhibitors do not induce significant liver vascular pathology and should be as effective as anti-Dll4 treatment at dysregulating tumor-related or ischemia-related angiogenesis, which can be beneficial in some therapeutic settings. It has also been shown that is possible to modulate the stability and pharmacokinetics of anti-Dll4 treatment to decrease its toxicity while maintaining its therapeutic and angiogenesis efficacy 5 . Our analysis also confirms the importance of Myc for the biology of ECs in the absence of Dll4. We previously reported that Myc loss rescues the ability of Rbpj KO or Dll4 KO ECs to form arteries 40 . Here, we show that Myc loss abrogates the generation of activated, proliferative and sprouting tip cells after Dll4 targeting, but surprisingly, this return to genetic and phenotypic quiescence is insufficient to prevent Dll4-targeting-induced vascular expansion, dysfunction and consequent organ pathology. In contrast, anti-VEGFA treatment did not completely abrogate the Dll4-targeting genetic program, but was able to prevent the associated vascular and organ pathology. However, this effect of anti-VEGFA treatment was not reproduced by inhibition of MAPK/ERK, Rac1, Pi3K/mTOR or NO signaling. This suggests a broader role for anti-VEGFA treatment in preventing pathological vascular enlargement and remodeling when combined with the anti-Dll4 antibody, that could be also related to its effect on liver EC survival. Our data suggest that the action of VEGF on vascular expansion and survival is independent of its direct effect on these signaling pathways 30 , 32 , 33 , 34 , 35 , 36 , and independent of cell proliferation and sprouting, as also previously proposed 26 , 43 . The sum of these findings also suggests that the recently developed bispecific antibody targeting both Dll4 and VEGF simultaneously (navicixizumab, OncXerna) may be less toxic than the use of anti-Dll4 treatment alone 44 . Altogether, the data obtained with several compound mutant and pharmacological approaches show that most of the transcriptional changes and angiogenic cell states elicited by targeting Dll4 correlate with, but do not cause, vascular pathophysiology (Extended Data Fig. 10 ). Therefore, vascular neoplasms are not the cause of the previously reported anti-Dll4 antibody toxicity 8 . Instead, we propose that the unrelated venous sinusoid enlargement and architecture abnormalization lead to vascular malfunction, blood accumulation, inflammation and hypoxia, altogether resulting in organ pathology. These data also raise questions about the general use of single-cell transcriptional or genetic states to describe and predict functional or dysfunctional vascular phenotypes and ultimately organ pathophysiology. A single-cell transcriptional state is only a small part of a cell’s phenotype and function. Methods Mice The following mouse ( Mus musculus ) lines and alleles were used and interbred: Tg(Cdh5-CreERT2) (ref. 45 ), Tg(iSuRe-Cre) (ref. 18 ), Dll1 flox/flox (ref. 46 ), Jag1 flox/flox (ref. 47 ), Jag2 flox/flox (ref. 48 ), Dll4 flox/flox (ref. 49 ), Notch1 flox/flox (ref. 50 ), Notch2 flox/flox (ref. 51 ), Notch4 KO (generated as described below), Rbpj flox/flox (ref. 52 ), Myc flox/flox (ref. 53 ), Cdkn1a(p21) KO (ref. 54 ), Rac1 flox/flox (ref. 55 ), Rosa26-EYFP (ref. 56 ), iChr-Mosaic (ref. 57 ) and iMb-Mosaic (ref. 57 ). To induce CreERT2 activity in adult mice, 20 mg or 10 mg of tamoxifen (Sigma-Aldrich, T5648) were first dissolved in 140 µl of absolute ethanol and then in 860 µl of corn oil (20 mg ml −1 or 10 mg ml −1 tamoxifen, respectively). From these stock solutions, dilutions were done and given to adult mice aged 2–5 months by intraperitoneal injection (total dose of 1 mg, 1.5 mg or 2 mg of tamoxifen per animal) every day for a maximum of 5 days. All mouse lines and primer sequences required to genotype these mice are provided in Supplementary Table 1 . Dll4/Notch signaling blockade in ECs was achieved using blocking antibodies to murine Dll4, developed by Regeneron (REGN1035) (ref. 58 ), or against Notch1 (anti-NRR1), developed by Genentech 4 . Mouse IgG (Sigma) was used in littermates as a control treatment. For the 48-h experiment, mice received a single intraperitoneal injection of 200 µl of IgG or anti-Dll4 antibody (20 mg kg −1 in PBS). For the 2-week blocking experiments, mice received anti-Dll4 antibody or anti-NRR1 antibody four times (day 1, day 4, day 8 and day 12) over 14 days at a concentration of 7.5 mg kg −1 or 10 mg kg −1 , respectively. For anti-VEGFA treatment experiments, mouse anti-VEGFA G6-31 antibody, developed by Genentech, was administered four times over 14 days at a concentration of 5 mg kg −1 . In mouse pups, anti-Dll4 antibody (REGN1035) was injected at 7.5 mg/kg or 20 mg/kg as indicated. The following inhibitors were injected intraperitoneally for 2 consecutive days in postnatal animals for retina analysis, or for 4–14 consecutive days in adult animals for liver analysis as indicated in the figures. γ-Secretase inhibitor DBZ (YO-01027; Selleck Chemicals, S2711) was injected at 30 µmol kg −1 in adult animals every day in the morning for 4 days, and 16 h before collection of the tissues. To inhibit MAPK/ERK phosphorylation, we injected 120 mg kg −1 SL327 (MEK inhibitor; Selleck Chemicals, S1066) every day, and 16 h before collecting the tissues for scRNA-seq. To inhibit Rac1, we injected NSC23766 at 3 mg kg −1 (Sigma, SML0952). To inhibit mTOR signaling, we injected rapamycin at 4 mg kg −1 (Enzo Life Sciences, BML-A275-0005). To inhibit NO synthase, we injected L-NIO at 30 mg kg −1 (R&D Systems, 0546). To inhibit Pi3K signaling, we injected alpelisib at 30 mg kg −1 (MedChemExpress, HY-15244). All mouse husbandry and experimentation was conducted using protocols approved by local animal ethics committees and authorities (Comunidad Autónoma de Madrid and Universidad Autónoma de Madrid CAM-PROEX 177/14, CAM-PROEX 167/17, CAM-PROEX 164.8/20 and PROEX 293.1/22 or Uppsala Committee permit number 5.8.18-03029/2020 or the Institutional Animal Care and Use Committee Protocol IS00013945). The mouse colonies were maintained in racks with individual ventilation cages according to current national legislation. Mice had dust-free and pathogen-free bedding, and sufficient nesting and environmental enrichment material for the development of species-specific behavior. All mice had ad libitum access to food and water in environmental conditions of 45–65% relative humidity, temperatures of 21–24 °C, and a 12 h/12 h light/dark cycle. In addition, to preserve animal welfare, mouse health was monitored with an animal health surveillance program, which follows the Federation of European Laboratory Animal Science Associations (FELASA) recommendations for specific pathogen-free facilities. We used mice with C57BL/6 or C57BL/6×129SV genetic backgrounds. To generate mice for analysis, we intercrossed mice with an age range of 7–30 weeks. Mice used for experiments were 2–5 months old. We do not expect our data to be influenced by mouse sex. To generate Notch4 KO mice, we used guide RNAs Notch4_1 (agggaccctcagagcccttg) and Notch4_2 (agggaatgatgccacgcata) to target mouse Notch4 in mouse eggs from the C57BL/6 genetic background. Injection mixture was composed by the described CRISPR RNA (crRNA; Integrated DNA Technologies) and trans-activating CRISPR RNA (tracrRNA; Integrated DNA Technologies, 1072533) at 0.305 μM and Cas9 nuclease (Alt-R S.p. HiFi Cas9 Nuclease V3, 100 µg, 1081060) at 20 ng µl −1 . Founders were screened by PCR with the primers below to confirm the genetic deletion. Immunofluorescence on cryosections Tissues were fixed for 2 h in 4% PFA in PBS at 4 °C. After three washes in PBS for 10 min each, organs were stored overnight in 30% sucrose (Sigma) in PBS. Organs were then embedded in OCT (Sakura) and frozen at −80 °C. Cryosections (35 μm) were cut on a cryostat (Leica), washed three times for 10 min each in PBS, and blocked and permeabilized in PBS containing 10% donkey serum (Millipore), 10% fetal bovine serum (FBS) and 1% Triton X-100. Primary antibodies were diluted in blocking/permeabilization buffer and incubated overnight at 4 °C. This step was followed by three 10-min washes in PBS and incubation for 2 h with conjugated secondary antibodies (1:200, Jackson Laboratory) and 4,6-diamidino-2-phenylindole (DAPI) in PBS at room temperature. After three washes in PBS, sections were mounted with Fluoromount-G (SouthernBiotech). All antibodies used are listed in Supplementary Table 2 . To detect Ki67 or c-Myc in the same section as ERG, we used rabbit anti-Ki67 or anti-c-Myc together with a Fab fragment Cy3 secondary antibody, which is compatible with the later use of rabbit anti-ERG conjugated to Alexa Fluor 647. Vibratome section immunofluorescence Tissues were fixed for 2 h in 4% PFA in PBS and washed as above. Organs were then embedded in 6% agarose low-melting gel (Invitrogen), and organ sections (100 μm) were cut on a vibratome. Sections were permeabilized for 1 h in PBS containing 1% Triton X-100 and 0.5% Tween 20. Sections were then blocked for 1 h in a PBS solution containing 1% Triton X-100, 10% donkey serum and 10% FBS. Primary antibodies were diluted in blocking buffer and incubated with sections overnight at 4 °C. This step was followed by six washes with 1% Triton X-100 in PBS for 15 min and incubation for 2 h with conjugated secondary antibodies (1:200, Jackson Laboratory) and DAPI in PBS at room temperature. After three 15-min washes in PBS, sections were mounted with Fluoromount-G (SouthernBiotech). All antibodies used are listed in Supplementary Table 2 . Whole-mount immunofluorescence of retinas For postnatal mouse retina immunostaining, eyes were collected and fixed in 4% PFA in PBS for 20 min at room temperature. After microdissection, retinas were fixed in 4% PFA for an additional 45 min, followed by two PBS washes of 10 min each. Retinas were blocked and permeabilized with PBTS buffer (0.3% Triton X-100, 3% FBS and 3% donkey serum) for 1 h. Samples were then incubated overnight at 4 °C in biotinylated isolectin B4 (diluted 1:50; Vector Laboratories, B-1205) and primary antibodies (Supplementary Table 2 ) diluted in PBTS buffer. After five washes of 20 min each in PBTS buffer diluted 1:2, samples were incubated for 2 h at room temperature with Alexa-conjugated secondary antibodies (Thermo Fisher). After three washes of 30 min each in PBTS buffer (diluted 1:2), and two washes of 10 min each in PBS, retinas were mounted with Fluoromount-G (SouthernBiotech). Immunofluorescence on paraffin sections The N1ICD epitope and the Jag1 ligand were detected with the tyramide signal amplification (TSA) kit (NEL774) procedure in paraffin sections after antigen retrieval. In brief, sections were dewaxed and rehydrated, followed by antigen retrieval in sub-boiling sodium citrate buffer (10 mM, pH 6.0) for 30 min. The slides were cooled down to room temperature for 30 min, followed by incubation for 30 min in 3% H 2 O 2 in methanol to quench endogenous peroxidase activity. Next, slides were rinsed in double-distilled H 2 O and washed three times for 5 min each in PBS, followed by blocking for 1 h in PBS containing 3% BSA, 200 mM MgCl 2 , 0.3% Tween 20 and 5% donkey serum. Sections were then incubated with primary antibody in the same solution overnight at 4 °C. After washes, slides were incubated for 2 h with anti-rabbit-HRP secondary antibody at room temperature, and, after washing, the signal was amplified using the TSA fluorescein kit (NEL774). Sections were mounted with Fluoromount-G (SouthernBiotech). All antibodies used are listed in Supplementary Table 2 . In vivo EdU labeling and EC proliferation detection To detect EC proliferation in adult livers, 20 μg per g body weight EdU (Invitrogen, A10044 ) was injected intraperitoneally into adult mice 5 h before dissection. Livers were isolated for cryosection analysis. EdU signals were detected with the Click-iT EdU Alexa Fluor 647 or 488 Imaging Kit (Invitrogen, C10340 or C10337 ). In brief, after all other primary and secondary antibody incubations, samples were washed according to the immunofluorescence staining procedure and then incubated with Click-iT EdU reaction cocktail for 40 min, followed by DAPI counterstaining. Image acquisition and analysis Immunostained organ sections were imaged at high resolution with a Leica SP5, SP8 or SP8 Navigator confocal microscope fitted with a ×10, ×20 or ×40 objective for confocal scanning. Individual fields or tiles of large areas were acquired from cryosections, vibratome or paraffin sections. Large Z -volumes of the vibratome samples were imaged for 3D representation. All images shown are representative of the results obtained for each group and experiment. Animals were dissected and processed under exactly the same conditions. Comparisons of phenotypes or signal intensity were made with pictures obtained using the same laser excitation and confocal scanner detection settings. Fiji/ImageJ was used to threshold, select and quantify objects in confocal micrographs. Both manual and automatic ImageJ public plug-ins and custom Fiji macros were used for quantification. Latex perfusion and CUBIC clearing Mice were euthanized in a CO 2 chamber. The abdominal cavity was opened, and the liver portal vein was exposed. With the help of a dissection microscope, latex (Injection Medium, Latex, Red, Laboratory Grade, Carolina, 868703) was injected in the portal vein with a 40G needle as previously described 59 . Perfusion was stopped as soon as latex was visually detectable in the liver surface vessels. Liver dissection was performed only 15 min after the perfusion to ensure latex solidification. The liver was then washed in PBS and put in PFA 4% in PBS at 4 °C overnight. After, three PBS washes for 15 min each were done at room temperature. To clear the organ, livers were incubated at 37 °C in CUBIC1 (ref. 60 ) solution (25 wt% urea, 25 wt% N′-Tetrakis(2-hydroxypropyl)ethylenediamine, 15 wt% Triton X-100) for a total 4 days, with the solution being exchanged every day. After clearing, liver images were captured with an Olympus camera connected to a Leica dissection scope with retroillumination. A magnification of ×0.8 was used. Western blot analysis For the analysis of protein expression, livers were transferred to a reagent tube and frozen in liquid nitrogen. On the day of immunoblotting, the tissue was lysed with lysis buffer (Tris-HCl pH 8, 20 mM, EDTA 1 mM, DTT 1 mM, Triton X-100 1% and NaCl 150 mM, containing protease inhibitors (Sigma, P-8340), phosphatase inhibitors (Calbiochem, 524629) and orthovanadate-Na 1 mM) and homogenized with a cylindrical glass pestle. Tissue and cell debris were removed by centrifugation, and the supernatant was diluted in loading buffer and analyzed by SDS–PAGE and immunoblotting. Membranes were blocked with BSA and incubated with the primary antibodies listed in Supplementary Table 2 . EC isolation for transcriptomic and proteomic analysis The following methods were used to isolate ECs for bulk RNA-seq, and proteomics analysis. At day 14 after the first tamoxifen injection, heart, lungs, liver and brain were dissected, minced and digested with 2.5 mg ml −1 collagenase type I (Thermo Fisher), 2.5 mg ml −1 dispase II (Thermo Fisher) and 50 ng ml −1 DNase I (Roche) at 37 °C for 30 min. Cells were passed through a 70-µm filter. Erythroid cells were removed by incubation with blood lysis buffer (0.15 M NH 4 Cl, 0.01 M KHCO 3 and 0.01 M EDTA in distilled water) for 10 min on ice. Cell suspensions were blocked in blocking buffer (DPBS containing no Ca 2+ or Mg 2+ and supplemented with 3% dialyzed FBS; Thermo Fisher). For EC analysis, cells were incubated at 4 °C for 30 min with APC-conjugated rat anti-mouse CD31 (1:200; BD Biosciences, 551262). DAPI (5 mg ml −1 ) was added to the cells immediately before fluorescence-activated cell sorting (FACS), which was performed with FACSAria (BD Biosciences) or Synergy 4L cell sorters. For bulk RNA-seq experiments, approximately 10,000–20,000 cells for each group of DAPI − APC-CD31 + ECs (for Dll4 loss of function and control) and DAPI − APC-CD31 + /MbTomato + ECs (for Rbpj loss of function and control) were sorted directly to RLT buffer (RNeasy Micro Kit, Qiagen). RNA was extracted with the RNeasy Micro Kit and stored at −80 °C. For proteomic analysis, approximately 3 × 10 6 DAPI − APC-CD31 + ECs per group were sorted directly to blocking buffer. Cells were spun down for 10 min at 350 × g , and the pellet was stored at −80 °C. To isolate ECs for scRNA-seq experiments, 1.5 mg of tamoxifen was injected on 4 consecutive days. At day 14 after the first tamoxifen injection, livers were dissected, minced and digested for 30 min with prewarmed (37 °C) dissociation buffer (2.5 mg ml −1 collagenase I (Thermo Fisher, 17100017), 2.5 mg ml −1 dispase II (Thermo Fisher, 17105041), 1 μl ml −1 DNase in PBS containing Ca 2+ and Mg 2+ (Gibco)). The digestion tube was agitated every 3–5 min in a water bath. At the end of the 30-min incubation, sample tubes were filled up to 15 ml with sorting buffer (PBS containing no Ca 2+ or Mg 2+ and supplemented with 10% FBS (Sigma, F7524)) and centrifuged (450 × g , 5 min, 4 °C). The supernatant was aspirated, and cell pellets were resuspended in 2 ml of 1x Red Blood Cell (RBC) Lysis Buffer (BioLegend, 420301) and incubated for 5 min on ice. We added 6 ml of sorting buffer to each sample, and samples were then passed through a 70-μm filter. Live cells were counted in a Neubauer chamber using trypan blue exclusion. Cells from each condition (4 × 10 6 per condition) were collected in separate tubes, and cells were incubated for 30 min with horizontal rotation in 300 µl of antibody incubation buffer (PBS + 1% BSA) containing 1 µl of CD31-APC, 1 µl of CD45-APC-Cy7, and 1 µl of hashtag oligo (HTO) conjugated antibodies (BioLegend). HTOs were used to label and distinguish the different samples when loaded on the same 10x Genomics port, thus also guaranteeing the absence of batch effects. After antibody incubation, samples were transferred to 15-ml Falcon tubes, 10 ml of sorting buffer were added, and samples were centrifuged (450 × g , 5 min, 4 °C). The supernatant was aspirated, pellets were resuspended in 1.5 ml of sorting buffer and transferred to Eppendorf tubes, and the resulting suspensions were centrifuged (450 × g , 5 min, 4 °C). The resulting pellets were resuspended in 300 µl of sorting buffer containing DAPI. Cells were sorted with a FACSAria Cell Sorter (BD Biosciences), and CD31 + CD45 − MbTomato + cells were sorted. BD FACSDiva v8.0.1 and FlowJo v10 were used for FACS data collection and analysis. Next-generation sequencing sample and library preparation Next-generation sequencing experiments were performed in the Genomics Unit at Centro Nacional de Investigaciones Cardiovasculares (CNIC). For bulk RNA-seq, control and Dll4 iDEC EC samples, 1 ng of total RNA was used to amplify the cDNA using the SMART-Seq v4 Ultra Low Input RNA Kit (Clontech-Takara) following manufacturer’s instructions. Then, 1 ng of amplified cDNA was used to generate barcoded libraries using the Nextera XT DNA Library Preparation Kit (Illumina). For control and Rbpj iDEC EC samples, between 400 pg and 3,000 pg of total RNA were used to generate barcoded RNA-seq libraries using the NEBNext Single Cell/Low Input RNA Library Prep Kit for Illumina (New England Biolabs) according to manufacturer’s instructions. For control and anti-Dll4 antibody-treated ECs, libraries were generated using the Ovation Single Cell RNA-Seq System (NuGEN) following manufacturer’s instructions. All libraries were sequenced on a HiSeq 2500 (Illumina). For scRNA-seq experiments, single cells were encapsulated into emulsion droplets using the Chromium Controller (10x Genomics). scRNA-seq libraries were prepared according to manufacturer’s instructions. The aim for target cell recovery for each port was in general 10,000 cells, with a target cell recovery of 2,000–2,500 cells per experimental condition labeled with a given hashtag antibody. Generated libraries were sequenced on a HiSeq 4000 or NextSeq 2000 (Illumina). Transcriptomic data analysis Transcriptomic data were analyzed by the Bioinformatics Unit at CNIC. For bulk RNA-seq, the number of reads per sample was between 12 million and 42 million. Reads were processed with a pipeline that assessed read quality using FastQC (Babraham Institute, ) and trimmed sequencing reads using cutadapt 61 , eliminating Illumina and SMARTer adaptor remains and discarding reads with <30 base pairs (bp). More than 93% of reads were kept for all samples. The resulting reads were mapped against the mouse transcriptomes GRCm38.76 and GRCm38.91, and gene expression levels were estimated with RSEM 62 . The percentage of aligned reads was above 83% for most samples. Expression count matrices were then processed with an analysis pipeline that used Bioconductor package limma 63 for normalization (using the trimmed mean of M values (TMM) method) and differential expression testing, taking into account only those genes expressed with at least 1 count per million (CPM) in at least two samples (the number of samples for the condition with the least replicates), and using a random variable to define blocks of samples obtained from the same animal. Changes in gene expression were considered significant if associated with a Benjamini and Hochberg–adjusted P value < 0.05. A complementary GSEA 64 was performed for each contrast, using the whole collection of genes detected as expressed (12,872 genes) to identify gene sets that had a tendency to be more expressed in either of the conditions being compared. We retrieved gene sets representing pathways or functional categories from the Hallmark, Kyoto Encyclopedia of Genes and Genomes (KEGG), Reactome, and BioCarta databases, and Gene Ontology (GO) collections from the Biological Process, Molecular Function and Cellular Component ontologies from MSigDB 65 . Enriched gene sets with a false discovery rate (FDR) < 0.05% were considered of interest. Data were analyzed with Python v2.7, using the Seaborn ( ) and Pandas ( ) libraries. The following pipeline was followed for scRNA-seq data processing and in silico EC selection. For alignment and quantification of gene expression, the reference transcriptome was built using mouse genome GRCm38 and Ensembl gene build v98 ( ). The phiYFP-sv40pA, MbTomato-2A-Cre-WPRE-sv40pa or CreERT2 transgene sequences expressed in the samples were added to the reference. Gene metadata were obtained from the corresponding Ensembl BioMart archive. Reads from hashtags and transcripts were processed, aligned and quantified using the Cell Ranger v4.0.0 pipeline. Single-cell analysis was based on Scater 66 and Seurat 67 packages. Low-quality cells were filtered out using the following criteria: total counts, >1,500 and <40,000; genes detected, >600; mitochondrial transcripts content, <25%; total counts/median, >0.1; hashtag counts, >100; hemoglobin transcripts, <0.1%; and percentage of counts in the top 50 genes, <65%. Cells were demultiplexed using the sample hashtag antibody signals (BioLegend). Counts were log-normalized and scaled, followed by principal component analysis (PCA) and clustering using the shared nearest-neighbors algorithm and Louvain clustering (settings as defaults except for the 1,000 most variable genes, 10 principal components, and a resolution of 0.5). Clusters and cells were classified based on the SingleR method 68 using Blueprint ENCODE and the Human Primary Cell Atlas cell-type profile collection. This identification was used to select ECs for the analysis and remove minor contaminants (T cells, B cells and monocytes). Hashtag-based doublets were removed, and only ECs were reclustered using the same procedure (with 2,000 variable genes, 7 PCs, a resolution of 0.3, and a random seed for uniform manifold approximation and projection (UMAP) = 123456) to get a final clustering that was later manually refined based on marker expression. Following cluster identification with the starting dataset, the remaining liver EC datasets were mapped using the FindTransferAnchors function from the Seurat R package using 30 PCA dimensions with the default settings. The following pipeline was followed for liver non-EC scRNA-seq. Cells were demultiplexed by applying the cellranger multi pipeline. The following quality-control steps were performed to minimize low-quality cells and improve posterior normalization and analysis: (1) a minimum of normalized counts per cell of 2,000 and a maximum of 30,000; (2) a minimum gene detection filter of 500 genes and a maximum of 6,000; (3) a maximum mitochondria content of 5%; (4) a maximum ribosomal content of 35%; (5) a maximum hemoglobin content of 1%; and (6) only single cells were selected, and doublets were filtered out in the cellranger multi demultiplexing step. Counts were log-normalized and scaled, followed by PCA and clustering using the shared nearest-neighbors algorithm and Louvain clustering (settings as defaults except for the 2,000 most variable genes, 30 principal components and a resolution of 0.8). Clusters and cells were classified based on the SingleR method using Blueprint ENCODE, Human Primary Cell Atlas, and mouse RNA-seq datasets available in the celldex package, as well as a recent liver single-cell dataset 69 , in order to classify each cluster to a different cell type. Final clustering was later manually refined based on marker expression. Liver EC proteomics Protein extraction from cell samples was carried out in the presence of SDS as described 70 . Protein concentration was determined by the RC DC Protein Assay (Bio-Rad Laboratories). Samples (100 μg) were subjected to overnight tryptic digestion using filter-aided sample preparation (FASP) technology (Expedeon) 71 . The resulting peptides were desalted on Oasis HLB C18 extraction cartridges (Waters Corporation) and dried down. The cleaned-up peptide samples were subjected to stable isotope labeling using isobaric tags for relative and absolute quantitation (iTRAQ 8-plex, AB Sciex) following the manufacturer’s instructions. The differentially tagged samples were then pooled and desalted on Oasis HLB C18 cartridges. A 100-μg aliquot of dried, labeled peptides was taken up in 0.1% trifluoroacetic acid and separated into five fractions by high pH reversed-phase peptide fractionation 72 . The bound peptides were eluted gradually with 12.5%, 15%, 17.5%, 20% and 50% acetonitrile, and the fractions obtained were vacuum-dried and stored at −20 °C for later use. The labeled peptide samples were taken up in 0.1% formic acid and analyzed on an EASY-nLC 1000 liquid chromatograph (Thermo Fisher Scientific) coupled to a Q Exactive HF mass spectrometer (Thermo Fisher Scientific). The peptide samples were loaded onto a C18 reversed-phase nano-precolumn (Acclaim PepMap 100; 75-μm internal diameter, 3-μm particle size and 2-cm length; Thermo Fisher Scientific) and separated on an analytical C18 nano-column (EASY-Spray column PepMap RSLC C18; 75-μm internal diameter, 3-μm particle size and 50-cm length; Thermo Fisher Scientific) using a linear gradient: 8–27% B for 240 min, 31–100% B for 2 min, 100% B for 7 min, 100–2% B for 2 min, and 2% B for 30 min (where A is 0.1% formic acid in high-performance liquid chromatography (HPLC)-grade water, and B is 90% acetonitrile, 0.1% formic acid in HPLC-grade water). Full MS spectra were acquired over the 400–1,500 mass-to-charge ( m / z ) range with 120,000 resolution, 2 × 10 5 automatic gain control, and 50-ms maximum injection time. Data-dependent tandem MS (MS/MS) acquisition was performed at 5 × 10 4 automatic gain control and 120-ms injection time, with a 2-Da isolation window and 30-s dynamic exclusion. Higher-energy collisional dissociation of peptides was induced with 31% normalized collision energy and analyzed at 35,000 resolution in the Orbitrap. Protein identification was carried out using the SEQUEST HT algorithm integrated in Proteome Discoverer v2.1 (Thermo Fisher Scientific). MS/MS scans were matched against a mouse protein database (UniProtKB release 2017-07) supplemented with pig trypsin and human keratin sequences. Parameters for database searching were as follows: trypsin digestion with a maximum of two missed cleavage sites allowed, precursor mass tolerance of 800 ppm, and a fragment mass tolerance of 0.02 Da. Amino-terminal and Lys iTRAQ 8-plex modifications were set as fixed modifications, whereas Met oxidation, Cys carbamidomethylation, and Cys methylthiolation were set as variable modifications. The identification results were analyzed with the probability ratio method 73 . An FDR for peptide identification was calculated based on searching against the corresponding inverted database using the refined method 74 , 75 after precursor mass tolerance postfiltering at 20 ppm. Quantitative information was extracted from the intensity of iTRAQ reporter ions in the low-mass region of the MS/MS spectra 76 . The comparative analysis of protein abundance changes relied on the weighted scan peptide–protein (WSPP) statistical model 77 by means of the SanXoT software package as described 78 . As input, WSPP uses a list of quantifications in the form of log 2 ratios (each cell sample versus the mean of the three wild-type cell samples) with their statistical weights. From these, WSPP generates the standardized forms of the original variables by computing the quantitative values expressed in units of standard deviation around the means (Zq). For the study of coordinated protein alterations, we used the Systems Biology Triangle (SBT) algorithm, which estimates functional category averages (Zc) from protein values by performing the protein-to-category integration, as described 76 . The protein category database was built up using annotations from the GO database. Statistical analysis and reproducibility All bar graphs show mean ± s.d. Experiments were repeated with independent animals. Comparisons between two groups of samples with a Gaussian distribution were by unpaired two-tailed Student’s t -test. Comparisons among more than two groups were made by one-way or two-way analysis of variance (ANOVA) followed by multiple comparison tests as indicated in the Source Data . All calculations were done in Microsoft Excel, and final data points were analyzed and represented with GraphPad Prism. No randomization or blinding was used, and animals or tissues were selected for analysis based on their genotype, the detected Cre-dependent recombination frequency, and the quality of multiplex immunostaining. Sample sizes were chosen according to the observed statistical variation and published protocols. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Data availability RNA-seq data can be viewed in the Gene Expression Omnibus (GEO) database under accession number GSE231613 (SuperSeries of GSE229793 and GSE231612 ). Instructions and code to reproduce all scRNA-seq results can be found at . Proteomics data can be found in the Proteomics Identifications (PRIDE) database under accession number PXD041349 . Unprocessed original photographs of the data are available upon request. All other data supporting the findings in this study are included in the main article and associated files.

Research by scientists at the Centro Nacional de Investigaciones Cardiovasculares (CNIC) has demonstrated that the on-target molecular and cellular effects of medicines used to modulate the formation of new blood vessels (angiogenesis) in cardiovascular disorders and cancer are not responsible for the toxicity and vascular pathology triggered by these drugs. The study is published in Nature Cardiovascular Research. Study leader Rui Benedito commented that, "Our results not only significantly increase our understanding of the biology of blood vessels, but will also help in the selection of the most effective and safe way to modulate angiogenesis in ischemic tissues or in cancer." The vascular system supplies oxygen and nutrients to the tissues and organs of the body. But the blood vessels do more than just conduct blood; they contribute actively to the physiology and homeostasis of all tissues and organs throughout life. Most blood vessels in the body are in an inactive state, which they maintain by expressing a large number of genes, including the genes of the signaling pathway mediated by Delta ligands and Notch receptors. Several drugs have been developed in recent years that either block or induce angiogenesis in cardiovascular disorders or cancer. A group of these compounds in clinical use inhibit different components of the Delta-Notch signaling pathway, which plays important roles in angiogenesis and in the maintenance of blood vessels in the inactive state. These compounds, by modulating the growth of blood vessels, efficiently block tumor growth. These compounds are also able to induce angiogenesis in ischemic tissues, thereby improving tissue regeneration and function. However, these drugs can also cause vascular injury in organs with no previous disease, including the liver and heart, and this has reduced clinical interest in their use. Until now, this vascular toxicity was thought to be due to the expression of genes that promote angiogenesis, leading to the appearance of neoplasms or tumors in the affected blood vessels. Thanks to the use of advanced genetic mouse models, high-resolution confocal microscopy, and single-cell sequencing and proteomics techniques, the team discovered that the vascular toxicity linked to these drugs is instead due to a change in the vascular architecture that impedes correct blood flow. "Our study shows that the vascular pathology that can result from treatment with these drugs is unrelated to the expression of genes involved in angiogenesis or the appearance of neoplasms," said Rui Benedito. The researchers showed that these changes happen even when they blocked cell activation and the expression of angiogenesis-related genes. Therefore, explained Rui Benedito, "although the neoplasms and the expression of genes associated with angiogenesis are associated with the change in vascular architecture, they are not the cause of this change." First author Macarena Fernández Chacón explained that, "after analyzing different genes and drugs targeting blood vessels, we have found new ways to control pathological angiogenesis without significantly affecting vascular architecture in other organs, thus avoiding toxicity."

10.1038/s44161-023-00272-4

Physics

New mechanism allows lower energy requirement for OLED displays

Selective triplet exciton formation in a single molecule, Nature (2019). DOI: 10.1038/s41586-019-1284-2 , www.nature.com/articles/s41586-019-1284-2 Journal information: Nature

http://dx.doi.org/10.1038/s41586-019-1284-2

https://phys.org/news/2019-06-mechanism-energy-requirement-oled.html

Abstract The formation of excitons in organic molecules by charge injection is an essential process in organic light-emitting diodes (OLEDs) 1 , 2 , 3 , 4 , 5 , 6 , 7 . According to a simple model based on spin statistics, the injected charges form spin-singlet (S 1 ) excitons and spin-triplet (T 1 ) excitons in a 1:3 ratio 2 , 3 , 4 . After the first report of a highly efficient OLED 2 based on phosphorescence, which is produced by the decay of T 1 excitons, more effective use of these excitons has been the primary strategy for increasing the energy efficiency of OLEDs. Another route to improving OLED energy efficiency is reduction of the operating voltage 2 , 3 , 4 , 5 , 6 . Because T 1 excitons have lower energy than S 1 excitons (owing to the exchange interaction), use of the energy difference could—in principle—enable exclusive production of T 1 excitons at low OLED operating voltages. However, a way to achieve such selective and direct formation of these excitons has not yet been established. Here we report a single-molecule investigation of electroluminescence using a scanning tunnelling microscope 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 and demonstrate a simple method of selective formation of T 1 excitons that utilizes a charged molecule. A 3,4,9,10-perylenetetracarboxylicdianhydride (PTCDA) molecule 21 , 22 , 23 , 24 , 25 adsorbed on a three-monolayer NaCl film atop Ag(111) shows both phosphorescence and fluorescence signals at high applied voltage. In contrast, only phosphorescence occurs at low applied voltage, indicating selective formation of T 1 excitons without creating their S 1 counterparts. The bias voltage dependence of the phosphorescence, combined with differential conductance measurements, reveals that spin-selective electron removal from a negatively charged PTCDA molecule is the dominant formation mechanism of T 1 excitons in this system, which can be explained by considering the exchange interaction in the charged molecule. Our findings show that the electron transport process accompanying exciton formation can be controlled by manipulating an electron spin inside a molecule. We anticipate that designing a device taking into account the exchange interaction could realize an OLED with a lower operating voltage. Main A scanning tunnelling microscope (STM) combined with an optical detection system provides atomically precise spectroscopy for investigating both optical and electron-transport processes at the nanometre scale, and has elucidated fundamental exciton physics in well defined molecular systems 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 . Here we applied this technique to a PTCDA molecule (Fig. 1a ) adsorbed on a three-monolayer-thick NaCl(100) film grown on a Ag(111) surface—we refer to this system as PTCDA/NaCl(3ML)/Ag(111). In an STM image of PTCDA/NaCl(3ML)/Ag(111) (Fig. 1b ), the molecules appear as double-lobed structures with a nodal plane along the long axis of the molecule 21 , 22 , 23 . The adsorption angle of PTCDA is tilted by 45° with reference to the [100] direction of the NaCl film, which is consistent with the adsorption structure 24 of PTCDA on bulk NaCl(100). Fig. 1: STM measurements of the PTCDA/NaCl(3ML)/Ag(111) system. a , Molecular structure of PTCDA (brown, C; red, O; beige, H) with long-axis direction. b , Constant-current STM image showing three PTCDA molecules on the NaCl(3ML) surface (25 nm × 25 nm, sample voltage V s = 1.0 V, tunnelling current I t = 10 pA). The directions of the long axes of the PTCDA molecules are indicated by the black double-headed arrows; white arrows indicate the [100] and [010] directions of the NaCl surface. c , The d I t /d V s spectrum of the PTCDA/NaCl(3ML)/Ag(111) system. d , STM images of PTCDA measured at the sample voltages indicated by the black arrows in c and shown at the bottom right of each image (4 nm × 4 nm, I t = 5 pA). Full size image To examine electron transport via PTCDA/NaCl(3ML)/Ag(111), we measured a differential conductance (d I t /d V s ) spectrum, where I t is the tunnelling current and V s is the sample bias voltage (Fig. 1c ). The d I t /d V s spectrum exhibits two resonant tunnelling peaks, at V s = −0.8 V and 1.1 V, and the peak onsets are at −0.5 V and 0.7 V, respectively. Double-lobed structures are observed in the STM images at both positive and negative sample voltages, and rounded structure is observed in the gap region of these peaks. The features in the d I t /d V s spectrum are similar to those in the previously reported 21 , 22 d I t /d V s spectrum of PTCDA/NaCl(2ML)/Ag(111), indicating that electron transport occurs by the same mechanism on NaCl films of different thicknesses. In the previous report 21 , 22 , PTCDA was considered to be negatively charged on NaCl(2ML)/Ag(111) owing to the high electron affinity of the molecule and the low work function of the substrate 21 , 22 , 23 . The authors 21 , 22 proposed that the lowest unoccupied molecular orbital (LUMO) was responsible for the resonant tunnelling channels observed at both positive and negative voltages. This interpretation is also valid for the PTCDA/NaCl(3ML)/Ag(111) system, and is supported by the similar double-lobed structure observed in the STM images at both voltage polarities (Fig. 1d ), which are almost identical to the spatial distribution of the LUMO determined from PTCDA/NaCl(3ML)/Au(111) (see Extended Data Figs. 1 , 2 and Methods ). In order to confirm exciton formation by charge injection, we performed scanning tunnelling luminescence (STL) measurements on PTCDA/NaCl(3ML)/Ag(111) (Fig. 2a ). The STL spectrum obtained with the tip located on a PTCDA molecule shows a broad peak ranging from 1.5 eV to 3.0 eV, and several sharp peaks at approximately 1.3 eV, 2.25 eV and 2.45 eV (Fig. 2b ). The broad peak is attributed to the radiative decay of the plasmon localized near the gap between the STM tip and the substrate 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 . To reveal the detailed structure of the sharp peaks, STL spectra were measured at a higher energy resolution in the ranges 2.37–2.52 eV (Fig. 2c ) and 1.25–1.40 eV (Fig. 2d ). In Fig. 2c , the main peak appears at 2.45 eV (corresponding to a wavelength of 506 nm), and smaller peaks are observed near the main peak. On the basis of good agreement with previous photoluminescence (PL) results 25 , the main peak at 2.45 eV is assigned to the 0–0 transition of fluorescence, which represents the transition between the vibrational ground states of the S 1 exciton state and the ground state (S 0 ). Because the positions of the small peaks near the 0–0 transition peak are almost identical to those in previous PL data (Extended Data Fig. 3 ), they are attributed to vibrational satellites of the fluorescence. Moreover, the sharp peak at 2.25 eV in Fig. 2b is also attributed to vibrational satellites. Fig. 2: STL fluorescence and phosphorescence spectra of the PTCDA/NaCl(3ML)/Ag(111) system. a , Schematic of STL measurement. b , STL spectrum of a PTCDA molecule ( V s = −3.5 V, I t = 50 pA, exposure time t = 60 s) at low energy resolution; see below for red and blue shaded regions. The tip position is shown as a red dot in the inset STM image (4 nm × 4 nm, V s = 1.0 V, I t = 5 pA). c , d , STL spectra at medium energy resolution ( V s = −3.5 V, I t = 50 pA, t = 180 s): c , photon energy 2.37–2.52 eV (blue region in b ); and d , photon energy 1.25–1.40 eV (red region). e , Peak widths (shown by horizontal arrows) of the 0–0 transitions ( I t = 50 pA, t = 600 s) of fluorescence (blue, V s = −3.5 V) and phosphorescence (red, V s = −2.5 V) at high energy resolution. Black lines show Lorentzian fitting results. Full size image In the lower-energy luminescence spectrum (Fig. 2d ), the main peak is at 1.33 eV (932 nm), and vibrational satellites are also observed. The positions of the vibrational satellites are in good agreement with those seen in the fluorescence spectrum (Extended Data Fig. 4 ). Because there have been no reports on the low-energy luminescence of PTCDA, we performed time-dependent density functional theory (TD-DFT) calculations. The energies of the S 1 and T 1 excitons (referred to hereafter as S 1 and T 1 ) were calculated to be 2.49 eV and 1.18 eV, respectively, which are in reasonable agreement with the observed luminescence peak positions. Thus, the results strongly suggest that the main peak in Fig. 2d originates from the 0–0 transition of phosphorescence. Next, we analysed the peak widths of the 0–0 transitions, because the peak width is dependent on the lifetime of the excited state (longer lifetimes provide narrower peaks). Figure 2e shows STL spectra measured at even higher spectral resolution. The 0–0 transition peak of fluorescence (blue dots) and the lower-energy luminescence peak (red dots) were fitted using Lorentzian functions (black lines), and the peak widths (full-width at half-maximum) were determined to be 5.11 meV and 0.63 meV, respectively. The observed peak width of 0.63 meV is considerably sharper than in the previously reported STL fluorescence spectra 15 , suggesting that this peak does not correspond to a fluorescence transition. Because, in general, the lifetime of T 1 is longer than that of S 1 , the sharpness of the low-energy luminescence peak also supports our interpretation that the peak at 1.33 eV originates from phosphorescence. On the basis of all the observations and discussions presented thus far, we conclude that charge-neutral excitons (S 1 and T 1 ) were formed by charge injection from the STM tip to a negatively charged PTCDA, and that they manifested themselves as fluorescence and phosphorescence in the STL spectra. In our measurement, phosphorescence as well as fluorescence are enhanced by the localized plasmon through exciton–plasmon coupling 26 , so substantial luminescence signals were observed (Extended Data Fig. 5 ). It should be noted that this is, to our knowledge, the first observation of phosphorescence from PTCDA, and also the first observation of phosphorescence in a single-molecule STL measurement. To investigate the exciton formation mechanism, we examined the voltage dependence of the STL spectra. Figure 3a, b shows a series of phosphorescence and fluorescence spectra measured at different sample voltages. As the sample voltage was decreased, the phosphorescence first appeared at −2.1 V, and the intensity of phosphorescence increased gradually with decreasing sample voltage. In a similar way, the fluorescence appeared at −3.3 V, and its intensity increased as the sample voltage was decreased. The photon intensities of phosphorescence and fluorescence were determined by Lorentzian peak fitting and are plotted as a function of the sample voltage in Fig. 3c , which clearly shows that the threshold voltages for phosphorescence ( \({V}_{{\rm{th}}}^{{\rm{p}}}\) ) and fluorescence ( \({V}_{{\rm{th}}}^{{\rm{f}}}\) ) are at approximately −2.1 V and −3.3 V, respectively. Therefore, PTCDA exhibits only phosphorescence between \({V}_{{\rm{th}}}^{{\rm{p}}}\) and \({V}_{{\rm{th}}}^{{\rm{f}}}\) , from −2.1 to −3.3 V. The value of \({V}_{{\rm{th}}}^{{\rm{p}}}\) is smaller than 2.45 V, which corresponds to the energy of S 1 (2.45 eV, determined from the fluorescence peak); this is clear evidence of the direct formation of T 1 without passing through S 1 at low applied voltage. To elucidate the origin of \({V}_{{\rm{th}}}^{{\rm{p}}}\) at −2.1 V, we conducted a d I t /d V s measurement in the negative sample voltage region (Fig. 3d ) 12 , 14 . In addition to the resonant features around the Fermi level (Fig. 1c ), the d I t /d V s spectrum shows a new strong resonant peak at −2.5 V with an onset at approximately −2.1 V. The correspondence between the onset voltage of the d I t /d V s peak and \({V}_{{\rm{th}}}^{{\rm{p}}}\) at −2.1 V indicates that the phosphorescence was triggered by a resonant tunnelling charge injection into a molecular orbital 10 , 11 , 12 , 13 , 14 . Although another peak is expected around V s = −3.5 V, we could not measure it because the high tunnelling current in a high applied voltage region often induced molecular movement during the d I t /d V s measurement. Fig. 3: Voltage dependence of STL spectra of the PTCDA/NaCl(3ML)/Ag(111) system. a , b , Series of STL spectra ( I t = 30 pA, t = 180 s): a , phosphorescence ( V s = −2.0 V to −2.4 V, labels on curves); and b , fluorescence ( V s = −3.1 V to −3.5 V). c , Sample voltage dependence of the intensities of fluorescence (blue circles, left-hand vertical axis) and phosphorescence (red squares, right-hand vertical axis). The threshold voltages for phosphorescence ( \({V}_{{\rm{th}}}^{{\rm{p}}}\) ) and fluorescence ( \({V}_{{\rm{th}}}^{{\rm{f}}}\) ) are shown by dashed lines. d , The d I t /d V s spectrum of the PTCDA/NaCl(3ML)/Ag(111) system. e , f , Schematic images of the exciton formation mechanism; the blue arrows represent electrons. (See main text for details.) e , V s < −2.1 V; and f , V s < −3.3 V . Full size image The proposed mechanism of the selective T 1 formation in the low applied voltage region is shown in Fig. 3e . PTCDA has a charge of −1 owing to adsorption on the NaCl(3ML)/Ag(111) surface, and rapidly transits between the −1 charged state and the neutral ground state (S 0 ) under the resonant tunnelling condition ( V s < −0.5 V) 21 , 22 . When a further voltage is applied ( V s < −2.1 V), the electron occupying the highest occupied molecular orbital (HOMO) can be removed from either S 0 or the −1 charged state. Assuming that electron removal occurs from the −1 charged state, the two electrons in the HOMO have different energies owing to the exchange interaction with the electron in the LUMO. If we increase the applied voltage in this situation, it is expected that the electron in the HOMO whose spin is ‘anti-parallel’ to the spin of the electron in the LUMO will be removed first at \({V}_{{\rm{th}}}^{{\rm{p}}}=-2.1{\rm{V}}\) , leading to the exclusive formation of T 1 (Fig. 3e ). When further voltage is applied, the removal of the other ‘parallel-spin’ electron in the HOMO becomes possible at \({V}_{{\rm{th}}}^{{\rm{f}}}=-3.3{\rm{V}}\) ; then both T 1 and S 1 are formed (Fig. 3e, f ). The difference between the threshold voltages for phosphorescence and fluorescence ( \({\rm{\Delta }}{V}_{{\rm{th}}}={V}_{{\rm{th}}}^{{\rm{p}}}-{V}_{{\rm{th}}}^{{\rm{f}}}\) ) is 1.2 V. This value of Δ V th corresponds to the S 1 –T 1 energy difference (ΔE ST ) or the exchange interaction energy, 1.12 eV, determined from the STL fluorescence (2.45 eV) and phosphorescence (1.33 eV) peaks. Therefore, the model process given in Fig. 3e can reasonably explain the voltage dependence results. The fact that the value of Δ V th is somewhat larger than ΔE ST is probably due to the potential drop in the NaCl film 27 , 28 . If we instead assumed that an electron is removed from the HOMO of S 0 and an electron is supplied to the LUMO, both T 1 and S 1 would be formed simultaneously, which is inconsistent with the observed STL voltage dependence. Furthermore, other reported mechanisms of exciton formation by energy transfer from a localized plasmon 8 , 13 , 15 or from an inelastic tunnelling electron 16 , 20 , 29 also cannot fully explain our observations (see Methods ). On the basis of the above discussions, we conclude that the removal of the ‘anti-parallel’ electron from the HOMO of the −1 charged state is the primary mechanism of T 1 formation this system. This proposed exciton formation mechanism is substantiated by theoretical simulations of the voltage dependence of the STL measurements and the d I t /d V s spectrum based on the many-body description using the Hubbard non-equilibrium Green’s function technique 30 (see Supplementary Information ). To reveal the submolecular spatial features of the single-molecule phosphorescence, an STL map was obtained at V s = -2.5 V (Fig. 4a ). Phosphorescence spectra were measured at selected pixels on and near the molecule, and the intensity of phosphorescence was determined by Lorentzian fitting at each pixel to create the phosphorescence STL map. There are two bright areas near the edges on the long axis of PTCDA; this is different from the STM topographic image obtained at V s = –2.5 V (Fig. 4b inset). Fig. 4: Phosphorescence STL map of the PTCDA/NaCl(3ML)/Ag(111) system. a , Phosphorescence STL spectra were measured at selected pixels in a 3.3 nm × 3.3 nm area (shown) with V s = −2.5 V, I t = 30 pA, t = 30 s per pixel. The intensities of the 0–0 peak of phosphorescence are measured by Lorentzian fitting and plotted (in counts; see colour scale at bottom right). Structure at top right shows molecular direction. b , d I t /d V s spectra obtained at four selected points (called 1, 2, 3 and 4) in PTCDA; spectra are labelled with point number. Tip positions are shown as coloured dots and numbered in the inset STM image ( V s = −2.5 V, I t = 10 pA). The d I t /d V s spectrum at point 4 is identical to the spectrum in Fig. 3d . Full size image On the basis of the T 1 creation mechanism (Fig. 3e ), it is expected that the bright spots in the phosphorescence map would correspond to the locations where anti-parallel electron removal occurs efficiently. To verify this, d I t /d V s spectra (Fig. 4b ) were measured at four selected points, which are shown in the inset to Fig. 4b . The substantial resonant feature at −2.5 V was only observed when the tip was located at position 4, where intense phosphorescence occurs. In contrast, the resonant feature was absent at the positions where the phosphorescence intensity is small (positions 2 and 3), supporting our expectation. The peak around −2.5 V can also be observed at the centre of the molecule (position 1), despite the very weak phosphorescence intensity at this position. This can be explained by considering the coupling strength between the transition dipole moment of PTCDA and the localized plasmon 11 , 13 , 14 . We conclude that the spatial distribution of the resonant tunnelling channel at around −2.5 V by which T 1 is created and the strength of the coupling between the transition dipole moment and the plasmon are responsible for the spatial distribution in the phosphorescence STL map. We have demonstrated selective T 1 formation by spin-selective electron transport through a single molecule that has an unpaired electron. Here, we prepared a negatively charged molecule by tuning the energy level alignment between the LUMO of the molecule and the Fermi level of the substrate. Such tunability indicates that spin-selective electron transport could in principle occur in various combinations of molecules and electrodes. For example, in addition to charged molecules, radical molecules, which also have an unpaired electron, could be suitable for implementing selective T 1 formation in OLEDs 7 . We note that selective T 1 formation based on an exchange interaction could be a valid explanation of the recently reported low-voltage operation of OLEDs 5 , 6 . The work we present here also proves the ability of our method to monitor transitions among spin states with spin angular momenta s of 0 (S 0 , S 1 ), 1/2 (charged state) and 1 (T 1 ). The combination of the versatile functions of the STM provides a unique and powerful experimental platform for quantum spintronics and excitonics at the single-atom and single-molecular level. Methods STM/STS observations All experiments were conducted using a low-temperature STM (Omicron) operating at 4.7 K under ultrahigh vacuum. Differential conductance (d I t /d V s ) spectra were measured using a standard lock-in technique with a bias modulation of 20 mV at 625.5 Hz. Preparation of sample and tip Clean Ag(111) and Au(111) surfaces were prepared by repeated cycles of Ar + ion sputtering and annealing. The deposition of NaCl onto Ag(111) or Au(111) was performed using a home-made evaporator heated to 850 K. PTCDA was deposited onto an NaCl/Ag(111) or NaCl/Au(111) surface, which was cooled to 4.7–10 K in the STM head, using another home-made evaporator heated to 620 K. The STM tip was prepared by electrochemical etching of a Ag wire in HClO 4 /C 2 H 5 OH electrolyte and conditioned by controlled-indentation and voltage pulse on the Ag(111) or Au(111) surface. STL measurement The STM stage was equipped with a lens (with a solid angle of about 0.5 sr). The emitted light was collimated using the lens and directed out of the ultrahigh vacuum chamber, where it was refocused onto a grating spectrometer (Acton SpectraPro 2300i, Princeton Instruments) with a charge- coupled-device photon detector (Spec 10, Princeton Instruments) cooled with liquid nitrogen. Gratings with either 50 grooves mm −1 (Fig. 2b , Extended Data Fig. 5 ), 300 grooves mm −1 (Fig. 2c, d , 3a, b , and Extended Data Figs. 3 , 4 ), or 1,200 grooves mm −1 (Fig. 2e ) were used for the optical measurements. In our sample, NaCl(3ML) films as well as NaCl(2ML) films were grown on the metal surface. As reported previously 31 , the luminescence intensity is strongly dependent on the thickness of the NaCl film. In the case of the PTCDA/NaCl(3ML)/Ag(111) system, phosphorescence intensity is 10 times stronger than that of PTCDA/NaCl(2ML)/Ag(111). Strong luminescence signals enable us to clearly determine the threshold voltage for luminescence and to obtain STL mapping. For this technical reason, we focus in this Letter on PTCDA adsorbed on NaCl(3ML) film. The typical tip position is shown as a red dot in the STM image of Fig. 2b inset. DFT and TD-DFT calculation First principles calculations based on density functional theory (DFT) were performed using the triple-zeta valence basis augmented with diffuse functions and double polarization functions (known as the 6-311++G(2d,2p) basis) implemented in the software package Gaussian 16 32 . Following the previous report 33 , the hybrid functional B3LYP was used. In order to obtain the structural information, the geometry of the ground electronic state of the PTCDA molecule was optimized, and the vibrational frequencies of the obtained geometry were computed to ensure that all positive frequencies were obtained. Furthermore, in order to investigate the luminescence spectra of the PTCDA molecule, geometry optimization and vibrational analysis for excited electronic states (S 1 and T 1 ) were performed using time-dependent density functional theory (TD-DFT) with the Tamm-Dancoff approximation at the B3LYP/6-311++G(2d,2p) level. The information obtained for the vibrational frequencies of the ground and excited electronic states was used to calculate the vibrational overlap integrals accompanying electronic transitions. The calculated energy for the transition between the ground vibrational states of S 1 and S 0 (T 1 and S 0 ) was 2.49 eV (1.18 eV). In the previous report, the vertical excitation energy from S 0 to S 1 was calculated as 2.39 eV by TD-DFT calculation at the B3LYP/6-311++G(d,p) level, which is almost identical to our result 33 . Analysis of single-molecule electroluminescence Theoretical analysis of electroluminescence from an isolated PTCDA molecule adsorbed on NaCl(3ML)/Ag(111) is conducted within the basis of the many-body state representation of the molecule. The Hubbard non-equilibrium Green’s function (NEGF) method, which was recently introduced by us 30 , 34 , was applied to simulate tunnelling current and electroluminescence in the system composed of the single molecule positioned between two metal electrodes. Details are given in the Supplementary Information . STM and d I t /d V s measurements of PTCDA on NaCl(3ML) A PTCDA molecule on the NaCl(3ML)/Ag(111) surface is considered to be negatively charged owing to the high electron affinity of the molecule and the low work function of the substrates 21 , 22 , 23 . The electron transport properties are affected by the charge state of the molecule. Therefore, we prepared neutral PTCDA and compared the results of neutral and charged molecules to understand the electron transport mechanism in the PTCDA/NaCl(3ML)/Ag(111) system. We deposited PTCDA onto a NaCl(3ML)/Au(111) surface, which has 35 , 36 , 37 , 38 , 39 a larger work function than that of NaCl(3ML)/Ag(111). The d I t /d V s spectrum of the PTCDA/NaCl(3ML)/Au(111) system exhibits two resonant tunnelling peaks with onsets at V s = −2.8 V and 0.9 V (Extended Data Fig. 1a ); these features are different from those in the d I t /d V s spectrum of the PTCDA/NaCl(3ML)/Ag(111) system (Fig. 1c ). STM images were obtained at the sample voltages indicated by the black arrows in Extended Data Fig. 1a . In order to prevent the STM tip from picking up the molecule, we conducted constant-height-mode scanning at −3.0 V and suppressed the magnitude of the tunnel current (Extended Data Fig. 1b ). STM images at 0.6 V and 1.0 V were obtained with the constant-current mode (Extended Data Fig. 1c, d ). An eight-lobed structure is observed in the STM image at the negative sample voltage, and a double-lobed structure is observed at the positive sample voltage. The voltage dependence of STM images for the PTCDA/NaCl(3ML)/Au(111) system is different to that for the PTCDA/NaCl(3ML)/Ag(111) system (shown in Fig. 1d ). Based on the good agreements with the spatial distribution of molecular orbitals calculated by DFT (Extended Data Fig. 1e ), the eight-lobed structure is attributed to the spatial distribution of the HOMO, and the double-lobed structure is attributed to the spatial distribution of the LUMO. From these results, it is concluded that PTCDA is neutral on the NaCl(3ML)/Au(111) surface, and that the peaks at V s = −2.8 V and 0.9 V in the d I t /d V s spectrum originate from resonant tunnelling through the HOMO and the LUMO, respectively. The spatial distribution of the LUMO in the PTCDA/NaCl(3ML)/Au(111) system (Extended Data Fig. 1d ) is almost identical to the double-lobed structures observed in the PTCDA/NaCl(3ML)/Ag(111) system (Fig. 1d ). Therefore, it is suggested that the LUMO is responsible for both the resonant tunnelling channels with onsets at −0.5 V and 0.7 V observed in the d I t /d V s spectrum of the PTCDA/NaCl(3ML)/Ag(111) system (Fig. 1c and Extended Data Fig. 2a ). Considering that PTCDA is negatively charged on the NaCl(3ML)/Ag(111) surface 21 , 22 , 23 , the following model can be proposed for the electron tunnelling. At negative sample voltage ( V s < −0.5 V), the pre-existing electron in the LUMO first moves to the tip, and the neutral state (S 0 ) is formed. The neutral state returns to the initial state (the −1 charged state) by means of an electron supplied from the substrate to the LUMO (Extended Data Fig. 2b ). Similarly, at positive voltage ( V s > 0.7 V), the −2 charged state is formed by injection of a second electron into the singly occupied LUMO, and the molecule returns to the initial state by passage of an electron to the substrate (Extended Data Fig. 2c ). A similar mechanism was proposed in previous reports dealing with the PTCDA/NaCl(2ML)/Ag(111) system 21 , 22 . In the previous reports, the authors speculated that the resonant tunnelling channel at negative sample voltage through the LUMO overlapped with another tunnelling channel through the HOMO. However, here we propose that only the LUMO is responsible for the resonant tunnelling channel in the range −2.1 V < V s < −0.5 V, according to the exciton formation mechanism (Fig. 3e ) that is based on the STL voltage dependence in combination with the d I t /d V s measurement results. The HOMO starts to contribute to the tunnelling at V s < −2.1 V in our model. Notation of molecular orbitals An electron occupies the LUMO of a PTCDA molecule owing to the molecule’s adsorption onto the NaCl(3ML)/Ag(111) surface 21 , 22 . This process can be described in another way: the LUMO becomes a singly occupied molecular orbital (SOMO). ‘SOMO’ is widely used to describe the character of electron transport in this type of junction 40 , 41 . However, in this Letter, we prefer to use the names of intrinsic molecular orbitals—HOMO or LUMO—for labelling, and we describe ‘SOMO’ as ‘LUMO with an unpaired electron’, because this description is useful for explaining the optical transitions. Discerning two exciton formation mechanisms Two exciton formation mechanisms have been reported in single-molecule electroluminescence measurements using an STM 7 . One is exciton formation by charge injection to the molecular orbitals 10 , 11 , 12 , 13 , 14 , 31 , and the other is exciton formation by energy transfer from a localized plasmon or from an inelastic tunnelling electron 12 , 13 , 15 , 16 , 42 . In the former mechanism, the exciton is created by resonant tunnelling to a molecular orbital 10 , 11 , 12 , 13 , 14 , 31 , and a clear peak originating from resonant tunnelling is therefore expected in the d I t /d V s spectrum. As clearly shown in previous reports 12 , 14 , the position of the peak onset in the d I t /d V s spectrum corresponds to the threshold voltage for luminescence ( V th ) in this mechanism. In contrast, in the latter mechanism, electronic transition owing to energy transfer from a localized plasmon or from an inelastic tunnelling electron is responsible for exciton formation 8 , and V th thus corresponds to the excitation or absorption energy of the molecule. In particular, when the Stokes shift (the energy difference between absorption and emission) of the molecule is quite small, V th corresponds to the emission energy of the molecule, as observed in previous reports 12 , 15 , 42 . In this energy transfer mechanism, the molecule is excited by the inelastic tunnelling electron itself 17 or by the localized plasmon generated by inelastic tunnelling 8 , 15 , 42 . Therefore, no apparent peak is expected 12 , 14 in the d I t /d V s spectrum around V th , because this process is based on inelastic tunnelling which has a much smaller tunnelling probability than resonant (elastic) tunnelling. Comparison between V th and the onset voltage of a d I t /d V s peak therefore enables us to determine the exciton formation mechanism in the single-molecule electroluminescence. The exciton creation mechanism Here we consider the exciton creation mechanism in the PTCDA/NaCl(3ML)/Ag(111) system. For PTCDA in isolated form, a small Stokes shift (<1 meV) was confirmed by PL measurements 25 . Therefore, it is expected that V th corresponds to the emission energy if a PTCDA molecule is excited dominantly by the energy transfer mechanism. However, the \(\left|{V}_{{\rm{th}}}^{{\rm{p}}}\right|\) value of 2.1 V is substantially larger than 1.33 V (determined from the energy of phosphorescence, 1.33 eV) and the \(\left|{V}_{{\rm{th}}}^{{\rm{f}}}\right|\) value of 3.3 V is also larger than 2.45 V (determined from the energy of fluorescence, 2.45 eV), indicating the energy transfer mechanism has a negligible effect in this system. In our experiment, the threshold voltage for phosphorescence was measured as \({V}_{{\rm{t}}{\rm{h}}}^{{\rm{p}}}=-2.1\,\,{\rm{V}}\) (Fig. 3c ), and the d I t /d V s spectrum exhibited a resonant tunnelling peak at −2.5 V with an onset at −2.1 V (Fig. 3d ). The correspondence between the peak onset voltage in the d I t /d V s spectrum and \({V}_{{\rm{th}}}^{{\rm{p}}}\) at −2.1 V suggests that the phosphorescence from PTCDA was triggered by resonant tunnelling. On the basis of all the results and discussions, we conclude that the primary exciton formation mechanism in the PTCDA/NaCl(3ML)/Ag(111) system is resonant tunnelling of the ‘antiparallel-spin’ electron from the −1 charged PTCDA molecule, by which T 1 is selectively created. Data and code availability The data that support the findings of this study and associated codes are available from the corresponding authors (Y.K., M.G. and H.I.) on reasonable request.

Scientists from RIKEN and the University of California San Diego, in collaboration with international partners have found a way to significantly reduce the amount of energy required by organic light emitting diodes (OLEDs). OLEDs have attracted attention as potential replacements for liquid crystal diodes, since they offer advantages such as being flexible, thin, and not requiring backlighting. The group achieved the advance, published in Nature, by developing a new way to manipulate the "excitons"—pairs of electrons and holes—that are key to the transport of electrons within OLEDs. Essentially, current passing through the device creates such pairs, which then change to a lower energy level and emit visible light in the process. Normally, the excitons in OLEDs arise in two patterns, with the spins being either the same or opposite, and the ones with same spins—known technically as triplet excitons—are three times more common. However, the singlets, which are created along with the triplets, require more energy, and though they can be converted into triplets it still means that the device as a whole requires the energy to create them in the first place. In the current work, the group found a way to lower the voltage so that only triplets are formed. The work began with fundamental research to understand the basic physics behind the creation of excitons using precise single-molecule electroluminescence measurements using a scanning tunneling microscope (STM) combined with an optical detection system. They prepared a model system based on an isolated molecule of 3, 4, 9, 10-perylenetetracarboxylicdianhydride (PTCDA), an organic semiconductor, adsorbed on a metal-supported ultrathin insulating film. They used a special technique to impart a negative charge to the molecule. Then, they used the current from an STM (scanning tunneling microscope) to induce luminescence in the molecule, and monitored what type of exciton was created based on the emission spectrum. The measurements showed that at low voltage, only triplets were formed. Theoretical calculations by Kuniyuki Miwa and Michael Galperin at UC San Diego confirmed the experimental results and substantiated the mechanism. "We believe," says Kensuke Kimura of the RIKEN Cluster for Pioneering Research, "that we were able to do this thanks to a previously unknown mechanism, where electrons are selectively removed from the charged molecule depending on their spin state." "It was very exciting to discover this new mechanism," says Yousoo Kim, leader of the Surface and Interface Science Laboratory in the RIKEN CPR, "We believe that these findings could become a general working principle for novel OLEDs with low operating voltage."

10.1038/s41586-019-1284-2

Biology

Stitching together the structure of the DNA replication toolbelt

Claudia Lancey et al, Structure of the processive human Pol δ holoenzyme, Nature Communications (2020). DOI: 10.1038/s41467-020-14898-6 Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-020-14898-6

https://phys.org/news/2020-05-dna-replication-toolbelt.html

Abstract In eukaryotes, DNA polymerase δ (Pol δ) bound to the proliferating cell nuclear antigen (PCNA) replicates the lagging strand and cooperates with flap endonuclease 1 (FEN1) to process the Okazaki fragments for their ligation. We present the high-resolution cryo-EM structure of the human processive Pol δ–DNA–PCNA complex in the absence and presence of FEN1. Pol δ is anchored to one of the three PCNA monomers through the C-terminal domain of the catalytic subunit. The catalytic core sits on top of PCNA in an open configuration while the regulatory subunits project laterally. This arrangement allows PCNA to thread and stabilize the DNA exiting the catalytic cleft and recruit FEN1 to one unoccupied monomer in a toolbelt fashion. Alternative holoenzyme conformations reveal important functional interactions that maintain PCNA orientation during synthesis. This work sheds light on the structural basis of Pol δ’s activity in replicating the human genome. Introduction Three DNA polymerases (Pols), α, δ, and ε replicate the genomic DNA in eukaryotes, with the latter two possessing the proofreading exonuclease activity required for high fidelity DNA synthesis 1 , 2 . Pol δ replicates the lagging strand and may share a role with Pol ε in replicating the leading strand 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 . In contrast to the continuous leading strand synthesis, the lagging strand is synthesized discontinuously in ~200 nucleotide (nt)-long Okazaki fragments, which are then ligated to form the contiguous lagging strand 2 . Synthesis of each Okazaki fragment starts with the low fidelity Pol α synthesizing a ~30 nt RNA/DNA initiator primer. Replication factor C (RFC) then loads the homotrimeric clamp PCNA at the primer/template (P/T) junction 11 . PCNA encircles the duplex DNA and tethers Pol δ to the DNA enhancing its processivity from few nucleotides to hundreds of nucleotides per DNA binding event 12 , 13 , 14 . PCNA has also been shown to increase the nucleotide incorporation rate of Pol δ 15 . Additionally, the interaction of PCNA with Pol δ is critical for coordinating its transient replacement by other PCNA partner proteins. In the maturation of Okazaki fragments, Pol δ invades the previously synthesized Okazaki fragments to gradually displace the RNA–DNA primers for their removal by the PCNA-bound FEN1 16 . In translesion DNA synthesis, Pol δ is transiently replaced by a PCNA-bound translesion DNA polymerase to ensure the continuation of DNA replication 17 . Mammalian Pol δ consists of a catalytic subunit and three regulatory subunits (Fig. 1a ). The catalytic subunit (p125) harbours the polymerase and exonuclease activities, and a metal-binding C-terminal domain (CTD). The regulatory subunits (p50, also referred to as the B-subunit, p66 and p12) are required for optimal activity of the holoenzyme 18 and there is evidence of different context-specific subassemblies of Pol δ in vivo 19 , 20 , 21 . In particular, DNA damage or replication stress triggers the degradation of the p12 subunit, resulting in the formation of a three-subunit enzyme with an increased capacity for proofreading 20 , 21 . While mammalian Pol δ was identified more than forty years ago 22 , the architecture of this essential enzyme and its interaction with PCNA are still poorly understood. The recently published cryo-EM structure of Saccharomyces cerevisiae ( Sc ) heterotrimeric Pol δ bound to (P/T) DNA elucidated the interactions between the catalytic subunit (Pol3, homologous to human p125) and the two regulatory subunits (Pol31 and Pol32, homologous to human p50 and p66, respectively), showing a unique molecular arrangement 23 ; the p12 subunit is absent in Sc Pol δ. However, how eukaryotic Pol δ achieves processive DNA synthesis and how it cooperates with PCNA and other factors during Okazaki fragment processing remains unknown. Fig. 1: Cryo-EM structure of the processive Pol δ–DNA–PCNA complex. a Domain organization of the four subunits of human Pol δ and amino acid sequence of PCNA-interacting (PIP-box) motifs. CTD C-terminal domain, OB oligonucleotide binding domain, PDE phosphodiesterase domain. b Gold-standard Fourier shell correlation for the Cryo-EM reconstruction of the Pol δ–DNA–PCNA complex, showing the resolution estimation using the 0.143 criterion. c Cryo-EM density map of the Pol δ–DNA–PCNA complex colored by domain. d Structure of the Pol δ–DNA–PCNA complex colored by domain and sequence of the DNA primer/template substrate. The region of the substrate that was modelled is boxed. Full size image To gain insight into these molecular mechanisms, we used cryo-EM single-particle reconstruction and determined the structure of the human Pol δ heterotetramer bound to (P/T) DNA and PCNA at 3.0 Å resolution (Fig. 1b ). We show that this complex exists in alternative conformations where key interactions regulating the holoenzyme activity are lost. In addition, we present the structure of the Pol δ–DNA–PCNA–FEN1 complex at 4.0 Å resolution. We discuss these structures in the context of Pol δ function in Okazaki fragment synthesis and maturation. Results and discussion Pol δ holoenzyme architecture and interaction with PCNA We firstly reconstituted a replication-competent holoenzyme comprising Pol δ, PCNA, a 25/38 (P/T) DNA bearing a 3′-dideoxy chain terminator in the primer strand and dTTP as an incoming nucleotide. A multi-subunit system was used to optimize the coexpression of human Pol δ’s heterotetrameric complex in baculovirus-infected Sf9 insect cells. We determined the cryo-EM structure of the Pol δ–PCNA–DNA–dTTP complex at 4.1 Å resolution (Supplementary Figs. 1 , 2 ). Reconstitution of this complex in the presence of FEN1 followed by gel filtration (Supplementary Fig. 3 ) led to an analogous map albeit at higher resolution (3.0 Å; Supplementary Figs. 4 , 5 ) in which FEN1 is invisible, as well as to a 4.0 Å resolution map in which FEN1 is visible and could be modelled (see below in the text). We therefore discuss the structure of the Pol δ–PCNA–DNA–dTTP complex based on the 3.0 Å reconstruction (Fig. 1c, d ) (Supplementary Data Table 1 ). The structure has approximate dimensions of 150 Å × 130 Å × 100 Å and displays the catalytic domain of p125 on top of the front face of PCNA in an open configuration, while the CTD of p125 and the p50, p66, and p12 regulatory subunits are positioned laterally (Fig. 1c, d ). This arrangement allows PCNA to thread and stabilize the duplex DNA exiting the catalytic cleft. The architecture of the catalytic and p50–p66 regulatory subcomplex of Pol δ is conserved between human and yeast 23 (RMSD Cα ~2 Å) (Supplementary Fig. 6 ). Importantly, our map allowed us to identify and model the p12 subunit of Pol δ (absent in yeast), showing that it bridges the exonuclease domain, the CTD and the oligonucleotide binding (OB) domain of the p50 subunit (Fig. 1c, d ). In addition, the critical region of the CTD interacting with PCNA, encompassing the C-terminus of the thumb domain and the CysA motif, was invisible in the cryo-EM map of the isolated yeast holoenzyme 23 , while it becomes visible in the human processive complex (Figs. 1c , 2a ). Pol δ is tethered to only one of the three PCNA monomers. Two main interaction sites are observed: (a) a short parallel β-sheet involving Cys991, Thr993, Leu995 of the CTD of p125 forming four main-chain hydrogen bonds with residues Asp120 and Asp122 on the IDCL of PCNA (Fig. 2 , Inset 1) and (b) a one-turn α-helix formed by p125 residues 1001–1005 which inserts into the canonical PIP-box hydrophobic cleft of PCNA through a two-fork plug made of conserved Leu1002 and Phe1005 (Fig. 2 , Inset 2). The CTD residues connecting the two interacting regions (residues 997–1000) are invisible, suggesting that they remain flexible. This is a mode of binding to PCNA not observed before. In particular, the newly identified p125 PIP-box ( 1001 GLLAFA, Fig. 1a ), which highly diverges from the conserved PIP-box (with strict consensus sequence Qxxhxxaa , where h is a hydrophobic, a is an aromatic, and x is any residue), adds to the notion that PCNA can bind a diverse range of sequences broadly grouped as PIP-like motifs 24 . Fig. 2: Cryo-EM density and model of selected regions of the Pol δ–DNA–PCNA complex. a Map regions showing the critical interactions tethering the polymerase to PCNA. Models are colored by domain. Inset 1: main-chain hydrogen bonds between the CTD of p125 and the IDCL of PCNA, indicated as red dotted lines. Residues involved in the interactions are labeled and colored by domain. Amino acid side chains are not shown. Inset 2: interaction between the p125 PIP-box and PCNA. b Sequence alignment of the CTD of human and Saccharomyches cerevisiae Pol δ and motifs. Asterisks correspond to conserved residues. Conserved cysteines in CysA and CysB motifs are highlighted in magenta. c Map region and model of the CTD of p125. Interfacial residues are boxed and colored by domain. Residues participating in polar interactions are connected by double-headed arrows. Some of the amino acid side chains are omitted for clarity. d Model region showing the pocket between the p125 and p50 subunits, where the CTD is inserted. p125 and p50 domains are shown as surfaces, and the CTD as a ribbon. The FeS cofactor and zinc ion in the CTD are shown as spheres. The p12 subunit was removed for clarity. e Map region and model of the p12 subunit of Pol δ, and p12 amino acid sequence and motifs. The segment of the p12 sequence that was modelled is boxed. Interfacial residues are boxed and colored by domain. Residues participating in polar interactions are connected by double-head arrows. f Model region showing the position of p12 relative to the holoenzyme. p12 and CTD are shown as ribbons and the latter is shown with enhanced transparency. p125 and p50 domains are shown as surfaces. g Model of human Pol δ in ribbon representation, showing the p12 subunit connecting the catalytic and regulatory modules. Full size image We first performed a primer extension assay to probe the role of the interaction site involving the IDCL of PCNA on the activity of Pol δ. We showed that a PCNA variant in which the key interacting residues were simultaneously mutated (D120A, L121A, D122A, and V123E) maintained the trimeric structure (Supplementary Fig. 7 ) but severely compromised the activity of Pol δ (Supplementary Fig. 8a ), demonstrating their indispensable role for binding to Pol δ. The binding of PCNA to only one Pol δ PIP-box in our structure is a striking observation since functional assays in Sc Pol δ showed that three PIP-boxes located in the Pol3, Pol31, and Pol32 subunits are required to achieve processive DNA synthesis 25 . However, the putative PIP-box in Pol31 ( 321 DKSLESYFNG) that is non-conserved in human, is located in a structured region, and would be positioned >40 Å away from the closest binding site on PCNA, suggesting that it may not take part in the interaction with PCNA. Besides the PIP-box on the p125 subunit, human Pol δ has two additional PIP-boxes in the p66 and p12 subunits (Fig. 1a ), both of which are able to interact with PCNA 26 , 27 , 28 .The p66 and p12 PIP-boxes are located at the extreme C-terminus and N-terminus of the subunit, respectively (Fig. 1a ), and are both in flexible regions invisible in the cryo-EM map. The last visible residue of p66 is Ala144 and that of p12 is Glu42, at approximate distances of 50 and 36 Å from the nearest PIP-box site, respectively (Supplementary Fig. 9 ). Therefore, both p66 and p12 PIP-boxes may possess a capture radius able to reach one or more of the PCNA binding sites. We performed activity assays to assess the role of each of the three PIP-boxes of human Pol δ (Supplementary Fig. 10 ). We found that mutation of the key residues in the PIP-box of p125 (L1002A and F1005A) severely reduced the activity of Pol δ, while mutating the key residues in the PIP-box of p66 (F462A and F463A) or p12 (I7A and Y11A) have minimal effect (Supplementary Fig. 10a ). To investigate the role of the three PIP-box motifs in processivity of Pol δ, we added heparin sulphate at the start of the reaction to trap dissociated DNA polymerase molecules. We observed a severe defect in the case of the PIP-box mutant of p125 followed by the PIP-box mutant of p12 and no detectable defect in the case of the PIP-box mutant of p66 (Supplementary Fig. 10b ). The mechanism by which the p12 PIP-box contributes to processivity remains unclear. It is highly likely that the polymerase occasionally falls off the DNA during synthesis and primarily employs the long-reach PIP-box in the p12 subunit to prevent its dissociation into solution, in agreement with previous findings 21 . However, we cannot rule out the possibility of a transient interaction during DNA synthesis where the p12 PIP-box swaps with the p125 PIP-box before the p125 PIP-box regains binding. The binding of Pol δ to only one PIP-box site during synthesis may explain the dynamic processivity of the replisome, where Pol δ from solution can exchange the replisome-associated Pol δ 29 . The conserved CysA motif in the CTD of p125, previously reported to be important for the processivity of Sc Pol δ holoenzyme 30 , connects to the PIP-box and folds into a zinc finger situated next to one of the outer β-sheets of the PCNA ring (Fig. 2 ). Consistent with these results, mutation of the PCNA residues closest to CysA (L66A and A67E; Fig. 2 , Inset 2) did not affect the trimeric structure of PCNA (Supplementary Fig. 7 ) but reduced the activity of Pol δ (Supplementary Fig. 8b ). This suggests that the CysA motif acts as an outrigger that stabilizes the holoenzyme structure and orients its interaction with PCNA. Role of the CysB motif and p12 subunit in human Pol δ The CysB motif in the CTD of the p125 subunit (Fig. 2b ), which contains an iron-sulfur (FeS) cluster 31 , is well resolved in the map (Fig. 2c ). We confirmed the presence of the FeS cluster in our Pol δ preparation as described previously 30 , 32 . A freshly purified Pol δ exhibited a yellow-brownish color and a broad absorption band that peaks at ~410 nm (Supplementary Fig. 11 ). The extinction coefficient at 410 nm yielded a value of 13,576 ± 1981 M −1 cm −1 demonstrating the contribution of 3.4 ± 0.5 [Fe]/[Pol δ] (Supplementary Fig. 11 ). The FeS cluster therefore is properly incorporated during expression and is maintained to a good extent during purification. The CysB motif contains the four conserved cysteines coordinating the FeS center and is connected to the zinc finger in the CysA motif by two antiparallel α-helices (helix α1 and helix α2), which are inserted between the catalytic domain and the p50 subunit (Fig. 2c, d ). Interestingly, the CTDs of Pol α and ε form equivalent complexes with their respective B-subunits but are larger and only coordinate divalent Zn 2+ ions 33 , 34 , 35 , 36 , 37 . The different size of the CTD and orientation relative to the B-subunit may correlate with the distinct functional dependencies of the three polymerases on PCNA (Supplementary Fig. 12 ); PCNA inhibits the primer extension activity of Pol α 38 , 39 and modestly increases the activity of Pol ε through relatively weak interactions 40 , 41 . Our results confirm the important structural and functional role of the FeS cluster previously shown in both yeast and human Pol δ 30 , 31 . Jozwiakowski and co-workers showed that, differently from the scenario in yeast 30 , disruption of the FeS cluster does not inhibit the assembly of the human holoenzyme but rather partially destabilizes the tetrameric structure, and this effect is fully reversed by the presence of PCNA 31 . In agreement with this, our structure shows that PCNA critically stabilizes the CTD, and also suggests that the p12 subunit may stabilize the FeS-deficient holoenzyme by providing additional inter-subunit contacts (see below in the text). Loss of FeS leads to strong defects in DNA synthesis and exonucleolytic activity in human Pol δ, coupled to an impaired ability to bind dsDNA 31 . This is consistent with our structure showing the FeS cluster inserting into a deep cavity between the thumb (which anchors the duplex portion of the DNA substrate) and the exonuclease domain (Fig. 2d ), and contacting both domains through a set of long-side chain residues forming polar and Van der Waals interactions (e.g., R1074, D1075, Y1080, K1084) (Fig. 2c ). The [4Fe4S] 2+ cluster in yeast Pol δ can be reversibly oxidized to [4Fe4S] 3+ resulting in a significant slowing of DNA synthesis 42 , and this may be used by the polymerase to sense oxidative stress and stall replication under mutagenic conditions. How this DNA-mediated redox signal is transferred to the FeS cofactor (situated >25 Å away from the DNA) and how it may decrease the synthetic activity of Pol δ remains unclear. Possibly, [4Fe4S] 2+ oxidation results in a long-range conformational change in the thumb domain which hinders the translocation of the DNA substrate through PCNA. The CTD connects the catalytic domain of p125 to the p50-p66 N subcomplex, which forms an elongated prawn-like structure that projects sideways relative to p125 with no direct interaction with either PCNA or DNA (Fig. 1c ). The CysB motif, helix α1 and helix α2, and a loop at the C-terminus of helix α2 interact with the OB domain and the inactive phosphodiesterase domain (PDE) of p50, burying ~2200 Å 2 at the interface (Fig. 2c, d ). The main interface is conserved in the human and yeast complex 23 , and the residues involved (e.g., Glu1048, Arg1060, and His1066 in CysB and Glu136, Asp137, and Glu138 in p50) agree with previous biochemical and genetics studies 43 , 44 . However, the interaction involving the hydrophobic C-terminal tail of helix α2 (from Phe1099 to Trp1107), which inserts into a groove below p50 PDE (Fig. 2d ), is unique to the human holoenzyme, and may further stabilize the catalytic module relative to the regulatory one. Two loops in p50 OB (residues 109–1024) and p50 PDE (residues 255–270), absent in the crystal structure of the p50–p66 N complex 45 , face the p125 subunit but their weak density prevented de novo model building (Supplementary Fig. 13 ). It has been shown that DNA damage or replication stress induces the degradation of the p12 subunit, a 107-residue polypeptide of previously unknown structure, leading to a three-subunit enzyme with enhanced proofreading activity 19 , 20 , 21 . We identified p12 in the cryo-EM map as the cuboidal-shaped density stitched across the p125 and p50 subunits (Fig. 1c ), and we could model the p12 C-terminal portion spanning residues 42–107 (Figs. 1d , 2e, f ). The p12 N-terminus, which contains the PIP-box, is instead invisible, suggesting that it is flexible and projects away from the holoenzyme. The p12 fold is rather compact, with approximate dimensions of 30 Å × 20 Å × 20 Å, and consists of three short α-helices and loops interacting with the p125–p50 subcomplex, burying a total of ~1200 Å 2 at the interface. The hydrophobic C-terminal loop of p12 (from Leu101 to Leu107) inserts like a hook into a pocket between the exonuclease domain and the CysB motif, establishing several contacts (Fig. 2e, f ). Additional polar interactions are observed between Gln51 on p12 helix α1 and Gly500 on the exonuclease domain, and between Thr62 on p12 helix α2 and Asp1068 in the CysB motif (Fig. 2e ). The interactions with the p50 OB are instead rather sparse (e.g., Gln68 and Arg72 on p12 helix α2 interacting with Tyr388 on p50). Collectively, these results suggest a role for p12 as a scaffold stabilizing the catalytic module relative to the regulatory module of human Pol δ, enhancing the synthetic activity of the holoenzyme 18 (Fig. 2g ). This stabilizing effect may regulate the small angular freedom between the two modules, which was observed in yeast Pol δ 23 and suggested to facilitate the transfer of a mismatched primer from the catalytic to the exonuclease active site, a process critical for high-fidelity synthesis 46 and predicted to involve a large rearrangement of the thumb and exonuclease domains 47 , 48 . Thus, degradation of p12 in response to DNA damage may slow down DNA synthesis, and the consequent increase of Pol δ flexibility may be required due to the higher rate of nucleotide misincorporation and frequent transfer of the primer between the pol and exonuclease sites. Interactions with DNA and implications for Pol δ activity The catalytic module of human Pol δ is composed of N-terminal, palm, thumb and exonuclease domains (Fig. 1d ) and is structurally homologous to the catalytic module of yeast Pol3 49 . The positions of the P/T DNA and the incoming nucleotide in the polymerase active site are also analogous to those observed in the crystal structure of Sc Pol3–DNA–dCTP ternary complex 49 , with Watson-Crick base pairing between the terminal adenine in the template strand and the incoming dTTP (Fig. 3 , Inset 1). The density ascribed to the 5′-overhang of the template strand is compatible with three bound nucleotides that are sharply bent by a 90° angle relative to the duplex DNA and face the unbound protomer of PCNA (Figs. 1d , 3 ). Therefore, the holoenzyme was captured in the act of synthesis. Twenty-three base pairs (bp) of the DNA substrate could be modelled and thread through the PCNA hole in a central location, almost perpendicular to the ring plane (α TILT ~ 4°) (Fig. 3 ). The duplex DNA is in B-form and is kept in position by the thumb domain up to the 10th bp from the active site, while the remaining bases are stabilized by the constraint of the PCNA closed topology and interactions with basic residues in the clamp channel (Fig. 3 ). In the absence of PCNA, the dsDNA portion below the thumb domain would be flexible and invisible in the map, as observed in the isolated Sc Pol δ holoenzyme bound to (P/T) DNA 23 . PCNA interacts with DNA via multiple polar residues approaching DNA phosphates within a coulombic interaction distance (< 6 Å) (Fig. 3 , Inset 2). Most of the interactions are established with one DNA strand, involving PCNA residues which follow the strand helix. A similar kind of electrostatic interactions were observed in the lac repressor binding to a non-specific DNA sequence 50 . The importance of the PCNA–DNA interactions for human Pol δ activity is supported by previous mutagenic studies, where a single mutation among PCNA basic residues K20, K77, K80, R149, and K217 at the sliding surface severely reduced Pol δ ability to incorporate an incoming nucleotide at the initiation of DNA synthesis 51 . Thus, such point mutations are expected to disrupt the observed pattern of interactions with DNA, resulting in a DNA tilt that is incompatible with DNA synthesis. Fig. 3: Interaction of the processive Pol δ holoenzyme with DNA. Map region around dsDNA and model, showing that DNA is held in place by the polymerase thumb domain, and stabilized by the PCNA central channel. The N-terminal, palm and fingers domains of Pol δ and the PCNA monomer in the foreground are removed for clarity. Inset 1: Cryo-EM map region and model of primer/template DNA in the Pol δ active site. The terminal adenine in the primer strand and paired incoming dTTP are labeled. Inset 2: PCNA interactions with DNA. PCNA subunits are shown as cartoons and colored in different shades of blue. DNA is shown as a yellow ribbon. DNA phosphates within a coulombic interaction distance (<6 Å) from PCNA residues are shown as spheres. Interacting phosphates on the template and primer strands are shown in red and yellow, respectively. Phosphates closer than 5 Å are shown with a larger sphere diameter. PCNA interacting residues are shown as sticks and labeled. The asterisk indicates that the side chain of R149 is flexible, but could approach a phosphate group at a distance <6 Å. Full size image The Pol δ–DNA–PCNA complex exists in multiple conformers The local resolution map of the Pol δ holoenzyme showed reduced resolution of the PCNA component, particularly for the two unbound protomers, suggesting conformational heterogeneity in the complex (Supplementary Figs. 1 , 4 ). Such heterogeneity was especially evident in the particle dataset from the sample fast-frozen immediately after mixing of the components (Supplementary Figs. 1 , 2 ), while it was significantly less prominent when the sample was separated by gel filtration prior to freezing (Supplementary Figs. 3 – 5 ), suggesting that the complex conformation may stabilize overtime. In agreement, 3D classification of the first dataset identified different holoenzyme conformers in which the orientation of the PCNA ring plane varies in a range up to ~20° (Fig. 4a and Supplementary Fig. 5 ). Importantly, the tilting of PCNA disrupts the interactions involving residues 121–123 of the IDCL of PCNA and the CTD of p125 (Fig. 2a ) that are critical for the polymerization activity (Supplementary Fig. 8b ), while the PIP-box interaction is maintained. The interaction via the p125 PIP-box is clearly seen in Class 2. However, the lower resolution in Class 3 made it difficult to conclusively attribute the bound PIP-box to p125, raising the possibility that other PIP-boxes may transiently swap with the p125 PIP-box in Conformer 3. The PCNA tilting also modulates the distance between a loop on the polymerase thumb domain and a loop on the PCNA front face (Fig. 4b ). Mutation of the highly conserved D41 residue on the PCNA loop (Supplementary Fig. 8c ) or K931A on the thumb loop (Supplementary Fig. 10a ) reduced the activity of Pol δ and severely compromised its processivity (Supplementary Fig. 10b ). This is consistent with previous findings showing the corresponding loop in Sc PCNA to be critical for enhancing the catalytic rate of the polymerase 15 . The contact between the two loops is visible in the lowest PCNA tilt conformer (Class 1, 4.3 Å resolution), which is the most populated (65%). However, the density at this contact is weak suggesting that the two loops interact in multiple orientations. Fig. 4: Alternative conformers of Pol δ holoenzyme. a Three representative EM 3D classes showing increasing tilting of the PCNA ring relative to the polymerase. The percentage values indicated next to the EM reconstructions correlate with the proportion of particles assigned to the three 3D classes and represent an estimate of the prevalence of each conformation in the data set. b Overlay of the models corresponding to the EM reconstructions shown in a . Inset: close up of the EM map of Class 1 showing the contact between the Pol δ thumb domain and PCNA loops. Residues K931 in the Pol δ loop and D41 in the PCNA loop are labelled. Density of K931 side chain is missing, indicating that K931 participates in a flexible interaction with PCNA. c PCNA interactions with DNA in Class 2. PCNA subunits are shown as cartoons and colored in different shades. DNA is shown as a yellow ribbon. DNA phosphates within a coulombic interaction distance (<6 Å) from PCNA residues are shown as spheres. Interacting phosphates on the template and primer strands are shown in yellow and red, respectively. Phosphates closer than 5 Å are shown with a larger sphere diameter. PCNA interacting residues are shown as sticks and labeled. d Side-view of the processive complex and Class 2 structures aligned using PCNA. The polymerase component is hidden for clarity. PCNA subunits are shown in different shades of blue and the subunit in the foreground is hidden. DNA molecules in the processive and Class 2 complexes are shown as purple and yellow ribbons, respectively. Phosphates interacting with PCNA residues are shown as spheres. Terminal bases of the DNA substrates are labelled. Full size image We propose that dynamic interactions among PCNA loops and both the thumb and the CTD of p125 may keep PCNA in a position that is competent for DNA polymerization. In the absence of such interactions (Class 2 and 3), the DNA tilts towards the PCNA inner rim and bends by ~10° to avoid clashing (Fig. 4b, c ). Interestingly, the DNA in Class 2 and 3 is translated vertically by one half of the major groove width relative to the processive holoenzyme (Fig. 4d ). As a consequence, the DNA strand mainly interacting with PCNA in the tilted conformers is the primer strand, while in the processive complex is the template strand. It is possible that this change in DNA positioning occurs frequently and disrupts the interactions of PCNA with the thumb and CTD. The binding via the PIP-box is integral for re-establishing the PCNA interactions with the thumb and CTD. Structure of the Pol δ–PCNA–DNA–FEN1 toolbelt The Pol δ holoenzyme architecture strongly suggests that other PCNA-interacting partners may bind to the PCNA monomers not occupied by the polymerase. Indeed, the strand displacement activity of Pol δ and 5′ flap cleavage by FEN1 in yeast act processively through iterative 1–3 nt steps of nick translation to ultimately remove the RNA/DNA primer 16 . While these data indirectly support the “toolbelt” mechanism in nick translation, a polymerase–DNA–PCNA–FEN1 complex has not been observed before in eukaryotes. To test this hypothesis, we reconstituted a complex comprising human Pol δ, P/T DNA and dTTP, PCNA, and FEN1, isolated the complex by gel filtration (Supplementary Fig. 3 ), and analyzed it by cryo-EM (Supplementary Figs. 14 , 15 ). We obtained a reconstruction at 4.0 Å resolution (Fig. 5a ) in which the mutual positions of Pol δ, DNA, and PCNA are analogous to those of the processive Pol δ holoenzyme while FEN1 fills the space between the N-terminal and palm domains of the p125 subunit, where the ssDNA template enters the active site, and the unoccupied PIP-box site on PCNA. Previous crystallographic data showed that FEN1 interacts with PCNA in multiple orientations through a single C-terminal PIP-box connected to the FEN1 core by a flexible hinge encompassing residues 333–336 52 . Flexible-tethering of FEN1 to PCNA in the holoenzyme is suggested by the cryo-EM local resolution map showing lower resolution in the FEN1 region (6–9 Å, Supplementary Fig. 14 ). Rigid-body fitting positions FEN1 in an upright configuration (Fig. 5b, c ) that is different from the three conformers observed in the FEN1–PCNA crystal structure, but that can be obtained by rotation and translation of any of these conformers via the flexible hinge. Density corresponding to the cap-helical gateway of FEN1, which forms a hole through which the 5′ flap threads and is guided into the active site, is missing, suggesting that the cap-helical gateway is flexible and points outward from the polymerase. The reconstruction fitting a defined FEN1 conformer comprises ~20% of the particle dataset, while the reconstruction from the entire dataset shows no significant density at the FEN1 binding site (Supplementary Fig. 4 ), suggesting that FEN1 may be absent from the complex in a major subset of particles. Fig. 5: Cryo-EM structure of the Pol δ–DNA–PCNA–FEN1 complex. a Cryo-EM map of the complex colored by domains. b Complex model colored by domains. c Close-up showing the map region around FEN1 filtered to 6 Å resolution. d Proposed toolbelt model for Pol δ and FEN1 bound to PCNA processing an Okazaki fragment. Full size image The Pol δ–DNA–PCNA–FEN1 complex structure reveals various aspects of the mechanism of maturation of Okazaki fragments. FEN1 sits across the template strand and in a proper orientation to bind the downstream duplex DNA when Pol δ encounters the previously synthesized Okazaki fragment (Fig. 5b ). Interestingly, the 90° bent angle of the template strand (Fig. 3 , Inset 1) is similar to that required for FEN1 activity 53 . This suggests that Pol δ may handoff an already bent nick junction to FEN1 and help in positioning the 5′ flap to thread into the cap-helical gateway (Fig. 5d ). However, handing off a bent DNA is not a must for FEN1 activity since FEN1 can actively bend the nick junction in diffusion-limited kinetics 54 , 55 . Nick translation occurs at rates that are 10-fold faster than nick DNA product release by FEN1 16 , 55 , 56 , 57 . This suggests that Pol δ and FEN1 must be actively handing off their products during nick translation. Our structure suggests that the proximity of FEN1 to the template strand and the potential interaction between FEN1 and Pol δ may facilitate their products handoff during nick translation. The preferential binding of FEN1 to the more exposed PIP-box site on PCNA (Fig. 5a ) indicates that DNA Ligase 1 (Lig1) would be sterically excluded during nick translation reaction and that its binding requires Pol δ to move out of the way. The alternate conformers of PCNA may also facilitate the exposure of the third PIP-box site. This is consistent with the finding that Lig1 acts in a distributive manner during maturation of Okazaki fragments 16 . It is unclear how Pol δ–PCNA–FEN1 would signal the binding of Lig1. Interestingly, FEN1 releases the nick product in two steps, where it binds it briefly in a bent conformer that is followed by a lengthy binding in an extended conformer 55 . It is possible that the bent conformer is more compatible with Pol δ binding while the extended one sequesters the nick until Pol δ moves out of the way and Lig1 is recruited from solution. In yeast, acute depletion of Lig1 allows nick translation to proceed up to the dyad of the pre-assembled nucleosome, and the Okazaki fragment termini are enriched around known transcription factor binding sites 58 . Thus, nucleosomes and DNA binding factors may trigger the stalling of Pol δ and its dissociation from PCNA and/or DNA, allowing Lig1 to bind and seal the nick of a mature Okazaki fragment. Methods Plasmid construction MultiBac™ expression system (Geneva Biotech) was used to express human Pol δ in Sf9 insect cells. For this purpose, the genes encoding p125 (accession no. NP02682) and p50 (accession no. NP006221) were amplified and cloned into pACEBac1 Gent + separately. The p50-encoding cassette along with promoter and terminator was excised with I-Ceu1 and BstX1 and ligated into p125 containing pACEBac1 linearized with I-Ceu1 . After cloning wild-type p125 and p50, L1002A and F1005A (p125 PIP box) or K931A (p125) mutations were introduced by polymerase chain reaction (PCR) separately. The gene sequence encoding p66 (accession no. NP006582) was amplified and cloned into pIDC Cm + while that of p12 (accession no. NP066996) was amplified along with an N-terminal 8xHis-tag and cloned into pIDS Spect + . After cloning wild-type p66 and p12, mutations of F462A and F463A (p66 PIP box) and I7A and Y11A (p12 PIP box) were carried out by PCR separately. Finally, the single transfer vectors with different subunit assemblies were generated using Cre recombinase according to the MultiBac™ expression system user manual and the resulting constructs were named hereafter as Pol δ, Pol δ-p125 PIP box mutant, Pol δ-p66 PIP box mutant, Pol δ-p12 PIP box mutant, and Pol δ-K931A mutant. At the end, the recombinant transfer vector encoding all four Pol δ subunits was transformed into DH10MultiBac™ cells to transpose the gene expression cassettes into MultiBac baculoviral DNA and the resulting bacmid DNA was then isolated. Full-length human PCNA (accession no. NM182649) was cloned in pETDuet-1 MCS1 (Novagen) Amp + to obtain 6× His N-terminally-tagged protein. After cloning wild-type PCNA (WT-PCNA), L66A and A67E (hereafter named LA-PCNA) and D120A, L121A, D122A, and V123E (hereafter named DLDV-PCNA) and D41A (hereafter named D41A-PCNA) mutations were carried out by PCR. Human RPA clone (pET11d-tRPA) was a generous gift of Professor Marc S. Wold (University of Iowa College of Medicine, Iowa City, Iowa). A 6× His-tag was introduced at the C-terminus of RPA1 subunit by PCR. The catalytically inactive FEN1-D181A was constructed as described previously 54 . Briefly, the gene sequence encoding WT human FEN1 was cloned via Gibson assembly into a pE-SUMO-pro expression vector (Lifesensors). An additional 6× His-tag was introduced by PCR at the N-terminus of the fusion. D181A mutation was introduced by QuikChange II Site-Directed Mutagenesis Kit (Agilent). For the human full-length RFC3 and RFC5 subunits, the expression vector pCDFK-RFC5/3 was a generous gift of Dr. Yuji Masuda 59 . pET E. coli expression vector with BioBrick polypromoter restriction sites (14-A) was a gift from Scott Gradia (Addgene plasmid # 48307) 60 . pCDFK-RFC5/3 was modified by inserting a 6× His-tag at the N-terminus of RFC3. The sequences encoding RFC2, RFC4, and a truncated version of RFC1, with the first 550 N-terminal amino acids deleted, were codon optimized and synthesized by IDT as gBlocks. These three sequences were amplified by PCR with primers containing LICv2 sequences and cloned independently by Ligation Independent Cloning (LIC) into 14-A vectors. The three resulting expression cassettes were assembled together into one 14-A vector by BioBrick sub-cloning as instructed in the MacroBac manual 60 . This resulting co-expression plasmid is denoted from here onwards as p14A-ΔNRFC1/4/2. Protein expression and purification For Pol δ, Pol δ-p125 PIP box mutant, Pol δ-p66 PIP box mutant, Pol δ-p12 PIP box mutant or Pol δ-K931A mutant expression, Sf9 cells were cultured in ESF 921 medium (Expression Systems). To prepare the baculovirus, bacmid DNA containing all four subunits was transfected into Sf9 cells using FuGENE® HD (Promega) according to manufacturer’s instructions. This baculovirus prep was then amplified twice to obtain a higher titer virus (P3 virus). The expression of Pol δ then proceeded by transfecting a 4 L Sf9 suspension culture at a density of 2 × 10 6 cells/mL with P3 virus. Cells were harvested at 72 h post-infection by centrifugation at 5500× g for 10 min and then re-suspended in 3 mL per 1 g of wet cells in lysis buffer [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 5 mM β-Mercaptoethanol (BME), 0.2% NP-40, 1 mM PMSF, 5% (v/v) Glycerol and EDTA-free protease inhibitor cocktail tablet/50 mL (Roche, UK)]. Cells were sonicated and debris was removed by centrifugation at 95,834× g for 1 h at 4 °C. The cleared lysate was then adjusted to 40 mM imidazole final concentration and directly loaded onto a HisTrap HP 5 mL affinity column (GE Healthcare) pre-equilibrated with buffer A [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 5 mM BME and 5% Glycerol] containing 40 mM imidazole. After loading, the column was extensively washed first with 50 mL of buffer A containing 40 mM imidazole followed by washing with buffer A containing 100 mM imidazole to remove the nonspecifically-bound proteins. Finally, the column was washed with 50 mL of buffer B [50 mM Tris-HCl (pH 7.5), 50 mM NaCl, 40 mM Imidazole, 5 mM BME and 5% Glycerol] to reduce the salt concentration in preparation for next chromatographic step. The bound proteins were eluted with 50 mL linear gradient against buffer C [50 mM Tris-HCl (pH 7.5), 50 mM NaCl, 600 mM Imidazole, 5 mM BME and 5% Glycerol]. Fractions that contain all Pol δ subunits were combined and loaded onto an anion exchanger, Mono Q 5/50 GL (GE Healthcare) column pre-equilibrated with buffer D [50 mM Tris-HCl (pH 7.5), 50 mM NaCl, 5 mM BME and 5% Glycerol]. The column was then washed with 20 mL of buffer D containing 150 mM NaCl. Pol δ was eluted with a 20 mL of gradient from 150 mM NaCl to 1 M NaCl in 50 mM Tris-HCl (pH 7.5), 5 mM BME and 5% Glycerol. The fractions that contained all Pol δ subunits were combined and concentrated to 1 mL and loaded onto HiLoad 16/600 Superdex 200 pg (GE Healthcare) pre-equilibrated with gel filtration buffer [50 mM Tris-HCl (7.5), 200 mM NaCl, 1 mM DTT and 5% Glycerol]. Protein fractions were pooled, flash frozen and stored at −80 °C. To express PCNA and its mutants, either WT-PCNA, LA-PCNA, DLDV-PCNA, or D41A-PCNA plasmid was transformed into BL21 (DE3) E. coli cells and grown at 37 °C in 2YT media supplemented with ampicillin to an OD 600 of 1.2. The cultures were then induced with 0.5 mM Isopropyl β- d -1-thiogalactopyranoside (IPTG) and continued to grow for 19 h at 16 °C. Cells were then harvested by centrifugation at 5500× g for 10 min. The resulting pellets were re-suspended in 3 mL/1 g of wet cells in lysis buffer [50 mM Tris-HCl pH (7.5), 750 mM NaCl, 20 mM imidazole, 5 mM BME, 0.2% Nonidet P-40, 1 mM PMSF, 5% Glycerol and EDTA-free protease inhibitor cocktail tablet/50 mL]. The cells were lysed with 2 mg/mL of lysozyme at 4 °C for 30 min followed by sonication. Cell debris was removed by centrifugation (22,040× g , 20 min, 4 °C) and the cleared supernatant was directly loaded onto HisTrap HP 5 ml affinity column pre-equilibrated with buffer A [50 mM Tris-HCl (pH 7.5), 500 mM NaCl, 20 mM Imidazole, 5 mM BME, and 5% Glycerol]. The column was washed with 50 ml of buffer A followed by washing with 50 mL of buffer B [50 mM Tris-HCl (pH 7.5), 100 mM NaCl, 20 mM Imidazole, 5 mM BME and 5% Glycerol] to reduce the salt concentration. The bound PCNA, or its mutants, were eluted with 50 ml of gradient with buffer C [50 mM Tris-HCl (pH 7.5), 100 mM NaCl, 500 mM Imidazole, 5 mM BME and 5% Glycerol]. The eluents were pooled and directly loaded onto HiTrap Q HP 5 mL anion exchange column (GE Healthcare) pre-equilibrated with buffer D [50 mM Tris-HCl (pH 7.5), 100 mM NaCl, 5 mM BME and 5% Glycerol]. After loading, the column was washed with 50 mL of buffer D followed by elution with 50 mL gradient with buffer E [50 mM Tris-HCl (pH 7.5), 1 M NaCl, 5 mM BME and 5% Glycerol]. Eluents were pooled and concentrated to 1.5 ml and loaded onto HiLoad 16/600 Superdex 200 pg pre-equilibrated with gel filtration buffer [50 mM Tris-HCl (pH 7.5), 150 mM NaCl, 1 mM DTT and 5% Glycerol]. Protein fractions were pooled, flash frozen, and stored at −80 °C. Recombinant PCNA used for the cryo-EM study was produced as described 61 . Human RPA was expressed and purified as described previously 62 . Briefly, the plasmid was transformed into BL21 (DE3) E. coli cells and cultured in 2YT media at 37 °C. The expression was induced with 0.5 mM IPTG when the OD 600 of the culture reached a value of 0.7. Buffer A is defined for all the rest of the RPA purification scheme as containing [50 mM Tris-HCl pH 7.5, 10% (v/v) glycerol and 10 mM β-mercaptoethanol (BME)]. The cells were collected and resuspended in lysis buffer [buffer A + 35 mM Imidazole and 750 mM NaCl]. RPA was first purified over a HisTrap HP 5 ml column pre-equilibrated with lysis buffer. The bound proteins were eluted with a linear gradient of buffer A + 250 mM Imidazole and 150 mM NaCl. From this step onward the purification followed the one described previously 63 with few modifications as following. The collected fractions containing all the three RPA subunits were then loaded onto a HiTrap Blue 5 mL column (GE healthcare) pre-equilibrated with buffer A + 150 mM NaCl and washed extensively with buffer A + 1 M NaCl. The bound protein was then eluted with a linear gradient of buffer A + 1.5 M NaSCN. Finally, RPA was purified over a HiLoad Superdex-200 pg size exclusion column with buffer containing [50 mM Tris-HCl pH 7.5, 10% glycerol, 1 mM DTT, 500 mM NaCl and 0.1 mM EDTA]. The fractions containing pure stoichiometric RPA were concentrated, flash frozen in liquid nitrogen, and stored at −80 °C. For the purification of RFC complex with an N-terminally truncated RFC1 (hereafter named as ΔNRFC), the two plasmids, p14A-ΔNRFC1/4/2 and pCDFK-RFC5/3, were co-transformed into BL21(DE3) E. coli cells and colonies were selected on agar plates containing two antibiotics (Kan + Amp). ΔNRFC was overproduced by growing the transformed cells in 10 L TB media containing the two antibiotics. Cells were grown at 20 °C to an OD 600 of 0.8 and then expression was induced by the addition of 0.1 mM IPTG and continued at 15 °C for an additional 24 h. Cells were collected by centrifugation and re-suspended in buffer A [50 mM HEPES-NaOH (pH 7.5), 10 mM BME, 5% Glycerol and 1 mM PMSF] containing 500 mM NaCl and one EDTA-free protease inhibitor cocktail tablet/50 mL. Cells were lysed by lysozyme treatment and sonication and cleared by centrifugation at 95,834× g for 45 min. The collected supernatant was adjusted to 10 mM imidazole and loaded onto a HisTrap HP 5 mL column. The salt concentration was lowered on-column to 250 mM NaCl and the protein complex was eluted with a linear gradient between Buffer A and Buffer A + 300 mM imidazole containing 250 mM NaCl. Fractions containing all the subunits of ΔNRFC were combined and loaded onto a HiTrap Heparin HP 1 mL column and eluted with linear gradient between Buffer A + 250 mM NaCl and Buffer A + 1 M NaCl. Fractions that contained all the ΔNRFC subunits were collected, concentrated, and loaded onto HiLoad Superdex 16/600 200 pg gel filtration column equilibrated with storage buffer [50 mM HEPES-NaOH (pH 7.5), 300 mM NaCl, 1 mM DTT, 0.5 mM EDTA, and 10% glycerol]. Fractions containing ΔNRFC with the correct stoichiometry were collected, concentrated, flash frozen, and stored at −80 °C. FEN1-D181A was expressed and purified as described previously 63 . Briefly, the plasmid was transformed into E. coli BL21 (DE3) cells and cultured in 2YT media. Expression was induced by the addition of 0.2 mM IPTG when the OD 600 of the culture reached a value of 0.8. Cells were harvested and then lysed in buffer A [50 mM Tris–HCl pH 7.5, 5% (v/v) glycerol, 750 mM NaCl, 10 mM β-mercaptoethanol (BME), and 30 mM imidazole] by a combination of lysosome treatment and sonication. Purification was carried out by two sequential Ni-NTA columns, separated by SUMO protease cleavage. Proteins were then concentrated and further purified over HiLoad Superdex-75 pg size exclusion columns using a buffer containing [50 mM HEPES-KOH pH 7.5, 500 mM NaCl, 2 mM dithiothreitol (DTT), and 10% (v/v) glycerol]. Fractions containing pure D181A FEN1 were collected, flash frozen, and stored at −80 °C. All protein concentrations were calculated by measuring their absorbance at 280 nm using the extinction coefficients calculated from the amino acid composition (Pol δ: 195,390 M −1 cm −1 , PCNA: 47,790 M −1 cm −1 , RPA: 87,210 M −1 cm −1 , FEN1: 22,920 M −1 cm −1 and ΔNRFC: 157,000 M −1 cm −1 ). To check the trimeric structure of the PCNA mutants, a HiLoad 16/600 Superdex 200 pg gel filtration column was equilibrated with PCNA storage buffer. Two milligram of WT-PCNA, LA-PCNA and DLDV-PCNA were independently diluted in 250 μL of PCNA Storage Buffer. In parallel, a mixture containing 4 mg of carbonic anhydrase (27 kDa) and Conalbumin (75 kDa) was prepared in 250 μL of PCNA storage buffer. Each sample was loaded onto the gel filtration column and ran for 1.5 column volumes at a flow-rate of 1 mL/min. The elution chromatograms were fit, using the cftool of MATLAB, to a single Gaussian in case of PCNA and to a sum of two Gaussians in case of the marker proteins. In both cases a linear baseline was added to the fit to correct for imperfect initialization of the UV 280 reading. All three PCNA forms eluted between 71 and 72 mL (Supplementary Fig. 7a–c ). The two molecular markers eluted at ~77 mL and ~89 ml, respectively (Supplementary Fig. 7d ), at a later retention volume compared to PCNA. Since the PCNA monomer has a molecular weight of ~29 kDa, these results show that in all three forms, PCNA is larger than the largest of the two molecular markers of 75 kDa, therefore it must exist in its homotrimeric form. DNA substrates M13mp18 single-stranded DNA (NEB N4040S) and Cy5 labelled HPLC purified primer (5′-/Cy5/TAA GGC CAG GGT TTT CCC AGT CAC G-3′) were annealed by mixing both templates in 1:1 molar ratio in T-50 buffer [50 mM Tris-HCl (pH 7.8), and 50 mM NaCl] and heating at 75 °C for 15 min that was followed by slow cooling to room temperature. For the substrate used in EM, a template strand (5′-CTGCACGAATTAAGCAATTCGTAATCATGGTCATAGCT-3′) was annealed to a primer containing a 3′ dideoxycytosine chain terminator (5′-AGCTATGACCATGATTACGAATTG[ddC]−3′) to form the P/T substrate. The strands were mixed in an equimolar ratio in the presence of 20 mM Tris pH 7.5 and 25 mM NaCl. They were then annealed by incubating at 92 °C for 2 min followed by slow cooling to room temperature overnight. All oligonucleotides were purchased from Sigma Aldrich. Primer extension assay Primer extension assays were carried out in reaction buffer consisting of [40 mM Tris-HCl (pH 7.8), 10 mM MgCl 2 , 1 mM DTT, 0.2 mg/mL bovine serum albumin, 50 mM NaCl, 1 mM ATP and 500 μM each of dGTP, dATP, dTTP, and dCTP]. Five nanomolar of primed-M13 substrate was initially mixed with 150 nM PCNA (WT or mutants), 40 nM ΔNRFC, and 500 nM RPA on ice and then pre-incubated at 30 °C for 2 min prior to the addition of different concentrations of Pol δ and further incubation for 10 min in case of DLDV-PCNA and 2.5 min for D41A-PCNA and LA-PCNA; the reaction volume was 30 µL. Primer extension assays for Pol δ and its mutants in the presence of WT-PCNA were carried out at 30 °C for 2.5 min or for 5 min with or without 575 ng/mL heparin. Reactions with or without heparin were initiated with 500 μM dTTP. The reactions were terminated with 6 µL of stop buffer [180 mM NaOH, 6 mM EDTA and 7.5% Ficoll] followed by heating at 75 °C for 5 min and cooling on ice for 2 min. All products were resolved by running on 1% alkaline agarose gel for 20 h at 10 V and visualized using Typhoon Trio (GE Healthcare). UV–Vis analysis of Fe-S in purified human Pol δ The absorption spectra of freshly purified Pol δ at four different concentrations were acquired at room temperature using a microplate spectrofluorometer (TECAN infinite M1000) in Pol δ storage buffer. Absorption spectra were collected from 260 to 600 nm in Corning 96-well UV-Plates. As per manufacturer’s instructions, 370 μL of each sample were used in order to generate an optical path length through the sample of 1 cm. Absorption spectra were also collected for background absorption using storage buffer as a blank. Criteria of 2-nm wavelength step size along with 100 flashes per step were used. The background absorption was then subtracted from the sample absorption spectra. For each of the four concentrations, four repeats of spectra were acquired. The final plotted spectra are the averages of the four individual repeats. To quantify the presence of Fe–S cluster, the extinction coefficient at 410 nm was monitored 30 , 32 , 37 , 64 , 65 . It was previously indicated that the protein absorption peak, centered at ~280 nm, and the Fe–S absorption peak, centered at ~410 nm, are separated enough to be considered independent for quantification analysis 64 , 65 . Taking into account the Beer-Lambert law at each of these wavelengths, the following system of equations holds: $$\left\{ {\begin{array}{*{20}{c}} {A_{280} = \varepsilon _{280} \ast c \ast d} \\ {A_{410} = \varepsilon _{410} \ast c \ast d} \end{array}} \right.,$$ (1) where A 280 and A 410 are the measured absorbances of the sample at 280 nm and 410 nm, ε 280 and ε 410 are the molar extinction coefficients of the sample at 280 and 410 nm, c is the concentration of the sample and d is the optical path length inside the sample. Given that for each sample the concentration and the optical path length are wavelength-independent, the above equations can be divided to remove the dependence on c and d , giving \(\frac{{A_{410}}}{{A_{280}}} = \frac{{\varepsilon _{410}}}{{\varepsilon _{280}}}\) . Multiplying both sides of this equation by ε 280 we obtain the final equation of interest: $$\varepsilon _{410} = \varepsilon _{280} \ast \frac{{A_{410}}}{{A_{280}}},$$ (2) ε 280 of Pol δ was calculated by Prot-Param 66 to be 198,140 M −1 cm −1 , based on its amino acids sequence and considering that all cysteines are in reduced form. Sixteen ε 410 values were determined from the absorption spectra of the four repeats for each of the four Pol δ concentrations using the formula described above. The mean ε 410 extinction coefficient together with its standard deviation (SD) were calculated from the sixteen values. It was previously established that each iron atom contributes with ~4000 M −1 cm −1 to this extinction coefficient when coordinated inside an Fe–S cluster 64 , 65 . Therefore, the average and SD of the number of iron atoms per Pol δ are obtained by dividing the mean and SD of ε 410 by 4000 M −1 cm −1 . Cryo-EM grid preparation and data collection For all complexes, UltrAuFoil® R1.2/1.3 Au 300 grids were glow discharged for 5 min at 40 mA on a Quorum Gloqube glow-discharge unit, then covered with a layer of graphene oxide (Sigma) prior to application of sample. For preparation of the Pol δ complex without FEN1, the P/TddC substrate was mixed, in order, with Pol δ, PCNA and dTTP. Final concentrations were 1.0 μM Pol δ, 0.8 μM PCNA trimer, 2.1 μM DNA and 20 μM dTTP. The buffer this was performed in comprised 25 mM HEPES-KOH (pH 7.5), 100 mM potassium acetate, 10 mM calcium chloride, 0.02% NP-40, 0.4 mM biotin and 1 mM DTT. 3 μl of this sample were applied to the grid, blotted for 3 s at blot force 10 and plunge frozen into liquid ethane using a Vitrobot Mark IV (FEI Thermo Fisher), which was set at 4 °C and 100% humidity. For the Pol δ–DNA–PCNA–FEN1 complex, a 40 μl inject containing 3.75 μM P/TddC, 1.9 μM Pol δ, 1.5 μM PCNA trimer, 20 μM dTTP, and 3 μM FEN1, was loaded onto a Superdex 200 increase 3.2/300 column (GE Life Sciences) equilibrated with the same buffer as above (Supplementary Fig. 3 ). 3 μl of a fraction corresponding to the first peak was applied to a grid, prepared in the same way as above for data collection. Cryo-EM data for all samples were collected on a Thermo Fisher Scientific Titan Krios G3 transmission electron microscope at the LISCB at the University of Leicester. Electron micrographs for the complex with multiple PCNA conformers were recorded using a Falcon III direct electron detector (FEI Thermo Fisher) at a dose rate of 0.7 e-/pix/sec for 60 s and a calibrated pixel size of 1.08 Å. These data were collected using a Volta phase plate with EPU 1.9, and focusing performed at every hole using a nominal value of −0.6 μm. For the toolbelt structure, electron micrographs were recorded using a K3 direct electron detector (Gatan Inc.) at a dose rate of 11 e-/pix/sec and a calibrated pixel size of 0.87 Å. Focusing was performed over a range between −2.3 and −1.1 μm, in 0.3 μm intervals. Cryo-EM image processing Pre-processing of the processive Pol δ complex (first dataset) was performed as follows: movie stacks were corrected for beam-induced motion and then integrated using MotionCor2 67 . All frames were retained and a patch alignment of 4 × 4 was used. Contrast transfer function (CTF) parameters for each micrograph were estimated by Gctf 68 . Integrated movies were inspected with Relion-3.0 69 for further image processing (3575 movies). Particle picking was performed in an automated mode using the Laplacian-of-Gaussian (LoG) filter implemented in Relion-3.0. All further image processing was performed in Relion-3.0. Particle extraction was carried out from micrographs using a box size of 360 pixels (pixel size: 1.08 Å/pixel). An initial dataset of 2.8 × 10 6 particles was cleaned by 2D classification followed by 3D classification with alignment. 3D refinement and several rounds of polishing and per-particle CTF refinement yielded a 4.09 Å structure of the Pol δ–PCNA complex. Masked refinement of the Pol δ and PCNA components yielded reconstructions at 3.88 and 6.70 Å, respectively. Classification of the Pol δ holoenzyme showing PCNA in different orientations was performed as follows: particles previously aligned on the Pol δ component using masked refinement were subsequently 3D-classified without alignment. Five 3D classes were generated with populations corresponding to 3, 4, 6, 22, and 65%. These classes were aligned on the Pol δ component and sorted for the tilting of the PCNA ring plane relative to the polymerase. 3D refinement was performed on the 3D classes corresponding to populations 65, 22, and 6%, and yielded reconstructions at 4.3, 4.9, and 8.1 Å resolution. Pre-processing of the Pol δ–DNA–PCNA–FEN1 complex data (second dataset) was performed as follows: movie stacks were imported in super resolution mode, then corrected for beam-induced motion and integrated using Relion’s own implementation, using a binning factor of 2. All frames were retained and a patch alignment of 5 × 5 was used. Contrast transfer function (CTF) parameters for each micrograph were estimated by CTFFIND-4.1 70 . Integrated movies were inspected with Relion-3.1 for further image processing (5071 movies). Particle picking was performed in an automated mode using the Laplacian-of-Gaussian (LoG) filter implemented in Relion-3.1. Further image processing was performed in Relion-3.1, taking two separate paths to generate the reconstructions with and without visible FEN1. The former was processed as follows: Particle extraction was carried out from micrographs using a box size of 400 pixels (pixel size: 0.87 Å/pixel). An initial dataset of 2.5 × 10 6 particles was cleaned by 2D classification followed by 3D classification with alignment. 3D refinement and several rounds of polishing and per-particle CTF refinement yielded a 4.05 Å structure of the Pol δ–PCNA-FEN1 toolbelt complex, with clear density corresponding to FEN1. The data without visible FEN1 was processed as follows: An initial dataset of 2.5 × 10 6 particles was cleaned by 2D classification followed by 3D classification with alignment. 3D refinement yielded a map of the holoenzyme at 4.6 Å resolution. Further 3D classification was then performed without alignment, followed by 3D refinement and several rounds of polishing and CTF refinement, yielding a 3.0 Å structure of the holoenzyme. Molecular modelling The homology model for the catalytic domain of human Pol δ was built with Phyre2 71 based on the crystal structure of the catalytic domain of Saccharomyces cerevisiae ( Sc ) Pol δ (PDB entry 3IAY) 49 , rigid-body docked and jiggle-fitted using Coot 72 . For enhanced fitting precision, the chain of the human Pol δ catalytic domain was partitioned into 12 segments (residues 2–83, 84–161, 162–288, 289–352, 352–371, 372–408, 409–496, 497–575, 576–638, 639–756, 757–873, and 874–908) corresponding to the homologous segments in Sc Pol δ undergoing TLS vibrational motion as determined from the Sc Pol δ crystal structure 49 with TLS Motion Determination (TLSMD) 62 , and the segments were fitted in the cryo-EM map as individual rigid bodies. The homology model for the zinc finger in the CTD of Pol δ was built with HHpred 73 , based on the crystal structure of Spt4/5NGN heterodimer complex from Methanococcus jannaschii (PDB entry 3LPE) 74 , and edited with Coot. Main-chain tracing was aided by ARP/wARP 75 . The homology model for CysB motif of human Pol δ was built with Phyre2 71 based on the cryo-EM structure of Saccharomyces cerevisiae ( Sc ) Pol δ (PDB entry 6P1H) 23 . Modelling of the DNA and dTTP in the catalytic site of Pol δ was guided by the crystal structure of the Sc Pol δ–DNA–dCTP complex 49 . DNA was built with Coot and 3D-DART 76 . The model of the p12 subunit of Pol δ was built using ARP/wARP 75 , and edited with Coot. Rigid-body docking of the human p50–p66N complex (PDB entry 3E0J) 45 and PCNA (PDB entry 1U76) 27 was performed in UCSF Chimera 77 . The model of the processive Pol δ complex was subjected to real-space refinement in Phenix 78 . FEN1 core (chain Y in PDB entry 1UL1 52 ) was rigid body fitted in the EM map of the Pol δ–DNA–PCNA–FEN1 complex using Chimera and FEN1 hinge (residues 333–336) was edited with Coot to have the C-terminal PIP-box anchored to its binding site on PCNA. The Pol δ component in Conformers 1, 2, and 3 was built from that of the processive complex, while PCNA was rigid-body fitted in the corresponding maps, and the region of the CTD connecting the PIP-box to the CysA motif was edited with Coot. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The maps of the Pol δ–DNA–PCNA complexes and Pol δ–DNA–PCNA–FEN1 complex have been deposited in the EMBD with accession codes EMD- 10539 , EMD- 10080 , EMD- 10081 , EMD- 10082 , and EMBD-10540 and the atomic models in the Protein Data Bank under accession codes PDB 6TNY , 6S1M, 6S1N, 6S1O, and 6TNZ.

The molecular machinery responsible for replicating human DNA has come into sharper focus and deepened understanding of this fundamental biological process and how it can malfunction. An international team led by researchers at KAUST has used cryo-electron micrography to study the structure of the human Pol δ-DNA-PCNA complex. During replication, the two strands in the DNA helix are unwound, and each serves as the template for a new copy. Pol δ is one of the enzymes responsible for copying one of the strands. It is tethered to the DNA by a molecule known as PCNA, forming a Pol δ-DNA-PCNA complex at the heart of replication. The team found that the catalytic core of Pol δ—the part that carries out the synthesis reaction—sits on top of PCNA and that its regulatory subunits project to the side. The DNA is threaded through a central channel in PCNA, enabling the molecule to stabilize DNA exiting the reaction area. Further analysis revealed locations in PCNA that are required for the interaction, and results showed that mutating those sites severely reduced the activity of Pol δ. Muhammad Tehseen, one of the lead authors of the study, explains that "Professor Hamdan's lab has been systematically reconstituting and studying the activities of these proteins by single-molecule imaging, recording molecular movies of these reactions." Tehseen purified the molecular complex so that the team could create higher resolution static images. "Combining the dynamic information from single-molecule imaging with static images from cryoEM is like bringing an image to life," he explains. The structure also clarified the role of some of the Pol δ subunits, shedding light on how it interacts with PCNA and DNA. For example, Pol δ interacts with only one monomer of PCNA, and sits in a manner that exposes the two other monomers, freeing them to interact with other proteins that act with Pol δ. One such protein is FEN1, which processes the short fragment strands of DNA, after they are synthesized by Pol δ, so that they can be stitched together to form a continuous mature strand." The researchers also studied the structure of the Pol δ-DNA-PCNA-FEN1 complex. The interactions they discovered support a 'toolbelt' model in which PCNA acts as a platform to which different processing enzymes bind, similar to a toolbelt with an array of tools. Working out these structures and interactions enables researchers to better understand the dynamics of DNA replication and how it goes wrong, such as mutations in Pol δ linked with tumors.

10.1038/s41467-020-14898-6

Physics

A new beat in quantum matter

M. Di Liberto et al. Non-Abelian Bloch oscillations in higher-order topological insulators, Nature Communications (2020). DOI: 10.1038/s41467-020-19518-x Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-020-19518-x

https://phys.org/news/2020-11-quantum.html

Abstract Bloch oscillations (BOs) are a fundamental phenomenon by which a wave packet undergoes a periodic motion in a lattice when subjected to a force. Observed in a wide range of synthetic systems, BOs are intrinsically related to geometric and topological properties of the underlying band structure. This has established BOs as a prominent tool for the detection of Berry-phase effects, including those described by non-Abelian gauge fields. In this work, we unveil a unique topological effect that manifests in the BOs of higher-order topological insulators through the interplay of non-Abelian Berry curvature and quantized Wilson loops. It is characterized by an oscillating Hall drift synchronized with a topologically-protected inter-band beating and a multiplied Bloch period. We elucidate that the origin of this synchronization mechanism relies on the periodic quantum dynamics of Wannier centers. Our work paves the way to the experimental detection of non-Abelian topological properties through the measurement of Berry phases and center-of-mass displacements. Introduction The quest for topological quantization laws has been a central theme in the exploration of topological quantum matter 1 , 2 , which originated from the discovery of the quantum Hall effect 3 . In the last decade, the development of topological materials has led to the observation of fascinating quantized effects, including the half-integer quantum Hall effect 4 and the quantization of Faraday and Kerr rotations 5 in topological insulators 6 , 7 , as well as half-integer thermal Hall conductance in spin liquids 8 and quantum Hall states 9 . In parallel, the engineering of synthetic topological systems has allowed for the realization of quantized pumps 10 , 11 , and revealed quantized Hall drifts 12 , 13 , 14 , circular dichroism 15 , and linking numbers 16 . In this context, Bloch oscillations (BOs) 17 , 18 , 19 , 20 , 21 , 22 have emerged as a powerful tool for the detection of geometric and topological properties in synthetic lattice systems 23 , 24 , 25 , 26 , hence providing access to quantized observables. Indeed, transporting a wavepacket across the Brillouin zone (BZ) can be used to explore various geometric features of Bloch bands, including the local Berry curvature 23 and the Wilson loop of non-Abelian connections 25 . This strategy has been exploited to extract the Berry phase 27 , 28 , the Berry curvature 29 , 30 , the Chern number 12 , 13 , 14 , and quantized Wilson loops 31 in ultracold matter and photonics. The Wilson loop measurement of ref. 31 highlighted a fundamental relation between two intriguing properties of multi-band systems: the quantization of Wilson loops, a topological property related to the Wilczek-Zee connection 32 , and the existence of “multiple Bloch oscillations”, which are characterized by a multiplied Bloch period 31 , 33 , 34 , 35 , 36 , 37 , 38 , 39 . The effect investigated in ref. 31 was eventually identified as an instance of “topological Bloch oscillations”, whose general framework was proposed in ref. 40 based on the space groups of crystals and its implications on the quantization of geometric quantities (Zak phases differences). The Bloch period multiplier appears as a topological invariant, protected by crystalline symmetries, thus making multiple BOs genuinely topological. Furthermore, when a Wannier representation of the bands is possible, Zak phases correspond to the positions of charges within the unit cell, namely the Wannier centers. As a consequence, a Zak-Wannier duality allows to connect BOs to the relative phases acquired by charges within a classical point-charge picture. More recently, BOs displaying topologically protected sub-oscillations have also been found in periodically driven systems in the context of quantum walks 41 . In this work, we identify a distinct topological effect that manifests in the BOs of higher-order topological insulators (HOTIs). These newly discovered systems belong to the family of topological crystalline insulators 42 , 43 , 44 , 45 , 46 , i.e., gapped quantum systems characterized by crystal symmetries; they are characterized by quantized multipole moments in the bulk and unusual topologically protected states (e.g., corner or hinge modes) on their boundaries; see refs. 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , 55 , 56 , 57 , 58 , 59 , 60 . Considering the prototypical Benalcazar-Bernevig-Hughes (BBH) model 47 , 48 , we unveil a phenomenon by which multiple BOs take the form of an oscillating Hall drift, accompanied with a synchronized inter-band beating, for special directions of the applied force, as summarized in Fig. 1 a. Although the Hall motion is attributed to the finite non-Abelian Berry curvature of the degenerate band structure, the inter-band beating captured by the Wilson loop is shown to be topologically protected by winding numbers. The synchronization of real-space motion and inter-band dynamics is elucidated through a quantum Rabi oscillation of Wannier centers. Finally, we observe that detached helical edge states are present on specific boundaries, compatible with the special symmetry axes associated with the topological BOs. A topological transition signaled by the sign change of the identified winding numbers and by the corresponding appearence/disapperance of these states is identified. Fig. 1: Schematics of the non-Abelian topological BOs. a The center-of-mass position \(\left\langle {\bf{r}}\right\rangle\) of a gaussian wavepacket experiences a sign-changing Hall drift under the applied force F , while displaying a synchronized beating within two occupied bands. This synchronized effect, represented by the metronome, is topologically protected by the winding number w . b BBH model: a square lattice with π flux and staggered hopping amplitudes J 1 and J 2 . Vertical bonds with arrows correspond to a Peierls phase π in the hopping amplitudes. c Band structure of the model. Each band is twofold degenerate. d Brillouin zones and paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) exhibiting topological BOs. Full size image Overall, our results demonstrate the rich interplay of non-Abelian gauge structures and winding numbers in the topological BOs of HOTI’s, but also establish BOs as a powerful probe for non-Abelian topological properties in quantum matter. Results Model and symmetries We consider the BBH model, as introduced in refs. 47 , 48 . It consists of a square lattice with alternating hopping amplitudes J 1 and J 2 in the two spatial directions and a π flux per plaquette, as depicted in Fig. 1 b. We have introduced the flux by Peierls phases on the vertical links but other conventions can be used without affecting the results of this work. The model is represented by a chiral-symmetric Hamiltonian of the form $$\hat{H}({\bf{k}})=\sum_{i = 1}^{4}{d}_{i}({\bf{k}}){\Gamma }^{i} ,$$ (1) where the 4 × 4 Dirac matrices are written in the chiral basis Γ i = − σ 2 ⊗ σ i for i = 1, …, 3 and \({\Gamma }^{4}={\sigma }_{1}\otimes {\mathcal{I}}\) . This model has twofold degenerate energy bands E ( k ) = ± ϵ ( k ) with \(\epsilon ({\bf{k}})=\sqrt{{\left|{\bf{d}}({\bf{k}})\right|}^{2}}\) . The eigenfunctions of the lowest two bands read \(\left|{u}_{{\bf{k}}}^{1}\right\rangle =\frac{1}{\sqrt{2}\epsilon }{\left({d}_{1}-i{d}_{2},-{d}_{3}-i{d}_{4},0,i\epsilon \right)}^{T}\) and \(\left|{u}_{{\bf{k}}}^{2}\right\rangle =\frac{1}{\sqrt{2}\epsilon }{\left({d}_{3}-i{d}_{4},{d}_{1}+\, i{d}_{2},i\epsilon ,0\right)}^{T}\) . The full expressions of the d i ( k )’s for the isotropic BBH model is \({d}_{1}({\bf{k}})=({J}_{1}-{J}_{2})\sin ({k}_{y}/2)\) , \({d}_{2}({\bf{k}})=-({J}_{1}+{J}_{2})\cos ({k}_{y}/2)\) , \({d}_{3}({\bf{k}})=({J}_{1}-{J}_{2})\sin ({k}_{x}/2)\) and \({d}_{4}({\bf{k}})=-({J}_{1}+{J}_{2})\cos ({k}_{x}/2)\) and the corresponding energy dispersion is displayed in Fig. 1 c. Here, we take the periodicity d = 2 a = 1, where a is the lattice spacing. The ordering of the unit cell sites chosen to represent the model in Eq. ( 1 ) is indicated in Fig. 1 b. Notice that the chosen basis takes into account the geometric shape of the unit cell. As a consequence, the Hamiltonian is not Bloch invariant, namely H ( k + G ) ≠ H ( k ), with G a reciprocal lattice vector. The presence of time-reversal symmetry \(\hat{T}\) , with \({\hat{T}}^{2}=1\) and chiral symmetry \(\hat{S}\) , represented by \({\Gamma }^{0}={\sigma }_{3}\ \otimes \ {\mathcal{I}}\) , sets the model into the BDI class 1 . Moreover, several crystalline symmetries are also present: two non-commuting mirror symmetries with respect to the x and y axis, namely \({\hat{M}}_{x}={\sigma }_{1}\otimes {\sigma }_{3}\) and \({\hat{M}}_{y}={\sigma }_{1}\otimes {\sigma }_{1}\) , respectively; and a π /2 rotation symmetry \({C}_{4}=\left(\begin{array}{ll}0&{\mathcal{I}}\\ -i{\sigma }_{2}&0\end{array}\right)\) . The two mirror symmetries guarantee that inversion ( C 2 ) is also a symmetry of the model, with \({\hat{C}}_{2}={\hat{M}}_{x}{\hat{M}}_{y}\) . Moreover, the presence of mirror and rotation symmetries allows us to define a pair of mirror symmetries with respect to the diagonal axes of the lattice, \({\hat{M}}_{xy}={\hat{M}}_{y}{\hat{C}}_{4}\) and \({\hat{M}}_{x\bar{y}}=-{\hat{M}}_{x}{\hat{C}}_{4}\) . This is one of the central ingredients allowing for topologically protected BOs, as we will show below. Owing to the two non-commuting mirror symmetries \({\hat{M}}_{x}\) and \({\hat{M}}_{y}\) , the BBH model is a quadrupole insulator that has a quantized quadrupole moment in the bulk, vanishing bulk polarization and corner charges 47 . The non-commutation of the mirror symmetries also provides a non-vanishing non-Abelian Berry curvature \({\Omega }_{xy}({\bf{k}})={\partial }_{{k}_{x}}{A}_{y}-{\partial }_{{k}_{y}}{A}_{x}-i[{A}_{x},{A}_{y}]\) of the twofold degenerate lowest (or highest) bands 61 , 62 , where A is the non-Abelian Berry connection 32 . However, the total Chern number of the degenerate bands remains zero due to time-reversal. It is then possible to define Wannier functions \(|{\nu }_{x,{k}_{y}}^{\alpha }\rangle\) and \(|{\nu }_{y,{k}_{x}}^{\alpha }\rangle\) , with α = 1, 2 numbering the bands below the energy gap, which are eigenstates of the position operators \(\hat{P}\hat{x}\hat{P}\) and \(\hat{P}\hat{y}\hat{P}\) projected onto the lowest two bands, respectively 47 , 48 . The non-commutation of the mirror symmetries (and therefore of the projected position operators) forces the use of hybrid Wannier functions, namely Wannier states that can only be maximally localized in one direction 63 , 64 . Furthermore, it provides a necessary condition to have gapped Wannier bands, namely Wannier centers that are displaced from each other at every momentum k x or k y . In ref. 47 , 48 , the Wannier gap has been exploited to define a winding of the Wannier states (nested Wilson loop) as a condition to have a quantized quadrupole moment in the bulk, which can be revealed from the Wannier-Stark spectrum 65 . Winding numbers We now prove that a nontrivial topological structure captured by novel winding numbers characterizes the BBH model along the diagonal paths of the BZ, \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) , which are shown in Fig. 1 d. In order to emphasize the generality of these results, we hereby consider a generic Dirac-like model [Eq. ( 1 )] without specifying the components of the d ( k ) vector. We assume that all previously discussed symmetries are satisfied with the additional constraint that each d i function only depends on one component of the momentum k , namely we assume that d 1 = d 1 ( k y ), d 2 = d 2 ( k y ), d 3 = d 3 ( k x ), d 4 = d 4 ( k x ). Such constraint is satisfied by the BBH model. Mirror symmetries impose that d 1 and d 3 are odd functions whereas d 2 and d 4 are even. We then find that along \({\mathcal{C}}\) (i.e., for k = k x = k y ), the diagonal mirror symmetry represented by the operator \({\hat{M}}_{xy}\) requires d 3 ( k ) = d 1 ( k ) and d 4 ( k ) = d 2 ( k ), while along \(\bar{{\mathcal{C}}}\) (i.e., for k = k x = − k y ), the symmetry operator \({\hat{M}}_{x\bar{y}}\) requires d 3 ( k ) = − d 1 ( k ) and d 4 ( k ) = d 2 ( k ). We then conclude that only two components of d are independent and we therefore define the vector \(\tilde{{\bf{d}}}(k)\equiv ({d}_{1}(k),{d}_{2}(k))\) . After writing the Hamiltonian in its chiral representation \(\hat{H}({\bf{k}})=\left(\!\!\begin{array}{ll}0&Q({\bf{k}})\\ Q{({\bf{k}})}^{\dagger}&0\end{array}\!\!\right)\) , where \(Q({\bf{k}})={d}_{4}({\bf{k}}){\mathcal{I}}+i{d}_{i}({\bf{k}}){\sigma}^{i}\) , we obtain the following result $${w}_{{\mathcal{C}}(\bar{{\mathcal{C}}})} \equiv \frac{i}{2\pi }{\int}_{{\mathcal{C}}}{\rm{d}}{\bf{k}}\ \cdot {\rm{Tr}}\left[Q{({\bf{k}})}^{-1}{\sigma }_{3(1)}\ {\partial }_{{\bf{k}}}Q({\bf{k}})\right]\\ =-\frac{1}{\pi }\mathop{\int}\nolimits_{0}^{2\pi }{\rm{d}}k\ {\varepsilon }^{ij}\frac{{\tilde{d}}_{i}{\partial }_{k}{\tilde{d}}_{j}}{| \tilde{{\bf{d}}}{| }^{2}}={\rm{sign}}({J}_{1}^{2}-{J}_{2}^{2}),$$ (2) where ε 12 = − ε 21 = 1, and where we used the \(\tilde{{\bf{d}}}\) vector of the BBH model in the last step. We therefore conclude that the quantities \({w}_{{\mathcal{C}}}\) and \({w}_{\bar{{\mathcal{C}}}}\) count how many times the vector \(\tilde{{\bf{d}}}\) winds over the closed paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) , respectively. The quantized windings \({w}_{{\mathcal{C}}}\) and \({w}_{\bar{{\mathcal{C}}}}\) are here protected by the crystalline symmetries \({\hat{M}}_{x}\) , \({\hat{M}}_{y}\) , and \({\hat{C}}_{4}\) , as shown in Supplementary Note 1 . These symmetries also imply that the two topological invariants are not independent. We point out that similar winding numbers have been introduced in chiral-symmetric one-dimensional topological superconductors 66 . Finally, the sign change of the winding numbers at the gap closing point J 1 = J 2 signals a phase transition. We will show below that the transition corresponds to the appearance of detached helical edge states. Let us now focus on the BOs of the BBH model and the role played by the quantized winding numbers discussed above. Topological BOs: band-population dynamics We consider a wavepacket obtained as a superposition of the lowest two bands and centered at k , which we write as \(\left|{u}_{{\bf{k}}}(t)\right\rangle ={\eta }_{1}(t)\left|{u}_{{\bf{k}}}^{1}\right\rangle +{\eta }_{2}(t)\left|{u}_{{\bf{k}}}^{2}\right\rangle\) with \(\eta ={({\eta }_{1},{\eta }_{2})}^{T}\) . Owing to the degeneracy of the states \(\left|{u}_{{\bf{k}}}^{1,2}\right\rangle\) , other parametrizations can be chosen. As shown in Supplementary Note 2 , this gauge ambiguity can be removed by weakly breaking time-reversal symmetry, thus splitting the two states in energy. Under an applied homogeneous and constant force F , which makes the crystal momentum change linearly in time, \(\dot{{\bf{k}}}={\bf{F}}\) , the bands occupation evolves according to 67 $$\dot{\eta }=-i{\epsilon }_{{\bf{k}}}\eta +i{\bf{F}}\cdot {\bf{A}}\eta,$$ (3) where the matrix elements of the Berry connection are defined as \({A}_{i}^{\alpha \beta }=i\langle {u}_{{\bf{k}}}^{\alpha }| {\partial }_{{k}_{i}}| {u}_{{\bf{k}}}^{\beta }\rangle\) . Here, the force is assumed to be weak enough so that transitions to upper bands are neglected. We can formally solve Eq. ( 3 ) as \(\eta (t)=\exp (-i\mathop{\int}\nolimits_{0}^{t}{\rm{d}}t\ {\epsilon }_{{\bf{k}}})W\ \eta (0)\) . The Wilson line operator W is defined as \(W={\mathcal{T}}\exp (i\mathop{\int}\nolimits_{0}^{t}{\rm{d}}t\ {\bf{F}}\cdot {\bf{A}})={\mathcal{P}}\exp (i\mathop{\int}\nolimits_{{{\bf{k}}}_{i}}^{{{\bf{k}}}_{f}}{\bf{A}}\cdot {\rm{d}}{\bf{k}})\) , where we have denoted as k i and k f the initial and final momenta of the BO, respectively. For a closed path \({{\mathcal{C}}}_{0}\) with k f = k i + G , where G is a reciprocal lattice vector, the bands population dynamics is determined by the Wilson loop matrix \({W}_{{{\mathcal{C}}}_{0}}={\mathcal{P}}\exp (i{\int}_{{{\mathcal{C}}}_{0}}{\bf{A}}\cdot {\rm{d}}{\bf{k}})\) . Importantly, the winding numbers \({w}_{{\mathcal{C}}(\bar{{\mathcal{C}}})}\) that we have previously introduced appear in the Wilson loops defined along the diagonal paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) , as \({W}_{{\mathcal{C}}(\bar{{\mathcal{C}}})}=\exp (i(2\pi /4){w}_{{\mathcal{C}}(\bar{{\mathcal{C}}})}{\sigma }_{1(3)})\) , with \({w}_{{\mathcal{C}}(\bar{{\mathcal{C}}})}=\pm\! 1\) . From this, we obtain that BOs require four loops in momentum space in order to map the wavefunction back to itself, namely \({[{W}_{{\mathcal{C}}(\bar{{\mathcal{C}}})}]}^{4}={\mathcal{I}}\) . Notice that the degeneracy of the bands brings a trivial dynamical phase that does not influence the internal band-population dynamics. According to the classification of topological BOs discussed in ref. 40 , rotational symmetries \({\hat{C}}_{n}\) can quantize BOs with a force applied orthogonal to the rotational symmetry axis. This is (partially) the case here, with \({\hat{C}}_{4}\) providing period-four BOs. However, \({\hat{C}}_{4}\) symmetry alone is not sufficient to quantize the BOs. Additional symmetries, namely \({\hat{M}}_{x}\) and \({\hat{M}}_{y}\) , are required in order to have a protected winding number along the paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) , see Supplementary Note 1 . In the general framework presented in ref. 40 , a “Wannier-Zak” relation is demonstrated when mirror symmetries commute. As a consequence, the Zak phase winding that appears in the Wilson loop has a one-to-one correspondence with the position of the Wannier centers. This is well described by independently evolving point charges within a classical picture. In our case, such a direct correspondence is not possible owing to the non-vanishing Berry curvature and we will see below that the physical consequences of this feature appear on the real-space motion of the wavepacket. Topological BOs: real-space dynamics Let us now consider the real-space motion of the wavepacket’s center-of mass, which satisfies the following semiclassical equations 67 $$\dot{x}= \, {\partial }_{{k}_{x}}{\epsilon }_{{\bf{k}}}-{F}_{y}{\eta }^{\dagger }{\Omega }_{xy}\eta ,\\ \dot{y}= \, {\partial }_{{k}_{y}}{\epsilon }_{{\bf{k}}}+{F}_{x}{\eta }^{\dagger }{\Omega }_{xy}\eta .$$ (4) Here, Ω x y denotes the SU(2) Berry curvature, whose components are shown in Fig. 2 a for the BBH model; they satisfy the following conditions: \({\Omega }_{xy}^{11}=-{\Omega }_{xy}^{22}\) , \({\rm{Re}}\ {\Omega }_{xy}^{12}={\rm{Re}}\ {\Omega }_{xy}^{21}\) and \({\rm{Im}}\ {\Omega }_{xy}^{12}=-{\rm{Im}}\ {\Omega }_{xy}^{21}\) . One anticipates from the accumulation of Berry curvature near the M point of the BZ that the paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) may display nontrivial features also in the real-space dynamics and not only in the band population beating discussed above. As we shall explain in detail below, the wavepacket experiences a transverse Hall drift that changes sign after each BO, thus bringing the center-of mass position back to its initial point after two BOs. This behavior is synchronized and tightly connected with the band-population dynamics captured by the Wilson loop. Fig. 2: Time dynamics of multiple BOs. a Non-Abelian Berry curvature Ω x y ( k ) for the BBH model. b Band occupation dynamics along the \({\mathcal{C}}\) path for θ (0) = 0 and ϕ (0) = 0. Here J 2 = 0.3 J 1 and \(| F| =0.2\sqrt{2}{J}_{1}\) . c Real-space wavepacket trajectory. d Comparison of the wavepacket \(\left\langle x\right\rangle\) position for (solid line) numerical real-space evolution of a wavepacket with width σ = 0.15 a −1 and momentum grid-spacing k pts = 20, (dashed line) exact evolution, (dots) semiclassical evolution. e Orthogonal displacement Δ r ⊥ after one BO along the paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) . Full size image The twofold degeneracy of the bands allows us to parametrize the evolving state \(\left|{u}_{{\bf{k}}}(t)\right\rangle\) on the Bloch sphere as \({\eta }_{1}(t)=\cos \theta (t)\) and \({\eta }_{2}(t)=\sin \theta (t){e}^{i\phi (t)}\) . We can therefore rewrite the anomalous velocity as $${\eta }^{\dagger }{\Omega }_{xy}\eta = \, (| {\eta }_{1}{| }^{2}-| {\eta }_{2}{| }^{2}){\Omega }_{xy}^{11}+\sin 2\theta \cos \phi \ {\rm{Re}}\ {\Omega }_{xy}^{12}\\ \, -\sin 2\theta \sin \phi \ {\rm{Im}}\ {\Omega }_{xy}^{12} .$$ (5) On the \({\mathcal{C}}\) path, the angle ϕ is a constant of motion, namely \(\dot{\phi }=0\) . This means that the Bloch vector is confined to a meridian of the Bloch sphere. Moreover, since \({\rm{Re}}\ {\Omega }_{xy}^{12}=0\) on \({\mathcal{C}}\) , only the first and the last term of Eq. ( 5 ) are relevant. After one BO, the two bands populations exchange, symmetrically with respect to the M point, as displayed in Fig. 2 b [i.e., \({W}_{{\mathcal{C}}}\propto {\sigma }_{1}\) ]. Let us consider the case with ϕ = 0 and θ (0) = 0, where only the first term in Eq. ( 5 ) matters. The Berry curvature has a node and the band population starts with η 1 (0) = 1. Near the M point and before crossing it, the occupation of band 1 is larger than the occupation of band 2. The Berry curvature \({\Omega }_{xy}^{11}\) is positive and the Hall displacement in the x direction is therefore negative (see the minus sign in the first of Eq. ( 4 )). Once the path has crossed the M point, the occupations are flipped but so is also the sign of the Berry curvature, thus the Hall displacement continues with the same sign until the wavepacket reaches the Γ point. As the bands occupations have exchanged, a second BO will experience an opposite Hall drift and bring back the wavepacket to its initial position, as shown in Fig. 2 c, d. From this, we obtain that the real-space motion is a witness of the non-Abelian band dynamics. For θ (0) ≠ 0, the off-diagonal component of the Berry curvature also contributes, and it fully suppresses the Hall displacement when θ (0) = π /4 since the anomalous Hall velocity vanishes identically: the two bands are equally populated, no band exchange takes place and therefore the positive and negative deflections compensate each other. On the \(\bar{{\mathcal{C}}}\) path, the Berry curvature has only off-diagonal components and the Hall dynamics is determined by the relative phase ϕ , whereas θ is a constant of motion. In this case, the populations of the two bands do not exchange over time but the relative phase does by an angle π . The transverse displacement as a function of θ (0) and ϕ (0) for the two paths is shown in Fig. 2 e. We have thus found that the center-of mass of the wavepacket displays a period-two BO instead of a period-four one. This difference with respect to the Wilson loop analysis occurs because after two BOs, the wavefunction has picked up an overall phase 2 × ( π /2), which does not appear in observables \(\left\langle \hat{O}\right\rangle\) , such as for the center-of mass position. Atomic limit: single-plaquette dynamics In order to elucidate the role of Wannier functions in the BOs analyzed here and the synchronization between real-space and band-population dynamics, we consider the instructive atomic limit with J 2 = 0, where we can study the time dynamics of a single plaquette. The lowest energy eigenstates read \(\left|{u}^{1}\right\rangle ={(1/2,1/2,0,1/\sqrt{2})}^{T}\) and \(\left|{u}^{2}\right\rangle ={(1/2,-1/2,1/\sqrt{2},0)}^{T}\) . We construct the position operator \(\hat{{\bf{r}}}={\sum }_{i}({{\bf{r}}}_{i}-{{\bf{r}}}_{0})\left|{{\bf{r}}}_{i}\right\rangle \left\langle {{\bf{r}}}_{i}\right|\) by setting the spatial origin at the plaquette center. We obtain the matrices \(\hat{x}/a={\rm{diag}}(1/2,-1/2,-1/2,1/2)\) and \(\hat{y}/a={\rm{diag}}(1/2,-1/2,1/2,-1/2)\) . Let us now call \(\hat{P}\) the projector operator on the states \(\left|{u}^{1}\right\rangle\) and \(\left|{u}^{2}\right\rangle\) , from which we can construct the projected position operators \(\hat{P}\hat{x}\hat{P}\equiv {\hat{x}}_{P}={\sum }_{\alpha ,\beta = 1,2}\left|{u}^{\alpha }\right\rangle \left\langle {u}^{\alpha }| \hat{x}| {u}^{\beta }\right\rangle \left\langle {u}^{\beta }\right|\) and \(\hat{P}\hat{y}\hat{P}\equiv {\hat{y}}_{P}={\sum }_{\alpha ,\beta = 1,2}\left|{u}^{\alpha }\right\rangle \left\langle {u}^{\alpha }| \hat{y}| {u}^{\beta }\right\rangle \left\langle {u}^{\beta }\right|\) . It follows that \([{\hat{x}}_{P},{\hat{y}}_{P}]\,\ne\, 0\) , whereas \(\{{\hat{x}}_{P},{\hat{y}}_{P}\}=0\) . The eigenfunctions of the projected position operators are the Wannier functions \(|{\nu }_{x,y}\rangle\) and the corresponding eigenvalues are the Wannier centers 48 , 61 , 68 , which read here ν x = ν y = ± a /4. In the presence of an external tilt (or electric field), the perturbative Hamiltonian governing the dynamics for small values of the force F reads $${\hat{H}}_{F}= \, {\bf{F}}\cdot {\hat{{\bf{r}}}}_{P}={F}_{x}\ {\hat{x}}_{P}+{F}_{y}\ {\hat{y}}_{P}\\ = \, \frac{a}{4}({F}_{x}+{F}_{y}){\sigma }_{1}+\frac{a}{4}({F}_{x}-{F}_{y}){\sigma }_{3}.$$ (6) The projected Hamiltonian reveals how the external force induces a quantum dynamics between the eigenstates of non-commuting position operators, in the form of a Rabi oscillation. This result is in sharp contrast with the classical point-charge picture introduced in ref. 40 , which is valid for commuting position operators. We can now diagonalize \({\hat{H}}_{F}\) and we find the spectrum \(E=\pm\! Fa/2\sqrt{2}\) , with \(F=\sqrt{{F}_{x}^{2}+{F}_{y}^{2}}\) . The Rabi period can be easily obtained as \({T}_{R}=2\pi /(Fa/\sqrt{2})\) . This solution is general and it does not depend on the direction of the force. Besides, we can always rotate the coordinate system in order to have one axis parallel to the force and one axis orthogonal to it, r ∥ and r ⊥ , and reduce the Hamiltonian to \({\hat{H}}_{F}={F}^{\parallel }\ {\hat{r}}_{P}^{\parallel }\) . Then, the corresponding time dynamics can be represented by the eigenstates of \({\hat{r}}_{P}^{\perp }\) , namely the Wannier functions obtained by diagonalizing \({\hat{r}}_{P}^{\perp }\) . As a consequence, we observe a transverse dynamics compared with the direction of the applied force F , as shown in Fig. 3 a, b. Fig. 3: Single-plaquette dynamics. a Comparison between (solid line) the exact dynamics and (dashed line) the projected model. b Full real-space evolution. c Hybrid Wannier centers obtained from diagonalizing the Wilson loops orthogonal to the \({\mathcal{C}}\) path away from the atomic limit for J 1 = 0.3 J 2 . Here the force is aligned along the diagonal F x = F y . Full size image However, the Wannier centers dynamics is not directly connected to the BOs and its period does not have to be the same as the Rabi period of the Wannier centers. For example, let us consider a BO with F y = 0. The periodicity of the BO occurs at the discrete times T B = 2 π n / d F x , for \(n\in {{\mathbb{Z}}}_{+}\) where d = 2 a . There is no solution that satisfies T B = T R . However, if we take F x =± F y we find that n = 2 provides T B = T R . Therefore, a force oriented along the diagonal axes allows to synchronize the Wannier centers dynamics with the BOs, whereas the other directions yield out-of-sync oscillations that does not bring the wavepacket back to its initial position at integer multiples of the fundamental Bloch period. Away from the atomic limit, we can still use Wannier functions as a complete basis to express the wavepacket. A direct calculation (see Fig. 3 c) shows that along the paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) , the Wannier centers remain gapped and their spectrum flat, namely they are equispaced along the entire path. We interpret this fact as a witness that the Wannier centers can be thought as oscillators with the same oscillation frequency (i.e., displacement), as in the atomic limit represented by Eq. ( 6 ), thus keeping the same oscillatory motion while changing the momentum \({k}_{{\mathcal{C}}}\) or \({k}_{\bar{{\mathcal{C}}}}\) . In conclusion, Wannier centers perform a quantum periodic dynamics where their transverse motion with respect to the applied force is periodic and synchronized with the BO period. Edge states The quantized winding numbers \({w}_{{\mathcal{C}}}\) and \({w}_{\bar{{\mathcal{C}}}}\) , which we have previously identified along the paths \({\mathcal{C}}\) and \(\bar{{\mathcal{C}}}\) , indicate that a topological transition takes place when J 1 = J 2 . Here, we show that an open system with edges along the diagonals x ± y displays detached helical edge states. From Fig. 4 a, we notice that near the atomic limit J 2 → 0, the bulk has gapped states at energies \({E}_{b} \sim \pm\! \sqrt{2}{J}_{1}\) . The edge displays disconnected single sites at energy E s ~ 0 and trimers, with energies E t 1 ~ 0 and \({E}_{t2}=\pm\! \sqrt{2}{J}_{1}\) . Thus, a pair of zero energy ( E s and E t 1 ) modes exists at the edge. Fig. 4: Open lattice and edge states. a Open lattice in the atomic limit respecting mirror M x , y and C 4 symmetries. Highlighted in blue the edge sites displaying zero modes and in red the unit cell for the stripe geometry. b Energy spectrum obtained by imposing periodic boundary conditions along \(\hat{y}^{\prime}\) and J 1 = 0.8 J 2 . Each edge hosts a pair of detached helical edge modes. c Positive energy edge state for \({k}_{y^{\prime} }={k}_{{\mathcal{C}}}=\pi /2\) . Full size image Near the gap closing point, J 2 = (1 + m ) J 1 with ∣ m ∣ ≪ 1, we construct an effective continuum theory 69 for two (pseudo)-spins satisfying $$\left(2m+\frac{1}{2}{\partial }_{x^{\prime} }^{2}\right){\sigma }_{2}{\psi }_{\uparrow ,\downarrow }(x^{\prime} )=\pm\! {\partial }_{x^{\prime} }{\psi }_{\uparrow ,\downarrow }(x^{\prime} ).$$ (7) These equations provide two independent zero-energy solutions $$\psi (x^{\prime} )= \, \left(\begin{array}{c}{\chi }_{-}\\ 0\end{array}\right){e}^{-2x^{\prime} }\left({e}^{2mx^{\prime} }-{e}^{-2mx^{\prime} }\right),\\ \psi (x^{\prime} )= \, \left(\begin{array}{c}0\\ {\chi }_{+}\end{array}\right){e}^{-2x^{\prime} }\left({e}^{2mx^{\prime} }-{e}^{-2mx^{\prime} }\right),$$ (8) where σ 2 χ η = η χ η and η = ±1, which are localized at \(x^{\prime} =0\) and exist only for m > 0, namely when J 1 < J 2 (see Methods section). To compute the dispersion relation of the edge modes, it is convenient to consider a cylindrical geometry. In this case, we find that the edge modes become helical, see Fig. 4 b. An example of such states is shown in Fig. 4 c. Discussion In this work, we have shown a new type of multiple BOs that is connected to the quantum beating of Wannier centers and we have identified HOTIs as a model where this effect can be observed. By studying the BBH model, we have shown that the Wilson loop imposes period-four oscillations and the center-of-mass motion displays an anomalous Hall displacement over one period of oscillation. We have connected these features to the crystalline symmetries of the model and we have identified quantized winding numbers that protect the topological BOs. Moreover, we have shown that detached helical edge states emerge in an open system with the required symmetries. Our results can be observed with cold atoms 70 , 71 , where flux engineering can be achieved through time-dependent protocols 72 , 73 and where the staggered hopping amplitudes requires a bipartite lattice 27 , 74 . Interferometric and tomographic methods can be exploited to measure the Wilson loop winding 25 , 31 , 75 and real-space cloud imaging makes possible to measure the center-of-mass displacement 76 . A fundamental question concerns the preparation of the initial state, owing to the degenerate nature of the bands. As shown in Supplementary Note 2 the bands can be split by slightly breaking time-reversal symmetry. In this case, it is possible to prepare a non-degenerate Bose-Einstein condensate (BEC) at the Γ point. When projected onto the eigenstates of the BBH model, this state is peaked at specific values of θ and ϕ . One can then obtain the desired superposition of the two zero-momentum modes (the BEC and the gapped mode) by a coherent coupling through an external driving. The subsequent BOs require that the applied force has a magnitude that is larger than the band separation to effectively recover the band degeneracy during the BOs. In the context of photonics, our results can be investigated by using optical waveguides 77 , where it has been recently possible to realize synthetic π flux 78 , 79 . In this platform, the input laser profile can be inprinted in order to map the degenerate manifold of states at the Γ point that are parametrized by the angles θ and ϕ . It is then possible to reconstruct the Wilson loop dynamics by measuring the output field phase profile, whereas the Hall displacement is obtained from the spatial profile of the field intensity. As a perspective of our work, it would be interesting to generalize our results to other two- and three-dimensional topological crystalline insulators and consider corrections to the semiclassical equations, e.g., involving the quantum metric once an inhomogeneous electric field or a harmonic trap potential are introduced 80 , 81 . Finally, given the role played by the initial state in the observation of the anomalous Hall displacement, BOs can be thought as a tool to witness the phenomenology of symmetry-broken condensates where the ground state degeneracy has been removed by interactions (Di Liberto et al., in prepration). Methods Berry connection and curvature For the BBH model introduced in the main text, the corresponding matrix elements of the non-Abelian Berry connection, defined as \({A}_{i}^{\alpha \beta }=i\left\langle {u}_{{\bf{k}}}^{\alpha }| {\partial }_{{k}_{i}}| {u}_{{\bf{k}}}^{\beta }\right\rangle\) , read $${A}_{x}^{11}= -\!{A}_{x}^{22}=-\frac{{J}_{1}^{2}-{J}_{2}^{2}}{4{\epsilon }_{{\bf{k}}}^{2}} ,\\ {A}_{x}^{12}= \,{({A}_{x}^{21})}^{* }={e}^{-i\frac{{k}_{x}+{k}_{y}}{2}}\frac{({e}^{i{k}_{y}}{J}_{1}+{J}_{2})({J}_{1}-{e}^{i{k}_{x}}{J}_{2})}{4{\epsilon }_{{\bf{k}}}^{2}} ,\\ {A}_{y}^{11}= -\!{A}_{y}^{22}=\frac{{J}_{1}^{2}-{J}_{2}^{2}}{4{\epsilon }_{{\bf{k}}}^{2}}\ \\ {A}_{y}^{12}= {({A}_{y}^{21})}^{* }={e}^{-i\frac{{k}_{x}+{k}_{y}}{2}}\frac{({e}^{i{k}_{y}}{J}_{1}-{J}_{2})({J}_{1}+{e}^{i{k}_{x}}{J}_{2})}{4{\epsilon }_{{\bf{k}}}^{2}}.$$ (9) The SU(2) Berry curvature, defined as \({\Omega }_{xy}({\bf{k}})={\partial }_{{k}_{x}}{A}_{y}-{\partial }_{{k}_{y}}{A}_{x}-i[{A}_{x},{A}_{y}]\) , reads $${\Omega }_{xy}^{11}= \, -\!{\Omega }_{xy}^{22}={J}_{1}{J}_{2}({J}_{1}^{2}-{J}_{2}^{2})\frac{\sin {k}_{x}+\sin {k}_{y}}{4{\epsilon }_{{\bf{k}}}^{4}} ,\\ {\Omega }_{xy}^{12}= \, {({\Omega }_{xy}^{21})}^{* }=-i({J}_{1}^{2}-{J}_{2}^{2}){e}^{-i\frac{{k}_{x}+{k}_{y}}{2}}\frac{{e}^{i{k}_{y}}{J}_{1}^{2}-{e}^{i{k}_{x}}{J}_{2}^{2}}{4{\epsilon }_{{\bf{k}}}^{4}}.$$ (10) Real-space wavepacket dynamics To validate the semiclassical real-space dynamics, we have numerically simulated the evolution of a real-space wavepacket using a finite size L x × L y lattice. We start by constructing a Gaussian wavepacket centered at k 0 = Γ = (0, 0) of the form $$\left|{\psi }_{{\bf{r}},\mu }(t=0)\right\rangle =\sum \limits_{{\bf{k}}}{e}^{-{k}^{2}/2{\sigma }^{2}}{e}^{i{\bf{k}}\cdot ({{\bf{r}}}_{\mu }-{{\bf{r}}}_{0})}\left|{u}_{{\bf{k}},\mu }(0)\right\rangle ,$$ (11) where μ indicates the sublattice degree of freedom in the unit cell and r μ is its spatial position. In the simulations we take a grid of k pts × k pts points in k space within an interval k ∈ [−3 σ , 3 σ ] × [−3 σ , 3 σ ]. We evolve the state with the real-space Hamiltonian \(\hat{H}\) and calculate the observable \({{\bf{r}}}^{{\rm{num}}}(t)=\left\langle \psi (t)| \hat{{\bf{r}}}| \psi (t)\right\rangle\) , with \(\hat{{\bf{r}}}\equiv {\sum }_{{\bf{r}}}({\bf{r}}-{{\bf{r}}}_{0})\ \left|{\bf{r}}\right\rangle \left\langle {\bf{r}}\right|\) . We also study the exact evolution of the Bloch wave vector, by considering the velocity operator \(\hat{{\bf{v}}}({\bf{k}})=\partial \hat{H}({\bf{k}})/\partial {\bf{k}}={\sum }_{i}{\partial }_{{\bf{k}}}{d}_{i}({\bf{k}}){\Gamma }^{i}\) . At each time t the velocity reads \({\bf{v}}({\bf{k}}(t))=\left\langle \psi ({\bf{k}}(t))| \hat{{\bf{v}}}| \psi ({\bf{k}}(t))\right\rangle\) , where \(\left|\psi ({\bf{k}}(t))\right\rangle ={\mathcal{T}}\exp [-i\mathop{\int}\nolimits_{0}^{t}{\rm{d}}t\ \hat{H}({\bf{k}}(t))]\left|{u}_{\Gamma }(0)\right\rangle\) and k ( t ) = k (0) + F t . This method corresponds to solving the Schrödinger equation in k space. We find the displacement by integration \({{\bf{r}}}^{{\rm{exact}}}(t)=\mathop{\int}\nolimits_{0}^{t}{\rm{d}}t\ {\bf{v}}(t)+{{\bf{r}}}_{0}\) . Edge states Here we derive the effective theory at the edge by considering periodic boundary conditions along the \(y^{\prime}\) direction (see Fig. 5 ) for J 1 ≈ J 2 . In this stripe geometry, we have to double the unit cell to correctly represent the lattice periodicity, which reads \(d^{\prime} =2\sqrt{2}a\) . In chiral form, the Hamiltonian reads $${H}_{{\rm{s}}}({\bf{k}})=\left(\begin{array}{rc}0&{Q}_{s}({\bf{k}})\\ {Q}_{s}{({\bf{k}})}^{\dagger}&0\end{array}\right),$$ (12) where $${Q}_{s}({\bf{k}})=\left(\begin{array}{cccc}-{J}_{1}&-{J}_{1}&-{J}_{2}&-{J}_{2}{e}^{-i{k}_{x}}\\ {J}_{1}&-{J}_{1}&{J}_{2}{e}^{-i{k}_{y}}&-{J}_{2}{e}^{-i({k}_{x}+{k}_{y})}\\ -{J}_{2}{e}^{i({k}_{x}+{k}_{y})}&-{J}_{2}{e}^{i{k}_{y}}&-{J}_{1}&-{J}_{1}\\ {J}_{2}{e}^{i{k}_{x}}&-{J}_{2}&{J}_{1}&-{J}_{1}\end{array}\right).$$ (13) Here, we are taking units \(d^{\prime} =1\) and we are also using the convention that the lattice points within the unit cell are all sitting in the center of the unit cell. The Hamiltonian H s ( k ) is therefore in Bloch form. Fig. 5: Edge states. a Stripe geometry with periodic boundary conditions along \(\hat{y}^{\prime}\) . b Unit cell choice used to develop the continuum theory. Full size image In order to build an effective theory near zero energy 69 , let us take J 2 = (1 + m ) J 1 , with ∣ m ∣ ≪ 1. We can then split the Hamiltonian into \({\hat{H}}_{{\rm{s}}}({\bf{k}})={\hat{H}}_{{\rm{s}}}({\bf{k}}=0)+{\hat{V}}_{{\rm{s}}}({\bf{k}})\) , where \({\hat{V}}_{{\rm{s}}}({\bf{k}})\) is expanded to lowest order in k . The zeroth order term \({\hat{H}}_{{\rm{s}}}({\bf{k}}=0)\) can be diagonalized and we find four eigenvectors \(\left|{v}_{0}^{i}\right\rangle\) with energy \(E=\pm\! \sqrt{2}m{J}_{1}\) , which we use as a basis for the effective theory, and four high-energy states \(\left|{v}_{e}^{i}\right\rangle\) that we neglect. We can then construct the projection operator \({\hat{P}}_{s}={\sum }_{i}\left|{v}_{0}^{i}\right\rangle \left\langle {v}_{0}^{i}\right|\) to obtain at lowest order $${\hat{H}}_{s}^{{\rm{eff}}}({k}_{x^{\prime} }) = {\hat{P}}_{s}[{\hat{H}}_{{\rm{s}}}({\bf{k}}=0)+{\hat{V}}_{{\rm{s}}}({k}_{x^{\prime} },{k}_{y^{\prime} }=0)]{\hat{P}}_{s}\\ = \left(\begin{array}{cc}{H}_{\uparrow }({k}_{x^{\prime} })&0\\ 0&{H}_{\downarrow }({k}_{x^{\prime} })\end{array}\right),$$ (14) where we have rearranged the order of the components to have the Hamiltonian in block-diagonal form and we have defined $${H}_{\uparrow ,\downarrow }({k}_{x^{\prime} })=\left(-\frac{2m}{\sqrt{2}}+\frac{1+m}{2\sqrt{2}}{k}_{x^{\prime} }^{2}\right){\sigma }_{3}\mp \frac{1+m}{\sqrt{2}}{k}_{x^{\prime} }{\sigma }_{2} ,$$ (15) in units where J 1 = 1. We can now substitute \({k}_{x^{\prime} }\to -i{\partial }_{x^{\prime} }\) and use m ≪ 1 to obtain the coupled equations $$\left(2m+\frac{1}{2}{\partial }_{x^{\prime} }^{2}\right){\sigma }_{2}\psi (x^{\prime} )=\pm \!{\partial }_{x^{\prime} }\psi (x^{\prime} ) .$$ (16) We use standard procedures to solve these equations, namely we take \(\psi (x^{\prime} )\) as an eigenstate of σ 2 , i.e., we decompose it as \(\psi (x^{\prime} )=\varphi (x^{\prime} ){\chi }_{\eta }\) , where σ 2 χ η = η χ η with η = ±1. After taking the ansatz \(\varphi (x^{\prime} )\propto {e}^{-tx^{\prime} }\) , we find that the following algebraic equations must be satisfied $${t}^{2}\pm 2\eta t+4m=0 ,$$ (17) for \({H}_{\uparrow }({k}_{x^{\prime} })\) and \({H}_{\downarrow }({k}_{x^{\prime} })\) , respectively. Let us focus on the solution for \({H}_{\uparrow }({k}_{x^{\prime} })\) , namely the one with plus sign. We find \({t}_{\uparrow }=-\eta \pm \sqrt{1-4m}\approx -\eta \pm (1-2m)\) . For η = −1, we can construct a solution \(\varphi (x^{\prime} )={c}_{1}{e}^{-{t}_{\uparrow }^{+}x^{\prime} }+{c}_{2}{e}^{-{t}_{\uparrow }^{-}x^{\prime} }\) that is exponentially localized for m > 0 and that vanishes at \(x^{\prime} =0\) , namely c 1 = − c 2 . The solution constructed for η = 1 does not satisfy these requirements for any value of m . A similar reasoning can be repeated for \({H}_{\downarrow }({k}_{x^{\prime} })\) , where we have to take the solution with η = 1 in this case and the solution only exists for m > 0. We end up with the two zero-energy solutions $$\begin{array}{rcl}{\psi }_{\uparrow }(x^{\prime} )&=&\left(\begin{array}{c}{\chi }_{-}\\ 0\end{array}\right){e}^{-2x^{\prime} }\left({e}^{2mx^{\prime} }-{e}^{-2mx^{\prime} }\right),\\ {\psi }_{\downarrow }(x^{\prime} )&=&\left(\begin{array}{c}0\\ {\chi }_{+}\end{array}\right){e}^{-2x^{\prime} }\left({e}^{2mx^{\prime} }-{e}^{-2mx^{\prime} }\right) ,\end{array}$$ (18) that are localized at the edge \(x^{\prime} =0\) and that exist for m > 0, namely for J 2 > J 1 . Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Code availability The code that supports the plots within this paper are available from the corresponding author upon reasonable request.

Oscillatory behaviors are ubiquitous in nature, ranging from the orbits of planets to the periodic motion of a swing. In pure crystalline systems, presenting a perfect spatially-periodic structure, the fundamental laws of quantum physics predict a remarkable and counter-intuitive oscillatory behavior: when subjected to a weak electric force, the electrons in the material do not undergo a net drift, but rather oscillate in space, a phenomenon known as Bloch oscillations. Ultracold atoms immersed in a light crystal, also known as optical lattices, are one of the many systems where Bloch oscillations have been observed. In general, the motion of particles is affected by the presence of forces, such as those generated by electromagnetic fields. In certain crystals, emergent fields reminiscent of electromagnetic fields can also exist as an intrinsic property of the material and they can potentially affect Bloch oscillations. From a mathematical point of view, these intrinsic fields can take various forms. Of particular interest are those fields represented by mathematical quantities that do not commute, namely for which the product 'a x b' is not equal to 'b x a'. These mathematical quantities, and the corresponding physical properties, are commonly called "non-Abelian". In nature, generalized non-Abelian forces are required to describe the weak or strong nuclear forces, whereas electromagnetism is more simply described by Abelian (commuting) ones. Writing in Nature Communications, M. Di Liberto, N. Goldman and G. Palumbo (Science Faculty, ULB) demonstrate that intrinsic non-Abelian fields can generate a novel type of Bloch oscillations in crystals. This exotic oscillatory phenomenon is characterized by a multiplication of the oscillation period, as compared to the fundamental period set by the crystal geometry. This multiplication factor has a profound origin, as it stems from the symmetries of the crystal and can be attributed to a topological invariant (a numerical quantity that is robust under small deformations of the crystal). Furthermore, these exotic Bloch oscillations are shown to be perfectly synchronized with a beating of internal states of the crystal. This work sheds new light on topological quantum matter with non-Abelian properties.

10.1038/s41467-020-19518-x

Space

The origin of the Moon and its composition

"A primordial origin for the compositional similarity between the Earth and the Moon," Alessandra Mastrobuono-Battisti, Hagai B. Perets & Sean N. Raymond, Nature, 2015 April 9, DOI: 10.1038/nature14333 . On Arxiv: arxiv.org/abs/1502.07346 Related: A new view of the Moon's formation: Crucial difference in 'fingerprints' confirms explosive, interconnected past Journal information: Nature , arXiv

http://dx.doi.org/10.1038/nature14333

https://phys.org/news/2015-04-moon-composition.html

Abstract Most of the properties of the Earth–Moon system can be explained by a collision between a planetary embryo (giant impactor) and the growing Earth late in the accretion process 1 , 2 , 3 . Simulations show that most of the material that eventually aggregates to form the Moon originates from the impactor 1 , 4 , 5 . However, analysis of the terrestrial and lunar isotopic compositions show them to be highly similar 6 , 7 , 8 , 9 , 10 , 11 . In contrast, the compositions of other Solar System bodies are significantly different from those of the Earth and Moon 12 , 13 , 14 , suggesting that different Solar System bodies have distinct compositions. This challenges the giant impact scenario, because the Moon-forming impactor must then also be thought to have a composition different from that of the proto-Earth. Here we track the feeding zones of growing planets in a suite of simulations of planetary accretion 15 , to measure the composition of Moon-forming impactors. We find that different planets formed in the same simulation have distinct compositions, but the compositions of giant impactors are statistically more similar to the planets they impact. A large fraction of planet–impactor pairs have almost identical compositions. Thus, the similarity in composition between the Earth and Moon could be a natural consequence of a late giant impact. Main Successful models for Moon formation typically require a relatively low-velocity, oblique impact 1 between the proto-Earth and a planetary embryo of up to a few tenths of an Earth-mass ( M ⊕ ). Such Moon-forming impacts typically occur at the late stages of planetesimal accretion by the terrestrial planets 2 , 3 . A circum-terrestrial debris disk is formed from material ejected during these impacts. The composition of the disk, in a typical impact, is dominated by material from the impactor mantle (>60 weight per cent 1 , 4 , 5 ) with a smaller contribution (typically about 20%) from the proto-Earth. More material can be extracted from the proto-Earth when a slightly sub-Mars-sized body hits a fast-spinning planet that is later slowed down by resonances 16 . The spin should be close to the break-up velocity. Another possible channel producing such mixing of material from both the planet and impactor is the rare collision between two embryos of comparable mass, in which both masses are about half of the Earth’s mass 17 . Although these new models can potentially solve some of the composition issues raised by the giant-impact scenario, they do require ad hoc assumptions and pose several difficulties (see ref. 3 for a discussion). Here, we focus on typical giant-impact events, in which the Moon aggregates mostly from material originating from the impactor. Lunar meteorites and rock samples returned by the Apollo mission have a very similar composition to that of the Earth’s mantle, across a variety of different isotopes 6 , 7 , 8 , 9 , 10 , 11 . Combining these with the giant-impact simulation results, one can infer that the Moon-forming impactor and the Earth should have had a similar composition. This poses a fundamental difficulty with the giant-impact model for the origin of the Moon, since analysis of material from other Solar System bodies have shown them to differ significantly (>20 σ ) in composition from that of the Earth (see refs 12 and 13 , and ref. 14 for a review). This would suggest that the composition of the Moon-forming impactor should similarly have differed from that of the Earth, in contrast with the giant-impact basic prediction. Here, we analyse the results of extensive N -body simulations of terrestrial planet formation to show that the Earth–Moon composition challenge can be addressed. In particular, we show that the compositions of a large fraction, 20% to 40%, of giant impactors are consistent with being similar to that of the planets they impact. More generally, late giant impactors have compositions more similar to those of the planets they impact than do other planets in the same system, showing large differences (the probability of these two distributions of coming from the same parent distribution is between 1.3 × 10 −9 and 6.7 × 10 −8 ). To study the compositions of planets and their impactors we analysed 40 dynamical simulations (from ref. 15 , using the Mercury code 18 ) of the late stages of planetary accretion, following the formation of Jupiter and Saturn, and after all the gas in the protoplanetary disk has been dissipated and/or accreted to the gas giants (see ref. 19 for a recent review). Each simulation started from a disk of 85–90 planetary embryos and 1,000–2,000 planetesimals extending from 0.5 astronomical units (1 au is the Earth–Sun distance) to 4.5 au . Jupiter and Saturn are fully formed and have different orbits and inclinations in different sets of simulations (detailed descriptions can be found in ref. 15 and in the Methods). Within 100–200 million years, each simulation typically produced 3–4 rocky planets formed from collisions between embryos and planetesimals. Each of these planets accreted a large number of planetesimals during its evolution. All collisions were recorded and provide a map of each planet’s feeding zone (see also ref. 20 ). Assuming that the initial composition of material in the protoplanetary disk is a function of its position in the initial protoplanetary disk, one can compare the compositions of different bodies formed and evolved in the simulations. Previous studies explored the compositions of the different planets formed in similar simulations. However, the compositions of impactors on formed planets have hardly been explored. Pahlevan and Stevenson 21 analysed a single statistically limited simulation, which included a total of about 150 particles, and compared the compositions of any impactors on any planets during the simulation (not only giant impacts, due to small number statistics) to the compositions of the planets. They 21 concluded that the scatter among the compositions of the various impactors is comparable to the observed differences between the planets. In particular, they found that none of the planetary impactors in the simulation they analysed had an isotopic composition similar enough to the final planet to yield a composition similarity such as that of the Earth and the Moon. Using the data from our large set of high-resolution simulations, we compare the composition of each surviving planet with that of its last giant impactor, that is, the last planetary embryo that impacted the planet (typical impactor-to-planet mass ratio in the range 0.2–0.5; see Table 1 ). We include only the 20 cases where both the impactor and planet are composed of at least 50 particles each, so as to have sufficient statistics. Analysis of the additional data for impactors composed of a smaller number of particles (and hence having smaller statistics for the specific composition) are consistent with the higher-resolution cases discussed here, as shown in the Methods. The comparison is then done as follows. First we compare the feeding zones of the planet and impactor, as shown in the examples in Fig. 1 (the cumulative plots for these and all other cases can be found in the Methods). We calculate the probability P that the feeding zones of the impactor and the planet are drawn from the same distribution, using a two-group Kolmogorov–Smirnov test (probabilities shown in the plots and in Table 1 ). In 3 out of 20 cases the feeding zones contributing to the Moon and those contributing to the planet are consistent with being drawn from the same parent distribution. In other words, the Moon’s feeding zones, if derived solely from the impactor, are consistent with the Earth’s in 15% of the impacts. The consistency further improves if we assume that a fraction of the proto-Earth was mixed into the Moon (as suggested by detailed collision simulations showing a 10%–40% contribution from the proto-Earth 14 ). For the typical 20% mix of proto-Earth material with the impactor material forming the Moon (as found in simulations), 35% of cases are consistent with their feeding zones being drawn from the same parent distribution, and the success rate increases further for a higher mass contribution from the proto-Earth (see Table 1 and Extended Data Figs 1 and 2 ). While this shows that the proto-Earth and the Moon-forming impactor may have had similar feeding zones, it does not yet quantitatively guarantee that the composition is as similar as that of the Earth–Moon system. Table 1 The modelled planet–impactor systems and observations of Solar System bodies Full size table Figure 1: The distribution of planetesimals composing the planet and the impactor. a , A case where the origins of the planetesimals composing the planet (red) and the impactor (blue) are consistent with being sampled from the same parent distribution for the expected typical 20% contribution of planetary material in moon-forming impacts (Kolmogorov–Smirnov test probability >0.05). b , A case where the planet and impactor compositions are inconsistent ( P < 0.05), but become consistent once a large (40%) contribution of material from the planet is considered. The lower plots in each panel show the results when different contributions from the planet are assumed (four cases are shown 10%; 20%; 30% and 40%). The cumulative distribution for these cases as well as all other planet–impactor pairs in Table 1 can be found in the Methods. PowerPoint slide Full size image We therefore further explore the compositional similarities with the Earth–Moon system, and calculate the oxygen isotope ratios of our simulated planets. We assume that a linear gradient existed in the 17 O isotopic composition in the initial protoplanetary disk of the Solar System. Following ref. 21 we calibrate the initial 17 O isotopic composition in each of our simulations using the measured compositions of Earth and Mars (see ref. 21 and the Methods, where we also discuss the sensitivity of the results to the calibration used, as well as the criteria for which planet–impactor pairs are considered in the analysis. We find qualitatively similar results when using different criteria and calibrations, as we discuss in detail in the Methods). Given this calibration we assign each planetesimal and planetary embryo a specific initial 17 O isotopic abundance based on its initial orbit, and then average the contribution of all accreted planetesimals, while weighting each accreted planetesimal/embryo according to its mass, to obtain the 17 O of the planets and impactors and derive the offset between them. The 17 O isotope is chosen for comparison since it provides the most stringent constraint on the Earth–Moon similarity, and its abundances were measured across a variety of Solar System bodies (enabling the best opportunity for calibration). The measured difference between the 17 O abundances of Mars and Earth (used for calibration) is Δ 17 O Mars = +321 ± 13 parts per million (p.p.m.) 12 (a similarly large difference was found for the composition of 4 Vesta asteroid derived from HED (howardite–eucrite–diogenite) meteorites; −250 ± 80 p.p.m.; ref. 13 ). The difference between the Earth and Moon is just Δ 17 O = 12 ± 3 p.p.m. (ref. 11 ). In Table 1 we show the Δ 17 O differences between the impactors and the planets in our simulations. We adopt the same Earth–Mars composition difference calibration as used by ref. 21 . The results are linearly dependent on the adopted calibration; see Table 1 . The calibration factor is defined as C cal = (Δ 17 O 4 − Δ 17 O 3 )/Δ 17 O Mars , where Δ 17 O 4 and Δ 17 O 3 refer to the compositions of the fourth and third planet in the simulations (unless only three planets formed, in which case the third and second planet were taken for calibration purposes). In 20% of the cases the impactors and planets have absolute offsets comparable or smaller than the measured absolute offset in the Earth–Moon system, that is, smaller than the 1 σ limit estimated using the lunar samples (<15 p.p.m.). Taking into account the 1 σ uncertainty calculated for the Δ 17 O in the simulated systems, the fraction of consistent pairs can be as large as 40% of the whole sample. This fraction becomes larger when partial mixing of Earth material is allowed, as observed in simulation data (it increases to 50% for a 20% (and to 55% for a 40%) contribution from the planet; see Table 1 and Fig. 2 ). Even planet–impactor pairs with statistically different feeding zones have Δ 17 O offsets significantly (see Table 1 ) smaller than those found for Mars and Vesta, in most cases. More generally, planet–impactor pairs are robustly more similar in composition than are pairs of surviving planets in the same system (see Fig. 2 and Extended Data Figs 3 , 4 and 5 , as well as the Supplementary Table , for the Δ 17 O difference distribution for planets and impactors). No less importantly, the differences between the planets are of a similar order to those found between the Earth, Mars and Vesta, that is, consistent with the observations of the Solar System. Interestingly, a small fraction of the planets do have very similar composition (small Δ 17 O difference), suggesting the possibility of the existence of Solar System bodies with similar compositions to the Earth besides the Moon. As shown in Extended Data Table 2 and in the Methods, this result still holds when considering lower thresholds for the minimal number of particles composing the planet or impactor (between 1 to 40). In particular, the mean fraction of compatible planet–impactor pairs falls between 10% and 20% for all cases. The fractions become even higher (20%–40%) when accounting for the 1 σ uncertainties. Figure 2: The cumulative distribution of the absolute Δ 17 O differences between planets and their last giant impactors (blue), compared with the differences between planets in the same system (red). Panels a , b and c correspond to the cases of zero, 20% and 40% contribution of material from the planet to a Moon formed from these impacts, respectively. The vertical lines depict the Δ 17 O difference of the Earth–Moon system (dashed lines for ± σ around the mean value; central dashed line). The differences between the planet–impactor pairs are systematically smaller than those found between different planets (the same parent distribution for the two groups can be excluded with high confidence; Kolmogorov–Smirnov probability 6.7 × 10 −8 , 1.1 × 10 −8 and 1.3 × 10 −9 for the zero, 20% and 40% cases, respectively). PowerPoint slide Full size image The Earth–Moon composition similarity poses a major challenge to the standard model of the giant-impact scenario because it conflicts with the predominant derivation of the Moon composition from the mantle of the impacting planet 14 . A wide range of alternative impact scenarios were therefore investigated 14 , 16 , 17 , 21 , 22 , 23 , 24 . However, all such models suffer from potentially considerable difficulties and/or require fine-tuned conditions (see ref. 14 for a review). Our analysis of Solar-System-like planet formation scenarios potentially offers a solution to the major composition-similarity obstacle to the standard giant-impact scenario. We find that a significant fraction of all planetary impactors could have had compositions similar to the planets they struck, in contrast with the distinct compositions of different planets existing in the same planetary system. Note that the solution, suggested by our results, of the impactor and planet having a similar composition is also applicable to the origin of the Δ 17 O similarity between the Earth and the Moon, and may similarly apply for the isotopic similarities of silicon and tungsten 25 . However, it is debatable whether even a similar impactor–planet composition could resolve the compositional similarity of silicon, given that the Earth’s silicate mantle may reflect the consequences of silicon sequestration by a core formed at high temperatures on a large planetary body 25 , 26 , that is, larger than the typical impactors considered. We conclude that our findings can potentially resolve the apparent contrast between the observed similarity of the Earth and the Moon composition and its difference from that of other Solar System bodies. This primordial composition similarity solution may therefore remove the main obstacle to the standard giant-impact origin of the Moon, as well as ease some of the difficulties for the alternative giant-impact scenarios suggested in recent years 14 . Methods In the following we supply additional data on the compositions of planet–impactor pairs and the compositions of different planets at the same systems. We also provide more detailed information on the methods used, as well as discuss the sensitivity of our results to the various criteria and calibrations which we applied. The full simulation data used in this work can be provided by the authors upon request. Initial configuration of Jupiter and Saturn The simulations analysed here are described in detail in ref. 15 . The various simulations explore a range of different initial conditions for the gaseous planets. In particular, in the cjs and cjsecc simulations Jupiter and Saturn are placed on orbits with semimajor axes of 5.45 au and 8.18 au and mutual inclination of 0.5°. In cjs the orbits are circular while in cjsecc they are eccentric with e Jupiter = 0.02 and e Saturn = 0.03. In eejs, Jupiter and Saturn are placed in their current positions (5.25 au and 9.54 au ), with mutual inclination of 1.5° and larger eccentricities than the observed ones ( e Jupiter = e Saturn = 0.1 or e Jupiter = 0.07 and e Saturn = 0.08). In ejs the orbits of Jupiter and Saturn have similar parameters to those observed ( a Jupiter = 5.25AU and e Jupiter = 0.05, a Saturn = 9.54 and e Saturn = 0.06) with mutual inclination of 1.5°. Finally, in jsres, Jupiter and Saturn are placed at a Jupiter = 5.43 au and a Saturn = 7.30 au with e Jupiter = 0.005 and e Saturn = 0.01 and with mutual inclination of 0.2°. The composition difference between planets in the same system The Supplementary Table shows the Δ 17 O differences between the different planets (in each system) and the impacted planets analysed in the main text. The full cumulative distribution for these data can be seen in Fig. 2 . Note that in the cases where two impacted planets were analysed in the same system, the differences are shown with respect to both the first and second planets. The cumulative composition distribution of planet–impactor pairs We used the following procedure to calculate the spatial distribution of the feeding zones (used for Fig. 1 , Extended Data Figs 1 and 2 and the Kolmogorov–Smirnov probabilities in Table 1 ). We extracted the record of planetesimals that constitute the planet and the impactor before the last Moon-forming impact, as well as the planet composition after the collision. To account for the different contributions of particles of different masses we replicated n i times each particle, where n i is the ratio between the mass of the i th particle and the minimum mass of the planetesimals. The planetesimal record was then used to produce the distribution of the feeding zones used in our analysis. In cases where contribution of planetary material to the Moon composition was considered, we randomly chose particles from the planet and added them to the planetesimals composing the impactor, where appropriate numbers of particles were taken so as to produce the relevant fractional contribution (for the cases of 10%, 20%, 30% or 40% contribution). We then repeated the same analysis as done for the impactor, with these new mixed impactors. Δ 17 O calibration To calculate the Δ 17 O (≡ δ 17 O − 0.52δ 16 O) for the planet and impactor pairs we followed the procedure described by ref. 21 . To assign specific values of Δ 17 O to each particle in the simulation we calibrated our simulations with the Solar System observations. We assume a linear gradient of Δ 17 O with heliocentric distance r where the two free parameters in equation (1) were calibrated by assuming that the third planet formed in the system has the composition of the Earth (Δ 17 O = 0‰) and the fourth one has the composition of Mars (Δ 17 O = +0.32‰). In cases where only three planets formed (marked with an asterisk in Table 1 ), we assigned the composition of the Earth and of Mars to the second and third planets in the simulation, respectively. We then mass-averaged over all the Δ 17 O( r ) of the planetesimals that accreted to form the Earth. We used the heliocentric distance r as the initial position of each body. We did the same for the planetesimals composing the simulated Mars. In this way we have the system of equations where r i , ⊕ and m i , ⊕ ( r i ,Mars and m i ,Mars ) are the initial position and mass of the i th planetesimal composing the Earth (Mars) and M ⊕ ( M Mars ) is the final total mass of the Earth (Mars). In this way it has been possible to evaluate the Δ 17 O value for each planetesimal in the system and thus for all the planets in each system. Given this calibration we evaluate the Δ 17 O of the planet (Δ 17 O P ) and of the last impactor (Δ 17 O I ), as the average of the Δ 17 O of all their respective components, as well as calculated the 1 σ s.e.m. for each of these values ( σ P and σ I ). To check whether or not the planet–Moon system is consistent with the Earth–Moon system we evaluated the difference Δ 17 O I − Δ 17 O P and the relative error . To calculate the Δ 17 O when a fractional contribution from the planet is included, we added the average Δ 17 O of the planet and the impactor, each weighted according to the appropriate fractional contribution considered. To study the sensitivity of our results to the Earth–Mars composition difference calibration used, we also considered lower and higher calibrations, between 0.5 and 1.5 times the Earth–Mars Δ 17 O difference. We re-analysed the fraction of compatible planet–impactor pairs (that is, producing planet–moon pairs with a composition difference equal or smaller than the Earth–Moon composition difference) for these different calibration factors. The results are summarized in Extended Data Table 1 . We find that although the fraction of consistent pairs decreases with the use of a larger difference calibration, as expected, difference calibrations as much as 1.5 times larger than the Earth–Mars difference still give rise to a mean 5% of planet–impactor pairs with similar composition (and 40% within 1 σ uncertainty in the simulation compositions), rising to a mean 10% to 20% for the cases of 20% and 40% mixing of the planetary material, respectively. In other words, the results are generally robust to that level and do not dependent on a fine-tuned calibration. Dependence on the criteria for the chosen planet–impactor pairs To verify the robustness of our results to different criteria for the choice of planet–impactor pairs used in our analysis we studied two different sets of criteria: (1) Use all planet–last-impactor pairs, considering smaller thresholds (that is, smaller than the 50-particles threshold considered in the main text) for the number of composing particles (and corresponding masses), but requiring an impactor mass of at least 0.5 M Mars to ensure a Moon-forming impact. Taking a threshold of 1, 10, 20, 30 and 40 as the minimal number of particles, we find that the general conclusion is unchanged; impactors have compositions more similar to the planets they impact than to other planets in the system. The mean fraction of planet–impactor pairs with similarity comparable to that of the Earth–Moon system is between 10% and 20% in all cases (and up to 40% when considering the 1 σ uncertainties), as shown in Extended Data Table 2 . Extended Data Figs 3 , 4 and 5 show the cumulative distributions of the compositions of planets and last impactors for all the systems, regardless of the number of particles that contributed to their formation, and for a minimum of 10, 20, 40 and 50 particles composing the planet and last impactor. Extended Data Figs 3 , 4 and 5 are shown for 0%, 20% and 40% mixing between the material of the planet and impactor. (2) Consider only last impactors on the third planet. Once we require the impactor to have at least 0.5 M Mars and be composed of a significant (>20) number of particles, the statistics become too small. When we do not consider a minimal threshold for the number of composing particles, we find that 2 out of 18 (11%; up to about 30% with the 1 σ uncertainty in the simulation compositions) planet–impactor pairs with a composition difference equal or smaller than the Earth–Moon system.

The Moon is thought to have formed from the debris of a small planet that collided with the Earth. Since the composition of other planets in the solar system differs from that of the Earth, it was expected that the Moon's composition would also differ from that of the Earth. Surprisingly, the composition of the Earth and the Moon are very similar (no, the Moon is not made out of cheese), raising a major challenge to the "giant impact" origin of the Moon. A new study by researchers from the Technion and Nice University explains the origin of such compositional similarity and helps solve this conundrum The Moon has fascinated human kind since the earliest days of history. It has played a central role in the making of annual calendars in Muslim, Jewish and other cultures; and was considered one of the gods in many pagan traditions. Questions regarding the origin of the Moon, its shape and composition gave rise to myths and legends that have accompanied humanity for thousands of years, and even today many children ask themselves—and their parents—whether the Moon is made of cheese. In the modern era such millennium-old puzzles have been replaced by scientific exploration that raised no-less challenging questions, which continue to perplex us—even 40 years after man first landed on the Moon. Now, research done by Technion researchers sheds new light on the origin of the Moon and its composition. The research, published in Nature, was led by post-doctoral researcher Dr. Alessandra Mastrobuono-Battisti and her adviser Assistant Prof. Hagai Perets from the Technion, in collaboration with Dr. Sean Raymond from Nice University. "Many models for the Moon origin were suggested by scientists, but since the 1980s the scientific community has been focusing on the most promising model—the so called 'giant impact' paradigm," explains Perets. "According to this model, the Moon was formed following a collision between a small Mars-like planet (usually called Theia) and the ancient Earth. Some of the debris from the collision fell back to Earth, some was scattered far into space and the rest went into orbit around the Earth. This orbiting debris later coagulated to form a single object: the Moon." Based on complex simulations of such collisions, researchers have found out that most of the material that eventually forms the Moon comes from the impactor, Theia, and only a smaller fraction originates from the impacted body (in this case, the Earth). Measurements of the composition of other bodies in the solar system such as asteroids and Mars have shown that they have a very different composition from that of the Earth. Given that most of the Moon material came from another body in the solar system, it was expected that the composition of the Moon should be similarly very different from that of the Earth, according to the "giant impact" model. However, analysis of samples brought from the Moon by the Apollo missions showed otherwise—in terms of composition, the Earth and Moon are almost twins, their compositions are almost the same, differing by at most few parts in a million. This contradiction has cast a long shadow on the 'giant impact' model, and for some 30 years this contradiction was a major challenge to physicists grappling with the formation of the Moon. Now, Mastrobuono-Battisti, Perets and Raymond have suggested a new solution to this mystery. Simulations of the formation of planets in the solar system, showed that different planets indeed have distinct compositions, as found from the analysis of material from different planets in the solar system. Such studies have traditionally focused on studying only the compositions of the final planets. In the new research, Perets and collaborators have considered not only the planets, but also the composition of the impactors on these planets. Consequently they have discovered that in many cases, the planets and the bodies that collide with them share a very similar composition, even though they formed independently. Thus, conclude the researchers, the similarity between the Moon and Earth stems from the similarity between Theia—from which the Moon was formed—and Earth. "It turns out that an impactor is not similar to any other random body in the solar system. The Earth and Theia appear to have shared much more similar environments during their growth than just any two unrelated bodies," explains Mastrobuono-Battisti. "In other words, Theia and Earth were formed in the same region, and have therefore collected similar material. These similar living environments also led them eventually to collide; and the material ejected mostly from Theia, ultimately formed the Moon. Our results reconcile what has been perceived as a contradiction between the process whereby Moons are formed (from matter from the impacting body) and the similarity between Earth and the Moon." "The Earth and the Moon might not be twins born of the same body," summarizes Perets, "but they did grow up together in the same neighborhood."

10.1038/nature14333

Biology

Researchers find fat burning molecule in mice

Yue Li et al, Thioesterase superfamily member 1 undergoes stimulus-coupled conformational reorganization to regulate metabolism in mice, Nature Communications (2021). DOI: 10.1038/s41467-021-23595-x Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-021-23595-x

https://phys.org/news/2021-07-fat-molecule-mice.html

Abstract In brown adipose tissue, thermogenesis is suppressed by thioesterase superfamily member 1 (Them1), a long chain fatty acyl-CoA thioesterase. Them1 is highly upregulated by cold ambient temperature, where it reduces fatty acid availability and limits thermogenesis. Here, we show that Them1 regulates metabolism by undergoing conformational changes in response to β-adrenergic stimulation that alter Them1 intracellular distribution. Them1 forms metabolically active puncta near lipid droplets and mitochondria. Upon stimulation, Them1 is phosphorylated at the N-terminus, inhibiting puncta formation and activity and resulting in a diffuse intracellular localization. We show by correlative light and electron microscopy that Them1 puncta are biomolecular condensates that are inhibited by phosphorylation. Thus, Them1 forms intracellular biomolecular condensates that limit fatty acid oxidation and suppress thermogenesis. During a period of energy demand, the condensates are disrupted by phosphorylation to allow for maximal thermogenesis. The stimulus-coupled reorganization of Them1 provides fine-tuning of thermogenesis and energy expenditure. Introduction Intracellular triglycerides are stored within lipid droplets (LD) that are juxtaposed to mitochondria in the cytoplasm 1 . This close relationship facilitates the rapid transfer of fatty acids to mitochondria, which are generated by lipolysis in response to cellular energy demands 2 . One such demand is non-shivering thermogenesis, which is mediated by brown adipose tissue (BAT) in response to cold exposure to generate heat 3 . This occurs, at least in part, when mitochondrial β-oxidation is uncoupled from oxidative phosphorylation via uncoupling protein 1. The hydrolysis of LD-derived triglycerides is initiated following the release of norepinephrine from neurons, which binds to β3 adrenergic receptors on brown adipocytes to stimulate a cell signaling cascade that activates adenylyl cyclase to generate cAMP and activate protein kinase A (PKA) 4 . In brown adipocytes, the generation of fatty acids for mitochondrial oxidation requires PKA to phosphorylate/activate perilipin, a regulatory membrane-bound protein that encircles LD. The function of perilipin is to protect triglycerides within the LD from lipolysis, and its phosphorylation by PKA allows access of cytoplasmic adipose triglyceride lipase (ATGL), hormone-sensitive lipase (HSL), and monoacylglycerol lipase (MAGL), to generate free fatty acids through sequential lipolytic steps 5 . Once fatty acids are free in the cytoplasm, they are esterified with coenzyme A (CoA) by long-chain acyl-CoA synthetase 1 (ACSL1) to form fatty acyl-CoAs, which are transported into mitochondria via carnitine palmitoyl transferase 1 and metabolized to produce heat 6 . Fatty acyl-CoA molecules can also be hydrolyzed to fatty acids in the cytoplasm by acyl-CoA thioesterase (Acot) isoforms 7 , 8 . The fatty acids generated can be utilized to make additional fatty acyl-CoAs that are transported into mitochondria or can be returned to the LD for storage. Thioesterase superfamily member 1 (Them1), which is also known as brown fat inducible thioesterase or steroidogenic acute regulatory protein-related lipid transfer (START) domain 14 (StarD14)/Acot11, plays an important role in energy homeostasis 9 , 10 . Mice with the genetic deletion of Them1 exhibit increased energy expenditure, which results in reduced weight gain when challenged with a high fat diet despite high food consumption 10 . This is attributable to increased mitochondrial fatty acid oxidation, and mechanistic studies have demonstrated that Them1 functions to suppress energy expenditure by limiting triglyceride hydrolysis in BAT, thus inhibiting the mitochondrial oxidation of LD-derived fatty acids 10 , 11 . Them1 is comprised of tandem N-terminal thioesterase domains and a C-terminal lipid-binding steroidogenic acute regulatory-related lipid transfer (START) domain. Proteomics has revealed that there are a number of BAT-selective phosphorylation sites near the N-terminus of Them1 in vivo, most notable are serine residues at (S) 15, 18, and 25 12 . Here, we show experimentally that Them1 is phosphorylated at the N-terminus after stimulation in vitro and exists in two different phosphorylation-dependent conformational states: punctate and diffuse. Punctate refers to small, highly concentrated aggregations of Them1 that localize to regions of the cytoplasm that interface with LD and mitochondria, whereas diffuse refers to homogeneously distributed Them1 across the cell cytoplasm and nucleus. We demonstrate that β-adrenergic stimulation with NE, which is used to mimic cold exposure, activates a signaling cascade that results in Them1 phosphorylation and a diffuse localization. A functional analysis revealed that Them1 in punctate form suppresses fatty acid oxidation, whereas this suppression is abrogated when it is diffusely distributed in the cell cytoplasm. Overall, these findings highlight the importance of Them1 phosphorylation in the regulation of thermogenesis in BAT and lend support for targeting Them1 for the management of obesity-related disorders. Results N-terminal serine phosphorylation of Them1 To assess the BAT-selective phosphorylation of Them1, we consulted the phosphomouse database 12 , which revealed phosphopeptides containing phosphorylation events at serine (S) 15, 18, and 25 in the N-terminus of Them1 in vivo within BAT, which are highly conserved (Supplementary Fig. 1 ). To explore the role of Them1 S-phosphorylation in regulating metabolism, cultured immortalized brown adipose cells (iBAs) from mouse BAT were used as an in vitro model. Because iBAs did not express Them1 mRNA or protein, as was the case for primary cultured brown adipocytes 11 , Them1 expression required plasmid or adenoviral transfection (Supplementary Fig. 2a–c ) to study its biochemical and physiological characteristics in vitro. To examine whether Them1 is S-phosphorylated after stimulation, we determined the aggregate abundance of phosphopeptides in the N-terminus of Them1 by mass spectrometry after stimulation with phorbol 12-myristate 13-acetate (PMA; Fig. 1a, b ) normalized to hormone-sensitive lipase as a housekeeping phosphopeptide that does not change with PMA stimulation (Supplementary Fig. 3 ). This resulted in a linear increase in phosphorylation events over time after stimulation (Fig. 1a ), with serine phosphorylation events that could be measured at S25, S27, and S31 (Fig. 1c ). Taken together, these in vivo and in vitro results suggested that stimulus-mediated S-phosphorylation within the N-terminus could control Them1 metabolic activity in BAT. Fig. 1: Regulation of Them1 phosphorylation and its subcellular localization in iBAs. a LC-MS/MS data for iBAs expressing Ad-Them1-EGFP stimulated with PMA for 0–4 h. The data are presented as normalized aggregate abundance of N-terminal phosphopeptides using hormore sensitive lipase as a reference for normalization, which does not change after PMA stimulation (see Supplementary Fig. 3 ). Regression line indicates a positive and significant correlation between phosphorylation events at the N-terminus and time after stimulation. Data are means ± SE for n = 3 different experiments/timepoint. Statistical significance was determined by ANOVA on the regression line, where P < 0.001. Normality test passed ( P = 0.759) and constant variance test, via Spearman Rank Correlation, passed ( P = 0.460). b Schematic diagram of the experimental design in a . c N-terminal amino acid sequence of Them1 showing specific phosphorylation events at S25, S27, and S31 in the Them1 sequence (blue asterisks) as determined by LC-MS/MS. Phosphorylation events in vivo at S15, S18, and S25 are represented by red font. Full size image Phosphorylation-dependent changes in Them1 localization after stimulation In transfected cells, Them1 exhibited a punctate intracellular localization (Fig. 2a ). When cells co-stained for LD and mitochondria were reconstructed from 3-D stacks, Them1 was closely associated, but did not co-localize, with either of these organelles (Fig. 2b ). The punctate localization of Them1 was not due to EGFP aggregation, per se, as evidenced by the diffuse intracellular localization of EGFP in cells transfected with EGFP alone (Supplementary Fig. 4a ). Furthermore, Them1 formed puncta in transfected cells that co-localized with Them1-EGFP (Supplementary Fig. 4b ), and puncta were also present when Them1 was transfected without the EFGP tag (Supplementary Fig. 4c ). The antibody used to detect Them1 was made and characterized by us 11 , 13 and the newly prepared antibody, which was also affinity purified using a Them1 peptide fragment, was specific for Them1 in BAT and iBAs in immunoblots (Supplementary Fig. 5a ) and specifically labeled structures in BAT from wild-type but not in Them1-deficient mice (Supplementary Fig. 5b–d ). Fig. 2: Regulation of Them1 subcellular localization in iBAs. a Confocal microscopy of an iBAs expressing Them1-EGFP (green). LD lipid droplets (magenta), N nucleus (blue), P puncta (green). b 3-D image of an iBAs expressing Them1-EGFP (green), perilipin 1 to identify lipid droplets (LD, magenta), and Tomm20 to identify mitochondria (M, red). This image is shown in surface projection mode. c Schematic illustration of the putative Them1 regulatory pathway required for the phosphorylation and 3-D organization of Them1. β 3 -AR beta-3 adrenergic receptor, DAG diacylglyerol, NE norepinephrine. DAG and PKC may have an indirect effect on puncta dissolution (dashed lines). d PMA-induced activation of iBAs expressing Them1-EGFP (green) resulted in the diffusion of Them1, which can be blocked by the PKCβ inhibitor LY333531 (LY). e Live-cell imaging was used to record changes in Them1 localization over time after PMA treatment. LD lipid droplets, P1–P3 individual puncta (green). f After PMA-induced PKC activation, neither PKI or atglistatin (Atglst) blocked the diffusion of Them1 from puncta. g , h Whereas the activation of PKA and PKC by forskolin (FSK) or norepinephrine (NE) resulted in the diffusion of Them1, PKI, Atglst, and LY blocked the diffusion of puncta (P, green). Scale bars in a – h , 10 μm. i Box and whisker plot with individual cell data (green squares). For each box, a solid line represents the median and a dashed line is the mean. Outliers are shown at the 5th/95th percentiles from 3 to 5 different experiments with the following n : Them1-EGFP, n = 16; PMA, n = 8; PMA + PKI, n = 16; PMA + Atglst, n = 7; PMA + LY333531, n = 12; FSK, n = 9; FSK + PKI, n = 12; FSK + Atglst, n = 7; FSK + LY333531, n = 10. For PMA, * P = 0.024 PMA + LY vs PMA alone; † P < 0.001 PMA + LY vs PMA + PKI; ‡ P < 0.01 PMA + LY vs PMA + Atglst. For forskolin, * P < 0.001 forskolin + PKI vs forskolin alone; † P = 0.011 forskolin + Atglst vs forskolin alone; ‡ P = 0.006 forskolin + LY vs forskolin alone. ND no difference. Statistical significance as determined by a multiple comparisons ad hoc test (Dunns Method) after Kruskal–Wallis one-way ANOVA on ranks. Full size image We next explored the hypothesis that the stimulus-mediated phosphorylation of Them1 facilitates changes in its localization. When differentiated iBAs were incubated with PMA, which activates protein kinase C (PKC), and the localization of Them1 was examined in fixed or living cells, PMA initiated the dissolution of Them1 from puncta (Fig. 2c, d ). The dissolution of puncta resulted in Them1 uniformly distributed within the cell cytoplasm (Fig. 2d ), which occurred over a 4-h period after PMA stimulation (Fig. 2e ) but not with vehicle alone (Supplementary Fig. 6 ), as determined by using time-lapse live-cell microscopy. We further observed that the isoform-specific PKCβ inhibitor ruboxistaurin (LY333531 or LY) completely blocked the dissolution of puncta in PMA-treated cells (Fig. 2d ). We next sought to identify upstream signaling events and define the pathway leading to Them1 phosphorylation (Fig. 2c ). iBAs incubated with the PKA inhibitor PKI or the ATGL inhibitor atglistatin prior to stimulation with PMA had a diffuse Them1 distribution in the cytoplasm (Fig. 2f ). Similarly, iBAs stimulated with forskolin (Fig. 2g ), which activates adenyl cyclase and protein kinase A (PKA) upstream of PKC (Fig. 2c ), or with norepinephrine, which activates both PKA and PKC (Fig. 2c ) resulted in diffusely localized Them1, which was blocked by inhibiting downstream effectors with PKI, Atglistatin, or LY, respectively (Fig. 2g, h ). DAG released by ATGL is in the sn-1,3 or sn-2,3 stereo isoform, so that triacylglycerol-derived DAG cannot activate PKC without isomerization to sn-1,2 DAG 14 . This event was not investigated, but would be required to confirm the pathway details (Fig. 2c ). Overall, these results suggest that norepinephrine stimulation activates PKA, and the activation of PKC occurs downstream of PKA activation (Fig. 2c ). Measuring the volume of intracellular Them1 in iBAs transfected with Them1-EGFP with or without inhibitors (Fig. 2i , schematic; green EFGP fluorescence signal) further demonstrated that Them1 in puncta is concentrated and occupies significantly less volume than occurs when Them1-EGFP is diffusely distributed after stimulation with PMA or forskolin (Fig. 2i ). The dissolution of Them1 from puncta after stimulation with PMA or forskolin was not attributable to the synthesis of new Them1-EGFP protein, because inhibition of protein synthesis using cycloheximide during the 4 h period of stimulation had no effect on Them1 localization or protein concentration (Supplementary Fig. 7 ). The canonical PKC binding sequence, R/K-X-S-X-R/K where X at +1 is a hydrophobic residue, is not present within the N-terminal Them1 sequence (Fig. 1c ). However, our pathway suggested that PKC was involved in the dissolution of puncta. To examine this experimentally, we used an antibody specific for the canonical PKC binding sequence, which showed that no antibody recognition of PKC-mediated S-phosphorylated Them1 occurred (Supplementary Fig. 8 ). Instead, at least six proteins in iBAs were time-dependently phosphorylated by PKC after PMA stimulation (Supplementary Fig. 8 ). These results support that Them1 per se is not a PKC substrate but that PKC may be indirectly involved in the dissolution of puncta after PMA stimulation (Fig. 2c ). We next sought evidence that the stimulus-coupled changes in Them1 localization also occurred in brown adipose cells in vivo (Fig. 3 ). For this, mice were injected with saline (control) or with the β3 adrenergic receptor selective agonist CL316,243. In saline-injected mice, Them1 localized to distinct puncta near LD and were in the cytoplasm (Fig. 3a ). The strong fluorescence signal for Them1 at baseline was due to mild cold exposure by housing mice at 22 °C, which is below their thermoneutral zone. Tissues at 4 h after saline injection showed the same localization of Them1 in puncta near lipid droplets and within the cytoplasm; Them1-containing puncta at 4 h after saline injection occupied a small intracellular volume (Fig. 3b, c ). Fig. 3: Them1-containing puncta in brown adipose tissue (BAT) diffuse after stimulation in vivo. a BAT was excised from mice immediately after the administration of saline. These tissues consisted of large lipid droplets (LD; magenta) surrounded by perilipin and numerous Them1-containing puncta (P; green) that were localized in the cytoplasm or in association with perilipin at the lateral edge of LDs. N, nuclei (blue) were stained with Hoechst 33342 (HOE). Scale bar, 10 μm. n = 5 images taken randomly from n = 2 mice with similar results. b BAT tissues from mice injected with saline and then tissues collected 4 h later. The localization of lipid droplets (LD, magenta), puncta (P, green), and nuclei (N, blue) were the same as in a . n = 5 images taken randomly from n = 2 mice with similar results. Scale bar, 10 μm. c Quantification of the 3-D volume of Them1 from mice injected with saline and collected 4 h later (contr; control) or CL316,243, a β3 adrenergic receptor agonist, at 1, 2, or 4 h after stimulation. Regression line indicates a positive and significant correlation between the volume of Them1 and time; the volume of Them1 is small when Them1 localizes to puncta in control cells and large as Them1 is diffuse after stimulation with CL316,243 (see Fig. 2i ). Error bars represent mean ± SE from n = 5 images taken randomly from n = 2 mice/treatment and analyzed using Imaris software. Statistical significance was determined by ANOVA on the regression line, where P = 0.006. d BAT tissues from mice exposed to CL316,243 for 1 h. In these tissues, LD (magenta) were smaller than in saline-treated mice and Them1 (green) was more diffuse. N nuclei (blue). e By 2 h after CL316,243 administration, LD were less abundant and Them1 occupied a larger cytoplasmic volume. f BAT tissue from mice 4 h after the administration of CL316,243. LD (magenta) were not present, perilipin was cytoplasmic, and Them1 (green) was nearly completely diffuse in the cytoplasm. Over time (compare d – f ) nuclear (N) Them1 staining (blue) increased in intensity. d – f n = 5 images taken randomly from n = 2 mice/treatment with similar results. Scale bars, 10 μm. Full size image By 1 h after CL316,243 administration, LD were reduced in size and number due to lipid hydrolysis and Them1 showed a larger cytoplasmic volume (Fig. 3c, d ). By 2 h after CL administration, LD were less abundant and Them1 occupied a larger cytoplasmic volume compared to 1 h (Fig. 3c, e ). By 4 h after CL316,243 administration, there were few intracellular LD, as evidenced by the loss and disruption of perilipin 1 expression, and Them1 occupied a larger cytoplasmic volume yet (Fig. 3c, f ). To accompany the diffuse localization of Them1 from 1–4 h after CL administration was an increase in the nuclear localization of Them1 (Fig. 3d–f ). When taken together with observations in vitro using iBAs, these in vivo studies support the notion that reorganization of Them1 puncta occurs after stimulation that activates the β 3 adrenergic receptor (Fig. 2c ). S-phosphorylation of the N-terminus of Them1 drives localization after stimulation To develop direct evidence that phosphorylation of N-terminal serine residues of Them1 are involved in puncta formation and dissolution, we deleted the first 36 amino acids of Them1 to generate a construct (Δ1-36-EGFP) that did not contain the phosphorylation sites that were detected in vivo or following transfection and stimulation of iBAs (Fig. 4a, b ). The Δ1-36 mutant of Them1 exhibited a diffuse localization (Fig. 4c ). We then linked a peptide containing the first 36 amino acids of Them1 directly to EGFP (Fig. 4b ), which exhibited the same punctate localization as full-length Them1 (Fig. 4c ). These results strongly support that amino acids 1–36 direct Them1 localization to puncta and suggest that phosphorylation of Them1 at one or more serine residues in the N-terminal region drives the conformational change in Them1 after stimulation. Fig. 4: Phosphorylation at S15 and S25 in the N-terminus regulates the diffusion of Them1 in iBAs. a , b Illustration of the Them1 structure and its mutant constructs tagged with EGFP used in the experiments. a Them1, highlighting the N-terminus S phosphorylation sites, linked to EGFP at the carboxy terminus. b Δ1-36, deletion of amino acids 1-36 in the Them1 sequence; 1-36, a construct containing only the first 36 amino acids in the Them1 sequence; AAA, serine residues at amino acids 15, 18, and 25 were mutated to alanine (A); DDD, serine residues at amino acids 15, 18, and 25 were mutated to aspartic acid (D). All mutant Them1 constructs were linked to EGFP at the carboxy terminus. c the Δ1-36 Them1 construct diffuses in the cytoplasm (green) whereas the 1-36 construct forms puncta (green) around lipid droplets (LD), which were stained with Oil Red O stain (magenta). d Mutation of S15/S18/S25 to aspartic acid (DDD) results in diffuse Them1 localization (green). e S15/S18 or S15/S25 mutant constructs result in diffuse Them1 localization (green) whereas the S18/S25 construct expressed some puncta and some diffuse Them1 (green). f The single mutation at S15 or S25, but not S18, to aspartic acid resulted in the diffusion of Them1 (green). g Box and whisker plot with individual cell data (green squares) to quantify volume measurements for mutation of S15/S18/S25 (DDD), or single mutations at S15D, S18D, or S25D. For each box, a solid line represents the median and a dashed line is the mean. Outliers are shown at the 5th/95th percentiles. n = 2–5 cells each from two different preparations. Data were evaluated using a one-way ANOVA with all pairwise multiple comparison using the Holm–Sidak method. * P = 0.05, S18D vs S15D; † P = 0.011, S25D vs S18D. h The serine to alanine mutation at S15 results in Them1-containing puncta, which are not dispersed after activation by PMA, FSK, or NE. Scale bar in c – h , 10 μm. For each construct evaluated in Fig. 4 , the data are representative of n = 3 independent experiments with 3–5 individual cells photographed per treatment with similar results. Full size image To further explore this idea, we mutated each putative phosphorylation site by exchanging S for aspartic acid (S15-18-25D, or DDD), which mimicked the phosphorylated state of the S amino acid residues that are phosphorylated in vivo (Fig. 4b ). This construct resulted in diffuse Them1 localization (Fig. 4d ). When the S sites were mutated in pairs (Fig. 4e ), or individually (Fig. 4f ), S15D and S25D led to diffuse Them1 localization. This result was quantified as the volume of Them1 per total cell volume using the strategy in Fig. 2i (Fig. 4g ). By contrast, exchanging S for alanine at S15 or at S15, S18, and S25 (AAA, Fig. 4b ), a configuration that cannot be phosphorylated, resulted in the punctate localization of Them1 alone or after stimulation with PMA, FSK, or norepinephrine (Fig. 4h ). These results demonstrate that the phosphorylation of Them1 at S15 and/or S25 results in the diffuse localization of Them1 in iBAs. Puncta formed by Them1 are metabolically active structures Although plasmid-mediated transfection of Them1 constructs was sufficient for fluorescence-based imaging studies, the low transfection efficiencies prevented quantitative assessments of the impact of puncta formation on cellular metabolism. To obtain higher transfection efficiency, we constructed adenoviral vectors (Ad). These included wild-type Them1 or Them1 with AAA or DDD substitutions and EGFP fused at the carboxy terminus (Fig. 4a, b ) with Ad-EGFP used as a control. Them1 expression at an MOI of 1:40 was ~45%, which resulted in Them1 expression similar to tissue expression in BAT of cold-exposed mice (Supplementary Fig. 2b ). Both wild type (not shown) and AAA (Fig. 5a ) substituted Ad-Them1-EGFP formed puncta in iBAs. Using near super-resolution imaging via Zeiss Airyscan, it was clear from 3-D images that Them1 resides in discrete puncta that were closely associated with LD (Fig. 5a ). In contrast, DDD substituted Ad-Them1-EGFP exhibited a diffuse localization (Fig. 5b ). In the near super-resolution 3-D images, diffuse Them1 filled the entire volume of cytoplasm not occupied by cellular organelles (Fig. 5b ). These results overall were consistent with the results from our plasmid constructs. Fig. 5: The punctate localization of Them1 is metabolically active. a Ad-Them1-EGFP with S15, S18, and S25 mutated to alanine (Ad-AAA-EGFP, green) formed puncta near lipid droplets (LD, magenta). This close relationship was highlighted by using near super-resolution Zeiss Airyscan imaging and reconstructing cells in 3-D. b Phosphomimetic mutations at S15, S18, and S25 (Ad-DDD-EGFP, green) caused diffusion of Them1. In 3-D, Them1 filled the thickness of the cytoplasm (double white arrow) and could be found above and below the plane of organelles including LD (magenta), which excluded Them1. N nucleus (blue). Scale bars, 10 μm (low magnification images-left and center); 2.5 μm (high magnification image-right). Data are representative of n = 3 independent experiments with similar results. c iBAs transduced with Ad-Them1-EGFP (blue), Ad-AAA-EGFP (red), Ad-DDD-EGFP (violet), or Ad-EGFP (green) were stimulated with norepinephrine (NE). The response of OCR values to stimulation was measured as % of baseline. Data ( n = 66–72 wells from three independent experiments) were evaluated using mixed-effects analysis (group effect P = 0.001; interaction effect P < 0.001) followed by multiple comparison testing using the Tukey method (* P = 0.015, 0.034, 0.041, 0.039, 0.030, Ad-Them1-EGFP vs. Ad-EGFP; † P = 0.031, 0.040, 0.034, 0.018, 0.013, 0.009, 0.007, Ad-AAA-EGFP vs. Ad-DDD-EGFP). d OCR values for Ad-Them1 (blue) with or without PMA. Data were evaluated using mixed-effects analysis (group effect P < 0.001; interaction effect P < 0.001) followed by multiple comparison testing using the Sidak method (* P = 0.012, <0.001, <0.001, <0.001, 0 vs. 3 μM PMA). e OCR values for Ad-EGFP (green), Ad-AAA-EGFP (red), and Ad-DDD-EGFP (violet) with or without PMA. Data were evaluated using mixed-effects analysis (vehicle: group effect P = 0.050, interaction effect P < 0.001; PMA: group effect P = 0.060, interaction effect P = 0.015) with multiple comparison testing using the Holm-Sidak method (vehicle: # P = 0.005, 0.003, Ad-EGFP vs. Ad-AAA-EGFP, * P = 0.034, Ad-DDD-EGFP vs. Ad-AAA-EGFP; PMA: # P = 0.032, 0.024, 0.024, Ad-EGFP vs. Ad-AAA-EGFP). f Change in area under the curve (Δ AUC) for vehicle – PMA. Data ( n = 33–36 wells from three independent experiments) were evaluated using a one-way ANOVA ( P = 0.004) with multiple comparison testing using the Holm–Sidak method (* P = 0.046, Ad-DDD-EGFP vs Ad-AAA-EGFP, ## P = 0.002, Ad-EGFP vs Ad-AAA-EGFP. Data are mean ± SE. Full size image We next examined the role of Them1 in regulating the oxygen consumption rate (OCR) in differentiated iBAs in response to norepinephrine stimulation. OCR values are a surrogate measure of oxidation rates of fatty acids generated by hydrolysis of LD-derived triglycerides 11 . After optimizing conditions of cell density by calculating the total number of EGFP-containing cells/total cells per plate (Supplementary Fig. 9a ), OCR values were measured at baseline and after norepinephrine stimulation. Baseline and norepinephrine-stimulated OCR values were similar for Ad-DDD-EGFP and the Ad-EGFP control, whereas norepinephrine-stimulated values of OCR for Ad-Them1-EGFP and Ad-AAA-EGFP were reduced (Fig. 5c ). These findings suggest that Them1 is active in suppressing LD-derived FA oxidation when in its punctate configuration. Because the stimulated redistribution of Them1-EGFP-APEX2 in puncta fully diffuses over a period of 4 h (Fig. 2e ), increased OCR values within the first 1 h after norepinephrine exposure (Fig. 5c ) would be expected to reflect only partial dissolution of Them1-containing puncta. By pre-incubating iBAs expressing Ad-Them1-EGFP with PMA prior to norepinephrine exposure, which would result in the further dissolution of puncta, we observed an increase in OCR values relative to no PMA (Fig. 5d ). The addition of PMA to iBAs led to a progressive increase in values of extracellular acidification rate (ECAR) until the time of NE exposure in the absence of differences in OCR (Supplementary Fig. 9b ). This result suggests that PMA stimulated anaerobic glycolysis, as opposed to fatty acid oxidation, which increases OCR following NE exposure. The PMA-induced increases in anaerobic glycolysis led to observed reductions in fatty acid oxidation, as reflected by OCR in response to norepinephrine stimulation, expressed as a percentage of baseline (Fig. 5e ). When the effects of PMA treatment on OCR per se are taken into account, the NE-stimulated suppression of OCR on cells transduced with Ad-AAA was maintained compared to cells transduced with Ad-EGFP or Ad-DDD, (Fig. 5f and Supplementary Fig. 9c ). These results argue against downstream regulation of these protein constructs by PMA. These metabolic studies support the role of Them1 puncta in regulating oxidative metabolism in BAT, which were further supported by measuring fatty acid oxidation contained within endogenous triglycerides in iBAs by pulse labeling with 3 H-oleate and then determining triglyceride utilization during a chase period (Supplementary Fig. 9d ). During the pulse period, equal fatty acid uptake was observed in iBAs transduced with Ad-EGFP, Ad-AAA-EGFP, or Ad-DDD-EGFP, leading to equal cellular accumulation of triglycerides (Supplementary Fig. 9d ). By contrast, during the chase period triglyceride accumulated in iBAs transduced with Ad-AAA-EGFP, but not Ad-EGFP or Ad-DDD-EGFP (Supplementary Fig. 9d ), which reflected reduced rates of fatty acid oxidation in the presence of Ad-AAA-EGFP (Supplementary Fig. 9e ). Puncta are biomolecular condensates, or “membraneless organelles”, by correlative light/electron microscopy To examine the ultrastructural details of puncta in iBAs, we next used Them1-EGFP vector constructs that included an ascorbate peroxidase–derived APEX2 tag appended to the carboxy terminus to perform correlative light/electron microscopy (Fig. 6 ). The APEX2 tag was developed with diaminobenzadine, resulting in brown reaction product only in cells expressing Them1 (Fig. 6a ). This procedure allowed us to clearly determine which cells were transfected and expressing Them1. Them1 puncta were imaged using the AAA vector construct (AAA-EGFP-APEX2; Fig. 6b, c ), and diffuse Them1 was evaluated using the DDD vector construct (DDD-EGFP-APEX2; Fig. 6d ). Cells transfected with EGFP-APEX2 without Them1 showed pale staining uniformly throughout cells including the nucleus (not shown). Fig. 6: Correlative light and electron microscopy identifies puncta as biomolecular condensates. a Thick, 0.5 μm, plastic sections of iBAs in culture transfected with phosphorylation mutants AAA- or DDD-Them1-EGFP-APEX2. Cells stained brown (box) also have dark brown punctate structures in mutant AAA-Them1 expressing cells (arrows). For AAA-transfected cells, serial sections show two different levels of the same cell. No puncta are present in cells transfected with the DDD-Them1 mutant although Them1 is found in the nucleus (arrows). Scale bar, 25 μm. b In thin sections of the boxed cell in a , puncta 1 (P1) is localized near the nucleus (N) and puncta 2 (P2) is localized near a lipid droplet (LD). At high magnification, punctate structures, P1 and P2, have no surrounding membrane and consist of liquid droplets embedded in an amorphous matrix. Scale bars, 5 μm (low magnification; original magnification, ×4000) and 0.5 μm (high magnification; original magnification, ×60,000). c In thin sections of the boxed cell in a , this section shows smaller puncta in the cytoplasm (P3), puncta associated with lipid droplets (LD; P4), and puncta in close association with LD and mitochondria (M) in the cell cytoplasm (P5). Scale bars, 5 μm (low magnification; original magnification, ×4000) and 0.5 μm (high magnification; original magnification, ×60,000). d In thin sections from the boxed region in a , no puncta are found in cells transfected with the DDD-Them1 mutant. Them1 is distributed throughout the cytoplasm and in the nucleus (N; arrows). Scale bars, 10 μm (low magnification; original magnification ×5000) and 1 μm (higher magnification; original magnification, ×25,000). All images are representative of the results obtained from n = 3 independent experiments. Full size image Consistent with fluorescence images of iBAs transduced with Them1-EGFP (Fig. 2a ), puncta in electron micrographs of iBAs transduced with AAA-EGFP-APEX2 were found exclusively in the cytoplasm in close proximity to both LD and mitochondria (Fig. 6a–c ). Although the puncta resembled mitochondria from steroid-secreting cells 15 , there was no apparent membrane surrounding the structure. Instead, puncta were an aggregate of small round and elongate droplets with amorphous boundaries (Fig. 6b, c ). Diffusely localized Them1 in iBAs transduced with DDD-EGFP-APEX2 was not clearly associated with any cytoplasmic structure (Fig. 6d ). Because puncta were membraneless structures with characteristics similar to that described for biomolecular condensates within the cell cytoplasm 16 , we evaluated the Them1 primary sequence for characteristics that would facilitate a phase separation, aggregation, and/or the formation of puncta (Fig. 7 ). Bioinformatic analysis of the amino acid sequence for Them1 (Fig. 7a and Supplementary Fig. 10a ) showed a highly disordered region at the N-terminus spanning residues 20-45 as predicted by IUPRED (prediction of intrinsic disorder) curves (Fig. 7b ). There were no prion-like regions in the sequence as determined by PLD (Fig. 7c ) or FOLD curves (Supplementary Fig. 10b ). The N-terminus contains a high proportion of charged residues with a patch of basic residues followed by a patch of acidic residues (Supplementary Fig. 10c, d ), which could engage in non-covalent crosslinking with other proteins or itself to drive a phase transition 17 . However, the PScore, a pi-pi interaction predictor, did not reach significance (Fig. 7d ), which suggests that the sequence of Them1 does not have the propensity to phase separate based on pi–pi interactions. ANCHOR2 analysis, which predicts protein binding regions in disordered proteins, showed that the N-terminal 20 amino acids were a highly disordered binding region (Fig. 7e ). Interestingly, two of the phosphorylation sites, S15 and S18, lie in this region (Fig. 7a ), suggesting that the phosphorylation sites may regulate binding events. Because multivalent interactions between proteins can lead to a phase separation 16 , we also performed a motif scan using the Motif Scan server ( ). This analysis highlighted no strong association of the 1 st 20 amino acids with other proteins. Fig. 7: Bioinformatic analysis of Them1 suggests that a disordered region at the N-terminus spanning residues 20–45 may be involved in puncta formation. a Schematic diagram of the Them1 amino acid sequence aligned with the amino acid number from the N- to C-terminus. Thio indicates the position of thioesterase domains 1 and 2 and START indicates the position of the START domain. b Results from the IUPRED server, which predicts regions of disorder. The dashed line signifies the threshold for significance. The data indicate a large region of disorder in the first 50 amino acids (highlighted in gold). c Results from the Prion-like Amino Acid Composition server. The dashed line at 0.5 is a threshold for significance. The graph, red line, demonstrates no prion-like regions in the Them1 amino acid sequence. d Results from the PScore, which is a predictor of the propensity to phase separate based on pi–pi interactions. The dashed line is a threshold set at 4 to determine significance; the amino acid sequence of Them1 did not reach significance. e Results from an ANCHOR2 analysis, which predicts protein binding regions in disordered proteins. This analysis demonstrates that the N-terminal 20 amino acids of Them1 were strongly predicted to be a disordered binding region (highlighted in red). Two of the phosphorylation sites (red font) lie in this region (expanded region, also refer to Fig. 4b ), suggesting they could potentially regulate a binding event. Full size image Discussion Them1 was first identified as a gene that is markedly upregulated in response to cold ambient temperatures and is suppressed by warm temperatures 9 . Whereas this initially suggested that Them1 function was to promote thermogenesis 9 , homozygous disruption of Them1 in mice led to increased energy expenditure 10 , 11 . This result indicated that Them1 functions, on balance, to suppress energy expenditure. The current study helps to explain how a suppressor of energy expenditure can be induced by cold ambient temperatures, which also upregulates thermogenic genes; Them1 is phosphorylated and inactivated by the same adrenergic stimulus that promotes thermogenesis. Additionally, our studies have demonstrated phosphorylation-dependent intracellular redistribution of Them1 in response to thermogenic stimuli, suggesting a model for the regulation of energy expenditure within BAT. For this, Them1 activity is transiently silenced during peak thermogenesis by relocation away from its active site in puncta, which are located in close proximity to LD and mitochondria. Our prior observations indicate that the activity of Them1 in brown adipocytes is to suppress the mitochondrial ß-oxidation of LD-derived fatty acids 11 . In keeping with this possibility, we observed that Them1 localized to puncta in the absence of adrenergic stimulation both in vitro and in vivo. These structures are juxtaposed with mitochondria and LD such that they could be expected to regulate fatty acid trafficking between these two organelles. Moreover, experiments in iBAs transduced with recombinant adenovirus revealed that full-length Them1 in unstimulated cells and the AAA-Them1 mutant, which is unresponsive to phosphorylation, were localized to puncta and were active in suppressing norepinephrine-stimulated oxygen consumption. These results are similar to our previous findings in cultured primary brown adipocytes, where Them1 actively suppressed norepinephrine-stimulated oxygen consumption 11 , although puncta formation and dissolution were not investigated in that study. By contrast, our data demonstrated that Them1 in its diffuse configuration was ineffective at suppressing oxygen consumption. This result was consistent with its relocation away from LD and mitochondria and suggests that phosphorylation within the N-terminus amino acids 1–36, likely S15 phosphorylation, and to a lesser degree S25 phosphorylation, relocates and inactivates Them1 in vivo. Levels of S15 phosphorylation that were below detectable levels in our study could be due to the nature of peptide cleavage sites for mass spec analysis, glycosylation of nearby sites, or indicate that PMA-induced stimulation is not relevant to S15 phosphorylation, which may instead reflect norepinephrine stimulation. These interesting possibilities must be pursued in future studies. Phosphopeptide analysis identified other potential sites, including S27 and S31, the roles of which in Them1 localization remain unknown but also need to be studied experimentally. Our inhibitor data suggest that PKCβ plays some role in regulating Them1 metabolic function by changing the physical state of Them1. This may be by phosphorylating one or more Them-1-interacting proteins after stimulation or other yet undefined mechanisms. Interestingly, PKCβ appears to be involved in coordinating thermogenesis, as evidenced by increased rates of fatty acid oxidation in both BAT and white adipose tissue in mice globally lacking this kinase 18 . The important role played by Them1-containing puncta in inhibiting the mitochondrial oxidation of LD-derived fatty acids begs the question about the composition of puncta, including how they are formed and why they disappear after phosphorylation. Our data using correlative light and electron microscopy suggest that Them1 constitutively resides in biomolecular condensates, also known as membraneless organelles, much like the nucleoleus, stress granules, and p-bodies 19 , 20 . Biomolecular condensates form by liquid-liquid phase separation or condensation 21 , which favors aggregation. The resulting structure often resembles an aggregation of round liquid droplets 16 , which is the result we found here for Them1 in puncta by electron microscopy. The driving force underlying the formation of biomolecular condensates is the exchange of macromolecular/macromolecular and water/water interactions under conditions that are energetically favorable 19 . What these conditions might be for Them1 are unknown but may be related to temperature, localized pH changes, or co-solutes that are expressed. Typical phase diagrams can define the conditions needed for condensation, which would be required to fully understand the localization of Them1 to puncta in vitro and in vivo. One consistent signature of biomolecular condensates is the concept that their formation includes scaffold and client proteins 19 . Scaffold molecules are drivers of the condensation process and clients are molecules that partition into the condensate. It is not clear into which category Them1 fits, although the characteristics of its molecular structure suggest that it could be the primary scaffold protein. Phase separation of scaffold proteins and the subsequent partitioning of their clients are driven by network interactions for which two distinct classes of protein architecture are thought to be involved 19 . The first is multiple folded (SH3) domains that interact with short linear motifs 19 . Since Them1 does not have an SH3 domain, as determined by using the Simple Modular Architecture Research Tool web server 22 , this could not be the sequence element driving puncta formation. Second is intrinsically disordered regions 19 , for which Them1 has a strong candidate region near the amino terminus; amino acids 20–45 represent a highly disordered area that includes the S25 phosphorylation site and the N-terminal 20 amino acids are strongly predicted to be a disordered binding region that also includes phosphorylation sites at S15 and S18. Because the N-terminal 36 amino acids of Them1 are sufficient to drive puncta formation in iBAs, our data strongly support that this region is involved in puncta formation and should be further investigated. Alternatively, Them1 may function as a client. Clients often have molecular properties similar to scaffolds including intrinsic disorder 20 , also making this role for Them1 a distinct possibility. Dissolution of biomolecular condensates can be used to change the available concentration of cellular proteins to affect reaction efficiency 19 , and is likely important for Them1 to reduce its local concentration at LD and mitochondria subsequently increasing fatty acid availability for β-oxidation in mitochondria and increasing thermogenesis. The phosphorylation state of proteins regulates their ability to assemble in biomolecular condensates, with threonine phosphorylation resulting in condensation and serine phosphorylation reducing retention in a dose-dependent manner 16 . Phosphorylation events were shown to regulate phase separation, aggregation, and the retention of mRNA’s in FUS proteins, which are neuronal inclusions of aggregated RNA-binding protein fused in sarcoma 23 , 24 . It is thus phosphorylation at S15 and S25, which are highly conserved residues in Them1 that reduces its retention in puncta as evidenced in inhibitor studies and by genetic manipulation of Them 1. Norepinephrine-induced protein phosphorylation events during cold exposure in addition to a 3-fold increase in Them1 protein concentration over time 11 , likely creates equilibrium between active Them1 in puncta and inactive Them1 that is diffusely localized in the cytoplasm. Our data thus suggest that the delicate balance of Them1 protein concentration, its phosphorylation status, and environmental factors that promote Them1 phase transitions dictate the rate of thermogenesis in BAT. Together, our results reveal an unexpected mechanism for the regulation of Them1, a protein that reduces fatty acid availability for β-oxidation in mitochondria and limits thermogenesis. Manipulation of Them1 localization in BAT may provide a therapeutic target for the management of metabolic disorders including obesity and non-alcoholic fatty liver disease. Methods Culture of brown adipocytes Immortalized brown preadipocyte cells (iBAs) from mouse brown adipose tissue 25 were obtained from Dr. Bruce Spiegelman. iBAs were cultured in Dulbecco’s Modified Eagle’s medium (DMEM; Gibco, Grand Island, NY), 20% fetal bovine serum (FBS; Gibco), 1% penicillin/streptomycin solution (P/S; Gibco), and 2 μM HEPES (Gibco). Confluent iBAs were induced to differentiate from preadipocytes by adding induction cocktail containing 5 μM dexamethasone (Sigma-Aldrich, St. Louis, MO), 125 μM indomethacin (Sigma-Aldrich), 0.02 μM insulin (Sigma-Aldrich), 500 μM isobutylmethylxanthine (Life Technologies, Carlsbad, CA), 1 nM triiodothyronine (Sigma-Aldrich), and 1 μM rosiglitazone (Sigma-Aldrich) in medium consisting of DMEM, 10% FBS, and 1% P/S. After 48 h of induction, the medium was replaced by maintenance medium consisting of DMEM, 10% FBS, 1% P/S with 0.02 μM insulin, 1 nM triiodothyronine, and 1 μM rosiglitazone. The maintenance medium was replaced every 2 days. For immunostaining and confocal applications, iBAs were plated on glass coverslips, which were pre-treated with 0.01% collagen type I (Sigma-Aldrich) in 0.01 M acetic acid. Production of DNA constructs The full-length cDNA for Them1 (from Mus musculus : Q8VHQ9) was cloned and linked to EGFP plus an engineered ascorbate peroxidase (APEX2) fusion protein the C-terminus. Codons for S15, S18, and S25 were mutated to alanine (A) or aspartic acid (D) using the QuikChange mutagenesis kit (Agilent, Santa Clara, CA). The Δ1-36 mutation includes deletion of the first 36 amino acids at the N-terminus of Them1, whereby Them 1 begins at amino acid 37 (M; AUG). The full-length cDNA for Them1 was cloned into a pVQAd CMV K-NpA vector (ViraQuest, North Liberty, IA) along with C-terminal EGFP downstream of the CMV promoter. The Them1 adenovirus was created by ViraQuest. Heterologous Them1 expression in cultured iBAs Plasmid transfection with Lipofectamine 3000 (Thermo Fisher Scientific, Waltham, MA) was conducted 48 h after iBAs were induced to differentiate. Purified DNA and P3000 reagent were mixed with Opti-MEM. The mixture was then added into Opti-MEM containing Lipofectamine 3000. The three reagents were allowed to incubate at room temperature for 10 min and then added to the cells. After 7 h of transfection, the medium was replaced by maintenance media. After 48 h, the cells were assayed. The same ratio of cells to DNA was used for each experiment but due to the very low transfection efficiency, it was not possible to assay Them1 concentration by immunoblotting. For experiments using adenoviral vectors to express Them1, recombinant adenoviruses including Ad-Them1-EGFP, Ad-AAA-EGFP, Ad-DDD-EGFP, or Ad-EGFP were added to the medium for 24 h after day 1 of differentiation. Adenoviral vectors were initially screened for MOI versus cell protein expression using our Them1 antibody 11 , mRNA expression using quantitative RT-PCR (Supplementary Fig. 3b, c ), and cell viability using the crystal violet assay 26 , which was nearly 100% with MOI’s ranging from 10 through 80 (not shown). Unless otherwise specified, an MOI of 40 was used because Them1 expression was most similar to that induced in BAT during cold exposure (Supplementary Fig. 3b ). LC-MS/MS for detecting Them1 phosphorylation Protein concentrations were estimated with a Bradford protein assay prior to acetone precipitation, done as described 27 . Pellets were resuspended in 8 M urea (Proteomics grade, PlusOne; GE Healthcare, Chicago IL), 5 mM TCEP (Bond-breaker; Thermo Fisher Scientific, Waltham, MA), 50 mM triethyl ammonium bicarbonate (TEAB, Sigma-Aldrich) and then alkylated with 100 mM iodoacetamide to a final concentration of 9 mM. Samples were then diluted with 100 mM TEAB (Sigma-Aldrich) followed by incubation with trypsin (Sigma-Aldrich) at a ratio 1 : 40 trypsin:total protein overnight at 37 °C. Samples were acidified to a final concentration 0.1% TFA (Sigma-Aldrich) and then desalted using HLB solid phase extraction plates (Waters, Milford, MA). Eluted peptides were dried in a centrifugal vacuum concentrator and then stored at −20 °C. Phosphorylated peptides were enriched with Fe-NTA magnetic beads (Cube Biotech, Cambridge, MA) using a 96-well plate magnet 28 . Beads were washed and the phosphorylated peptides eluted with 50% acetonitrile and 1% ammonium hydroxide. Eluted peptides were then acidified with formic acid and dried. Before LC-MS/MS analysis, samples were resuspended to yield a 0.3 μg/μL solution, assuming a 1 : 100 reduction in peptide amount. LC-MS/MS analysis was conducted with a 1200-bar nano HPLC (Thermo Easy 1200 nLC; Thermo Fisher Scientific) using a 15 cm × 75 μm ID, 1.6 μm C18 column (Ionoptickcs, Aurora series emitter; Fitzroy VIC, Australia). The temperature of the column was maintained with an integrated column oven (PRSO-V1; Sonation lab solutions, Biberach, Germany). Enriched phosphorylated peptides were chromatographed using a mixed linear gradient of buffer A containing 0.1% formic acid in water and buffer B containing 0.1% formic acid in 80% acetonitrile. Flow rates were maintained over the course of the HPLC run and the gradient applied was (Time:%B): 0:3, 1:7, 81:30, 91:40, 93:80, 96:80, 97:3. The LC system was interfaced to an Eclipse Tribrid Orbitrap mass spectrometer (Thermo Fisher Scientific). The positive spray voltage was set to 2–2.5 kV, ion transfer tube temp 305 °C, RF funnel 30%. The FAIMS Pro system was set to survey three compensation voltages of −40, −60, and − 80 V. The full profile MS spectra were collected at 120k resolution (200 m/z). The data-dependent acquisition used a monoisotopic peak determination filter set to peptides, intensity set to 5e3, charge states 2–6, dynamic exclusion of 60 s and a cycle time of 0.6 s for each FAIMS CV used, total cycle time 1.8 s. The MS/MS fragmentation was collected with high energy collision induced dissociation with the following settings; quadrupole isolation with 1.6 Da window, normalized collision energy set to 30%, detector set to ion trap with rapid scan rate and defined mass range of 200–1500, AGC 250% with a maximum injection time set to 35 ms data stored as centroid data. A pooled sample of enriched peptides from all of the samples was crudely fractionated with centrifugal columns following manufacturer’s instructions (Pierce High pH reversed-phase peptide fractionation kit (Thermo Fisher Scientific). This generated 8 fractions, a flow through, and wash fraction all were analyzed using the same method as the samples. The results were used to generate an in-depth library of phosphorylated peptides. Data were processed using Proteome Discoverer v2.4 (Thermo Fisher Scientific). Data-dependent data were searched with Sequest module using the following parameters; MS1 mass accuracy 10 ppm with MS/MS tolerance 0.6 Da, searched against the Mus musculus protein database Uniprot taxonomy ID 10090 v2017-10-25, trypsin set as the digestion enzyme with maximum of three missed cleavages. Dynamic modification included oxidation of methionine, phosphorylation of Ser and Thr, N-terminal pyro-glutamate formation, N-terminal Met loss and Met loss with acetylation, static modification was set to cabamidomethyl of Cys. Validation was done using the percolator processing node with FDR set to 0.01. Them1 (Q8VHQ9) phosphorylated peptides identified via Sequest were imported to Skyline for chromatographic extraction and integration. The signal intensity of Them1 peptides was normalized to the phosphorylated peptide RSS(phospho)QGVLHMPLYTSPIVK from isoform 2 of hormone-sensitive lipase (P54310-2). This peptide was chosen as a housekeeping phosphorylated peptide. The total amount of Them1 phosphorylation was normalized to this peptide. Fixation and Oil Red O staining for LD Differentiated iBAs were washed with phosphate buffered saline (PBS) and then fixed with 4% paraformaldehyde in PBS for 20 min at room temperature. Cells were washed twice before staining with 0.3% Oil Red O solution in 60% isopropanol (Sigma Aldrich) at room temperature. The cells were washed with PBS and then mounted with ProLong Diamond Antifade Mountant with DAPI (ThermoFisher) to identify nuclei. Immunostaining for LD, mitochondria, and plasmid-expressed Them1 Differentiated iBAs were fixed, as above, and then stained for perilipin (anti-Plin1) to identify LD, Tomm20 (anti-Tomm20) to detect mitochondria, or anti-EGFP to visualize the plasmid-transfected Them1-EGFP. For staining Them1 in paraffin sections of BAT, tissues were treated with Image-iT™ FX Signal Enhancer (Thermo Fisher Scientific, Hampton, NH) after antigen retrieval with citrate buffer, pH 6.0. The sections were stained with anti-Them1 antibody, anti-Plin1, and Hoechst 33342 to identify nuclei. Details of the antibodies: affinity purified rabbit anti-Them1 11 1:250 for immunostaining or 1:1000 for immunoblots (see Supplementary Fig. 5 ); anti-Tomm20, Santa Cruz Biotech (Dallas TX, USA), rabbit, sc-11415, 1:200; anti-Plin1, Fitzgerald (Acton MA, USA), guinea pig, 20R-PP004,1:500; anti-EGFP, Novus Biologicals (Littleton CO, USA), goat, NB100-1678, 1:500; anti-actin, Abcam (Cambridge, MA, USA), mouse, Ab3280,1:1000. Secondary antibodies labeled with HRP, AlexaFluor 488, 647, or Cy3 were purchased from Jackson ImmunoResearch (West Grove PA, USA). Confocal microscopy The localization of fluorescence signal in cultured iBAs was evaluated using a Zeiss LSM880 confocal microscope system with or without Airyscan. Images were acquired at high resolution as 2 µm z-stack slices through the thickness of cells and assembled in 3-D using Volocity image processing software. For confocal imaging of living cells, a time-lapse image was created to capture the dynamics of protein expression and localization, signal transduction, and enzyme activity. Cells grown on coverslips were placed in the incubator chamber of a Zeiss LSM880 inverted confocal microscope system to visualize EGFP expression (linked to Them1). The interior of the chamber was kept at a constant 37 °C, and set to a 5% CO 2 level and a 95% O 2 level. An image was captured at T0, and immediately following the application of a treatment. The microscope then automatically tracked, focused, and photographed cells every 30 min for 4–5 h. These images are also taken in the high resolution z-stack format and analyzed using Volocity software. Treatment of iBAs to explore metabolic pathways To explore the metabolic pathway leading to reorganization of intracellular Them1, various treatments were applied for 4 h to cultured iBAs as follows: Norepinephrine (1 μM; Sigma Aldrich; St. Louis, MO), a neurotransmitter, was used to mimic cold exposure. Forskolin (1 μM; Selleckchem, Houston, TX), a membrane permeable labdane diterpene produced from the Coleus plant, was used to activate adenylyl cyclase. This treatment, through the activation of cAMP, activates PKA. PMA (3 μM; LC Laboratories, Woburn, MA) was used to activate PKC. NE, adenylyl cyclase, and PKC were selected for activation as they represent three key control points in the pathway allowing the characterization of all upstream regulators. Pathway inhibitors were added 1 h prior to activation and were continued throughout the 4 h activation period as follows: PKI [14–22] myristoylated (0.5 μM; Invitrogen, Camarillo, CA), a synthetic peptide inhibitor of PKA, was added to inhibit PKA activation. Atglistatin (40 μM; Selleckchem, Houston, TX) was used to selectively inhibit the processing of triacylglycerol to fatty acids via the formation of diacylglycerol by ATGL. Ruboxistaurin, or LY333531, (2 μM; Selleckchem, Houston, TX), an isozyme-selective inhibitor of PKC that competitively and reversibly inhibits PKC-βI and PKC-βII, was added to inhibit PKCβ activity. Correlative light and electron microscopy of Them1 in iBAs Differentiated iBAs transfected with a plasmid containing Them1 linked to APEX2 were fixed and the APEX2 developed using an ImPACT peroxidase substrate kit (Vector Labs, Burlingame, CA). Cells were then frozen using a Wohlwend Compact 02 High Pressure Freezer. Frozen cells underwent super-quick freeze substitution and embedding using the protocol established by McDonald 29 . In 0.5-μm thick sections, positive cells stained brown by light microscopy. When positive cells were identified, four consecutive levels of serial ultra-thin sections were obtained with one thick section for light microscopy between levels. Light microscopy images, taken with an Axioimager (Zeiss) equipped with a color ccd camera were used to follow puncta in the sections and to determine the position of positive cells in ultra-thin sections. Images were taken on a JEOL 1400 transmission electron microscope (TEM) equipped with a Gatan (Pleasanton, CA) Orius SC1000 camera. Low magnification TEM images were taken at an original magnification of ×4000–5000, intermediate magnification TEM images at ×25000, and high magnification TEM images at ×60,000. O 2 consumption rates in iBAs Oxygen consumption rates (OCR) were measured in iBAs using a Seahorse XFe96 (Agilent Technologies, Santa Clara, CA, USA). Briefly, iBAs were seeded at a density of 1000 cells/well in collagen-coated Seahorse XF96 cell culture microplates 2 days before induction. On day 1 after induction, iBAs were incubated with Ad-Them1-EGFP (MOI 40), Ad-AAA-EGFP (MOI 50), Ad-DDD-EGFP (MOI 60), or Ad-EGFP (MOI 40). In preliminary experiments these MOIs led to similar Them1 and EGFP protein expression. One day after adenovirus infection, the medium was changed to the maintenance medium supplemented with 1 μM NE. After 2 days, CO 2 was withdrawn from the cultures for 1 h at 37 °C in Krebs-Henseleit buffer (pH 7.4) containing 2.5 mM glucose, 111 mM NaCl, 4.7 mM KCl, 2 mM MgSO 4 -7H 2 O, 1.2 mM Na 2 HPO 4 , 5 mM HEPES, and 0.5 mM carnitine (denoted as KHB). Basal OCR was measured in KHB for 18 min and then 10 μM NE was injected through the Seahorse injection ports to a final concentration of 1 μM in each well. NE responses were measured after injections as previously described 11 . Relative transfection efficiency was assessed by counting the number of EGFP expressing cells relative to the number of total viable cells. Data were normalized to the number of viable cells, determined using the NucRed™ Live 647 ReadyProbes™ assay (Thermo Fisher Scientific, Waltham, MA, USA) and the color read using a Spectramax i3x (Molecular Devices, San Jose, CA, USA) visualized and calculated using the Spectramax i3x. Fatty acid oxidation Rates of oxidation of triglyceride-derived fatty acids were determined by pulse-chase methods as described by Cooper et al. 30 . In brief, iBAs were cultured, differentiated, and transfected in 12-well plates as described above. On day 2 prior to pulse, the media was changed to pre-labeling medium including DMEM, 10% FBS, 1 g/l glucose, and 10 mM HEPES. Cells were pulsed for 1.5 h with pre-labeling media containing 1 mM carnitine and 500 µM sodium oleate (Sigma-Aldrich) containing 2 µCi/ml [9,10- 3 H]oleate (Perkin Elmer, Waltham, MA) conjugated to BSA 31 . Cells were chased by washing with PBS containing 1% fatty acid-free BSA and then incubated for 1.5 h with pre-labeling media containing 1 mM carnitine, 500 µM oleate, and 5 µM NE. At the end of the chase period, cells were washed with PBS containing 1% fatty acid-free BSA prior to analysis. Relative transfection efficiency was determined by immunoblot analysis of Them1 and EGFP. Lipids were extracted using the Folch method and then separated by thin-layer chromatography using hexane:ethyl-ether:acetic acid (80:20:1, v/v/v) to determine radiolabeled triglycerides 31 . The concentrations of total triglycerides were determined enzymatically (Wako Diagnostics, Mountain View, CA) 32 . Rates of fatty acid oxidation were determined from utilization of triglycerides during the chase period and expressed relative to Ad-EGFP. Activation of BAT in mice Male C57BL/6J mice (7 w) were purchased from the Jackson Laboratory (Bar Harbor, ME) and housed in a specific pathogen-free AAALAC International Facility with a standard 12 h alternate light/dark cycle at an ambient temperature of 22 ± 2 °C and 30–70% humidity, at Weill Cornell Medical College. Animal use and euthanasia protocols were approved by the Institutional Animal Care and Use Committee at Weill Cornell Medical College. For the experiment, mice were moved to Promethion metabolic cages (Sable Systems International, Las Vegas, NV) with one mouse/cage. Mice were housed at 22 °C for the first 24 h, but were changed to 30 °C overnight prior to experiments. Room temperature, or 22 °C, is below the thermoneutral zone for mice and the mild cold exposure stimulates Them1 expression in BAT 11 . Experimental mice were injected subcutaneously with one dose of CL316,243 (Sigma-Aldrich, 1 mg/kg body weight) in saline, which is a β3 adrenergic receptor agonist. Control mice were injected with saline alone. Subscapular BAT was excised from mice after euthanasia at T0, 1, 2, or 4 h after injection. Tissues were fixed overnight in 10% neutral buffered formalin and then processed. Paraffin sections were used for H&E stained samples and for immunocytochemistry as described above. Image processing and quantification of fluorescence signal For experiments aimed to quantify changes in intercellular Them1 localization after stimulation, image segmentation was used to calculate the volume of Them1 by thresholding the green (EGFP) pixels of specified intensity that represented Them1-EGFP minus background intensity using either Volocity software (Quorum Technologies, Ontario, Canada) or Imaris software (Oxford Instruments, Concord, MA, USA). The strategy used for both iBAs and for tissue Them1 expression was that puncta represent concentrated Them1, which occupies a small intracellular volume whereas diffuse Them1 is distributed evenly throughout the entire cell volume resulting in a large intracellular volume (Fig. 1k , schematic diagram). Them1 fluorescence signal, which represented pixel intensity in the EGFP (green) channel, was measured by using a normalized histogram at each 12-bit (0–4096) intensity value. For iBAs, background signal was determined from cells that were not transfected and the fluorescence threshold set to exclude this low-intensity background signal. For paraffin sections of brown adipose tissue, the background fluorescence intensity was determined using Them1-deficient mice and green pixels in this range were excluded from the experimental data. This strategy did not stratify the signal into intense (as found in puncta) or weak (as found in the diffuse state), but instead identified the volume of total intracellular Them1 fluorescence irrespective of intensity. Bioinformatics The amino acid sequence of Them1 from Mus musculus was acquired from UniProt 33 and analyzed with multiple servers. The IUPRED2A server was used to predict intrinsic disorder and ANCHOR binding regions 34 , 35 , 36 , 37 . The Prion-like Amino Acid Composition (PLAAC) server was used to predict prion-like regions through the prion-like domains (PLD) and FOLD curves 38 . The propensity of Them1 to phase separate through long-range planar pi-pi contacts was predicted using a server generated by the laboratory of Professor Julie Forman-Kay (PScore) 39 . The Motif Scan server ( ) was utilized to examine the first 20 amino acids for known motifs 40 . Statistics and reproducibility Data were analyzed using SigmaPlot software with a P value < 0.05 considered significant. Prior to analysis, data was subjected to an outlier test (Grubbs test; GraphPad on-line calculator ) and individual data points were excluded that were extreme outliers, alpha value of 0.01 or less. For comparison of data from many groups, one-way analysis of variance was used to determine statistical significance. If variances were not normal, non-parametric ANOVA was performed using ranks. For experiments with multiple times and treatments, 2-way analysis of variance was used to determine statistical significance. Post-hoc analyses were done to determine differences between groups. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The authors declare that the main data supporting the findings of this study are available within the article and its Supplementary Information files. Adenoviral vectors or plasmid constructs are available upon reasonable request to the corresponding authors. The raw data files for LSMS were deposited at the Center for Computational Mass Spectrometry, Center for Computer Science and Engineering, University of California, San Diego. The dataset files can be accessed under the dataset ID MSV00008675 [ ]. Source data are provided with this paper.

Linked to serious health problems including cancer, diabetes and cardiovascular disease, obesity affects more than a third of adults in the United States. Presently, there are few safe and effective nonsurgical therapeutic interventions available to patients with obesity. Now, a multi-disciplinary team of researchers has demonstrated that a metabolic regulatory molecule called Them1 prevents fat burning in cells by blocking access to their fuel source. Led by microscopy experts at Beth Israel Deaconess Medical Center (BIDMC) and metabolism experts at Weill Cornell Medicine and NewYork-Presbyterian, the study may contribute to the development of a new type of obesity treatment. The team's findings were published June 9 in Nature Communications. To help explain how the protein Them1 turns off heat production, BIDMC's cell biology and microscopy expert, Susan Hagen, Ph.D., associate vice-chair for research in the Department of Surgery at BIDMC, and Yue Li, Ph.D., a postdoctoral researcher in her laboratory, used light and electron microscopy to observe Them1 in action in mouse brown fat cells grown in the laboratory. "Them1 is an interesting molecule," said Hagen. "If you inhibit or block its expression, metabolism increases and that reduces body weight." The experiments showed that when the cells are stimulated to burn fat, a chemical modification causes Them1 molecules to spread out, or diffuse, throughout the cell. This frees the cellular powerhouses called mitochondria to efficiently turn the cell's fat stores into energy. But when the stimulation stops, Them1 molecules quickly reorganize into a structure called a biomolecular condensate. Situated between the mitochondria and the fats they use as fuel, the condensed Them1 molecules limit energy production. "It turned out to be so incredibly interesting," said Hagen, who is also director of Microscopy and Histology Core Facilities at BIDMC and associate professor of surgery at Harvard Medical School. "We asked other microscopy experts whether they had ever seen anything like the unusual images we found in resting cells. Using very sophisticated electron microscopy techniques, we were able to show—for the first time, as far as we know—what the bimolecular condensate looks like in electron microscopy." "The study explains a new mechanism that regulates metabolism," said David Cohen, chief of the Division of Gastroenterology and Hepatology at Weill Cornell Medicine and NewYork-Presbyterian/Weill Cornell Medical Center and the Vincent Astor Distinguished Professor of Medicine at Weill Cornell Medicine. "Them1 hacks the energy pipeline and cuts off the fuel supply to the energy-burning mitochondria. Humans also have brown fat and produce more Them1 in cold conditions, so the findings may have exciting implications for the treatment of obesity." Cohen and Hagen, both members of the Harvard Digestive Diseases Center, have been collaborators since 1983. The current study—supported in part by a five-year, multi-PI grant from the National Institutes of Health—also included collaborators with expertise in structural biology from Emory University. "This was the most fun I have ever had in science in my life," Hagen added. "Including multiple primary investigators with different expertise gives you the power of doing things that you could never do on your own."

10.1038/s41467-021-23595-x

Physics

Transparent iron? For the first time, an experiment shows that atomic nuclei can become transparent

DOI: 10.1038/nature10741

http://dx.doi.org/10.1038/nature10741

https://phys.org/news/2012-02-transparent-iron-atomic-nuclei.html

Abstract The manipulation of light–matter interactions by quantum control of atomic levels has had a profound impact on optical sciences. Such manipulation has many applications, including nonlinear optics at the few-photon level 1 , 2 , 3 , slow light 4 , 5 , lasing without inversion 6 , 7 , 8 and optical quantum information processing 9 , 10 . The critical underlying technique is electromagnetically induced transparency, in which quantum interference between transitions in multilevel atoms 11 , 12 , 13 , 14 , 15 renders an opaque medium transparent near an atomic resonance. With the advent of high-brilliance, accelerator-driven light sources such as storage rings or X-ray lasers, it has become attractive to extend the techniques of optical quantum control to the X-ray regime 16 , 17 . Here we demonstrate electromagnetically induced transparency in the regime of hard X-rays, using the 14.4-kiloelectronvolt nuclear resonance of the Mössbauer isotope iron-57 (a two-level system). We exploit cooperative emission from ensembles of the nuclei, which are embedded in a low-finesse cavity and excited by synchrotron radiation. The spatial modulation of the photonic density of states in a cavity mode leads to the coexistence of superradiant and subradiant states of nuclei, respectively located at an antinode and a node of the cavity field. This scheme causes the nuclei to behave as effective three-level systems, with two degenerate levels in the excited state (one of which can be considered metastable). The radiative coupling of the nuclear ensembles by the cavity field establishes the atomic coherence necessary for the cancellation of resonant absorption. Because this technique does not require atomic systems with a metastable level, electromagnetically induced transparency and its applications can be transferred to the regime of nuclear resonances, establishing the field of nuclear quantum optics. Main The basic requirement to observe electromagnetically induced transparency (EIT) is a three-level system represented by the ground state, |1〉, and two upper states, |2〉 and |3〉, with respective energies E 2 and E 3 > E 2 , where a strong laser field with Rabi frequency Ω C induces an atomic coherence between states |2〉 and |3〉. This leads to a Fano-type quantum interference 18 if a (weak) probe laser field is tuned across the resonant transition |1〉 → |3〉, rendering the medium almost transparent in a narrow window around the exact resonance frequency. The degree of transparency is limited by the dephasing of the atomic coherence caused by the decay of state |2〉. Thus, maximum transparency is observed if |2〉 can be considered metastable, that is, if it has a decay width, γ 2 , that is negligibly small relative to the coherent decay width, γ 3 , of the state |3〉. Quantitatively, the spectral response of an ensemble of N atoms under these conditions can be described in terms of the linear susceptibility, χ : Here Δ is the detuning of the probe field from the exact resonance and g is the atom–field coupling constant 10 . The susceptibility approaches zero at exact resonance ( Δ = 0) as γ 2 → 0. This is the phenomenon of EIT. Here we extend the concept of EIT into the regime of hard X-rays by using the Mössbauer isotope 57 Fe, which is a two-level system (neglecting the nuclear hyperfine interaction) with a transition energy of 14.4 keV and a natural linewidth of Γ 0 = 4.7 neV. It is not immediately clear how to achieve EIT without a proper nuclear three-level system. In fact, nuclear three-level systems with a metastable level together with a properly synchronized two-colour X-ray/X-ray or X-ray/light source are not available for use to establish conventional EIT schemes in the nuclear regime. Therefore, the possibility of using a single field of hard X-rays for EIT with a nuclear two-level system is highly desirable. This would open the way to exploring quantum optical concepts and nonlinear effects at very short wavelengths, which is particularly appealing in view of the X-ray laser sources in development. The key to the realization of nuclear EIT described here is cooperative emission from ensembles of Mössbauer nuclei that are properly placed in a planar cavity for hard X-rays. The physics of cooperative emission from atoms in cavities has many interesting phenomena even in the linear regime where the atom–cavity interaction can be treated in the weak-coupling limit, which is typically the case at X-ray wavelengths. Owing to the high resonant cross-section of 57 Fe, its 14.4-keV transition is a two-level system well suited to such studies. This isotope was recently used to study superradiant emission and the collective Lamb shift for a single ensemble of atoms located at an antinode of the field within a planar cavity 19 , 20 . Figure 1a shows the energy spectrum of the reflectivity of such a cavity excited in its third-order mode at a grazing angle of ϕ = 3.5 mrad. The spectrum, calculated using a transfer matrix algorithm for resonant X-ray scattering from layered media 21 , shows the superradiant enhancement of the decay width, Γ N , together with the collective Lamb shift, L N . Figure 1: Calculated reflectivity spectra of different cavity configurations. Sample geometry (top row) of planar cavities for X-rays, containing 2-nm-thick layers of 57 Fe nuclei (dark grey), and spectrally resolved reflectivity (bottom row) around the 14.4-keV nuclear resonance energy. The cavities are excited in the third-order mode under grazing angle ϕ = 3.5 mrad. The graphs in the top row show the standing wave intensity of the electromagnetic field in the cavities. a , For a single layer in the centre of the cavity, the collective decay width, Γ N , is broadened owing to superradiant enhancement and shows a shift, the collective Lamb shift L N (refs 19 , 20 ). b , If the cavity contains two 57 Fe layers placed respectively at a node and an antinode of the standing wave field, we observe a pronounced dip in the spectral response, indicative of an EIT window. c , The transparency window vanishes if the 57 Fe layers are arranged in the sequence antinode–node as viewed from the top of the planar cavity. d , Subtracting the energy spectrum in b from that in a reveals an asymmetric line profile resembling that of a Fano resonance. PowerPoint slide Full size image A qualitatively new situation is encountered when two resonant 57 Fe layers instead of one are placed in a cavity. A pronounced dip in the spectral response appears when one of the 57 Fe layers is placed at a node of the standing wave in the cavity and the other is placed at an antinode ( Fig. 1b ). This dip is very reminiscent of the transparency dip observed in EIT. The appearance of this feature sensitively depends on the separation and the location of the two resonant layers within the cavity. For example, EIT completely vanishes if the two layers are arranged in the sequence antinode–node instead of node–antinode as seen from the top surface of the cavity ( Fig. 1c ). To investigate this effect quantitatively, we performed a perturbation expansion of the cavity reflectivity in powers of the grazing-incidence nuclear resonant scattering amplitude, f N ( ϕ ), for a single Lorentzian resonance line, f N = f 0 ( ϕ )/( x − i ), where f 0 is the scattering amplitude at resonance ( Supplementary Information , section C), x = Δ / γ 0 and γ 0 = Γ 0 /2. We obtain the following expression for the reflected amplitude as function of energy detuning, Δ (details of the derivation are given in Supplementary Information , sections A and C): Here the quantities E 2−+ , E 2−− and E 1+− are elements of the transfer matrices that describe the propagation of the photon fields in the transmitted (+) and reflected (−) directions in the unperturbed cavity, and d 1 and d 2 are the respective thicknesses of the two resonant layers. Equation (2) is basically identical to equation (1) for the complex susceptibility in EIT with γ 2 = γ 0 , γ 3 = γ 0 (1 + d 2 f 0 E 2−− ) and . This result admits the following interpretation: the two atomic ensembles, one at the node of the standing wave field and one at the antinode, experience two significantly different photonic densities of states, leading to two different collective decay rates, γ 2 and γ 3 . This effectively converts the nuclei in the cavity into three-level systems with two degenerate upper levels represented by the states |2〉 and |3〉 in the level scheme sketched in Fig. 2 . The term , where , takes the role of the Rabi frequency, Ω C , of the EIT control field. Here it scales with the two transfer matrix elements E 2−+ and E 1+− , which are proportional to the two counter-propagating fields in the cavity at the position of the two resonant layers. That is, here the control field with Rabi frequency Ω C arises from the radiative coupling of the two resonant layers ( Fig. 3 ). The two excited states, |2〉 and |3〉, are coupled through their common ground state, |1〉, by the vacuum field of the cavity, which effectively establishes a control field between the two upper states. This scheme bears some resemblance to the recently reported effect of vacuum-induced transparency, whereby the probe field generates its own control field 22 . Because γ 3 γ 2 , the probe field dominantly couples the ground state, |1〉, to the excited state |3〉. The resulting arrangement of levels and their coupling in Fig. 2 closely resembles a Λ-type level scheme. Figure 2: Effective EIT level scheme of the nuclei in the cavity. Two ensembles of resonant 57 Fe atoms are respectively located at a node and an antinode of a low-finesse cavity. Owing to the significantly different photonic density of states of the atoms at the antinode of the cavity relative to that of those at the node, the radiative widths of the two ensembles differ considerably ( ), such that |2〉 can effectively be considered a metastable state. The cavity field thus causes each nucleus to act as a three-level system. The two upper states are coupled through their common ground state by the vacuum field of the cavity, effectively establishing the control field between them. We note that the cavity width ( ∼ 100 eV) is much larger than the decay width of the nuclei ( γ 3 ≈ 250 neV). PowerPoint slide Full size image Figure 3: Origin of the coherent control field in the cavity. The Rabi frequency of the control field is given by Ω C = idf 0 γ 0 ( L 2++ + L 2−+ )( L 1−− + L 1+− ) with ( Supplementary Information , equation (27)). The graphical representation of the scattering amplitudes, and , for the fields propagating in the transmitted (+) and reflected (−) directions at the positions of the resonant layers supports the interpretation that Ω C arises from the radiative coupling between the two resonant layers in the cavity. PowerPoint slide Full size image Cooperative emission is critical to EIT in this system. Whereas one of the atomic ensembles undergoes single-photon superradiant enhancement leading to a decay width of Γ N = 2 γ 3 = d 2 f 0 Re[ E 2−− ] Γ 0 and a collective Lamb shift of L N = − d 2 f 0 Im[ E 2−− ] Γ 0 /2, the decay width, 2 γ 2 , of the other ‘subradiant’ ensemble is given by just the natural linewidth, Γ 0 , such that γ 3 ≈ 50 γ 2 in the example shown in Fig. 1b . Thus, in the presence of a strong superradiant enhancement of state |3〉, state |2〉 is relatively long lived and therefore can be considered metastable. This is an important condition if a pronounced EIT effect is to be observed. If γ 2 = γ 3 , the response of the system given by equation (1) is merely a sum of two Lorentzians without any destructive interference between them. In an earlier investigation of nuclear resonant EIT, a reduction in absorption of a few per cent was observed on the basis of nuclear level anticrossing in a FeCO 3 crystal 23 , 24 , 25 with γ 3 ≈ 2 γ 2 . The magnitude of Ω C relative to the decay rates γ 2 and γ 3 determines the emergence of Fano interference as the basic signature of EIT. When the control field is resonantly applied to the |2〉 → |3〉 transition, the excited state splits into the dressed states , which are separated by an energy Ω C . If this splitting is smaller than the linewidth γ 3 , Fano-type interference 18 occurs between the two indistinguishable quantum mechanical paths between the dressed states and the ground state. Conversely, if Ω C is much larger than γ 3 then Fano interference is negligible and two well-separated Lorentzians are obtained in the spectral response, equivalent to an Autler–Townes doublet 26 in atomic absorption profiles for strong driving fields. Thus, for EIT, Ω C must not exceed γ 3 . However, Ω C has to be large enough to overcome decoherence effects, such as non-radiative decay, that result in a non-zero value of γ 2 . Evaluation of equation (1) at resonance ( Δ = 0) reveals that EIT is still observable even for non-zero values of γ 2 if | Ω C | 2 is greater than γ 2 γ 3 . Overall, we find that the criterion must be obeyed to obtain a pronounced EIT effect. The right-hand side of this inequality is to a very good approximation equivalent to df 0 | E 1+− E 2−+ / E 2−− | > 1. A numerical investigation reveals that | E 1+− E 2−+ / E 2−− | is of order one. With f 0 ( ϕ ) = 2.3 nm −1 for metallic 57 Fe and ϕ = 3.5 mrad, this condition is fulfilled for layer thicknesses d ≥ 2 nm, as used in the calculations the results of which are shown in Fig. 1 . In the example in Fig. 1b , we find that , where is Planck’s constant divided by 2π, such that the left-hand side of the EIT criterion above is also satisfied. To analyse the spectral shape of the transparency window, we subtract the spectrum in Fig. 1b from that of a single layer at an antinode ( Fig. 1a ). We obtain a spectral shape ( Fig. 1d ) that is indicative of a Fano resonance profile. If levels |2〉 and |3〉 were energetically degenerate, the transparency window would be of Lorentzian shape 27 . In our system, however, the degeneracy is lifted by the collective Lamb shift, L N γ 3 , leading to the asymmetry of the transparency window. As emphasized above and illustrated in Fig. 1 , the EIT effect is not symmetric with respect to the sequence of the resonant layers at the nodes and antinodes of the cavity field. To understand this, we use the definitions of E 1+− and E 2−+ ( Supplementary Information , section A) to write Ω C = df 0 Γ 0 ( L 1−− + L 1+− )( L 2++ + L 2−+ ). The quantities L nμν ( μ , ν ∈ {+, −}) are the amplitudes for the scattering of fields from the incoming direction ( ν ) into the outgoing direction ( μ ) after interaction with the atomic ensembles located at the positions z n ( n ∈ {1, 2}) in the cavity. The graphical representation of these amplitudes ( Fig. 3 ) supports the interpretation that Ω C is determined by the radiative coupling between the two ensembles in the cavity. It can be shown ( Supplementary Information , section A) that L n ++ + L n− + vanishes at the nodes of the cavity field and that L n −− + L n +− does not. Therefore, if the second layer is located at a node of the field, Ω C = 0 and there is no EIT effect. This means that the effective magnitude of Ω C can be controlled not only by the thicknesses of the resonant layers but also by their placement within the wave field in the cavity. To verify EIT experimentally, we prepared two planar X-ray cavities consisting of Pt(3 nm)/C(38 nm)/Pt(10 nm) sandwich structures, each containing two 57 Fe layers occupying a node and, respectively, an antinode of the cavity field in the sequences shown in Fig. 1b, c . To achieve a maximum EIT effect without perturbing the cavity field too much, we chose each 57 Fe layer to be 3 nm thick. At this thickness, the 57 Fe layers order ferromagnetically with the magnetization confined to the plane of the films. The magnetic hyperfine interaction lifts the degeneracy of the nuclear magnetic sublevels, leading to four allowed dipole transitions for the given scattering geometry, where the magnetization is aligned parallel to the wavevector of the incident photons, k 0 . Despite the magnetic level splitting, the basic physics of the EIT effect as discussed above remains unaffected because the separation of the lines is large enough to consider each of the four resonance lines separately. The experiments were performed at the PETRA III synchrotron radiation source (DESY, Hamburg) using the method of nuclear resonant scattering ( Supplementary Information , section D). This technique relies on the pulsed broadband excitation of nuclear levels followed by the time-resolved detection of coherently scattered, delayed photons that are emitted on a timescale of after resonant excitation 28 . To determine the energy spectrum of the cavity reflectivity from the time-resolved data, we used a technique based on stroboscopic detection of the delayed response of the sample after passing through a resonant energy analyser 29 ( Supplementary Information , section E). As an analyser, we used a 1-μm-thick stainless steel foil with a 95% isotopic enrichment in 57 Fe, providing a single-line transmission function with a spectral width of ∼ 10 neV. The foil was mounted on a Doppler drive that provided a periodic energy detuning, Δ , as a function of which single photons were counted using a fast avalanche photodiode detector. This set-up is similar to that used for detection of the collective Lamb shift 19 . Single-photon events were registered by recording their arrival time after excitation together with the velocity of the Doppler drive at the moment of detection. From data sets of about 10 7 such events, we extracted the reflectivity, | R ( Δ )|, of the samples as a function of Δ by applying the procedures described in ref 29 . The results are shown in Fig. 4 for the samples with the 57 Fe layers arranged in the sequence node–antinode (sample A) and antinode–node (sample B). The red solid lines are simulated spectra based on the structural data obtained from X-ray reflectivity and nuclear hyperfine interaction parameters obtained from conversion electron Mössbauer spectroscopy, using no adjustable parameters. Figure 4: Observation of nuclear resonant EIT. a , b , Measured spectral response (reflectivity) of sample A ( a ) and sample B ( b ). Sample A shows evidence for EIT, as predicted, with two strong transparency dips at Doppler detunings of Δ = −51 Γ 0 and 51 Γ 0 (dashed vertical lines). The EIT effect vanishes in sample B. Solid red lines are calculations taking into account the stroboscopic detection procedure applied here 29 . The difference in the baselines of the experimental spectra of the two samples at large detunings is a feature of the stroboscopic detection technique. c , d , The central areas of the measured spectra ( Δ = −70 Γ 0 to 70 Γ 0 ), however, closely resemble the calculated spectra of sample A ( c ) and sample B ( d ). e , Difference spectrum of the measured spectra in a and b in the Δ / Γ 0 range between −70 and −30. The solid red line is a guide to the eye based on a Fano resonance line shape. Error bars are estimated from photon counting statistics. PowerPoint slide Full size image Sample A shows evidence of EIT, as predicted. Its spectral response has transparency dips that are particularly pronounced at the outer (and strongest) resonance lines of the hyperfine-split spectrum, at detunings of Δ = ±51 Γ 0 ( Fig. 4a–d , vertical dashed lines). The EIT dips in the centre are less resolved because they are much sharper than the outer ones and thus exceed the resolution limit of the stroboscopic detection method. In sample B, the EIT vanishes owing to the reversal of the sequence of the two layers in the cavity field. As a result, the spectral response is dominated by superradiantly broadened lines without transparency dips. Owing to the stroboscopic detection process, the off-resonance baseline is located at different levels in the two spectra. This is fully described by the simulation and does not affect the spectral signature of EIT observed here. To characterize the spectral shape of the measured EIT window, we subtracted the spectrum of sample A from that of sample B over an energy range from Δ / Γ 0 = −70 to −30. The spectral profile ( Fig. 4e ) shows an asymmetric shape typical for a Fano resonance, as indicated by the red solid curve. Its width, of ∼ 10 Γ 0 , is consistent with the value of Ω C estimated above. The depth of the strong transparency dips of sample A corresponds to a reflectivity that is reduced to a level of | R | 2 = 0.10 relative to | R | 2 = 0.45 in sample B. The degree of transparency would be much larger if the nuclear resonance were a single line, as illustrated in Fig. 1 , where | R | 2 = 0.03 in the transparency window. In forthcoming experiments, it should be possible also to detect the transmitted field that leaves the cavity after propagating along its axis. This brings the production of slow light in this spectral regime into reach. We estimate group velocities in the range of 30 m s −1 ( Supplementary Information , section B). To produce such group velocities, it is necessary to adjust the width of the transparency window to allow propagation of transform-limited pulses along the axis of the cavity. The width can be controlled by adjusting the thickness, d 1 , of the layer located at the node of the cavity field or by placing several layers in the multiple nodes of high-order cavity modes. Moreover, we predict that drastic changes will occur in the linear refractive index at resonance if the positions of the resonant layers in the cavity are changed by only 1 nm. This effect is analogous to the cross-phase modulation due to giant Kerr nonlinearity 2 as it makes it possible to control, for example, the phase accumulation of the probe field on propagation along the cavity. Self-phase modulation of the probe field under nuclear EIT conditions might become relevant at intensities achieved at X-ray laser facilities in the near future ( Supplementary Information , section B). We emphasize that this technique, based on coherent light scattering, can be generally applied to any ensembles of resonant emitters (for example atoms, ions or quantum dots) properly placed in optical cavities. Although in our experiment the EIT effect relies on the spatial modulation and interference of single-photon cooperative emission from two ensembles of many emitters in a cavity field, the same effect can be expected for two single emitters that are sufficiently well confined in a cavity or microcavity. With a large Purcell enhancement of the spontaneous emission from an antinode of the cavity field and inhibited emission from a node, the condition for our cavity EIT scheme will be equally well fulfilled for single emitters as for many. The fact that this scheme works with two-level systems extends EIT and its applications to systems that do not have a metastable level, facilitating the transfer of EIT and its applications to the nuclear realm. Moreover, the radiative coupling of separated atomic ensembles in a cavity provides a very sensitive tool to probe the properties of cooperative emission such as superradiance and the collective Lamb shift.

At the high-brilliance synchrotron light source PETRA III, a team of DESY scientists headed by Dr. Ralf Röhlsberger has succeeded in making atomic nuclei transparent with the help of X-ray light. At the same time they have also discovered a new way to realize an optically controlled light switch that can be used to manipulate light with light, an important ingredient for efficient future quantum computers. The research results are presented in the current edition of the scientific journal Nature. The effect of electromagnetically induced transparency (EIT) is well known from laser physics. With intense laser light of a certain wavelength it is possible to make a non-transparent material transparent for light of another wavelength. This effect is generated by a complex interaction of light with the atomic electron shell. At DESY's X-ray source PETRA III, the Helmholtz research team of Röhlsberger managed to prove for the first time that this transparency effect also exists for X-ray light, when the X-rays are directed towards atomic nuclei of the Mössbauer isotope iron-57 (which makes up 2% of naturally occurring iron). Quite remarkably, only very low light intensities are needed to observe this effect, in contrast to standard EIT experiments. How does the experiment work? The scientists positioned two thin layers of iron-57 atoms in an optical cavity, an arrangement of two parallel platinum mirrors that reflect X-ray light multiple times. The two layers of iron-57 atoms, each approximately three nanometres thick, are precisely kept in position between the two platinum mirrors by carbon, which is transparent for X-ray light of the wavelength used. This kind of sandwich with a total thickness of only 50 nanometres is irradiated under very shallow angles with an extremely thin X-ray beam from the PETRA III synchrotron light source. Within this mirror system, the light is reflected back and forth several times, generating a standing wave, a so-called resonance. When the light wavelength and the distance between both iron layers are just right in proportion, the scientists can see that the iron becomes almost transparent for the X-ray light. In order for this effect to occur, one iron layer must be located exactly in the minimum (node) of the light resonance, the other one exactly in the maximum. When the layers are shifted within the cavity, the system immediately becomes non-transparent. The scientists attribute this observation to a quantum-optical effect, caused by the interaction of atoms in the iron layers. Unlike single atoms, the atoms in an optical cavity together absorb and radiate in synchrony. In the geometry of this experiment their oscillations mutually cancel each other, as a result of which the system appears to be transparent. In contrast to previous experiments in the optical regime, only few light quanta are necessary to generate this effect. "Our result of achieving transparency of atomic nuclei is virtually the EIT effect in the atomic nucleus," Röhlsberger describes the experiments. "Undoubtedly, there is still a long way to go until the first quantum light computer becomes reality. However, with this effect, we are able to perform a completely new class of quantum-optical experiments of highest sensitivity. With the European XFEL X-ray laser, currently being built in Hamburg, there is a real chance to control X-ray light with X-ray light." This experiment definitely means considerable technical progress for quantum computing: apart from the basic possibility to make materials transparent with light, the intensity of light is decisive for a future technical realisation as well. Every additional quantum of light produces additional waste heat; this would be reduced by the use of the presently discovered effect. For the continuation of these experiments and the optimal utilisation of the extremely small X-ray beam size of the highly brilliant X-ray source PETRA III, a new coating facility will be installed at DESY for the production and optimisation of these optical cavities. The experiments of the DESY scientists also showed another parallel to the EIT effect: the light trapped in the optical cavity only travels with the speed of a few metres per second – normally it is nearly 300 000 kilometres per second. With further experiments, the scientists will clarify how slow the light really becomes under these circumstances, and whether it is possible to use this effect scientifically. A possible application and at the same time an important building block on the way to light-quantum computers is, for example, the storage of information with extremely slow or even stopped light pulses.

10.1038/nature10741

Earth

Researchers measure a record concentration of microplastic in Arctic sea ice

Ilka Peeken et al, Arctic sea ice is an important temporal sink and means of transport for microplastic, Nature Communications (2018). DOI: 10.1038/s41467-018-03825-5 Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-018-03825-5

https://phys.org/news/2018-04-microplastic-arctic-sea-ice.html

Abstract Microplastics (MP) are recognized as a growing environmental hazard and have been identified as far as the remote Polar Regions, with particularly high concentrations of microplastics in sea ice. Little is known regarding the horizontal variability of MP within sea ice and how the underlying water body affects MP composition during sea ice growth. Here we show that sea ice MP has no uniform polymer composition and that, depending on the growth region and drift paths of the sea ice, unique MP patterns can be observed in different sea ice horizons. Thus even in remote regions such as the Arctic Ocean, certain MP indicate the presence of localized sources. Increasing exploitation of Arctic resources will likely lead to a higher MP load in the Arctic sea ice and will enhance the release of MP in the areas of strong seasonal sea ice melt and the outflow gateways. Introduction Marine debris is a growing environmental concern as recent reports indicate that increasing quantities of litter disperse into secluded environments, including Polar Regions 1 , 2 , 3 and the deep ocean floor 4 . Plastic accounts for 73% of marine debris globally 5 , and it has been estimated that about 8 million tons of plastic move from land into the ocean each year 6 . However, only 1% of this has been accounted for in terms of small plastic debris 7 , highlighting that some of the major sinks of oceanic plastic litter remains to be identified. The Arctic Ocean is now in a state of rapid transition that is best exemplified by the marked reduction in age, thickness and extent of the sea ice cover 8 . The European Arctic margin is influenced by drift ice formed on the Siberian shelves and carried to the Fram Strait via the Transpolar Drift 9 . In contrast, the Fram Strait is the gateway that transports warm Atlantic water, via the West Spitsbergen Current to the Central Arctic 10 , containing an anthropogenic imprint 11 . It is well known that regions of the Arctic Ocean are highly polluted owing to local sources and long-range atmospheric input 12 . In this context sea ice has been identified early on as a major means of transport for various pollutants 13 , 14 , with north and east Greenland as well as the Laptev Sea, being especially prone to contamination from several sources 15 . A useful method to study sea ice drift pattern is by using passive microwave satellite images combined with the motions of sea ice buoys 16 , 17 , which highlight the role of sea ice, e.g. spreading oil spills 18 . Recent studies stress the changes caused by the shift to first year ice resulting in the tendency of sea ice floes to diverge from the main drift pattern 19 such as the Transpolar Drift, with complex effects on exchange processes of any contaminants between the exclusive economic zones (EEZ) of the various Arctic nations 20 . Despite the scant knowledge of Arctic ecosystems, the trend towards thinner sea ice and ice-free summers in the future has already stimulated increasing exploitation of its resources in terms of shipping, tourism, fisheries and hydrocarbon exploration 21 . Plastic degrades into smaller fragments under the influence of sunlight, temperature changes, mechanic abrasion and wave action 22 . Particles < 5 mm are called microplastics (MP) and have recently featured in a strongly growing number of studies and publications 23 . MP raise particular concerns because plastic in this size category can be taken up by a much wider range of organisms with currently largely unknown health effects on marine life and humans 24 , 25 . Over the past decade, MP were identified from numerous marine ecosystems globally 5 , including the Arctic 2 , 26 and the Southern Ocean 3 . High concentrations of MP also occur in the surface waters south-west of Svalbard 26 , but overall, MP origins, pathways and their compositional framework within Arctic sea ice remains unclear. Here, we analyzed the content and composition of MP from sea ice cores at five different locations along the Transpolar Drift to assess if sea ice is a sink and transport vector of MP. Ice cores were taken from one land-locked and four drifting ice floes to distinguish between local entrainment of MP and long-distance transport. MP composition of the cores was analyzed by focal plane array detector-based micro-fourier-transform infrared imaging 27 , 28 (Imaging FTIR) and compared to a previous study with respect to MP in Arctic sea ice cores. Analyses of discrete ice core horizons allowed us to assess the spatial variability within sea ice and to reconstruct the location of MP incorporation. By computing drift trajectories, coupled to a thermodynamic ice growth model 29 , possible source regions of MP entrainments during ice growth were identified. Results Microplastics in entire sea ice cores To quantify the MP concentration and composition we obtained ice cores during expeditions of the German research ice breaker Polarstern in spring 2014 and summer 2015 in the Fram Strait and Central Arctic (Fig. 1a , Table 1 ). The highest MP particle concentration ((1.2 ± 1.4) × 10 7 N m −3 ) was detected in an ice core taken in the pack ice of Fram Strait (core B; Fig. 1b ). The MP concentrations of all sea ice cores were highly variable with the second highest MP concentration found in the land-fast ice of the Fram Strait (core A; (4.1 ± 2.0) × 10 6 N m −3 ). The MP load of core C and E, collected north of Svalbard and in the Nansen Basin respectively, varied between (2.9 ± 2.4) × 10 6 N m −3 and (2.4 ± 1.0) × 10 6 N m −3 . The lowest concentration was found in core D (1.1 ± 0.8) × 10 6 N m −3 ; Fig. 1b ) from north of Svalbard. The values recorded in this study are two to three orders of magnitude higher than in a previous study from the Central Arctic 2 (1.3–9.6) × 10 4 N m −3 , values exclude rayon, for further details see method section), which can largely be explained by the different methodology used. In the previous study 2 , the filter area was first inspected by light microscopy and suspected MP particles were then analyzed individually by Fourier-transform infrared (FTIR) microscopy. In contrast, we used Imaging FTIR 27 , 28 , where entire areas were scanned. This excluded the human bias introduced by visual selection of particles (Fig. 2a ). Imaging FTIR includes the far more informative infrared region of the spectrum from the very onset of the analysis and enables the detection of very small particles (down to 11 µm), which are most likely overlooked by visual inspection and therefore not included in the majority of the previously published studies. By using this approach, we were able to show that most of the MP particles identified in the sea ice cores were smaller than 50 µm. On average 67% of the particles were within the currently smallest detectable size class of 11 µm (Fig. 3 ). Such small particles were not considered in the previous study on MP in sea ice by Obbard et al. 2 . Concerning the extremely high error values calculated for the individual samples it should be noted, that they result from the analyses of three different areas per filter indicating an unequal particle distribution (Fig. 2b ). In contrast to our previous studies 30 , 31 , the samples were not macerated beforehand and only H 2 O 2 was applied after filtration to remove natural organic residues. MP particles tend to form hetero-aggregates with microalgae or natural organic matter 32 and therefore it is likely that these aggregates were printed on the membranes by filtration. Thereafter the organic matrix was removed by wet oxidation before the Imaging FTIR measurement. However, it should be noted that, to the best of our knowledge, our study is the first to indicate and document an uneven MP particle distribution on filters overall. With respect to the early days of direct bacterial counting, where similar observations were made, future MP studies should address this problem by improving the general sample preparation or by applying a statistically valid recording approach as used in direct bacterial counting 33 . We provide our measurements as particle count per volume for consistency with previous studies. However, we suggest that future studies also consider polymer-specific MP mass per volume data 34 to allow for calculation of fluxes or total load of synthetic polymers (independently of the degree of fragmentation). Fig. 1 Pathway and microplastic content of sea ice cores in the Central Arctic. a Sampling position of sea ice cores (A–E) obtained during three Polarstern expeditions overlaid with the sea ice concentration (June 2014) and a schematic view of the major cold and warm water currents. Blue arrows indicate the inflow of Pacific water. For comparison, previously sampled sea ice cores 2 are included (Ha–Hd); b Total microplastic (MP) particle load m −3 of the various sea ice cores (this study) and data reproduced from Fig. 2 of Obbard et al. 2 *; c Average % composition of polymers (polyethylene (PE), varnish (including polyurethanes and polyacrylates), polyamide (PA), ethylene vinyl acetate (EVA), cellulose acetate (CE-Alkylated), polyester (PES) and polypropylene (PP) and others) from the entire core (this study) and digitized data of figure two from Obbard et al. 2 *, acrylic equals varnish (others include acrylonitrile butadiene, chlorinated polyethylene, nitrile rubber, polycaprolactone, polycarbonate, polylactic acid, polyimide, polystyrene, polyvinyl chloride, rubber); d Drift trajectories of sea ice cores, except for land-fast ice station of Greenland (A) and the sample originating from the Chukchi Shelf Ha. The map was created using ArcGIS 10.3 and based on the General Bathymetric Chart of the Oceans (GEBCO)-08 grid, version 20100927, , with permission from the British Oceanographic Data Centre (BODC). *The polymer rayon was excluded Full size image Table 1 Sea ice core sampling and accompanying information table Full size table Fig. 2 Images of the microplastic analysis. a Overview image collected by the fourier-transform infrared imaging (FTIR) microscope prior to measurement. b Polymer dependent false-colour image of an exemplary measurement field after FTIR measurement and automated analysis Full size image Fig. 3 Size classes of observed microplastic particles. Box and whiskers plot of percentage (%) shares of MP numbers in different size classes in all sea ice cores. The boundary of the box closest to zero indicates the 25th percentile, the line within the box marks the median and the boundary of the box farthest from zero indicates the 75th percentile. Whiskers (error bars) above and below the box indicate the 90th and 10th percentiles. Blue lines indicate the mean and black circles indicate outliers Full size image In total, 17 different polymer types were identified (Supplementary Fig. 1 ; Supplementary Table 1 ), with polyethylene (PE), varnish (including polyurethanes and polyacrylates), polyamide (PA) also called nylon, ethylene vinyl acetate (EVA), cellulose acetate (CE alkylated), polyester (PES) and polypropylene (PP) contributing on average between 48% (PE) and 1.65% (PP) to the total measured MP composition. The following were on average below 1%: nitrile rubber, rubber, polystyrene (PS), polylactic acid (PCA), polyvinyl chloride (PVC), chlorinated polyethylene (PE-Cl), polycarbonate (PC), polycaprolactone (PCL), acrylonitrile butadiene (AB) and polyimide (PI) (summarized as others in Figs. 1 and 4 ). Overall, the MP composition of the sampled sea ice cores was variable with PE being almost exclusively found in the upper horizons of core B sampled in the Fram Strait (above 90%; Fig. 1c ). PE, which is among the economically most important polymers 35 , also dominated the other core from the Fram Strait (A) and the cores retrieved north of Svalbard (core C, D; Fig. 1c ). In these cores, PA, which is usually associated with fishing gear 35 , accounted for 6 and 22% of the MPs. Both cores also contained varnish, which dominated core E taken in the Nansen Basin (Fig. 1c ) and was present in the land-fast sea ice (core A, Fig. 1c ). The polymer type varnish includes the previously described acrylic polymer type, known to account for up to 10% of the MP in marine systems 35 . Core A and E also shared a relatively high proportion of EVA (up to 10%). The latter core was further characterized by almost 9% of CE-Alkylated (Fig. 1c ), which is indicative of cigarette filters and commonly found in ocean debris 22 . Except for PE, overall the sea ice cores only partly reflect the composition of the globally produced polymers, which are dominated by PE, followed by PP, PVC, PS, PUR and polyethylene terephthalate (PET) 35 . Fig. 4 Vertical distribution of ancillary data in sea ice cores. a Refers to the salinity (PSU), b refers to temperature (°C), c refers to particulate organic carbon content (POC; mg×L −1 ) and d refers to chlorophyll a concentration (µg×L −1 ) for each core. Steps indicate the sampling horizons taken for each core Full size image Sea ice trajectories and MP comparison to a previous study Large portions of sea ice are formed on the Siberian shelves 9 . Depending on individual ice floe drift patterns, they pass through different regions of the Central Arctic Ocean but are eventually carried to the Fram Strait via the Transpolar Drift 9 , 36 . To determine drift trajectories and source areas of sampled sea ice we tracked sampled sea ice backwards in time using low-resolution ice drift and concentration products from passive microwave satellites 17 . This back-tracking approach 17 showed that the sea ice samples originated from different source areas, namely the Amerasian and Eurasian Basins (Fig. 1d ; Table 1 ). In particular core B can be retraced to the Makarov Basin, while the cores within the Eurasian Basin originated from the Laptev Sea (core E), near Franz Josef Land (core D) and the deeper Nansen Basin (core C). Except for the land-fast ice, east of Greenland (core A), all ice cores encountered the main path of the Transpolar Drift pack ice 9 , 36 . To enable a comparisons with Obbard et al. 2 , we applied the back-tracking approach to the four sea ice cores from that study (Fig. 1d ). The origin of the sea ice cores from the earlier study 2 only overlap in the region of the Laptev Sea (core Hd) with one of our cores (E), while the other cores can be related to the Beaufort Gyre (core Hc), the Chukchi Shelf (core Ha) and the East Siberian Sea (Fig. 1d ). The strikingly high contribution of varnish (58%) in the sea ice cores originating from the Laptev Sea in our study (core E; Fig. 1c ) did not feature at all in the earlier core Hd. However, even in the previous study, acrylic (similar type as varnish) was present in one core (Hb) and the contribution was in the same order of magnitude as described in other studies 35 . PA seemed to be a common compound observed in eight out of the nine investigated cores in both studies (Fig. 1c ) and was particularly dominant in the core from the Chukchi Shelf (core Ha). Ice cores analyzed by Obbard et al. 2 also showed high concentrations of PET (Fig. 1c ), which might be partly reflected in our polymer type PES. The large difference between both studies might be related to our exclusion of fibres. Alternatively, it might reflect the increased effort to recycle this particular compound 37 . Ancillary variables and MP along sea ice cores Since the formation and growth of sea ice is a process in space and time, it is most likely, that the MP composition of the seawater–ice interface reflects the respective water body, which is in contact with the drifting sea ice floes as shown by the drifting path of the sampled ice floes (Fig. 1d ; Supplementary Fig. 2 ). Given our current lack of knowledge about MP contamination of surface waters from the Central Arctic the mere consideration of the complete core is probably not sufficient with respect to the spatiotemporal history of the sea ice floe. In order to consider this spatiotemporal history we studied the vertical distribution of MP in the context of environmental and biological variables (ancillary variables) in each horizon (~20–30 cm sections) of the sea ice cores (Figs. 4 and 5 ). Temperature and salinity measurements suggest that the sampled sea ice consisted of first and second year ice (Fig. 4 ). Second year ice was indicated by a low salinity in the surface (core C & E). With a maximum of 1.67 m (core E), sea ice thickness was very low for spring conditions, but is in agreement with the observed reduction of sea ice in the Fram Strait 36 and Central Arctic 38 . The highest concentrations of ice algal biomass, as indicated by chlorophyll a, were mostly found in the lowest core horizons (Fig. 4 ; max. 6.64 µg L −1 ), with a general L-shaped distribution pattern. Only the land-fast-ice core (A) had a maximum concentration at the top. Particulate organic carbon (POC) had a maximum of 1.34 mg L −1 and also showed an L-shaped distribution pattern (Fig. 4 ). Only the land-fast ice core (A) and one core sampled north of Svalbard (D) had their POC maximum in the middle of the core. Fig. 5 Vertical distribution of microplastic in sea ice cores. a refers to the concentration of microplastic particles (in N×L −1 ) for each core. b refers to the polymer composition for each core: polyethylene (PE), varnish (including polyurethane and polyacrylate), polyamide (PA), ethylene vinyl acetate (EVA), cellulose acetate (CE-Alkylated), polyester (PES) and polypropylene (PP) and others. c refers to polymer richness (N), and d refers to the Shannon–Wiener index (H'). Steps indicate the sampling horizons taken for each core Full size image In contrast to these generally L-shaped vertical profiles of biological variables, MP quantities in the various sea ice core horizons were extremely variable with concentrations ranging between 33 and 75,143 L −1 (Fig. 5 , units are adjusted to environmental data (L −1 )). The latter concentration was found in the surface horizon of core B originating from the Makarov Basin (Fig. 1d ). Another high surface value with 6421 L −1 was observed in core E, while all other cores contained relative low MP concentration in the top 20 cm (max. of 992 MP L −1 ). The core originating from the Laptev Sea (E) displayed another MP maximum in the bottom horizon at the seawater–ice interface (4437 L −1 ), while core C was characterized by a sub-surface maximum (Fig. 5 ). All other cores were characterized by variable concentrations reaching a maximum concentration of particles throughout the middle of the ice core, between 4159 and 13,794 L −1 (Fig. 5 ). The synthetic polymer composition of the individual horizons showed strong differences within a single sea ice core (Fig. 5 ). PE dominated the synthetic polymers in the top 100 cm of the sea ice core B originating from the Makarov Basin (Fig. 5 ). PE was also present in all other cores but with highly variable contributions in the different horizons. In the core originating from the Makarov Basin (B), a change in salinity was associated with a drop in MP numbers and a shift to a more complex polymer composition, with high proportion of varnish in the bottom horizon (Figs. 4 and 5 ). However, only few significant correlations between MP numbers, MP derived diversity indices (N, H', Fig. 5 ), MP sizes and ancillary variables were found (Supplementary Table 2 ). Highest Spearman rank correlations (negatively correlated) were found between chlorophyll a, total MP numbers, numbers of PE particles and the numbers of particles in the six lower size classes (covering the sizes 11–125 µm). This underscores that sea ice does not have a uniform MP imprint throughout the whole ice body and that the MP numbers or the polymer composition cannot be explained by ancillary variables recorded in parallel in the different horizons. This finding might point to a patchy distribution of MP in Arctic waters where local MP populations of the respective waterbodies are archived when in contact with the ice–seawater interface. This is evident for sea ice cores C and D, which had extremely different polymer distributions even though both sampling locations were in close proximity north of Svalbard. Indeed, the back trajectories of these two cores revealed different sea ice origins and pathways along the Transpolar Drift (Fig. 1d ). The core originating from Franz Josef Land (core D), showed two horizons dominated by PE polymers, with varnish and PA distributed along the core, particularly in the two lowest horizons. In contrast, the core originating from the deeper Nansen Basin (C) showed a strong variability of polymer distributions in all horizons with a higher proportion of PE and varnish and with the occurrence of PA, EVA and PES (Fig. 5 ). The highest contribution of varnish in all sea ice core horizons was evident in the core originating from the Laptev Sea (E). The middle part of this sea ice core was further characterized by various proportions of CE-Alkylated. This compound was also found in one horizon of core B, together with PP, which was also present in few horizons from core C (Fig. 5 ). Overall results from the 1D sea ice growth model highlight localized polymer entrainment (Supplementary Fig. 3 ), whereby, e.g., PE is present in high concentrations associated with the Atlantic and Pacific inflow in the Central Arctic, while e.g. varnish and EVA are more concentrated in the eastern region of the Eurasian basin (for details see Supplementary Fig. 3 ). Cluster analyses of MP numbers, MP sizes and ancillary variables resulted in contrasting sample groups. No clear groups of the different cores or horizons were obtained for MP numbers and sizes (Supplementary Figs. 4 , 5 and 6 ). In contrast, the ancillary variables grouped cores A, B and E and C and D (Supplementary Fig. 5 ). These results were also supported by the SIMPROF analysis. The clear separation found in the ancillary data might be related to the drift trajectories of the floes, where cores A, B and E follow the main path of the Transpolar Drift 9 , while cores C and D drifted nearer the Atlantic water inflow 10 . Cores C & D (SIMPROF group c) displayed significantly higher Chl a concentrations, salinities and C/N-ratios compared to A, B and E (SIMPROF group d). Still, POC, PON and temperature were significantly higher in cores A, B and E (Supplementary Table 2 ). However, when superimposing the SIMPROF grouping of the ancillary variables on the polymer-specific MP numbers or MP derived diversity indices in the different core horizons, with the exception of nitrile rubber and PVC-particles there was no significant difference displayed in average between the two groups. This was also the fact for the MP numbers in all size classes (Supplementary Table 3 and Supplementary Data 1 and 2 ). Overall, it can be assumed that all environmental and biological variables are characterized by a strong seasonality 39 , 40 , while MP particles once they are incorporated into the sea ice during sea ice growth seem to be more stationary in the ice core matrix. Discussion Combination of Imaging FTIR 27 and an automated polymer identification approach 28 revealed that MP concentrations in Arctic sea ice are extremely high and therefore sea ice can be seen as a temporary sink for MP. However, even during winter large fractions of the sea ice are exported southward and eventually doomed to melt, with highest sea ice export fluxes out of the Fram Strait 17 . Hence, Arctic sea ice can be considered a temporary sink, a source and an important transport vector of MP, at least to the Fram Strait and North Atlantic. The role of sea ice to redistribute, e.g. coastal sediments 41 and contaminants 15 , 42 along the Transpolar Drift or into the Central Arctic has long been recognized. It could be shown that the particular region of the Fram Strait will always be reached by any contamination source from the distant Arctic. The estimated time between contamination and arrival in the Fram Strait ranges between two and four years for sources in the Laptev and Kara Seas, and up to six to eleven years from sources of the Amerasian Basin 15 . Although the recent circumpolar TaraOceans Expedition highlighted certain hotspots of floating plastics in the Arctic transported via the poleward branch of the Thermohaline Circulation 1 , nothing is known to date about the actual MP concentrations in the waters of the Central Arctic Ocean itself, where all sea ice cores were collected. Overall, the sources and sinks of MP are currently not very well understood 43 , but since the Fram Strait is one of the main inflow gateways 44 to the Central Arctic Ocean, MP may have been transported with the relatively highly MP-contaminated offshore North Atlantic waters 45 into to the Arctic Ocean. For MP, this hypothesis was the focus of a recent study with a detailed sampling of (sub-) surface water south and west of Svalbard 26 , but no clear transportation pathways of MP could be identified. In that study, water with a high melt fraction was rather low in MP particle load, suggesting a comparably low impact of sea ice originating MP. However, in our study we found several order of magnitude higher concentrations of MP in the sea ice and thus confirm the first study by Obbard et al. 2 . Particularly the abundance of the small particles (11 µm) are of concern, since they can be taken up within the microbial food web 25 . A recent modelling study, which suggested an active transport of plastic debris with the North Atlantic currents into the Central Arctic Ocean 46 corroborates the notion that polar waters can no longer be considered free of plastic litter. However, owing to the different analytical approaches applied in MP concentration estimates, a direct comparison of the studies is hampered by the current lack of standards in MP research and thus only general inherent trends can be compared. Terrestrial MP sources in these sparsely populated high-latitude regions seem to have a negligible contribution to the MP load of Arctic waters as Lusher et al. 26 reported extremely low MP concentrations close to shore. In contrast to other coastal (and more densely populated) areas, which are known to be much more contaminated with MP 24 , emissions from Arctic terrestrial sources may be considered to be low. In addition, nothing is known about the impact of MP coming from the large rivers entering the Arctic. However, results from a recent modelling study suggest that contaminates originating from the Lena and Mackenzie river mouths, lead to a rapid spreading of theses pollutants in the entire Arctic Ocean 15 . Cores from east of Greenland (core A), close to Franz Josef Land (core D) and the Makarov Basin (core B) were further characterized by an abrupt change of MP composition in the various horizons. This suggests that the ice floe had passed through regions with fast changing MP compositions during the sea ice growth process. In analogy to previously described sediment entrainment into sea ice particularly frazil ice scavenged sediment particles 41 . During sea ice growth, characteristic salt fingers are developed 40 and enriching the particle concentration in the brine channels. In addition, coagulation with exopolymer particles excreted from sea ice algae 47 might further enrich the concentrations of MP. An accumulation of sediment particles in the surface horizons, as has been described for multiyear ice 41 , due to constant surface melting, might explain the high MP concentration observed in the surface of the cores originating from the Makarov Basin and the Laptev Sea. It is however unlikely that this process redistributes MP in first year sea ice. The highest concentration of MP ever determined in sea ice was found in core B, originating from the Makarov Basin. The core contained concentrations comparable to those from South Korean waters 48 or the Skagerrak 49 , which are the highest hitherto reported values in terms of reports per volume unit ( ). This peak consisted almost entirely of PE. Since PE has a low density 35 , particles are likely to float over long distances at the sea surface with the ocean currents before they eventually sink through ballasting 50 . The Canadian Basin is supplied with water originating from the northeast Pacific and transported through the Bering Strait 51 . From the southern part of the Chukchi Sea sea ice has a direct path via the Central Arctic towards the Fram Strait 52 . We thus speculate that the high PE concentrations in the core from the Makarov Basin might reflect remains from the so-called North Pacific Garbage Patch 53 , transported with the incoming Pacific inflow. A recent study by Desforges et al. 54 showed quite high MP concentrations for the NE Pacific and highlights the role of oceanographic conditions for the accumulation patterns of MP’s 54 . Indeed, models project that on long time scales inter-ocean exchanges play a significant role in the distribution of marine debris enabling transport between accumulation areas 53 . Because of its widespread use and the low density, PE might nowadays, be considered to be a background MP even in surface water of the Central Arctic, analogous to the global distribution of certain persistent organic pollutants 55 . The high proportions of varnish are remarkable, particularly in the upper core from the land-fast ice (A) and the core originating from the Laptev Sea (E). In the latter, varnish is present throughout the entire core and often associated with a high proportion of cellulose acetate, indicative of cigarette filters 35 and to a lesser extent with ethylene vinyl acetate copolymer (EVA), a polymer, which is also used in antifouling paints for ships 56 . The occurrence of varnish and EVA can be attributed to ship traffic, which has increased between 2009 and 2014 in the Arctic 57 , 58 . PA, usually associated with fishing gear 35 , was found frequently in almost all sea ice cores. This polymer was the most abundant in the sea ice core originating from the Chukchi Shelf 2 . It accounted for 22% in the core originating from the deeper Nansen Basin, which is comparable to other reports (<20%) 35 . Overall, a high contribution of PA may be related to increasing commercial fishing efforts in the eastern Bering Sea, Barents Sea, north of Svalbard and north of Franz Josef Land 58 implying local input. Fisheries accounts for a major share of the ship traffic in the Arctic Ocean 57 , 58 and continuous reduction of sea ice is assumed to increase fisheries further, particularly in the high Arctic 59 . The above attribution of source regions is supported by the reconstruction of the location of MP incorporation into the ice by coupling the back-tracking approach with a one-dimensional thermodynamic sea ice growth model (Supplementary Fig. 3 ). Overall, this is the first detailed mapping of MP particle composition and size classes from sea ice cores obtained in the high Arctic with special emphasis on the vertical pattern of MP. The ice cores were characterized by various footprints of polymer composition resulting from different origins and pathways during the period of ice growth, although to prove this concept more data are needed. We identified unique footprints for different origin areas. This occurrence pattern is akin to data obtained for the presence of coloured dissolved organic matter 60 , reflecting different source areas of the water, e.g., input from Lena river water or Pacific water. It is likely that parts of the MP, which are embedded in the sea ice, were transported by currents into these regions and that different oceanic realms (Pacific versus Atlantic) currently still have specific MP imprints. However, these imprints are altered by localized dispersal of MP in the Arctic, which need to be considered for future budgeting of global MP sources and sink estimates. With respect to global climate change, large fractions of MP might be released from melting Arctic sea ice. Given a yearly melt of sea ice between 1.6×10 4 km 3 and 1.93×10 4 km 3 (PIOMAS 2011–2016 based on ref. 61 ) large fractions of these particles are released. Basic calculations show the potential release of MP between a minimum of 7.2 × 10 20 and a maximum of 8.7 × 10 20 particles per year between 2011 and 2016, assuming the here-observed average MP values. The maximum values can be attributed to the sea ice record minimum low found in 2012. Since currently only a few studies focus on the occurrence of MP in Arctic waters, it remains speculative whether these potentially released MP remain in Arctic waters or are transported to lower latitudes. On the other hand, due to the co-occurrence of sticky exopolymer particles in sea ice 62 , a formation of hetero-aggregates might occur 32 , resulting in a change of buoyancy of MP 63 and sedimentation to the seafloor. Indeed, very high numbers of MP were recently detected in deep-sea sediments of the HAUSGARTEN observatory in Fram Strait 30 . Highest MP concentrations occurred at the northern most stations, which are characterized by a long-lasting marginal sea ice zone area. Recent studies in the Central Arctic also showed that biogenic particles below 2 µm can contribute to the vertical flux 64 due to particle coagulation. The process of brine circulation due to melt progression 65 may redistribute MP in the Central Arctic Ocean as there is, for example, a high exchange of sea ice algae between the ice and the underlying water to depths of 40 m 66 . Many MP are in the same size range as sea ice algae, and may therefore also be transported far below the euphotic zone by brine convection. We conclude that the MP distribution in the Central Arctic is more complex than previously considered, assuming only transport with high MP loads from the urban areas into the remote Polar Regions, although this undoubtedly constitutes the main point of entry. Our results also point to localized MP sources, which might become more pronounced as the exploitation of the Arctic progresses. Methods Sea ice coring Sea ice sampling has been carried out during three cruises with the ice breaker Polarstern in the Fram Strait (PS85, FRAM; June/July 2014 67 ), the Barents Sea slope (PS92, TRANSSIZ; May/June 2015 68 ) and the Central Arctic (PS94, TransArc II; August-October 2015 69 ). At each station, a designated coring site was assigned and if present, the snow was removed before drilling the sea ice cores. Nitrile gloves were used and cores were drilled with a Kovacs 9 cm diameter corer (Kovacs Enterprise, Roseburg, USA). The microplastic cores were immediately transferred into plastic bags (polyethylene tube films (LDPE) by Rische and Herfurth) and stored at −20 °C. Back-tracking of sea ice To determine drift trajectories and source areas of sampled sea ice we tracked the sampled ice backward using low-resolution ice drift and concentration products from passive microwave satellites after Krumpen et al. 17 and Krumpen 29 . Sea ice concentration data used in this study were obtained from the National Snow and Ice Data Center (NSIDC). Ice drift data are provided by different institutions and have been widely used in various studies to investigate pathways and source areas of sea ice 17 , 70 , 71 . In this study, two different sets of ice drift products were used: During summer months (June–August), the Polar Pathfinder Sea Ice Motion product provided by the NSIDC provided on a 25 km grid 72 was applied. During the rest of the year, tracking is forced with sea ice motion data provided by the Center for Satellite Exploitation and Research (CERSAT) at the Institut Français de Recherche pour l’Exploitation de la Mer (IFREMER). Motion data are available with a grid size of 62.5 km, using time intervals of 3 days for the period between September and May 73 . The tracking algorithm works using motion and concentration data. A specific ice area is tracked backwards until: the ice reaches a position next to a coastline, the ice concentration at a specific location reaches a threshold value of <15% when ice parcels are considered lost, or the tracking time exceeds 4 years. To quantify uncertainties of estimated sea ice trajectories using satellite sea ice motion and concentration data, pathways of 39 buoys were re-tracked. Buoy data were obtained from the SeaIcePortal.de and followed from their deployment position in a forward direction. On average, the displacement of virtual buoys during the first 150 days (around 1000 km of ice drift) is around 35 km. After one year (ice drift of more than 2500 km), the average displacement is around 150 km. For general details of the AWI ICETrack tool (Antarctic and Arctic Sea Ice Monitoring and Tracking Tool), please see Krumpen 29 . 1D sea ice growth model Along the pathways derived from the back-tracking model sea ice trajectories, air temperatures (NCEP atmospheric reanalysis data) and snow depth (Warren climatology) and other atmospheric parameters were extracted as input for the thermodynamic ice growth model. The model was then used to estimate the location of MP incorporation into the respective ice sample (see below). To link the vertical distribution of MP within ice cores to the location, where MP particles were incorporated into the ice; we used a simplified one-dimensional thermodynamic model of sea ice growth. The model calculates sea ice growth based on surface air temperature, ocean heat flux and snow cover 74 , where the change in ice thickness Δ h /Δ t is given by: $$\frac{{{\mathrm{\Delta }}h}}{{{\mathrm{\Delta }}t}} = \frac{{ - 1}}{L} \cdot \left( {F_{\mathrm{OHF}} + \left( {T_{\mathrm{surf}} - T_0} \right)} \right) \cdot \frac{{\kappa _{\mathrm{i}} \cdot \kappa _{\mathrm{s}}}}{{\left( {\kappa _{\mathrm{i}}z_{\mathrm{s}}} \right) + \left( {\kappa _{\mathrm{s}}z_{\mathrm{i}}} \right)}}$$ (1) where the latent heat of fusion L , the thermal conductivities of ice κ i and snow κ s and the freezing point of seawater T 0 are set to literature values according to 74 . Surface air temperature is extracted along the ice trajectory from NCEP reanalysis 75 , while the ocean heat flux is assumed to be constant at 2 W/m 2 in agreement with Meyer et al. 76 and earlier work 77 . Ice thickness is calculated along the trajectories at daily increments and is in reasonably good agreement with ice core length for such a simple model. The model was also validated against an ice-mass-balance buoy 78 providing accuracy of few centimetres during the growth phase. While such simple models perform very well on simple ice growth of typical Arctic sea ice, they are not suited to model strong regional melting features. Sampling locations of ice cores C and D, taken on the TRANSSIZ expedition in 2015, are strongly affected by small-scale abnormal basal melting in the inflow area of Atlantic water just north of Spitsbergen. Meyer et al. 76 report tenfold increased heat fluxes for the sampling region with basal melt rates of up to 26 cm/day. This dramatic melting, which is strongly limited in time and space can generally not be reproduced by single point to point comparisons even with the most sophisticated thermodynamic sea ice models. The difference between our simple model and measured ice core lengths, on the order of 50 cm, can be caused by just a few days of melting. This does, however, not affect the growth phase of the respective sea ice, so that our simple thermodynamic model can still be used to estimate the location of MP incorporation into the ice. Using back-tracking (Supplementary Fig. 2 ) and the thermodynamic model we thus attributed a location to the incorporation of the different kinds of MP into the different ice samples, supporting our conclusions on the source regions of different MP species (Supplementary Fig. 3 ). Environmental and biological variables Handling the variables and measurements from sea ice cores was performed as described in previous studies 66 , 79 . MP sample preparation To prevent contamination of samples, handling and processing the sea ice cores was conducted under a clean bench (Labogene Scanleaf Fortuna, Lynge, Denmark). Ice cores were cut individually into horizons ranging from 10 to 35 cm using a bone saw. To exclude sample contamination from sampling the surface of the ice core horizon was removed by a stainless steel grater. Afterwards, it was washed with 1 L of MilliQ to remove the particles that would eventually adhere. Each horizon was weighed before being melted in glass-preserving jars at room temperature and then concentrated onto Anodisc filters (47 mm, Whatman, Freiburg, Germany). All samples were treated with of 35% H 2 O 2 (Roth, Karlsruhe, Germany, filtered over 0.2 µm Anodisc). After filtering the melt water the filter was overlaid with 40 mL H 2 O 2 and incubated at room temperature overnight. Lastly, the H 2 O 2 was drained and the filters were flushed with approx. 750 mL MilliQ water. To remove the adhering material, the filtration funnel was further flushed with 30% ethanol (VWR Chemicals, Darmstadt, Germany, filtered over 0.2 µm Anodisc) to reduce surface tension and thereby assure the concentration of all particles on the filter. The Anodisc filters were placed in glass petri dishes and dried at 30–40 °C in a drying cabinet (Memmert, Schwabach, Germany) overnight. Blank test For the blank tests, artificial ice cores were produced by freezing MilliQ water in a stainless steel beaker for one day. Afterwards, they were transferred into an ice core transportation bag and kept in the freezer for three days. Each day, the cores were rolled to simulate transport and then treated in the same manner as the sea ice core samples described above. Set up and operation of FTIR microscope (Imaging FTIR) For the particle measurements, a Hyperion 3000 microscope (Bruker Optics) attached to a Tensor 27 (Bruker Optics) spectrometer was used. The microscope features a FPA detector with 64 × 64 detector elements. With a visual objective of ×4 magnification the whole filter area was photo-documented to obtain a sample overview. Afterwards, the IR measurement was performed via two 15× magnification Cassegrain lenses. The measurements and analyses were performed with the OPUS 7.5 software (Bruker). In previous studies, the optimal settings for the measurements were evaluated 27 . The scan was run in transmittance mode with 6 co–added scans, a range of 3600–1250 cm −1 and a resolution of 8 cm −1 . A 4 × 4 binning was selected to balance the amount of data and the analysis time. The whole system was flushed with compressed dry air and with a flow rate of approx. 200 L h −1 to prevent signals caused by air humidity and CO 2 . After drying, the Anodisc filter was placed on the FTIR microscope. The background was measured on the Anodisc surface without sample impurities. As it was not possible to analyze the whole filter cake of (36 mm) in diameter in a single measurement, three separate fields on the concentrated sample were measured as technical triplicates. Three grids of 70 × 70 FPA fields each were placed on the filter equalling 3,763,200 single spectra. The measurement took 12.25 h per grid resulting in a total analysis time of 36.75 h per filter. Quantification and identification method Each measurement field was subjected to the automated analysis by Primpke et al. 28 . During this process, each spectrum was compared twice against a spectral library with different data handling. Each successful hit was stored together with the x , y and a quality factor into a csv file. Afterwards, the file was analyzed by image analysis to determine the polymer types, particle number per polymer and size distribution 28 . For each filter, the mean value N F and standard deviation of the three m technical triplicates were calculated. To extrapolate to the total area in contact with the sample, Eq. ( 2 ) was used: $$N = \frac{N_{\mathrm{F}}}{V_{\mathrm{F}}}$$ (2) The derived particle ( N ) numbers per litre melted ice were calculated from the mean value N F and the volume fraction of one measurement field V F from the overall volume. Particle numbers derived from blank samples were subtracted from N F . To estimate errors of this conversion an error propagation was performed. Error propagation for calculation of particle numbers ( N ) To estimate errors for the technical triplicate, as well as the overall calculation and individual particle numbers ( N ) values, Eq. ( 3 ) was used: $${\mathrm{\Delta }}N = \sqrt {\left( {\frac{1}{{V_{\mathrm{F}}}} \times \Delta N_{\mathrm{F}}} \right)^2 + \left( { - \frac{{N_{\mathrm{F}}}}{{V_{\mathrm{F}}^2}} \times \Delta V_{\mathrm{F}}} \right)^2}$$ (3) Δ N is the error of particle numbers per litre, Δ N F is the standard deviation of the technical triplicate, and Δ V F is the error of the conversion from full sample volume to volume per field. Δ V F was calculated individually for each filter based on Eqs. ( 4 ) and ( 5 ) $$V_{\mathrm{F}} = \frac{V}{{\mathrm{CF}}}$$ (4) $$\Delta V_{\mathrm{F}} = \sqrt {\left( {\frac{1}{{\mathrm{CF}}} \times {\mathrm{\Delta }}V} \right)^2 + \left( { - \frac{V}{{\mathrm{CF}^2}} \times {\mathrm{\Delta }}\mathrm{CF}} \right)^2}$$ (5) \(\mathrm{CF}\) is the calculation factor derived from the measurement field size and concentrated filter area, and the sample volume of the investigated sample fraction. For Δ V , an error of 0.01 L was estimated due to the gravimetric determination. For ΔCF, an error value of 0.267 was found that includes the errors from the measurement field size determination and filter cake diameter. Quality factor thresholds for image analysis To maintain a 95% confidence interval for the chosen type of sample purification, several fields were manually reanalyzed 28 . For image analysis, the following quality factor thresholds differing from 600 were used (points with lower hit qualities were excluded from analysis): Polyethylene type 1 = 1100, polyethylene type 2 = 1350, PE-Cl = 1310, PC = 700, PA = 1020, PVC = 800, PES = 800, quartz = 700, EVA = 900, rubber type 1 = 1190, rubber type 2 = 1300, polyethylene type 3 = 1070. In addition, polychloroprene had to be excluded due to the low hit qualities. Exclusion of rayon and fibres in this study The polymer r ayon is often included in the classification of MP found in the marine realm. For example, it accounted for up to 30% of the MP found in samples from the Arctic 2 , 26 . However, FTIR-based studies showed that 30% of suspected rayon fibres turned out to be cellulose, which is considered a natural product. Since cellulose and the semi-synthetic polymer rayon, have almost identical FTIR spectra 45 , we resigned from identifying this particular compound in our study. To compare ours with the previous study of MP in sea ice, we also excluded rayon from the MP identified in Obbard et al. 2 . We digitized the data from Fig. 2 (which according to the erratum are given as per liter) and up-scaled the numbers to N m −3 for Fig. 1b . The visible inspection classifies usually a large fraction of the MP as fibres 2 , 26 , which cannot be identified with our way of applying the FTIR spectra, since fibres are not distributed flat on the surface of the filter and regions out of focus of the IR-beam were hardly or not detectable. The typical diameter of fibres was around 10–20 µm and the increased scattering due to their shape was not suitable for identification with our analytical approach. Statistical analyses Multivariate analyses were performed by using the software package Primer 7.012 (Primer-E), and univariate analyses by using Statistica 11 (Statsoft). Polymer-specific MP numbers and numbers of particles in the different size classes were fourth root transformed; ancillary variables were normalized before multivariate analyses. For clustering (group average), Bray-Curtis similarities were used for MP-related data and Euclidean distances for ancillary variables. Calculation of diversity indices (richness, Shannon–Wiener H') were performed by using the PRIMER routine DIVERSE. The exploratory similarity profile test (SIMPROF) was applied to detect structures in the datasets. A significance level of 5% was used to test the SIMPROF statistic. SIMPROF groups (of ancillary variables; SIMPROF av ) were further used in ANOVAs for testing group differences of MP-related data and ancillary variables. Spearman rank correlations and ANOVAs were calculated by using non-transformed data. Data availability The authors confirm that all data underlying this study are fully available without restriction. All data can be downloaded from the public repository PANGAEA; .

Experts at the Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research (AWI) have recently found higher amounts of microplastic in Arctic sea ice than ever before. However, the majority of particles were microscopically small. Ice samples from five regions throughout the Arctic Ocean contained up to 12,000 microplastic particles per litre of sea ice. Further, the types of plastic showed a unique footprint in the ice, allowing the researchers to trace them back to possible sources. This involves the massive garbage patch in the Pacific Ocean; the high percentage of paint and nylon particles pointed to intensified shipping and fishing activities in some parts of the Arctic Ocean. The new study has just been released in the journal Nature Communications. "During our work, we realised that more than half of the microplastic particles trapped in the ice were less than a twentieth of a millimetre wide, which means they could easily be ingested by Arctic microorganisms like ciliates, but also by copepods," says AWI biologist and first author Dr. Ilka Peeken. The observation is a very troubling one because, as she explains, "No one can say for certain how harmful these tiny plastic particles are for marine life, or ultimately also for human beings." The AWI researcher team had gathered the ice samples in the course of three expeditions to the Arctic Ocean on board the research icebreaker Polarstern in the spring of 2014 and summer of 2015. They hail from five regions along the Transpolar Drift and the Fram Strait, which transports sea ice from the Central Arctic to the North Atlantic. Infrared spectrometer reveals heavy contamination with microparticles Microplastic refers to plastic particles, fibres, pellets and other fragments with a length, width or diameter ranging from only a few micrometres—thousandths of a millimetre—to under five millimetres. A considerable amount of microplastic is released directly into the ocean by the gradual deterioration of larger pieces of plastic. But microplastic can also be created on land—e.g., by laundering synthetic textiles or abrasion of car tires. The plastic initially floats through the air as dust, and is then blown to the ocean by the wind, or finds its way there through sewer networks. In order to determine the exact amount and distribution of microplastic in sea ice, the AWI researchers were the first to analyse the ice cores layer by layer using a Fourier transform infrared spectrometer (FTIR), a device that bombards microparticles with infrared light and uses a special mathematical method to analyse the radiation they reflect back. Depending on their makeup, the particles absorb and reflect different wavelengths, allowing every substance to be identified by its optic fingerprint. "Using this approach, we also discovered plastic particles that were only 11 micrometres across. That's roughly one-sixth the diameter of a human hair, and also explains why we found concentrations of over 12,000 particles per litre of sea ice—which is two to three time higher than what we'd found in past measurements," says Gunnar Gerdts, in whose laboratory the measurements were carried out. Surprisingly, the researchers found that 67 percent of the particles detected in the ice belonged to the smallest-scale category of 50 micrometres and smaller. Germany's research icebreaker POLARSTERN above the Lomonossov Ridge in the central Arctic Ocean. Credit: Alfred-Wegener-Institut/Ruediger Stein Ice drift and the chemical fingerprint offer clues to pollutants' regions of origin The particle density and composition varied significantly from sample to sample. The researchers determined that the plastic particles were not uniformly distributed throughout the ice core. "We traced back the journey of the ice floes we sampled and can now safely say that both the region in which the sea ice is initially formed and the water masses in which the floes drift through the Arctic while growing, have an enormous influence on the composition and layering of the encased plastic particles," relates Ilka Peeken. The researchers also learned that ice floes, which are driven in the Pacific water masses of the Canadian Basin, contain particularly high concentrations of polyethylene particles. Polyethylene is used in packaging material. As the experts write in their study, "Accordingly, we assume that these fragments represent remains of the so-called Great Pacific Garbage Patch, and are pushed along the Bering Strait and into the Arctic Ocean by the Pacific inflow." In contrast, the scientists predominantly found particles from ship's paint and nylon waste from fishing nets in ice from the shallow marginal seas of Siberia. "These findings suggest that both the expanding shipping and fishing activities in the Arctic are leaving their mark. The high microplastic concentrations in the sea ice can thus not only be attributed to sources outside the Arctic Ocean. Instead, they also point to local pollution in the Arctic," says Ilka Peeken. The researchers found a total of 17 different types of plastic in the sea ice, including packaging materials like polyethylene and polypropylene, but also paints, nylon, polyester, and cellulose acetate, the latter primarily used in the manufacture of cigarette filters. Taken together, these six materials accounted for roughly half of all the microplastic particles detected. According to Ilka Peeken, "The sea ice binds all this plastic litter for two to a maximum of 11 years—the time it takes for ice floes from the marginal seas of Siberia or the North American Arctic to reach the Fram Strait, where they melt." But conversely, this also means that sea ice transports large quantities of microplastic to the waters off the northeast coast of Greenland. The researchers can't yet say whether the released plastic particles subsequently remain in the Arctic or are transported farther south; in fact, it seems likely that the plastic litter begins sinking into deeper waters relatively quickly. "Free-floating microplastic particles are often colonised by bacteria and algae, which makes them heavier and heavier. Sometimes they clump together with algae, which makes them drift down to the seafloor much faster," explains AWI biologist and co-author Dr. Melanie Bergmann. The observations made by researchers at the AWI's deep-sea network HAUSGARTEN in the Fram Strait lend additional weight to this thesis. As Melanie Bergmann relates, "We recently recorded microplastic concentrations of up to 6500 plastic particles per kilogram of seafloor; those are extremely high values."

10.1038/s41467-018-03825-5

Medicine

Researchers identify potential anti–cancer target

S Vaidyanathan et al. In vivo overexpression of Emi1 promotes chromosome instability and tumorigenesis, Oncogene (2016). DOI: 10.1038/onc.2016.94 Journal information: Nature , Oncogene

http://dx.doi.org/10.1038/onc.2016.94

https://medicalxpress.com/news/2016-06-potential-anticancer.html

Abstract Cell cycle genes are often aberrantly expressed in cancer, but how their misexpression drives tumorigenesis mostly remains unclear. From S phase to early mitosis, EMI1 (also known as FBXO5) inhibits the anaphase-promoting complex/cyclosome, which controls cell cycle progression through the sequential degradation of various substrates. By analyzing 7403 human tumor samples, we find that EMI1 overexpression is widespread in solid tumors but not in blood cancers. In solid cancers, EMI1 overexpression is a strong prognostic marker for poor patient outcome. To investigate causality, we generated a transgenic mouse model in which we overexpressed Emi1. Emi1-overexpressing animals develop a wide variety of solid tumors, in particular adenomas and carcinomas with inflammation and lymphocyte infiltration, but not blood cancers. These tumors are significantly larger and more penetrant, abundant, proliferative and metastatic than control tumors. In addition, they are highly aneuploid with tumor cells frequently being in early mitosis and showing mitotic abnormalities, including lagging and incorrectly segregating chromosomes. We further demonstrate in vitro that even though EMI1 overexpression may cause mitotic arrest and cell death, it also promotes chromosome instability (CIN) following delayed chromosome alignment and anaphase onset. In human solid tumors, EMI1 is co-expressed with many markers for CIN and EMI1 overexpression is a stronger marker for CIN than most well-established ones. The fact that Emi1 overexpression promotes CIN and the formation of solid cancers in vivo indicates that Emi1 overexpression actively drives solid tumorigenesis. These novel mechanistic insights have important clinical implications. Introduction Cell cycle progression is regulated by the oscillation in expression and activity of various proteins, such as cyclins and cyclin-dependent kinases. 1 Proteasome-mediated protein degradation is critical in this process. The anaphase-promoting complex/cyclosome (APC/C) catalyzes the sequential proteolysis of substrates from late G 2 to early G 1 phases of the cell cycle. 2 At different stages, Cdc20 and Cdh1 act as co-factors required for APC/C activity. Several APC/C inhibitors, such as Mad2 and Emi1 (early mitotic inhibitor 1, also known as Fbxo5), compete with the APC/C for Cdc20 and Cdh1 binding, thereby stabilizing APC/C substrates. 2 In addition, Emi1 potently inhibits APC/C by blocking its catalytic site, by competitively preventing APC/C substrates from binding to APC/C co-receptors and by actively suppressing substrate mono- and polyubiquitination. 3 , 4 , 5 Emi1 is the principal inhibitor of APC/C activity from S phase to early mitosis, whereas Mad2 fulfills this role later during mitosis. 6 Defective cell cycle regulation is a hallmark of cancer. Consistently, APC/C substrates and regulators are often aberrantly expressed in tumors. 7 , 8 During mitosis, the mitotic checkpoint tightly regulates APC/C activity. It prevents precocious sister chromatid separation by ensuring that the APC/C does not prematurely target securin for degradation, thereby preventing chromosome missegregation. Thus, the aberrant expression of mitotic APC/C regulators can lead to chromosome instability (CIN), which may fuel cancer progression. The observations that CIN can promote tumorigenesis in mice and that at least two-thirds of human cancers are aneuploid indicate that CIN is an important driver of tumor evolution. 9 , 10 , 11 , 12 , 13 , 14 A number of studies have focused on unraveling the cell biological and in vivo function of Emi1 using loss-of-function and genetic ablation experiments, respectively. 5 , 15 , 16 , 17 , 18 , 19 , 20 This has provided invaluable new insights. For instance, it has become clear that Emi1-mediated stabilization of APC/C is required for mitotic progression during embryogenesis and prevents DNA re-replication and polyploidy. 16 , 17 , 18 However, it has remained unclear how Emi1 misexpression might be involved in the biology of cancer development in vivo . Here, we find that EMI1 is overexpressed in the vast majority of human solid tumors, but not blood cancers, and that this is a strong marker for poor clinical outcome. To mimic this expression status in human cancers, we generated an inducible Emi1 overexpression mouse model. We find that in vivo overexpression of Emi1 promotes CIN and the formation of a wide variety of aneuploid solid tumors due to delayed chromosome alignment followed by chromosome missegregation. This indicates that Emi1 overexpression is not a cancer-associated bystander but actively promotes CIN and tumorigenesis in vivo . Results EMI1 is overexpressed in a broad range of human cancers We assessed EMI1 mRNA expression in human cancers using 115 previously published unique expression analyses. 21 Together, EMI1 expression levels in 7403 tumor samples were compared with those in 1467 normal control samples of matched tissue type. Of the 115 studies, 102 (89%) showed that EMI1 is significantly overexpressed and 13 (11%) reported that EMI1 is underexpressed ( Figure 1a and Supplementary Table 1 ). EMI1 overexpression is observed in a range of cancers and tumor subtypes ( Figure 1b , Supplementary Table 1 and Supplementary Data set 1 ). Similarly, using the cancer genome atlas (TCGA) breast cancer data set, 22 we find that EMI1 levels are significantly increased in all four major subtypes ( P ⩽ 0.0019, t -test; Figure 1c ). Figure 1 EMI1 overexpression is a marker for poor prognosis in human solid but not blood cancers. ( a ) Fractions of 115 independent studies reporting under- and overexpression of EMI1 mRNA in cancer. Fractions are shown for all cancers and separately for solid and blood cancers. P value: Fisher’s exact test. ( b ) Box plot showing EMI1 expression levels in normal brain and glioblastoma (see also Supplementary Data set 1 for details). P value: t -test. ( c ) EMI1 expression levels in breast cancer per subtype in the TCGA data set. 22 Bottom P values compared with normal breast tissue. Top P values between groups: one-way analysis of variance Tukey’s multiple comparisons test. ^ P value<2.2 × 10 −16 . ( d ) Box plot showing reduced EMI1 expression in chronic lymphocytic leukemia (CLL) compared with normal control peripheral blood mononuclear cells (PBMCs; see Supplementary Data set 1 for details). P value: t -test. ( e ) Immunohistochemistry for EMI1 on five types of human solid cancers comparing matched normal control tissue to tumor tissue. Insets show enlargements of areas in the dashed boxes. For extended data, see Supplementary Figure 1 . Scale bar, 100 μm. ( f ) Quantification of extended EMI1 immunohistochemistry analysis using the multiplicative IHC quickscore as previously described. 45 See also Supplementary Figure 1 . P values above bars compared with normal, other comparisons as indicated by horizontal bars. All P values: t -test. ( g ) Western blot analysis on primary normal control and tumor breast tissue of various breast cancer subtypes (left) and quantification of actin-normalized EMI1 levels (right). Numbers above bars indicate fold increase compared with normal control expression level. P values: t -test. ( h and i ) Survival curves for patients with high (red) and low (green) EMI1 mRNA expression in breast and lung cancer, respectively. P values: log-rank test. See also Supplementary Table 2 . Full size image The etiology and genetic makeup of solid and hematological cancers differ considerably. 14 , 23 Interestingly, EMI1 mRNA expression is also misregulated differently in these types of cancer. Of the above 100 solid cancer studies, 95 report significant EMI1 overexpression, whereas this fraction is significantly lower for hematological cancers (7 out of 15, 47%; P <0.0001, Fisher’s exact test; Figures 1a and d , Supplementary Table 1 and Supplementary Data set 1 ). Thus, solid and blood cancers not only show differences in cancer genomics and pathology, but also in how EMI1 is misexpressed. Our primary interest in solid cancers led us to test whether EMI1 overexpression is sustained at the protein level in these cancers. We confirmed that this is the case using both western blot analysis on primary tumor samples and immunohistochemistry on tissue microarrays across a panel of five different solid cancer types ( Figure 1e–g and Supplementary Figure 1 ). The western blot analysis indicates that EMI1 is expressed approximately two- to six-fold higher in tumors compared with normal tissue ( Figure 1g ). EMI1 overexpression is a marker for poor prognosis in human solid cancers We next compared EMI1 mRNA expression levels in low-grade and high-grade cancers. In high-grade cancers, EMI1 is overexpressed to higher degrees than in low-grade cancers, as observed in 28 data sets for cancers that arise in 11 distinct tissue types ( Figure 1c and Supplementary Data set 2 ). To more directly assess the prognostic value of EMI1 mRNA overexpression for patient outcome, we performed survival analysis on 60 data sets. This revealed that EMI1 overexpression is a marker for poor prognosis for at least nine forms of solid cancers and for several types of patient survival ( Figures 1h and i , Supplementary Figure 2 and Supplementary Table 2 ). Using our largest data set ( n =3751 patients; Supplementary Table 2 ), we further assessed the prognostic strength of EMI1 mRNA overexpression relative to three well-established clinical markers. First, with a P value of 9.32 × 10 −13 (log-rank test) EMI1 performs nearly as well as the well-established prognostic marker KI67 ( P =4.15 × 10 −14 ). Also, when survival is adjusted for the Nottingham Prognostic Index or by Adjuvant! Online, which each predict patient outcome using several clinical parameters, as previously described, 24 , 25 it remains highly significant ( P =6.35 × 10 −5 and P =9.29 × 10 −5 ; Supplementary Table 3 ). Importantly, as observed by both immunohistochemistry and western blot analysis, these observations are sustained at the protein level with high-grade/metastatic tumors expressing higher EMI1 protein levels than low-grade cancers ( Figures 1f and g and Supplementary Figure 1 ). Thus, together these data indicate that EMI1 overexpression is a strong independent clinical prognostic marker for solid cancers. Prognostic significance of EMI1 overexpression is only partly cell cycle dependent Importantly, as EMI1 is expressed from S phase to early mitosis, it is possible that increased EMI1 expression is simply a reflection of its increased abundance in cycling cells. Consistent with this, EMI1 expression is associated with KI67 expression, a marker for cycling cells, in at least six different cancer types ( P ⩽ 0.0082, Spearman correlation; Supplementary Figure 3A ). However, for several reasons, we believe that EMI1 overexpression in tumors is only partly cell cycle dependent. First, in tumors, EMI1 levels increase more than KI67 levels increase, as observed by a significantly higher EMI1/KI67 expression ratio in tumors than in normal tissues ( P ⩽ 0.0241, Mann–Whitney test; Supplementary Figure 3B ). Second, in TCGA breast tumors that express normal levels of KI67, EMI1 expression is significantly higher than in normal breast tissue ( P =9.713 × 10 −12 , t -test; Supplementary Figure 3C ). Third, 57 out of 107 (53%) TCGA breast tumors with normal KI67 expression express EMI1 above normal levels ( P <0.0001, Fisher’s exact test; Supplementary Figure 3C ). Fourth, when we perform survival analysis adjusted for a previously described 42-gene cell cycle signature, 26 the prognostic strength of EMI1 overexpression remains highly significant ( P =0.0175, log-rank test), whereas that of the KI67 marker for cycling cells does not ( Supplementary Figure 3D ). Fifth, in 14 out of 102 data sets (13.7%) showing EMI1 overexpression, KI67 expression is not significantly increased ( Supplementary Table 4 ). For instance, one data set shows a 2.2-fold increase in EMI1 expression ( P =0.002, t -test), whereas KI67 levels are not significantly altered ( P =0.428; Supplementary Table 4 ). Sixth, in 9 out of 27 data sets comparing low-grade to high-grade cancers (33%), EMI1 is significantly overexpressed, whereas KI67 is not ( Supplementary Table 5 ). Finally, Emi1 is significantly underexpressed in 13 cancer data sets, including highly proliferative leukemias ( Supplementary Table 1 ) and in three blood cancer data sets reduced, rather than elevated, EMI1 expression significantly correlates with poor patient survival ( Supplementary Table 2 ). Thus, collectively these data indicate that EMI1 overexpression is a strong prognostic marker for human solid cancers and that this is not solely due to its periodic expression in cycling cells. Generation of an Emi1 overexpression mouse model Although the above data demonstrate a strong link between EMI1 overexpression and cancer patient prognosis, they are correlative and not causative. To study whether Emi1 overexpression can actively promote solid tumorigenesis in vivo , we generated a transgenic mouse model in which we could overexpress Emi1 in an inducible manner. First, we cloned the murine Emi1 cDNA into the pTRE vector, containing a CMV-based tetracycline-regulatory promoter, an N-terminal HA-tag and a poly-adenylation signal ( Figure 2a ). The small HA-tag readily facilitates detection of transgene expression (see below) and the addition of small or large tags to the Emi1 N-terminus, which is exposed even when bound to the APC/C, 3 does not interfere with Emi1 activity. 4 , 5 , 6 Transfection of this construct into HeLa-Tet-Off cells showed that exogenous Emi1 expression was doxycycline-repressible ( Supplementary Figure 4A ). We next released a fragment containing the inducible Emi1 transgene ( Figure 2a ) from the pTRE vector, injected it into fertilized oocytes and implanted the latter into pseudopregnant mice. Of 51 founder animals that were born, 11 (22%) tested positive for transgene integration, as determined by PCR on their genomic DNA ( Supplementary Figure 4B ). Southern blot analysis on the genomic DNA of these animals, as well as of selected mice that tested negative by PCR and wild-type (WT) control mice, confirmed that at least 9 out of the 11 positive animals (18% of 51 founders) carried the transgene ( Figure 2b and Supplementary Figure 4C ). Figure 2 Generation of an Emi1 overexpression mouse model. ( a ) DNA fragment used for the generation of Emi1 transgenic mice. The inducible promoter (P) is CMV-based and contains 7 tetracycline-response elements (TRE 7 ). ( b ) Southern blot showing genomic integration of the transgene in a subset of mice. Equal loading is shown in Supplementary Figure 4C and by the 1210-bp internal control band. The transgene band is 799 bp. ( c ) Semi-quantitative RT–PCR on RNA from tissues of bitransgenic Tet-On mice on a regular or doxycycline diet for 4 weeks after weaning, as indicated. In, intestine; Li, liver; Lu, lung; Ki, kidney; Th, thymus. ( d ) Western blots on tissue extracts from bitransgenic TRE-Emi1/rtTA (Tet-On) and TRE-Emi1/tTA (Tet-Off) mice on a regular or doxycycline diet for 4 weeks after weaning, as indicated. Full size image To be able to induce Emi1 transgene expression in vivo , we crossed the TRE-Emi1 transgenic mice with CMV-rtTA (Tet-On system) or CMV-tTA (Tet-Off system) transgenic animals, in which the CMV promoter drives the expression of the reverse tetracycline transactivator (rtTA) or the tetracycline transactivator (tTA), respectively. 27 This generated TRE-Emi1/CMV-rtTA and TRE-Emi1/CMV-tTA bitransgenic mice. In these contexts, Emi1 transgene expression could be induced or repressed with doxycycline, respectively, as observed in mouse tissues at the mRNA and protein levels by semi-quantitative reverse transcriptase–PCR (RT–PCR) and western blot analysis, respectively ( Figures 2c and d ). In vivo Emi1 overexpression promotes tumorigenesis To assess whether in vivo overexpression of Emi1 is sufficient to initiate tumor formation, we aged groups of TRE-Emi1/CMV-rtTA and TRE-Emi1/CMV-tTA bitransgenic mice. Each cohort included experimental and control animals in which Emi1 transgene expression was induced or not induced by feeding the mice a regular or a doxycycline-containing diet after weaning. The phenotypes that developed in the Tet-On and Tet-Off backgrounds were indistinguishable ( Supplementary Table 6 and Supplementary Figures 5A–C ). We therefore combined our analyses on a total of 45 Emi1-overexpressing (Emi1-OE) mice and 44 genetically identical non-Emi1-overexpressing control (Ctrl) mice, as well as 16 WT control mice. Emi1-OE mice developed a wide variety of cancerous lesions in a range of organs ( Figure 3a and Table 1 ). Although these developed at long latencies, Emi1-OE mice succumbed to cancer significantly earlier than bitransgenic Ctrl ( P <0.0001, log-rank test) and WT control mice ( P =0.0004; Figure 3b and Supplementary Figure 5 ). Emi1-OE mice also died earlier that Ctrl and WT mice irrespective of the cause of death ( P <0.0005; Supplementary Figure 6 ). In addition, 80% of the Emi1-OE mice developed tumors versus 50% of the bitransgenic Ctrl animals ( P =0.0028, Fisher’s exact test; Figure 3c ). Compared with Ctrl mice, Emi1-OE mice had a twofold increased tumor burden and their tumors proliferated more rapidly and were larger at the time of death ( P =0.0006, P =0.0089, P =0.0148, respectively, t -test; Figures 3d–f ). Figure 3 In vivo Emi1 overexpression promotes tumorigenesis. ( a ) Macroscopic and hematoxylin and eosin (H&E)-stained microscopic images of selected lesions from Emi1-overexpressing mice. Scale bars, 1 cm. ( b ) Tumor-free survival of Emi1-overexpressing (Emi1-OE, red), non-Emi1-overexpressing control (Ctrl, green) and wild-type control (WT, black) mice. P values: log-rank test. See also Supplementary Figure 5 . ( c ) Cancer penetrance. P value: Fisher’s exact test. ( d ) Tumor burden. Mean values ( μ ) are indicated±s.e.m. P value: t -test. ( e ) Tumor growth rate. Means±s.e.m. are shown. P value: t -test. ( f ) Time line of tumor development in Ctrl and Emi-OE mice. P value: t -test. ( g ) Cancer penetrance per major tumor type. Lymph, lymphoma; Sarc, sarcoma; Carcin, carcinoma; Aden, adenoma. P values: Fisher’s exact test; NS, not significant. ( h , i ) Carcinoma and adenoma burden as in d . ( j - l ) Incidences of tumor-bearing animals with metastases ( j ), mice with cancer-associated immunological abnormalities ( k ) and hyperplastic or neoplastic lesions ( l ). P values: Fisher’s exact tests. Full size image Table 1 Incidence of tumors in Emi1-overexpressing and Ctrl mice Full size table The cancer incidences by tumor cell of origin were also higher in Emi1-OE mice for all major tumor types: lymphoma, sarcoma, carcinoma and adenoma ( Figure 3g ). However, the differences were only statistically significant for carcinomas and adenomas ( P =0.0129 and P =0.0133, Fisher’s exact test; Figure 3g ). Similarly, the tumor burden was elevated for all four of these types of cancer, yet the differences were significant only for carcinomas and adenomas ( P =0.0056 and P =0.0073, t -tests; Figures 3h and i and Supplementary Figures 7A and B ). In addition, the fraction of animals with liver tumors significantly increased by more than 2.5-fold, from 11–29% ( P =0.0396, Fisher’s exact test; Supplementary Figure 7C ) and a significantly higher number of Emi1-OE female mice developed lesions in the ovaries ( P =0.0276, Fisher’s exact test; Supplementary Figure 7D ). However, as the liver tumors included sarcomas and carcinomas and the ovarian lesions comprised cysts, cystadenomas and hemangiosarcomas, they originated from different cell types. Primary cancers in Emi1-OE mice were more aggressive than in Ctrl animals, as evidenced by their increased proliferation rate ( Figure 3e ) and propensity to metastasize ( P =0.0343, Fisher’s exact test; Figure 3j ). The latter may in fact be an underestimation, because micro-metastases may have been missed. Consistent with an Emi1 overexpression-induced increase in malignancy, cancer-associated immunological abnormalities (other than lymphoma or leukemia), such as inflammation and lymphocyte infiltration, were also significantly elevated ( P =0.0003, Fisher’s exact test; Figure 3k ). Finally, Emi1-OE mice showed an increased incidence of hyperplasia and neoplasia, such as mammary and prostate intraepithelial neoplasias ( P =0.0009, Fisher’s exact test; Figure 3l ; not included in Table 1 ). Such abnormalities have the potential to develop into malignant lesions. Taken together, in vivo overexpression of Emi1 promotes the development of a broad range of solid cancers and cancer-associated abnormalities, such as metastasis and inflammation. Tumors in Emi1-overexpressing mice are aneuploid Using peri-centromeric probes against chromosomes 12, 16 and 17, we next performed interphase-fluorescent in situ hybridization (FISH) on the mouse tumor sections. This revealed that, in contrast to Ctrl tumors, tumors from Emi1-OE mice were highly aneuploid ( Figures 4a and b ). Figure 4 Tumors from Emi1-overexpressing mice are aneuploid. ( a ) Interphase-FISH on tumor tissue sections from mice, as indicated. 4,6-Diamidino-2-phenylindole (DAPI): white; chromosomes 12, 16 and 17: blue, green and red, respectively. ( b ) Quantification of the fractions of aneuploid cells in tumor tissue sections. Each bar reflects the average of at least 100 cells per tumor. P value: t -test. Full size image Tumors in Emi1-overexpressing mice show evidence of CIN To study tumor cell mitoses in vivo , we subjected tissue sections from Emi1-OE and Ctrl tumors to hematoxylin and eosin staining and immunofluorescence using markers for DNA, microtubules and centrosomes. Aside from normal anaphases, we observed anaphases with anaphase bridges, lagging chromosomes and chromosomes that did not segregate with the main chromosome masses—henceforth termed ‘not in mass’ ( Figures 5a and b ). Anaphase bridges did not occur at a significantly elevated frequency and the total number of anaphases per mitosis was also not increased in Emi1-OE tumors ( Figure 5b and Supplementary Figure 8A ). In contrast, the frequencies of lagging and ‘not in mass’ chromosomes were significantly higher than in Ctrl tumors ( t -test; Figure 5b ). Overall, 56% of Emi1-OE tumors exhibited mitotic abnormalities compared with 14% of Ctrl tumors ( P =0.0101, t -test; Figure 5b ). Emi1-OE tumors also showed a wider range of mitotic abnormalities with multiple types of abnormalities co-occurring in 10% of the cells ( Figure 5b ) and significantly more abnormalities per mitosis ( P =0.0047, t -test; Figure 5c ). In addition, the numbers of lagging chromosomes and chromosomes not segregating with the main chromosome masses per cell were significantly elevated ( P =0.0335 and P =0.0043, t -test; Figures 5d and e ). Emi1-OE tumor cells also showed a broader distribution of cells with multiple lagging and 'not in mass' chromosomes ( Figures 5f and g ). Figure 5 Tumors from Emi1-overexpressing mice show evidence of chromosome missegregation. ( a ) Microscopic images of mitotic cells in hematoxylin and eosin (H&E)-stained mouse tumor sections showing examples of normal and abnormal mitoses. Some cells simultaneously showed more than one abnormality, as exemplified in the image on the right with both an anaphase bridge—chromatin connecting the two main chromosome masses—and a chromosome separate from and not segregating with the two chromosome masses (‘not in mass’). Scale bar, 5 μm. ( b – e ) Quantification of mitotic abnormalities (Ctrl n =146, Emi1-OE n =114 mitoses): frequencies and types of mitotic abnormalities ( b ; P value summaries based on frequencies irrespective of whether abnormality occurred as single event or in conjunction with other abnormalities), numbers of abnormalities per mitosis ( c ), lagging chromosomes per cell ( d ) and chromosomes not segregating with the two chromosomal masses per cell ( e ). P values: t -test. ( f , g ) Distributions of the numbers of lagging chromosomes and chromosomes not in mass per cell, respectively. P values: t -test. ( h ) Fractions of cells in early mitosis. P value: t -test. NS, not significant; * P <0.05; ** P <0.01; **** P <0.0001. Full size image Aside from the above abnormalities, we noticed that mitotic cells in Emi1-OE tumors were often in early mitosis, as their chromosomes were condensed while their centrosomes had not yet separated ( Figure 5h and Supplementary Figures 8B and C ). Compared with Ctrl tumors, the frequency of cells in early mitosis increased 13-fold, from 2.6 to 35% ( P =0.0298, t -test; Figure 5h ). Thus, in vivo Emi1 overexpression promotes CIN by inducing lagging and missegregating chromosomes and these defects may have an early mitotic origin. Prometaphase and metaphase defects underlie EMI1 overexpression-mediated CIN To further investigate how EMI1 overexpression promotes CIN, we overexpressed EMI1 in Hela cells that stably expressed GFP-labeled histone H2B ( Supplementary Figure 9A ) and monitored cell division using live-cell microscopy. This unveiled several striking differences between EMI1-OE and control cells. First, EMI1-OE cells progressed through mitosis more slowly ( P =0.0004, t -test) due to delays in both chromosome alignment at the metaphase plate and subsequent anaphase onset ( P =0.0008 and P =0.0028, respectively; Figures 6a and b ). On average, the time interval between nuclear envelope breakdown and anaphase onset was ~6 min longer in EMI1-OE cells, whereas the time from anaphase onset to completion of cytokinesis was not affected ( Figure 6b ). Figure 6 In vitro EMI1 overexpression delays mitotic progression and promotes chromosome instability. ( a ) Time-lapse microscopic images of histone H2B-GFP-expressing Hela cells that do (EMI1-OE) or do not (Ctrl) overexpress EMI1. Time is indicated in hours:minutes with nuclear envelope breakdown (NEBD) set at time point zero. Note the difference in timing between the Ctrl and EMI1-OE cells. ( b ) Quantification of time-lapse microscopy analysis on 100 cells per condition. Time is shown in minutes. Error bars: s.e.m. P values: t -test. ( c ) Frequencies of cells with lagging chromosomes (top left image), anaphase bridges (bottom left image) or chromosomes not in mass. P -value: Fisher’s exact test. ( d – f ) Frequencies of cells that arrest in mitosis, die and show abnormal mitoses (not including mitotic delay), respectively. P values: Fisher’s exact test. See also Supplementary Figure 9 . Full size image EMI1-OE cells also showed significant increases in anaphase bridges, lagging or 'not in mass' chromosomes, mitotic arrest and cell death ( P =0.0048, P =0.0297 and P =0.0027, Fisher’s exact test; Figures 6c–e ). We also observed small increases in the formation of micronuclei, binucleate cells and tripolar spindles. However, those were not statistically significant ( Supplementary Figures 9B-D ). Overall, there was a profound increase in the frequency of abnormal mitoses ( P =0.0004; Figure 6f ). Taken together, EMI1 overexpression delays chromosome alignment at the metaphase plate and anaphase onset and this is followed by chromosome missegregation. EMI1 overexpression is a strong marker for CIN in human solid cancers Based on the expression levels of 70 genes, the CIN70 gene expression signature is a widely used measure of CIN in human tumor samples. 28 , 29 , 30 Consistent with previous studies, 29 we find that the CIN70 signature stratifies breast cancer subtypes according to clinical outcome ( Supplementary Figure 10 ). Interestingly, this pattern is similar to the one for EMI1 expression ( Figure 1c ). We therefore compared CIN70 score and EMI1 expression level directly. This revealed a strong linear correlation between these parameters ( r =0.7792, P <0.0001, Pearson’s correlation; Figure 7a ). Figure 7 EMI1 expression strongly correlates with a CIN signature in human cancers and is a stronger marker of CIN than most well-established ones. ( a – h ) Scatter plots of EMI1 expression levels and corresponding CIN70 scores for TCGA data sets of: ( a ) breast carcinoma, ( b ) ovarian cancer, ( c ) glioblastoma multiforme, ( d ) low-grade glioma, ( e ) colon, ( f ) kidney, ( g ) lung and ( h ) endometrium cancers. Sample size and r for linear regression are indicated. P values: Pearson’s correlation coefficient. ( i ) Top 15 genes most significantly co-expressed with EMI1 in breast cancer. CIN70 genes are highlighted. See Supplementary Data set 3 for extended list. ( j ) Performance of EMI1 as a marker for CIN in comparison to the 70 CIN70 genes. Full size image Using a total of 2099 samples, we also analyzed TCGA data sets from seven other forms of cancer. For these cancers, the EMI1 expression levels and CIN70 score relationships were also linear and highly significant (0.6349 ⩽ r ⩽ 0.9166, all P values <0.0001, Pearson’s correlation; Figures 7b and h ). Furthermore, 6 CIN70 genes ranked in the top 15 (40%) of most significantly co-expressed genes with EMI1 ( Figure 7i and Supplementary Data set 3 ). EMI1 is not among the 70 genes whose expression levels contribute to the CIN70 score. 28 To compare how EMI1 would rank among the CIN70 genes, we plotted the expression level of each of the CIN70 genes individually against the CIN70 score, as in Figures 7a–h for EMI1, and calculated each R 2 for best fit for four solid cancer data sets ( Supplementary Data set 4 ). Surprisingly, some CIN70 genes poorly correlated with the CIN70 signature with the average R 2 for the CIN70 genes ranging from 0.037 to 0.742. With an average R 2 of 0.562, EMI1 ranks in the 74.4th percentile in this range of the widely accepted top 70 markers of CIN ( Figure 7j and Supplementary Data set 4 ). This indicates that EMI1 is a stronger marker for CIN in human solid cancers than most well-established ones. Discussion We find that EMI1 overexpression is widespread in human solid cancers and strongly correlates with poor patient prognosis in a manner that is only partly cell cycle dependent. Consistent with this, Emi1-overexpressing mice develop a wide variety of solid cancer types, most significantly adenomas and carcinomas. Compared with isogenic control mice, Emi1 overexpression significantly increases cancer penetrance, tumor size, overall tumor burden, carcinoma and adenoma burden, metastatic potential, cancer-associated immunological changes and the incidence of pre-malignant lesions. Our data indicate that, rather than being a cell proliferation- or cancer-associated bystander, Emi1 overexpression actively promotes genomic instability and tumorigenesis in vivo . Interestingly, reduced EMI1 mRNA expression often predicts poor prognosis for hematological malignancies. Consistent with this, Emi1 overexpression in mice does not predispose to blood cancer development. These differences could provide an explanation for the widely different aneuploidy signatures and etiologies of these cancer types. 14 , 23 We note, however, that we have not demonstrated decreased EMI1 protein levels in human blood cancers or increased Emi1 levels in Emi1 transgenic mouse hematopoietic lineages. Hence, it remains possible that hematological malignancies do not develop in the mice, because Emi1 levels are not induced in these cells. Nonetheless, it would be interesting to see whether reduced Emi1 expression levels predispose to hematological cancers in vivo . Genomic amplifications and mutations involving EMI1 are rare in human solid cancers. As EMI1 is an E2F target gene, 15 it is likely that in many solid cancers EMI1 overexpression occurs through defective Rb pathway signaling. P53 pathway defects can also lead to overexpression of E2F target genes through crosstalk to the Rb pathway 31 and each of these pathways are impaired in up to 50% of human cancers. 32 Thus, aberrations in the two most frequently mutated pathways in human solid cancers can each contribute to EMI1 overexpression and, as we show here, contribute to ‘oncogene-induced mitotic stress' 13 , 33 leading to CIN and cancer progression. Emi1-overexpressing mice develop tumors at long latencies. This is characteristic for CIN mouse models. In the absence of cooperating mutations, most CIN models develop cancer at latencies of a year or more. 9 , 10 , 11 , 12 , 13 Consistent with recent work, 34 this suggests that cells may need time to acquire specific genetic or genomic aberrations before malignant transformation occurs. Our inducible overexpression model will also enable us to test whether Emi1 overexpression-induced tumors are ‘oncogene-addicted’, requiring continued Emi1 overexpression, or whether transient Emi1-induced CIN is sufficient to sustain tumor growth, as observed in another CIN model. 35 Emi1 overexpression-induced tumors show a significant increase in cells in early mitosis. Consistent with this, our live-cell imaging experiments show that Emi1 overexpression delays the timing between nuclear envelope breakdown and chromosome alignment. This is probably caused by chromosome misalignment. Previous studies have shown that Emi1 degradation in early mitosis is required for progression beyond prometaphase. 36 Elevated levels of Emi1 may therefore cause the prometaphase delay in vitro and explain the high number of early mitotic cells in transgenic mouse tumors in vivo . Emi1-overexpressing cells may subsequently arrest at metaphase. Alternatively, if anaphase ensues, their chromosomes are often lagging or do not segregate with the main chromosome masses, thus facilitating CIN. Since its definition, the CIN70 signature has often been used both to assess the level of CIN in human tumors and to predict clinical outcome. 28 , 29 , 30 For the TCGA breast cancer cohort, 22 we confirm that the CIN70 signature stratifies breast cancer subtypes according to clinical outcome. We demonstrate that EMI1 expression has a clinical stratification power similar to the 70-gene CIN70 signature. In addition, EMI1 overexpression is a stronger predictor of CIN in human cancers than most CIN70 genes. This highlights that EMI1 expression may be used as a strong diagnostic marker. Alternatively, inclusion of EMI1 in the CIN70 signature would increase its strength as a clinical prognostic tool. Finally, the current study’s identification of EMI1 overexpression as both a common phenomenon in human cancers and an in vivo driver of CIN suggests that it could be used as a means to induce excessively high levels of CIN in cancer cells, an approach that has previously been proposed for the eradication of tumor cells. 37 , 38 Thus, our findings not only provide novel mechanic insights into tumorigenesis, they also have important implications for cancer diagnosis and potentially therapy. Materials and methods Data set analysis and statistics For EMI1 expression analysis, data from previous studies were used and analyzed as described below and previously. 21 , 22 , 39 , 40 , 41 , 42 , 43 All data sets are publicly accessible (GEO: , TCGA: , DCC: ). Expression differences were determined by Fisher’s exact, one-way analysis of variance or t -tests, as indicated. Under- or overexpression of EMI1 and MKI67 were considered different if their levels were statistically significantly reduced or elevated ( t -test) by at least 1.5-fold compared with reference tissue. For TCGA expression and survival analysis, level 3 Agilent data were used. Expression outliers were identified and removed based on the standard box-whisker’s plot formulae that uses inter quartile range (IQR). Only expression data falling into the following range would be taken for further analysis to avoid the effect of skewed data: Q1–(1.5*IQR)>expression data<Q3+(1.5*IQR). For survival analysis, expression data were grouped into low and high as previously described. 44 R and shell scripts were used to segregate the data into two. Survival curves were plotted where the survival was significantly different between the two groups. In R , the library surv was used to fit ( survfit ) the survival curves and compute probability using the log-rank test ( survdiff ). All survival curves were re-plotted in GraphPad Prism (La Jolla, CA, USA). Adjustment of prognostic strength for Nottingham prognostic index, Adjuvant! and proliferation was performed as previously described. 24 , 25 , 26 Immunohistochemistry Tissue microarray slides with fixed paraffin-embedded human tumor tissue and matched normal control tissue sections (US Biomax, Rockville, MD, USA) were subjected to antigen retrieval in sodium citrate buffer, permeabilized in 0.01% Triton-X in phosphate-buffered saline (PBS) for 10 min, washed twice in PBS and blocked in 10% BSA in PBS for 1 h. Emi1 antibody (1:100; Santa Cruz Biotechnology, Dallas, TX, USA; sc-365212) incubation was performed overnight at 4 °C. Endogenous horseradish peroxidase was blocked in 0.3% H 2 O 2 in PBS for 10 min followed by two PBS washes. Goat anti-Mouse IgG HRP (Life Technologies, Carlsbad, CA, USA) was applied at 1:1000 for 1 h followed by two PBS washes. DAB (3,3'-diaminobenzidin; Biocare Medical, Concord, CA, USA) was applied for 2 min and tissues were counterstained with hematoxylin. Semiquantitative analysis was performed in a blinded manner using the IHC multiplicative quickscore method, as described. 45 Human primary tumor protein analysis With patient informed consent and Institutional Human Research Ethics Approval from the University of Queensland, de-identified frozen primary human tumor tissues were obtained from the Wesley Research Institute Tissue Bank (Toowong, QLD, Australia). Tissue samples were suspended in RIPA buffer (1:5 (w/v) ratio) and protein extracted using Precellys24 in tubes containing 1mm Zirconia beads (Daintree Scientific, St Helens, TAS, Australia) at 6500 r.p.m. for 3 × 15 s with 15 s gaps. Western blot analysis was performed as described below. Generation of transgenic mice Mouse experiments and procedures were approved by the institutional animal ethics committee. Inducible Emi1-overexpressing mice (in a mixed C57BL/6 x 129/Sv background) were generated by injecting a BsrBI -digested fragment (containing an inducible, CMV-based promoter with seven tetracycline-response elements; Figure 2a ) into fertilized oocytes, followed by implantation into pseudopregnant mice. Founder mice were tested for transgene genomic integration by PCR (F: 5′-AGCAGAGCTCGTTTAGGAACC-3′; R: 5′-ACTTCAAGCTCGGAAAGATCAG-3'; 29 cycles of (30 s at 94 °C, 30 s at 57 °C, 30 s at 72 °C)) and Southern blot analysis (see below). Two positive founder mice were crossed with CMV-rtTA (Tet-On) and CMV-tTA (Tet-Off) transgenic mice 27 to create bitransgenic mice. Transgene expression was regulated by feeding the mice a regular or doxycycline-containing diet (625 mg/kg; Specialty Feeds, Glen Forrest, WA, USA) after weaning. Southern blot analysis Mouse genomic DNA extraction and Southern blotting were performed according to standard methods. Briefly, for each sample, 8 μg of DNA was digested with Sac I and loaded on a 0.8% agarose gel. Following transfer to the blot, a probe of about 750 bp was released from pTRE-HA-Emi1 by Sac I digestion, purified, radioactively labeled and hybridized to the membrane. RNA and protein analysis RNA and protein were extracted from mouse tissues and analyzed for transgene and endogenous expression by semi-quantitative RT–PCR and western blot analysis as described. 31 For semi-quantitative RT–PCR, Superscript III First-Strand Synthesis System (Invitrogen, Carlsbad, CA, USA) and the following primers were used: GAPDH: Quantitect primer mix (Qiagen, Venlo, the Netherlands), Emi1 transgene: F: 5′-CTTACGATGTACCGGATTAC-3′, R: 5′-CCCACAATTGGTGAGTCGATG-3′. Antibodies used for western blot analysis were against HA (1:2000; Covance (Princeton, NJ, USA), HA.11, clone 16B12), mouse Emi1 (1:1000; Zymed (Carlsbad, CA, USA), 38-5000), human Emi1 (1:1,000; IHC antibody listed above) and actin (1:5000; Sigma-Aldrich (St Louis, MO, USA), A2066). Pathology and fluorescence in situ hybridization Mouse tumor volumes were measured following necropsy and calculated using the formula π /6x( L x W ) 3/2 , as described previously. 46 Where tumor sizes were compared between mice, only tumors with elliptic shape were included and if the mouse developed multiple tumors, only the largest tumor was included. Tumors were formalin-fixed and paraffin-embedded. Tissue sections were subjected to blinded pathological analysis after H&E staining and to interphase-FISH using fluorescently labeled peri-centromeric probes for chromosomes 12, 16 and 17 to count chromosome numbers per nucleus on at least 100 nuclei per tumor. Aneuploidy was defined as: (1) 0 or 1 FISH signal for each of the three chromosomes, as long as the total number of FISH signals is at least 1 or (2) ⩾ 3 copies of at least one chromosome. Immunofluorescence Formalin-fixed, paraffin-embedded tissues were sectioned at 4 μm, de-waxed 2 × 10 min in 100% xylene and rehydrated 2 × 5 min in 100% ethanol, 1 × 1 min each in: 90, 80, 70, 50, 30% ethanol, 1 × 5 min in distilled water. Antigen retrieval was carried out in 0.01 M sodium citrate, pH 6.0 with 0.05% Tween-20 at 105 °C in a decloaking chamber for 15 and 20 min cooling. Sections were washed 3 × 5 min in dH 2 O and blocked in 10% normal goat serum (Sigma-Aldrich) in PBS (blocking buffer) for 60 min at room temperature in a humidified chamber. Sections were incubated with primary antibodies in blocking buffer in a humidified chamber at 4 °C overnight. Primary antibodies were β-tubulin (Abcam (Cambridge, UK), 1:100), γ-tubulin (Sigma-Aldrich, 1:50), 4,6-diamidino-2-phenylindole (Sigma-Aldrich, 0.2 μg/ml). Sections were washed 3 × 20 min in PBS-0.5% Tween-20, incubated in secondary antibodies in blocking buffer for 1 h at room temperature in a dark humidified chamber. Secondary antibodies were Alexa Fluor-488 and Alexa Fluor-594-conjugated goat anti-rabbit and goat anti-mouse IgG (Invitrogen, 1:500). Sections were mounted in Vectashield (Vector Labs, Burlingame, CA, USA) and imaged using an Olympus Fluoview v4.2 confocal microscope and Olympus Fluoview FV1200 software (Olympus, Tokyo, Japan). Time-lapse microscopy Hela cells (American Type Culture Collection, Manassas, VA, USA) stably expressing histone H2B-GFP (B. Gabrielli, mycoplasma-negative) were transduced with a lentiviral vector containing the human EMI1 ORF. Following puromycin selection (3 days at 2 μg/ml), resistant cell populations and control cells were subjected to time-lapse microscopy. For that, 10 5 cells were seeded in a six-well plate in DMEM/F12 media supplemented with 10% FBS, 2 m M l -glutamine, 1 m M sodium pyruvate, 20 m M HEPES, 100 U/ml penicillin and 100 μg/ml streptomycin. Cells were maintained at 5% CO 2 and 37 °C and monitored with an Olympus time-lapse microscope and 100 cells in each condition were analyzed. CIN70 analysis Agilent probe-level expression data (log 2 signal intensity normalized to reference) were obtained from TCGA (BRCA, CRC, LUAD+LUSC, LGG, KIRP+KISC, UCEC, OV, GBM). 22 , 39 , 40 , 41 , 42 , 43 , 47 , 48 Expression data were obtained after background correction, normalization and loess transformation. Agilent platform data (G4502A_07_2, G4502A_07_3) were analyzed. Probes corresponding to the 70 CIN70 genes 28 and EMI1 were taken from the Agilent ADF Data sets. Probe sequences were megablasted against the human NR/NT database. Probes mapping to multiple or incorrect genes were excluded. These were: KIF4A : A_23_P148474, A_23_P148475; MT1JP : A_23_P414341, A_24_P74828, A_24_P74830; PTTG1 : A_23_P7636, A_24_P72112; UBE2C : A_23_P304770. Of 261 blasted CIN70 probes, 253 (69 genes) mapped to their corresponding genes with significant values to pass quality control. Probe level expression data were averaged for each gene and the average (CIN70 score) was calculated. EMI1 probes were similarly quality checked and expression data averaged. Quality-checked probes are listed in Supplementary Data set 5 . Data were plotted and statistically analyzed in GraphPad Prism.

University of Queensland researchers have discovered a key driver in the development of most cancers, including breast, lung, liver and ovarian cancers. UQ Diamantina Institute researcher Dr Pascal Duijf said the discovery could be the foundation for improved cancer diagnosis, new treatments and better assessment of a patient's prognosis. "My team has discovered excessively high levels of the protein EMI1 in cancer samples, including the aggressive brain cancer glioblastoma and tumours of the bone," Dr Duijf said. "High levels of EMI1 promote tumour development, increase the tendency of cancer cells to spread, and change immune responses, which fuel cancer progression," "This is associated with poor patient prognosis, particularly in breast cancer." Dr Duijf, a National Breast Cancer Foundation Career Development Fellow, said high levels of EMI1 disrupted normal cell division leading to new cells with abnormal chromosome numbers. "This process, referred to as chromosome instability, accelerates cancer progression and allows cancer cells to become resistant to cancer therapies." "Our findings indicate that high EMI1 levels are one of the strongest indicators of chromosome instability identified to date." Dr Duijf said the discovery was an exciting step forward. The next step was to determine whether tumours had to maintain high levels of EMI1 to survive. "If that is the case, it could present a promising anti-cancer target," he said. The study is published in the Nature journal Oncogene.

10.1038/onc.2016.94

Medicine

Zoom fatigue affects young women more than others

Nicola Döring et al, Videoconference Fatigue: A Conceptual Analysis, International Journal of Environmental Research and Public Health (2022). DOI: 10.3390/ijerph19042061 Journal information: International Journal of Environmental Research and Public Health

https://dx.doi.org/10.3390/ijerph19042061

https://medicalxpress.com/news/2022-05-fatigue-affects-young-women.html

Abstract Guidelines Hypothesis Interesting Images Letter New Book Received Obituary Opinion Perspective Proceeding Paper Project Report Protocol Registered Report Reply Retraction Short Note Study Protocol Systematic Review Technical Note Tutorial Viewpoint All Article Types Advanced Search Section All Sections Adolescents Aging Air Anthropogenic Circularity Biosafety Chemoenvironment Children's Health Climate Change Digital Health Disabilities Disaster Medicine Disease Prevention Emerging Contaminants Environment and Applied Ecology Environmental Analysis and Methods Environmental Chemistry and Technology Environmental Earth Science and Medical Geology Environmental Ecology Environmental Health Environmental Microbiology Environmental Remediation and Management Environmental Science and Engineering Exercise and Health Global Health Health Behavior, Chronic Disease and Health Promotion Health Care Sciences & Services Health Communication and Informatics Health Economics Health Informatics Health-Related Quality of Life and Well-Being Industrial Ecology Infectious Disease Epidemiology Injury Prevention and Rehabilitation Mental Health Nursing Occupational Safety and Health Oral Health Public Health Statistics and Risk Assessment Reproductive Health Skin Health Sport and Health Toxicology and Public Health Traumas Water Science and Technology Women's Health All Sections Special Issue All Special Issues 2nd Edition of Environmental Impact Assessment by Green Processes 2nd Edition of Health Literacy, Nutrition and Public Health 2nd Edition of Oral Inflammation and Chronic Autoimmune Diseases 2nd Edition of Protecting, Supporting and Promoting Appropriate Breastfeeding in the 21st Century 2nd Edition of Sleep Quality and Health-Related Outcomes 2nd Edition of Social Determinants of Health 2nd Edition of Treatment of Foot and Ankle Injury and Public Health 2nd Edition of Trends in Sustainable Buildings and Infrastructure 2nd Edition of Water Sports Implications for Training, Environment and Health 2nd Edition: Advances in Maternal and Child Healthcare 2nd Edition: Advances in Personalized Exercise Prescription for Chronic Disease Prevention and Rehabilitation 2nd Edition: Evidence-Based Nature for Human Health 2nd Edition: Movement Studies for Individuals with Visual Impairments 2nd Edition: Tobacco Smoke Exposure and Tobacco Product Use A More Sustainable and Healthier Future for All: What Works? Access, Health, Regulations, and Policy: Exploring Rural and Remote Drinking Water Supplies in North America Actions in EU MS for Combating Health Threats in the Maritime Transport Sector Active Commuting and Active Transportation Addiction Behavior Addictions and Cognitive Behavioral Therapy Approaches Addictions in Children and Adolescents Addressing Environmental Health Inequalities – Proceedings from ISEE Conference 2015 Adolescent Depression Prevention Advances in Big Data Analytics and Intelligence Advances in Earth System Science Advances in Endodontic Pain Control Advances in Environmental Biology Advances in Environmental Chemistry Advances in Environmental Economics Advances in Environmental Geotechnics Advances in Environmental Justice Advances in Environmental Modelling Advances in Environmental Neurotoxicology Advances in Environmental Sensor Networks Advances in Environmental Sustainability, Resilience, and Health Advances in Epidemiology Advances in Foot Disorders and Its Treatment Advances in Gastroenterology, Hepatology and Clinical Nutrition Advances in LGBTQ+ People's Health Advances in Mental Health, PTSD and Moral Injury Advances in Physical Diagnosis, Physiotherapy and Rehabilitation in Medicine and Dentistry Advances in Squamous Cell Cancer of the Head and Neck Advances in Substance and Drug Abuse Prevention Advances in Telehealthcare Advances in the Diagnosis and Management of Renal Diseases Advances in the Use of eHealth for Pain Management Advancing Health Equity for Sexual and Gender Minority Populations Ageing Well: The Role of Age-Friendly Environments Aging and Cognition Aging and Work Agrochemicals in the Agri-Food Chain Air Pollution and Carbon Dioxide Emissions Air Pollution Exposure and Health Risks Air Pollution Modeling Air Pollution, Climate Change, and Public Health: The Unavoidable Path towards Decarbonization Air Quality and Healthcare Associated Infections Alcohol Abuse: Newer Approaches to an Old Problem Alcohol and Drugs of Addiction, Aggression and Violence Alcohol and Health Alcohol and Public Health Alcohol Policy and Public Health Alcohol-Related Violence Alcoholism Allergic Disease Epidemiology Allergy and Environment Ambient Air Pollution and Health Vulnerability Anemia and Patients Blood Management Patient in Critical Care Animal Assisted Interventions and Activities for Health and Wellbeing Anti-cancer Activity for Cancer Prevention and Treatment Antibiotic Resistant and Pathogenic Bacteria in the Food Production Environment: Epidemiological Evolution and Control Antimicrobial Resistance Prevention and Control Antimicrobials and Antimicrobial Resistance in the Environment Application of Bio-Based Materials in Environmental Governance Application of Biostatistical Modelling in Public Health and Epidemiology Application of Robotic Devices for Neurologic Rehabilitation Application of Statistical Methods in Public Health and Medical Research Applications of Statistical and Data Science Methods in Public Health Applied Physiology and Exercise Testing in Endurance and Ultra-Endurance Sports Applying Clinical Psychology to Medical Conditions Arsenic Contamination, Bioavailability and Public Health Arsenic in Drinking Water: Current Perspectives and Future Directions Asbestos and Cancers: Exposure from Neighboring Factories and Mines, Contaminated Soil, and Slate Roofs Asbestos Exposure and Disease: An Update Asbestos Exposure and Health Impact Assessment of Public and Occupational Electromagnetic Field Exposure in New and Emerging Connectivity Exposure Scenarios Assessment, Management, and Policy-Making for Environmental Pollution Related to Land Use and Land Cover Atmospheric Change and Impact on Health Auditory Experience, Music and Voice across the Perinatal Period: Innovations in Theory, Research and Clinical practice Back and Neck Pain Bariatric Surgery: Nutritional, Metabolic and Public Health Aspects Bayesian Design in Clinical Trials Beyond Amyloid and Tau - Targeting Lipid Metabolism for Alzheimer’s Disease Biomarkers and Therapies Big Data and Mathematical Modeling in Biomedicine Biochemical and Genetics Tools for Monitoring Health and Risk of Disease Biodegradability and Environmental Sciences Biological and Agricultural Engineering Biological and Health Effects of Low Level Non-Ionising Electromagnetic Fields Biomarkers: Environmental Research and Public Health Bioprocesses for Air Pollution Control Biostatistics Reveals New Insights in Sports Science, Sports Medicine, and Health Birth Defect Prevention Blood Cells, Hematopoiesis, Molecules, Nanomedicine and Diseases Research Bone Health: Nutritional Perspectives Boredom in Health, Education and Sports Breastfeeding and Infant Health Building Related Illnesses Caffeine in the Diet: Health Implications, Safety Issues, and Molecular Mechanisms Cancer Prevention, Treatment, and Survivorship Cannabis, Cannabis-Based Products, and Cannabinoids for Medicinal Use Carbon Capture and Storage Carbon Dioxide (CO2), Emissions, Environmental and Public Health Impact Carbon Emissions and Environmental Protection Cardiac Rehabilitation in the Time of COVID-19 Cardiovascular Disease Self-Care Interventions Cardiovascular Diseases and Public Health Challenges for Health Inequalities Research during COVID-19 Pandemic Challenges in Positive Organizational Psychology Child Injury Prevention Child Injury Prevention 2015 Childhood Obesity: Novel Approaches to a Global Problem Childhood Obesity: Prevention and Treatment Children, Adolescents and Nutrition Children’s Exposure to Environmental Contaminants Chronic Diseases and Multimorbidity in Primary Care Circular Economy & Social Inequalities Climate Change and Human Health Impacts and Adaptation Climate Change and Human Health Impacts and Adaptation Climate Change Impacts on Vector-borne Diseases Clinical and Translational Pathways for Cardiovascular Research College Student Health Behaviors Impact Their Food Security Influenced by Their Environment Community Health Intervention to Reduce Chronic Disease Community Participation in Health Promotion: Challenges and Successes Community Resilience and Recovery for Public Health Emergencies Community-Based Health Promotion Community-Based Solutions for Injury and Violence Prevention Community-Engaged Research to Address Health and Healthcare Disparities Complex Interventions for Public Health Improvement Computational Modeling in Biology and Medicine Contemporary challenges in public health Correctional Health COVID-19 and Health Education COVID-19 and NCDs: Emerging Trends in the Tropics COVID-19 in Low and Middle Income Countries COVID-19 Outbreak and Beyond: Psychological and Behavioral Responses and Future Perspectives COVID-19 Pandemic and the Environment COVID-19 Research COVID-19: Public Health Response Creativity and School Functioning Crisis Management and the Circular Economy for Healthier, Smarter and More Sustainable Cities Critical Issue on Lower Urinary Tract Symptoms and Urinary Incontinence Culture of Evidence-Based Practice and Quality Improvement in Nursing Cumulative and Integrated Health Impact Assessment Cumulative Health Risk Assessment Current Issues in the Neurological Rehabilitation of Children and Adolescents Current Research Trends in Transgender Health Current Therapeutic Trends and Challenges in the Management of Patients with Type 2 Diabetes Mellitus Cyber Pathology: Cyber Victimization and Cyber Bullying Cycling Medicine Data Analytics and Statistical Approaches Applied in Injury Risk, Illness, Well-Being, and Exercise Monitoring Decarbonization and the Benefits of Tackling Climate Change Decision-Making in the Health Care of Older Adults Decompensated Heart Failure Dental Hygiene and Oral Health Research: Lessons and Challenges Design and Development of Rehabilitation in Individuals with Parkinson’s Disease Determinants of Health and Well-Being in at Risk Populations Development and Evaluation of New Tobacco Control Interventions Development, Implementation and Evaluation of Gender-Specific, Community Based Health Promotion Interventions Diabetes Prevention: Challenges and Opportunities Diagnosis and Healthcare for Dementias Diet, Adiposity and Metabolic Health in Pregnant Women Different Views on a Child's Motor Development and Motor Performance Digital Governance and Low-Carbon Development Disabilities – Constant Quest in Medical, Social and Political Records Disability and Public Health Discoveries in Oral and Maxillofacial Surgery Diseases Etiology and Management: Towards a Precision Medicine Approach Drinking Water and Health Drug Abuse and Addiction Drug Resistance Early exposure to endocrine disrupting chemicals Eating Behaviour and Food Safety, Physical Fitness and Health Ecological and Human-Health Effects of Pesticides in the Environment Economic Causes and Impacts of Diabetes and Diabetes-Related Conditions Economics of New Drug Development and Approval Economics of the Prevention and Treatment of Obesity Edentulism and Oral Health Problems among Older Adults Editorial Board Members' Collection Series: Environmental Ecology and Management Effect of Air Pollution Exposure on Children and Elderly’s Health and Neurological Functions Effect of Insomnia on Human Health Effect of Sport Activity on Health Promotion Effecting a Safe and Healthy Environment in Construction Effects of COVID-19: Issues on Health Economics and Education Effects of Environmental Factors on Autism Effects of Hyperoxic Training on Acute Responses and Exercise Performance Effects of Parental Stress on Child Development Effects of Physical Activity on Cognitive Function in Young People Effects of Physical Activity on Cognitive Function in Young People: 2nd Edition Effects of Resistance Training on Strength and Muscle Thickness Effects of Stress Exposure on Mental Health and Well-Being Efficient Use of Acute Hospital Services for Older People Electromagnetic Fields and Health Electromagnetic Fields and Health—Effects on the Nervous System Electronic Cigarette Use and Public Health Electronic Cigarettes as a Tool in Tobacco Harm Reduction Eliminating Health Disparities to Achieve Health Equity Emergency Medical System and Emergency Medicine in the Time of COVID-19 Emerging Biological Threats and Public Health Preparedness Emerging Contaminants in the Environment Emerging Issues in Air Quality: Pollutants, Sources, Concentrations, Chemistry, and Health Risks Emerging Technologies for Treating Emerging Contaminants in Water/Wastewater Emerging Transportation Solutions for Safer and Greener Future Mobility Emotional, Behavioral, and Psychological Impact of COVID-19 Pandemic Endocrine Disruptors and Human Health Endocrine Disruptors and Public Health Endocrine Disruptors Leading to Obesity and Related Diseases Energy Conservation Measures, Indoor Air Quality and Health Energy Use and Environmental and Public Health Energy-Saving Technology in the Urban Sewage Treatment Process Environment and Behavior Environment and Health - Bridging South, North, East and West: Proceedings from the ISEE, ISES and ISIAQ Conference 2013 Environment and Patient Safety in Intensive Care Units Environmental and Food Allergy Environmental Bacterial Pathogens and Human Health Environmental Carcinogens Environmental Chemical Mixture at Low Concentration and Children’s Health Environmental Determinants of Infectious Disease Transmission Environmental Diseases Environmental Fate and Effect of Nanoparticles and Nanomaterials Environmental Geochemistry and Health Environmental Health Risk Assessment Environmental Hygiene Environmental Hygiene and Health Promotion Environmental Impacts of Food Consumption and Nutrition Environmental Legislation and Public Health Environmental Nanoparticles and Their Toxicity Environmental Neurotoxicology Environmental Pollution and Human Health Risk Environmental Research and Public Health: Featured Review Papers Environmental Research on Alcohol: Public Health Perspectives Environmental Sustainability: Knowledge, Attitudes and Behavior Environmental Systems Engineering Epidemiology and Determinants of Dental Caries in Children Epidemiology as a Tool for Thinking Critically about the Cause of Disease and Ill-Health Epidemiology of Iron Deficiency in Children and Its Consequences Epidemiology of West Nile Virus Epidemiology, Prevention and Control of Legionellosis: New Trends and Perspectives Epidemiology, Prevention and Control of Malaria Equity of Health and Health Care in Aging Societies Ethics, Social Responsibility and Quality of Life in Times of Crisis Ethnicity and Religiosity as Risk Factors for Health Etiology, Diagnosis and Treatment of the Temporomandibular Joint and Masticatory Muscle Disorders in the Era of Emotional Challenges Evaluation of Rural Water Systems and Public Health Evidence for Incorporating Green Exercise into Clinical and Public Health Practice Evidence-Based Nature Therapy: Advances in Physiological Evaluation Evidence-Based Prevention of Childhood Malnutrition Evolutionary Medicine in Sport and Exercise: A Preventative Approach Exercise as a Polypill: The Effects of Physical Activity among Healthy Individuals and Patients Exploring Clinical Outcomes in Diabetes Patients Exposure and Health Effects of Secondhand Smoke Extreme Weather and Public Health Extreme Weather-Related Morbidity and Mortality: Risks and Responses Factors Affecting Human Fat Distribution and the Metabolic Health Implications Family Health Family Violence Financial Stress and Mental Health First International Conference on Environmental Sciences and Sustainability in University of Venda Food Addiction and Binge Eating Food Allergy, Genes and Environment Food Insecurity: Evidence to Move the Needle Food Safety Food Safety and Public Health Food Supply Chain, Consumer Choice and Obesity Foods for Plant-Based Diets: Innovations, Technologies and Applications Fourth Edition of Health Emergency and Disaster Risk Management (Health-EDRM) Frailty and Aging From Understanding Suicide Risk to Preventing Suicide Frontiers in Mental Health and the Environment Future Challenges in Emergency Response and Public Health in Compound Disasters Gender and Geoethics in the Geosciences Gender and Health Gene-Environment Interaction in Chronic Diseases Gene-Environment Interactions and Disease Gene-Nutrient-Environment Interactions Genetic Epidemiology Genetic Impact on the Development of Allergic Disease Geostatistics in Environmental Pollution and Risk Assessment Global Children’s Environmental Health Global Climate Change and Contaminants Global Environment and Children Health Global Epidemics of Zika? Implications for Public Health Global HIV Prevention and Treatment: Public Health Considerations Global Public Health and Epidemiology Green Transportation Green-Blue Space and Health: Advances in Methods, Technologies and Applications Greening Urban Spaces: A Healthy Community Design Groundwater Contamination in Urbanized Areas Gulf War Illness, A Drug and Environmentally-Triggered Condition Gut Microbiota in Health and Disease Hair Disorders Hazardous Waste and Human Health Hazardous Waste and Human Health-2015 Head and Neck Cancers: What's New? Headaches HEAL: Transformational Change for Environmental, Planetary, and Human Health Health and Health Care for Homeless People in Various Contexts Health and Healthcare for People with Diabetes Health and Wellbeing of Children, Adolescents and Young Adults Health Behaviour and Lifestyle Health Benefits of Nature Health Care for Old People Health Disparities Arising from Inequitable Access to Water and Sanitation Systems in Developed Countries Health Economics Health Effects of Waterborne Contaminants: A Focus on Emerging Concerns Health Emergencies and Disasters Preparedness Health Environment and Sustainable Development Health Geography and Its Relevance for Future Public Health Health Humanities: Humanism in Health and Healthcare Delivery for Migrant and Minority Populations Health Impact Assessment Health Impact Assessment: Realizing Its Potential Health Inequalities in Urban Areas—Factors, Processes, and Dynamics Health Inequality and Spatially Distribution Health Informatics and Public Health Health Literacy and Equity – Interdisciplinary Perspectives and Recent Trends in Health Literacy Research and Action around the World Health Literacy, Nutrition and Public Health Health Market: Incentives and Competition Health Promotion and Quality of Life Improvement Throughout the Life Cycle Health Promotion, Physical Activity and Health Behaviors among Teenagers and Young Adults Health Systems and Services Health Technology Assessment and Public Health: Relation, Potentialities and Evidence Generation Healthy Workplaces, Employment and Chronic Conditions in Europe: Answering the Hidden Emergency with Innovative Strategies Heated Tobacco Products Heavy Metals and Health Heavy Metals: Environmental and Human Health HIV Prevention and Mental Health Disparities HIV Status of the World after Three Decades: Progress, Prevention Efforts, Challenges, and Strategies for the Way Forward HIV/AIDS: Social Perspectives Holistic Approaches to Understanding and Caring for Vulnerable Populations Hospitals’ Contribution to the Geographies of Social and Economic Innovation Housing and Health Human Factors and Ergonomics: Bridging the Gap between Research and Practice in Occupational Safety and Health Human Health, Risk Analysis and Environmental Hazards Human Microphysiology Systems Human Performance in Extreme Environments ICERPH 2020—Addressing Environmental Threats to Human Health from Pregnancy to Senility IJERPH: 10th Anniversary IJERPH: 15th Anniversary Impact of Clear Aligners and Custom-Made Mechanics on Children’s and Adults’ Oral Health: New Frontiers, Indications, Limitations, and Patient Perspectives Impact of Coronavirus (COVID-19) Impact of Genes and the Environment on Obesity and Cardiovascular Disease Impact of Out-of-Home Meals on Public Health Implementation of Interventions in Public Health Implementation Research in Chronic Disease Prevention and Control Improving the Health of Individuals Who Inject Drugs In-Silico Medicine in Diagnosis, Prevention and Treatment of Musculoskeletal Disorders Independent Mobility: Exploring Children’s Freedom to Travel and Play without Adult Supervision Indigenous Health in Canada: Integration of Community-Based Environmental, Health, Medical, Natural and Social Sciences Research to Mitigate Disparities in Wellness Indigenous Health, Environments and Wellbeing in Canada Indoor activities and health risks/protection Indoor Air Pollution and Health Indoor Air Pollution and Human Health Indoor Environmental Quality: Exposures and Occupant Health Indoor Radon Risk Assessment and Remedial Actions Inequalities among Families Involved with Child Welfare Services Inequalities in Health Infections in Nursing Homes: Evidence-Based Interventions Influence of Dairy Consumption and Production on the Environment Influence of Expectations, Psychological, Cognitive, and Emotional Factors on Pain Perception Innovations in Biostatistical Methods and Data for Public Health Research Innovations in Exposure Assessment for the Sound Management of Chemicals Innovations in Remote Sensing and GIS for Environmental, Urban and Public Health Sciences Innovative Research in Health Communication Insights into Defeating HIV-Associated Neurocognitive Disorders Integrated Assessment of Artisanal and Small-Scale Gold Mining (ASGM) in Ghana International Symposium on Environmental Physiology and Medicine Internet Based Obesity Prevention in the Time of COVID-19 Interval Training: Different Approaches and Designs Applied to Health and Fitness Intimate Partner Violence: Predictor Factors and Consequences for the Victim and Their Children Invasive Procedures in the Hands of Respiratory Physicians in the Diagnosis, Staging, and Treatment of Thoracic Cancer: Why, How, and When? Is Exercise the Best Medicine during the COVID-19 Pandemic? Latest Insights and Research Perspectives ISEE Commentaries Job Burnout: A Deep-Rooted Issue in Ever-Changing Workplaces Job Stress and Health Latin American Study of Nutrition and Health (ELANS) Lead: Risk Assessment and Health Effects Leaving no one behind: Equity and Eye Health Leptospirosis in the Animal—Human-Ecosystem Interface Lessons Learned from Research on Rare Diseases: Ethical and Legal Challenges Leukemia Arising from Chemical Exposures and Chemotherapeutic Drugs Lifecourse Environmental Epidemiology Lifestyle and Environmental Interventions to Promote a Healthy Ageing Lifestyle and Risk of Depression Lifestyle Intervention for Chronic Diseases Prevention Livelihoods Resilience and Sustainable Rural Development Long COVID and Post-COVID-19 Syndromes Lung Diseases Associated with Environmental Pollutants Malnutrition and Public Health Management of the Circular Economy in the Productive Sectors: Innovative Practices for a More Sustainable World Mass Spectrometry and Environmental Analysis Maternal and Child Health Maternal and Child Health and Nutrition in Latin American Populations Measurement and Evaluation of Occupational and Behavioral Exposures Measuring Disability and Disability Inclusive Development Measuring Inequalities in Health: Innovative Approaches and Applications Medical Ethics in the Time of COVID-19: Challenges and Opportunities Medical Ontologies Melanoma Epidemiology: Analytical Studies, Systematic Reviews and Meta-Analyses Addressing Environmental Risk Factors and Preventive Measures Mental Health and Cognitive Function in Elite Athletes around the Globe Mental Health and Health Psychology Mental Health and Social Care and Social Interventions Mental Health and Well-Being on School Campus in the Post-pandemic Era Mental Health Care Mercury and Health: Current Perspectives and Future Directions Metabolic Syndrome and Its Association with Biomarkers Methodological Innovations and Reflections-1 Methodological Study in Environmental Health and Public Health Microplastics: Hazards to Environmental and Human Health Migraine: Symptoms, Causes, Diagnosis and Treatment Migrant Health Migrant Health 2012 Migrant Health Burden: Emerging Challenges and Future Solutions Migration and Migration Status: Key Determinants of Health and Well-Being Migration, Population Mobility and Public Health: Focus on Vulnerable and Marginalised Populations Migration, Resilience, Vulnerability and Migrants’ Health Mindfulness-Based Practice for Health Benefits Modeling Tools for Occupational Exposure Assessment Modern Occupational Medicine Molecular Mechanisms of Helicobacter pylori Pathogenesis and Host Factors Mosquito Control Innovations into The 21st Century Motor-Vehicle Crashes and Occupant Protection Mountain Sports Activities: Injuries and Prevention Multidimensional Aspects of Oral Health-Related Quality of Life Multimorbidity, Polypharmacy, and Medication Appropriateness: Public Health Challenges and Research Priorities Multiple Criteria Analysis and Artificial Intelligence for Multidimensional Risk Management with Applications in Healthcare, Supply Chain and Sustainability Musculoskeletal System in Exercise and Rehabilitation: From Lab to Clinical Practice Mycotoxins in the Agri-Food Chain Nano-Bio Interactions: Nanomedicine and Nanotoxicology Nanomaterials-Based New Techniques, New Drugs, and Antibacterial Reagents Nanomaterials-Based New Techniques, New Drugs, and Antibacterial Reagents: An Update Nanotechnology to the Benefit of Environment and Public Health Neglected Diseases: Public Health and Environmental Studies Neurodiseases and Public Health Neuromuscular Control of Human Movement Neurorehabilitation and Neuroeducation of Developmental Dyslexia in Healthcare Systems of Different Countries New Advances in Cardiovascular Diseases New Advances in Cervical Cancer New Advances in Neurological Physiotherapy and Rehabilitation New Advances in Palliative Care New Advances in the Diagnosis and Treatment of Congenital Heart Disease New and Re-emerging Pathogens New Challenges and Crucial Topics for 2030 Public Health New Challenges in Prehospital Emergency Care New Directions in Environmental Communication Research New Frontiers in Rehabilitation New Horizons in Cerebellar Research New Techniques, Technologies and Materials for Dentistry and Maxillofacial Surgery New Trends in Intelligent Quantitative Analysis on the Effects of Energy Consumption on Environment and Public Health New Trends in Research on Training, Performance, Conditioning, Coaching, Evaluation and Health in Basketball (Second Edition) New Trends in Virtual World Addictions and Problematic Internet Use Next Generation of Interoperability in Health Care: Standards and Application Nicotine, Electronic Cigarettes, and Alcohol Use and the COVID-19 Pandemic Noise and Quality of Life Noise and Sleep: 2nd Edition Noise Induced Hearing Loss and Tinnitus, Public Health and Sustainability Effects Noise Pollution and Health Non-cigarette Tobacco Product Dependence Novel Insights of the Etiological Factors Involved in Oral Conditions of Children and Teenagers Novel Methodologies in Drug Use and Drug Use Disorder Epidemiology Nursing and Public Health Nursing Care of Older Adults and People with Disability Nutrient Removal and Recovery Nutrition and Diabetes: A Health Issue Nutrition and Exercise in the Health Sciences Nutrition and Health Equity: Revisiting the Importance of Fruit and Vegetable Availability, Purchasing, and Consumption Nutrition and Public Health Nutrition in the First 1000 Days Nutrition, Diets and Public Health Nutritional and Non-nutritional Supplements in Sports: The Public Health Impact Obesity and Public Health Obesity Prevention: Systems change for healthy environments and healthy people Obstructive Sleep Apnea Syndrome: From Symptoms to Treatment Obstructive Sleep Apnea: diagnosis, prevalence, treatment, clinical presentation, comorbidities, and responsible mechanisms Obstructive Sleep Apnoea Occupational Cancer Occupational Health Occupational Safety and Related Impacts on Health and the Environment Occupational Sedentary Behaviour Occupational Stress in Planetary Health: Explaining Multilevel Systems and Effects of Work, Time and Place Occupational Stress, Human Health and Wellbeing Occupational Therapies and Human Well-Being Occupational Therapy: Neurorehabilitation of Children and Adults Occurrence of Oral Epidemiology and Its Determinants Open Innovation in the Research and Industry about Natural Environment and Public Health after Pandemic of COVID-19 Operational Research to Tackle Antimicrobial Resistance in Ghana Operations and Innovations for the Environment Opioids: A Challenge to Public Health Optimal Mental Health for Optimal Academic Performance in University Students Optimising Drug Prescribing and Improving Medication Management: What Can We Do? Oral Cancer: Public Health Concern Oral Diseases: Prevention, Diagnosis and Treatment Oral Health among the Older Population Oral Health and Host Related Environmental Factors in Periodontal Status Oral Health-Related Quality of Life and Periodontal Health Status among Different Age Groups Oral Hygiene and Function: Lessons and Challenges for General Health Oral Inflammation and Chronic Autoimmune Diseases Oral Microbiota and Oral Health Oral Pathologies and Their Impact on Public Health Outcomes of Joint-Preserving Surgery for Rheumatoid Forefoot Deformity Overweight and Obesity—Diagnosis and Treatment Pain Neuroscience Education, Chronic Pain, and Health Care Parents and Children during COVID-19 Pandemic Passion, Grit, Mindset, Achievement and Well-Being Passive Exposure to Conventional, Heat-Not-Burn and Electronic Smoking Products and Related Health Effects: Second Edition Patient and Consumer Engagement in Health Care and Wellbeing: Challenges and Opportunities for a Participatory Health Approach Patient Satisfaction with Different Service in Rare Diseases Pediatric Infectious Diseases Pesticides and Health Pharmaceutical Policy and Practice Research Photocatalysis Assists Carbon Neutrality Physical Activities in the Water Environment: Drowning, Prevention and Rescue Physical Activity and Dietary Behaviors among Underserved Groups Physical Activity and Exercise Programs in Older Adults Physical Activity and Health Behaviors Physical Activity and Nutrition Among Border and Other Marginalized Populations Physical Activity and Psychosocial Health: A Two-Way Path to Improve Youth Wellbeing Physical Activity and Public Health- Physical Activity and Sedentary Behavior: Trends and Impact during the Pandemic Physical Activity and Well-Being in School Setting Physical Activity Programmes for the Elderly and Their Health Implications Physical Activity, Fitness, and Physical Education Physical Activity, Sedentary Behavior and Sleep at Different Stages of Life Physical Fitness and Health Improvement Physical Fitness and Health Outcomes throughout the Life Span Physical Performance and Recovery during Exercise-Induced Muscle Damage Physical Well-Being and Motor Development over the Life Span Positive Development in Neuropsychology for Young Population Poverty and Child Well-Being Preparedness and Emergency Response Prescription Medication Addiction/Abuse and Public Health Consequences Prevention and Treatment of Cerebrovascular Diseases Prevention Better than Cure for Long Term Conditions? Prevention of Mental Health Disorders in Children and Adolescents Prevention, Treatment and Rehabilitation of Musculoskeletal Disorders Preventive Medicine Primary Care and Health Behavior Change Proceeding of 'Ecosystems and Human Well-Being in the Transition towards Green Economy' Proceedings from 2014 Global Land Project (GLP) Asia Conference Proceedings from 9th International Conference on Managing Fatigue Proceedings from the 11th International Symposium on Recent Advances in Environmental Health Research Proceedings from the 12th International Symposium on Recent Advances in Environmental Health Research Proceedings from the 13th International Symposium on Recent Advances in Environmental Health Research Proceedings from the 2014 Minority Health & Health Disparities Grantees’ Conference (MHHDGC) Proceedings from the 3rd International Symposium on Environment and Health Proceedings from the Eighth International Symposium on Recent Advances in Environmental Health Research Proceedings from the Fifth International Symposium on Recent Advances in Environmental Health Research Proceedings from the Ninth International Symposium on Recent Advances in Environmental Health Research Proceedings from the Seventh International Symposium on Recent Advances in Environmental Health Research Proceedings from the Sixth International Symposium on Recent Advances in Environmental Health Research Proceedings from the Tenth International Symposium on Recent Advances in Environmental Health Research Proceedings of 1st Latin American Congress of Clinical and Laboratorial Toxicology (Toxi-Latin 2014) Proceedings of 2nd International Symposium on Environment Science and Health Proceedings of ICOH Occupational Health for Health Worker (OHHW2019) Conference Proceedings of Research Centers at Minority Institutions (RCMI) Translational Science 2017 Conference Proceedings of the 2019 Research Centers in Minority Institutions (RCMI) Program National Conference Proceedings of the 2020 inVIVO Planetary Health Annual Conference: Project Earthrise Proceedings of the 2022 Research Centers in Minority Institutions (RCMI) Consortium National Conference Process-Oriented Data Science for Healthcare 2018 (PODS4H18) Processed Food: Nutrition, Safety and Public Health Promoting Health and Wellness: Implications for Physical Therapy Promotion of Big Data and Intelligent Transportation to Traffic Safety and Environment Promotion of Healthy Active Habits in Children, Adolescents and University Students Promotion of Healthy Foods: Effectiveness, Evaluation and Agenda Setting for Examining Intervention Techniques to Improve Dietary Intake Promotion of Healthy Work Providing Appropriate Health Care in Patients with Myelodysplastic Syndromes and Clonal Hematopoiesis Psychological Health and Benefits of Mindfulness-Based Interventions Psychological Impact of Stress- and Trauma-Related Events in Early Years of Parenthood Psychology of Addictive Behaviors Psychology of Eating: Understanding of Eating Behaviours Psychology of Learning in Higher Education Psychology, Behavior and Health Outcomes Psychology, Education and Sport in Children Psychosocial Aspects and Quality of Life in Bariatric Surgery: An Update and Directions for Future Research Psychosocial Factors and Health at Work: Evaluation and Intervention Public Health and Digital Technologies during Pandemic: A Multidisciplinary Perspective Public Health Consequences of Social Isolation and Loneliness Public Health Informatics Public Health Informatics Public Health: Feature Papers Public Health: How Safe Is Cardiac Imaging? Quality of Life and Mental Health of the Elderly in Nursing Homes Quality of Life in Orthopedic Diseases Quality of Life, Well-Being and Nurse-Patient Interaction in Late Life Quality of Life: The Interplay between Human Behaviour, Technology and the Environment Quantifying Atmospheric Ammonia and Its Impacts: Measurements, Modeling and Mitigation Radiation and Cancer Risk Real World Data for Population-Based Pediatric Studies Real-World Evidence for Resuscitation Science Reappraisal of Risk Factors for Cardiovascular Diseases Recent Advances and New Perspectives on the Multidisciplinary Management of COVID-19 and Long COVID in the Life Span Recent Advances in Environmental Research Recent Advances in Orthodontics and Clear Aligner Therapy Recent Advances in Public Health Recent Advances in the Management of Chronic Pain Recent Advances on Environmental and Toxicologic Pathology Recent Research in Health Psychology Recreation, Ecosystems and Social Wellbeing: Towards More Effective Synergy Reducing Exposure to Second-Hand Tobacco Smoke Regional Scale Industrial Contamination of Soils and Groundwater — From Risk Assessment to Risk Management Regulation of Muscle Mass, Exercise, Metabolism Rehabilitation in the COVID-19 Pandemic Remote Sensing, Crowd Sensing, and Geospatial Technologies for Public Health Research Workforce and Healthcare Disparities Resilience, Stress, and Risk Factors during the COVID-19 Pandemic Resistance Exercise/Training to Improve Physical Fitness and Health Responsible Risk Governance in Hazardous Industries Retail Strategies to Support Healthy Eating Risk Assessment and Preventive Child Health Care during the First 1000 Days from Conception Onwards Risk Factors for Oral Disease Road Safety: Public Health Challenge Roma Health Roma Health Disadvantage Routes to Improve Health Literacy during the Life-Course Salutogenesis and Coping: Ways to Overcome Stress and Conflicts Salutogenic Cities for Chronic Diseases Prevention Sarcopenia, Exercise and Quality of Life SARS-CoV-2 Variants, Where Is the End? School-Based Prevention Programs for Drug and Alcohol Misuse in Children and Adolescents School-to-Work Transitions: Developmental and Mental Health Outcomes Second Edition of Active Commuting and Active Transportation Second Edition of COVID-19: A Public Health Approach for Health Professionals Second Edition of Effects of Environmental Pollutants on Human Reproductive Health Second Edition of Recent Advances in Polycyclic Aromatic Hydrocarbons Research: Occurrence, Fate, Analysis and Risk Assessment Second Edition of Social-Emotional Development and Learning in Early Childhood across Cultures Second Edition of Teenage Reproductive Health: Pregnancy, Contraception, Unsafe Abortion, Fertility Second Edition of the COVID-19 Pandemic in Europe: Response to Challenges Second Edition of the Current Situation and Distribution of Rare Diseases: Challenges, Prevention, Healthcare, and Effects Second Edition of the Next Frontier in Health Geography: Context and Implications for Interventions Second Edition of the Nutrition Transition and Physical Inactivity and Health Outcomes thereof in Low- and Middle-Income Countries: From Preconception to Adulthood Second Edition of Urban Disaster Resilience and Sustainability Second Edition: Diagnosis and Treatment of ADHD in Adolescents Selected Paper from the 15th International Symposium on Recent Advances in Environmental Health Research Selected Papers from the 1st International Electronic Conference on Environmental Health Sciences Selected Papers from the 3rd International Electronic Conference on Environmental Research and Public Health--Public Health Issues in the Context of the COVID-19 Pandemic Self-Control, Compliance and Adherence to Health Prescriptions Self-Efficacy and Self-Management of Chronic Diseases Sepsis in a Changing Environmental and Technological Landscape Sexual and Domestic Violence and Adolescent Health Sleep Apnea Syndrome Sleep Health Sleep Medicine and Health Sleep Quality Research Small Solutions for Big Water-Related Problems—Innovative Microarrays and Small Sensors to Cope with Water Quality and Food Security Smart Coach and Injuries Prevention in Young Athletes Smoking and Tobacco Control Smoking Cessation Smoking Cessation in Pregnancy Social and Economical Determinants of Health Social and Environmental Determinants of Oral Health Social Determinants and Geographic Disparities in Health and Health Care Social Inequality and Health: Determinants, Mechanisms, and Consequences Social Justice and Administration and Public Health: Rights, Risks and Management Social Marketing’s Contribution to Public Health Social Network Analysis and Public Health Social Network Interventions for Health Behaviours Social Vulnerability and Frailty in Older People Societal Side Effects: The Wider Impact of Pharmaceuticals on Society Socioeconomic Circumstances and Mental Health Soil Pollution and Public Health Soil Pollution: Prevention and Mitigation Solving the Global Water Shortage Crisis: A Focus on Treatment and Reuse of Wastewater Sound and Health related Quality of Life Spatial Dimensions of Public Health: Identifying and Monitoring Vector-Borne Diseases and Their Geographic Diffusion Spatial Epidemiology Spatial Modelling for Public Health Research Spatio-temporal Frameworks for Infectious Disease Epidemiology Sport-Exercise and Stress: A Winning Combination Stress and Health Stress and Training Load Effects on Recovery, Well-Being and Sports Performance Stress Biomarkers Stress, Faith, Resiliency, and Health among Black Men Stroke: Athletes, Cardiac Risk, Physical Fitness, and Fatigue Studies and Advances to Evaluate the Impact of Epidemic Diseases in Modern Times Studies on Heavy Metals and Health Substance and Behavioral Addictions: Co-Occurrence and Specificity Substance and Drug Abuse Prevention Suicide Bereavement and Postvention: Advances in Research, Practice and Policy Suicide in Asia and the Pacific Suicide Prevention among Youth Suicide Prevention and Public Health Sunbathing Habits and Skin Cancer Sustainability of Wine Production and Food Systems in the Mediterranean Region Sustainability: Environmental Studies and Public Health Sustainable Healthy Working Life for All Ages—Work Environment, Age Management and Employability Sustainable Prosperity without Growth, or Damage to Nature Systematic Reviews and Meta-Analyses in Public Health Tackling Long-Term Care Needs in Ageing Societies in (Post) Pandemic Times: Multidisciplinary Research Approaches in an International Perspective Tactical Forces Injury Risk Management Teaching and Learning Process: Psychological Variables in Education, New Applied Technologies and Physical Activity Techniques in Renewable Energy Production and Their Future Aiming Sustainability and Commercialization Technology, Data, and the Assessment of Atmospheric Exposure on Finer Scales Test Your Limits: HIRT, HIIT, and Functional Training The (Un)Sustainable Side of Resilience in the Anthropocene The 2nd Edition: Land Use Changes and the Corresponding Ecological Risks The 2nd Edition: Mindfulness-Based Practice for Health Benefits The 2nd Edition: Stroke: Athletes, Physical Activity, and Resistance Training The 9/11 Disaster and Other Man-Made Trauma Events and Their Health Effects The Assessment of Alternative Interventions to Manage Health Emergencies The Burden of Orthopedic Surgery The Close Connection between Environmental Pollution and Medicinal Prescriptions The Combined Health Effects of Environmental Exposures The Complexity of Chronic Pain The Economics of Mental Illness The Effects of Advance Care Planning in Healthcare The Effects of Non-cigarette Tobacco Use on Health The Effects of Occupational Health and Safety Education and Training on Different Communities The Emerging Role of Sedentary Behaviour in the Health of Youth and Young Adults: Should There be a Recommended Threshold of Sedentary Behaviour Permissible? The Environment Risk of Autism The Environmental, Public Health, and Human Rights Impacts on Enhancing the Quality of Life of People with Intellectual Disability The Evolution of Dentistry in a Changing World between Technological Progress and Environmental Challenges The Evolving Relationship between Science and Disaster Risk Reduction The Health Consequences of Chronic Energy Imbalance The Health Effects of Water Fluoridation The Impact of Parasitology on Public Health The Impact of Sleep Loss on Human Behavior and Neural Activity The Impact of the COVID-19 Pandemic for Health Inequalities The Impact of the Gut Microbiota on Human Health The Impacts of the Built Environment on Public Health The Importance of Mentoring for Diversity, Equity and Inclusion The Importance of Statistical Analysis in the Field of Rehabilitation The Influence of Mediterranean Diet on Health and Environment The Injustices: Social Determinants of Vulnerability to Unhealthy Behaviours The Lived Experience of People Living with Dementia and Caregivers The Most Common Behaviors Associated with Substance Abuse The Negative Effects on Health Due to Noise Exposure The New Era of Treatment for Obesity The Nutrition Transition and Physical Inactivity and Health Outcomes thereof in Low- and Middle-Income Countries: From Preconception to Adulthood The Protection of Quiet Areas as a Public Health Aim Towards Sustainable Health: Approaches, Case Studies and Implementation The Relationship between Children’s Asthma and Air Quality The Role of Health Technology Assessment in Redesigning Chronic Disease Services The Role of Plants and Microorganisms in Ecological Restoration The Role of Science, Technology and Innovation in Ensuring Food Safety and Food Security The Role of Surgical Systems in Promoting Public Health The Social and Health Issues Facing HIV/AIDS Patients The Social Cost and Public Health Impact of Gambling and Online Game Playing Therapeutic Advances and Challenges in the Treatment of Multiple Sclerosis Thermal Comfort and Safety TikTok and Public Health Tobacco Control Tobacco Control 2015 Tobacco Control and Priority Groups Tobacco Control in Vulnerable Population Groups Tobacco Harm Reduction: Policy Considerations to Mitigate Risk to Youth and Adults Tobacco Smoking and Public Health Tobacco Smoking: Public Health, Science and Policy Tobacco Use and Treatment among Cancer Survivors Tobacco Use Research in Youth and Young Adults Tobacco-Related Diseases and Their Impact on Individual and Public Health Issues Together in the Fight against Arthropod-Borne Diseases: A One Health Perspective Topical Advisory Panel Members’ Collection Series: Environmental Science and Environmental Factors That Affect Health Towards More Sustainable Food Systems Toxicology of Xenobiotic Mixtures and Health Toxicology, Exposure Assessment and Epidemiology of Primary and Secondary Ultrafine Particles Traffic Safety and Injury Prevention Transport Impacts on Public Health Trauma, Addiction and Criminality Treating Alcoholism between Harm Reduction and Immediate Abstinence Ultrafine Particles and Potential Health Effects Urban Geochemistry and Human Health Urban Place and Health Equity Using Big Data to Advance Knowledge in Child Maltreatment Using Fuzzy Multi-Criteria Decision-Making Methods for Improving the Performance of Public Emergency Departments during the COVID-19 Outbreak UV-Radiation: From Physics to Impacts Vaccination and Health Outcomes Vaccine Safety and Public Health Violence against Women and Intimate Partner Violence Vitamin D and Public Health Water Desalination Water Microbial Pollution and Disinfection WHO Framework Convention on Tobacco Control. Are Countries Fully Implementing It? WHO Noise and Health Evidence Reviews Whole Systems Approaches to Process Improvement in Health Systems Winter Sports Implications for Training, Environmental and Health Women's Health and the Environment Work and Addictions: From Biology to Practice Work Engagement and Job Crafting Work Environment and Cardiovascular Diseases: From Evidences of Causal Associations to Workplace Interventions Workplace Aggression Workplace Interventions for the Prevention or Amelioration of Musculoskeletal Problems in the Working Population Youth and Child Development and Health Youth Sports, Young Athletes Evaluation, Implications for Performance and Health Youth Violence as a Public Health Issue All Special Issues Volume Issue Number Page Logical Operator Operator AND OR Search Text Search Type All fields Title Abstract Keywords Authors Affiliations Doi Full Text

A group of researchers, including Markus Fiedler at Blekinge Institute of Technology in Sweden has studied the phenomenon of video conference fatigue, or "Zoom fatigue," and concluded that both age, gender and personality play a role in how tired you get. For example, they note that young, introvert women are affected more than others. However, there are several ways to avoid this fatigue. Markus Fiedler, researcher at Blekinge Institute of Technology, has together with other researchers in media design, media psychology and media sociology published a study on the factors behind video conference fatigue. This phenomenon, commonly known as "Zoom fatigue" is something that many have experienced during the pandemic, but for which there is relatively little evidence. The researchers have identified several different factors behind Zoom fatigue. "We have found direct evidence that both age, gender and personality affect. For example, young introvert women are affected more than others. But Zoom fatigue depends on many factors, ranging from personal circumstances via the meeting's organization and technology, to the environment in which the video conference takes place," says Markus Fiedler. The researchers have also found a number of factors that they believe contribute to Zoom fatigue. "We assume that both the length and the number of meetings each day play a role, as well as whether you take breaks or not between meetings. It affects whether there are delays and disturbances in pictures and speech, so that you may talk in each other's mouths, or if you have to struggle to follow the conversation. We have also seen that the environment plays a role. You may be sitting at home and ashamed to show your messy environment," Markus Fiedler explains. However, the researchers believe that there are some simple things you can do to reduce Zoom fatigue: Do not schedule meetings one after the otherUse simple technologyRaise your hand if you want to speak during meetingsTurn off the microphone and camera when it is not needed The researchers state that more research and evidence is needed when it comes to the contributing factors behind Zoom fatigue. They also mean that research is needed on other ways of communicating, for example, immersive experiences and technology. The study is published in the online journal International Journal of Environmental Research and Public Health.

10.3390/ijerph19042061

Earth

Lakes on the Tibetan Plateau freeze later and melt earlier under changing climate, shows study

Yanhong Wu et al, Ice phenology dataset reconstructed from remote sensing and modelling for lakes over the Tibetan Plateau, Scientific Data (2022). DOI: 10.1038/s41597-022-01863-9 Journal information: Scientific Data

https://dx.doi.org/10.1038/s41597-022-01863-9

https://phys.org/news/2022-12-lakes-tibetan-plateau-earlier-climate.html

Abstract The Tibetan Plateau (TP) is a region sensitive to global climate change and has been experiencing substantial environmental changes in the past decades. Lake ice phenology (LIP) is a perceptible indicator reflecting changes of lake thermodynamics in response to global warming. Lake ice phenology over the Tibetan Plateau is however rarely observed and recorded. This research presents a dataset containing 39-year (1978–2016) lake ice phenology data of 132 lakes (each with area >40 km 2 ) over the Tibetan Plateau by combining the strengths of both remote sensing (MOD11A2, MOD10A1) and numerical modelling (air2water). Data validation shows that the ice phenology data derived by our method is highly consistent with that based on existing approaches (with R 2 > 0.75 for all phenology index and RMSE < 5d). The dataset is valuable to investigate the lake-atmosphere interactions and long-term hydrothermal change of lakes across the Tibetan Plateau. Measurement(s) lake ice phenology Technology Type(s) remote sensing and numerical modeling Background & Summary The formation and duration of ice cover for lakes in the Cryosphere plays an important role in thermodynamics of lakes. It interacts with lake water temperature and shapes aquatic ecosystems as well. Lake ice phenology (LIP, e.g., dates of freezing-up and breaking-up) are closely related to temperature variation 1 , 2 . Change of lake ice phenology in the Cryosphere is considered as one of the most reliable pieces of evidence reflecting global warming 3 , 4 , 5 . It could subsequently lead to substantial alteration in biogeochemical processes (e.g., influencing growth of plankton and causing anoxia in deep water 6 ). Research efforts on investigating changes of lake ice phenology are mainly based on ground observation, satellite observation or modelling 7 , 8 , 9 , 10 . Most of the research, however, are for lakes locating in North America or Northern Europe as there are relatively abundant in ground observations 10 . Lake ice phenology observations are not always widely available particularly for regions like the Tibetan Plateau (TP), where ground-based observation is challenging and costly. Observation from space therefore has been becoming more attractive in investigating the lake ice phenology. For instance, with stratified random sample from Sentinel-2 and PlanetScope, Pickens et al . (2021) reported that 41% of inland water area around the globe were seasonally frozen 11 . Though with uncertainties 12 from both data sources 13 , 14 , 15 and extraction methods, various satellite-based observations have shown their merits in tracking the freezing-thawing cycle of lake ice cover 16 . The most commonly used satellite data includes that from the optical sensors like Advanced Very High Resolution Radiometer (AVHRR) 17 and Moderate-resolution Imaging Spectroradiometer (MODIS) 18 , the passive microwave sensors like Scanning Multichannel Microwave Radiometer (SMMR) 3 , the Advanced Microwave Scanning Radiometer for EOS (AMSR-E) 19 and Special Sensor Microwave/Imager (SSM/I) 20 . Active microwave also shows capabilities in monitoring lake ice status based on sensors like European Remote Sensing Satellite (ERS)-1/2 synthetic aperture radar (SAR) 21 and Radarsat-1/2 SAR 22 , but the narrow swath width and the relatively low temporal resolution limited the application of the technology to monitor lake ice phenology at a daily scale and for a large area. It is also noted that the satellite-based observations of lake ice phenology mostly start from the late 1990s and could be with considerable data gaps. For example, the currently available dataset for lake ice phenology in the TP is that derived from microwave brightness temperature measured by AMSR-E, the Advanced Microwave Scanning Radiometer 2 (AMSR2) and Micro-Wave Radiation Imager (MWRI). The dataset however is limited to 51 lakes on the TP covering the period from 2002 to 2015 23 . In the literature of limnology, numerous research efforts also have been invested on developing mathematical models to reconstruct lake ice phenology for the historical period or to predict the response of lake ice phenology to a warming climate 5 , 24 . The model-based approach goes beyond reproducing data on lake ice phenology and can be used to quantify the response of lake ice phenology to climate change or climate variation. The models may cover different level of details on lake thermodynamics, subject to their research focuses 25 , 26 , 27 , 28 , 29 . It is recognized that a well simplified process-based model (like air2water 30 , 31 ) could perform as good as a more complicated model (like MINLAKE 26 , LIMNOS 27 , HIGHTSI 28 ) if the research interest is mainly on the timing of freezing and thawing of lake ice. The simplified model could be more competitive and applicable for regions with little in-situ observations (like the TP) as it requires much less data for running and calibration. The dataset herein provides complete, consistent and continuous time series (1978–2017) of reconstructed ice phenology for 132 lakes in the TP (Fig. 1 ) by combing the strengths of remote sensing and mathematical modelling forced by meteorological data. The lake surface water temperature (LSWT) derived from MODIS products (MOD11A2) 32 is applied to calibrate and validate the numerical model in reproducing the consecutive daily surface water temperature for the studied period, based on which algorithm has been developed to determine the ice phenology indices (including dates of freezing-up and breaking-up and durations). Lake ice phenology derived directly from satellite observations in this research and by other researchers is used to validate the reconstructed LIP 33 . Fig. 1 Locations of studied lakes in the Tibetan Plateau. Full size image The reconstructed LIP represents a unique dataset to investigate the long-term trends of lake ice phenology in the TP under a warming climate. The dataset can also be used for a variety of applications related to climate change, limnology, hydrology, and aquatic ecology. It is conductive to assess the impacts of climate warming on dynamics of water/heat budget and aquatic biota in lakes across the TP. Methods Figure 2 shows the general framework of our study in determining the lake ice phenology for the 132 largest lakes across the Tibetan Plateau. The satellite-based lake surface temperature is used to calibrate the modified air2water model. The air temperature data from meteorological station is the solely required input data to drive the model. The temperature threshold of lake ice phenology is determined with the aid of MODIS snow cover product. Details of the procedures and methods are described in the following sections. Fig. 2 Flowchart in producing the lake ice phenology (LIP) dataset. Full size image Data The modified air2water model requires solely air temperature as input. For the TP, the best publicly accessible ground-based observation of daily air temperature is that from the dataset of daily climate data from Chinese surface stations (V3.0) provided by China Meteorological Data Service Centre of National Meteorological Information Centre ( ). The daily air temperature from the dataset is only available at 31 meteorological stations located in the TP. The spread and density of the meteorological stations are big constraints for earth system research in the region. We noticed that quite a few research efforts have been invested to provide reliable reanalysis climate datasets for the TP (e.g., China meteorological forcing dataset). The public available climate reanalysis datasets however are found not necessarily of high good quality for our simulation when we compare them against the observations from the meteorological stations. Hence, we use the air temperature interpolated from the nearest station with altitude adjustment (by a lapse rate of 0.65 °C 100 m −1 ) 34 to provide air temperature input to drive the modified air2water model. The direct output of the model is lake surface water temperature. Hence, the model is firstly calibrated and validated against the satellite-based observation of lake surface water temperature deriving from the MOD11A2 32 (i.e., Terra product of Land Surface Temperature/Emissivity 8-day L3 Global 1 km). The surface water temperature of a specific lake is the lake-wide mean temperature based on all pixels within the lake. Boundaries of the lakes (shapefiles included in the dataset) are mainly from National Tibetan Plateau Data Center (TPDC, ) and are cross-checked by maps from HydroLAKES ( ) 35 and Google Earth. The lake-wide MODIS-based surface water temperature is the mean of daytime and night-time LST in the MOD11A2. All the data processing procedures are implemented via Google Earth Engine (GEE) 36 . Prior to determine the lake ice phenology given the modelled lake surface temperature, temperature thresholds should be defined, which is done herein with the aid of remotely sensed snow cover product, i.e., MOD10A1 37 . The MOD10A1 dataset is a snow cover product with a spatial resolution of 500 m containing daily, gridded snow cover and albedo derived from radiance data acquired by MODIS on board the Terra satellite. Snow cover is identified using the Normalized Difference Snow Index (NDSI) and a series of screens designed to alleviate errors and flag uncertain snow cover detections 37 . Lake surface temperature modeling We used a slightly modified air2water model to simulate the daily lake surface temperature for the period 1978–2017. The air2water 27 , 28 is a semi-physical model designed to simulate lake surface temperature principally based on lake surface energy balance expressed as: $$\rho {c}_{p}{V}_{s}\frac{d{T}_{w}}{dt}=A{H}_{net}$$ (1) where T w is water surface temperature at time t (day step in this study), H net is the net heat flux per unit surface, A is the surface area of the lake, V s is the volume of water involved in the heat exchange with the atmosphere, ρ and c p is the density and specific heat capacity of water. The model considers all contributions in the energy balance ( H net ) as a function of air temperature or is included in the parameters of the model. The original air2water model has simplified the lake energy balance by introducing eight calibratable parameters a 1–8 with air temperature as the only required input, which can be expressed as: $$\frac{d{T}_{w}}{dt}=\frac{1}{\delta }\left\{{a}_{1}+{a}_{2}{T}_{a}-{a}_{3}{T}_{w}+{a}_{5}\,{\cos }\left[2\pi \left(\frac{t}{{t}_{y}}-{a}_{6}\right)\right]\right\}$$ (2) $$\delta =\left\{\begin{array}{c}{\exp }\left(-\frac{{T}_{w}-{T}_{h}}{{a}_{4}}\right),\quad \quad \quad for\quad {T}_{w}\ge {T}_{h}\\ {\exp }\left(-\frac{{T}_{h}-{T}_{w}}{{a}_{7}}\right)+{\exp }\left(-\frac{{T}_{w}}{{a}_{8}}\right),\quad \quad for\quad {T}_{w}\le {T}_{h}\end{array}\right.$$ (3) where, T a is air temperature, t y is days of a year. δ = D/D r is normalized well-mixed depth, where D is the depth of well-mixed surface layer (the epilimnion thickness), D r is the maximum thickness. T h is the deep water temperature assumed to be 4 °C. The model has been tested efficient in different regions 38 , 39 . However, the original air2water is limited to simulate the surface temperature of open water, which results in difficulties for applications in lakes with long ice cover duration. For this sake, we assume that when the lake is completely covered by ice, the heat exchange between air and water is blocked and surface energy balance would be replaced by land surface energy balance without mixing depth in water 36 , which indicates δ = 1 and Eq. ( 2 ) be rewritten as: $$\frac{d{T}_{i}}{dt}={a}_{9}+{a}_{10}{T}_{a}-{a}_{11}{T}_{w}+{a}_{12}\,{\cos }\left[2\pi \left(\frac{t}{{t}_{y}}-{a}_{13}\right)\right]$$ (4) where T i is the ice surface temperature, a 9–13 have similar physical significance to a 1,2,3,5,6 . To represent the shifts of lake between open water and ice-covered, two additional parameters a 14 and a 15 are introduced to determine the states of the lake, which the lake surface water temperature ( LSWT ) can be expressed as: $$LSWT=\left\{\begin{array}{c}{T}_{w},\;for\;{T}_{L}\ge {a}_{15}\\ {T}_{i},\;for\;{T}_{L}\le {a}_{14}\\ (1-{K}_{ice}){T}_{w}+{K}_{ice}{T}_{i},\;for\;{a}_{15} > {T}_{L} > {a}_{14}\end{array}\right.$$ (5) where we set \({K}_{ice}=\sqrt{({a}_{15}-LSWT)/({a}_{15}-{a}_{14})}\) as the proportion of ice on the surface of the lake. Same as the original model, the model equations were solved numerically by using the Crank-Nicolson numerical scheme at daily time step. The model is calibrated against the LSWT derived from MOD11A2 using the PSO (Particle Swarm Optimization) approach 31 , 40 . The objective function of model calibration is the root mean square error (RMSE). To ensure the rationality of the calibrated model parameters, a physically consistent priori range are assigned to each parameter 31 and D r is bounded by the average depth of the lake obtained from Records of Lakes in China 41 and HydroLAKES 35 . The calibration period is 2000–2012, while 2013–2017 is for validation. Temperature thresholds for ice phenology The lake surface temperature output from air2water model is used to derive from lake ice phenology given prior defined temperature thresholds. The temperature thresholds are defined with the aid of the snow cover products (MOD10A1), where the MOD10A1 images with cloud proportion over the lake exceeding 20% are neglected. For each lake, the ratio of “inland water” ( R water ) to total lake area is firstly calculated as: \({R}_{water}={N}_{water}/\left({N}_{water}+{N}_{ice}\right)\) , where N water is the pixel number of “inland water”, while N ice is the pixel number of “ice”. The lake is assumed to be open when R water > 90% and frozen when R water < 10% 42 , 43 . As shown in Fig. 3 The temperature thresholds on determining lake ice phenology are then defined as the temperature of the date when water area percentage ( R water ) drops down below 10% or rises above 90%. The multi-year medians of the temperature thresholds are finally considered as the overall thresholds to determine the starting and ending dates of freezing-up ( T FUS and T FUE ) and breaking-up of lake ice phenology ( T BUS and T BUE ) separately for each lake. Fig. 3 Determination of lake surface water temperature (LSWT) thresholds for identifying the lake ice phenology indicators. T FUS, i T FUE,i , T BUS,i and T BUE,i are the temperature threshold cand i dates derived for each specific year i (=1,2,3). R water is the percentage of open water area within a lake. Data of this figure is from the Qinghai Lake for the period 2008–2009. Full size image Lake ice phenology determination The long-term lake ice phenology for the period 1978–2016 is derived from the simulated lake surface temperature given the temperature thresholds. The lake ice phenology considered herein consists of 6 indicators reflecting the timing or duration of different freezing/thawing phases. In addition to starting and ending dates of freezing-up (FUS and FUE) and breaking-up (BUS and BUE), the ice-covered duration (IceD, in days) and the fully ice-covered duration (CID, in days) are also presented in the dataset. The indicators are defined as: $$\left\{\begin{array}{c}FUS={\min }\left(t\right)s.t.\,LSW{T}_{t} < {T}_{FUS}\,and\,LSW{T}_{t-1} > {T}_{FUS}\\ FUE={\max }\left(t\right)s.t.\,LSW{T}_{t} < {T}_{FUE}\,and\,LSW{T}_{t-1} > {T}_{FUE}\\ BUS={\min }\left(t\right),s.t.\,LSW{T}_{t} > {T}_{BUS}\,and\,LSW{T}_{t-1} < {T}_{BUS}\\ BUE={\max }\left(t\right),s.t.\,LSW{T}_{t} > {T}_{BUE}\,and\,LSW{T}_{t-1} < {T}_{BUE}\\ IceD=BUE-FUS\\ CID=BUS-FUE\end{array}\right.$$ (6) Where LSWT t is the simulated lake surface temperature at date t . The date t is expressed as the day of a calendar year (starting at January 1 st ). Since the lake ice phenology could spans two consecutive years, we assign herein the value of t to range from 1 to 730, where t >365 represents the date in the next consecutive calendar year. Data Records The lake ice phenology data across Tibetan Plateau is archived and openly accessible via Figshare 44 via the link: . Description about the 132 lakes is shown in the lake_info.csv file, including lake ID, name, latitude, longitude, lake area, water level, depth, water volume, freezing property, and glacial replenishment. The information about the lakes is mainly collected from Records of Lakes in China 41 , HydroLAKES 35 , and Map of Glaciers and Lakes on the Tibetan Plateau and Adjoining Regions. The Phenology folder in the dataset consists of 132 CSV files. Each CSV file is named according to the lake ID presented in lake_info.csv. The columns of each CSV file are listed in Table 1 . Table 1 Files and formats of the dataset. Full size table Technical Validation Quality control The reconstructed LIP dataset is produced with strict quality control throughout the procedures. The boundaries of the lakes are examined via professional GIS tools (both ArcGIS and Google Map) to ensure the consistency of their geographic locations, and topography. Lake-wide mean surface temperature instead of that of lake centroid is derived from MOD11A2 for model calibration. The 132 lakes presented in the dataset are all with good model performance (Nash-Sutcliffe efficiency coefficient 45 above 0.6) when the simulated lake surface temperature is compared against the MODIS-based one. Suspicious outliers (i.e., abnormal extremely high or low values) are eliminated from satellite-based lake surface temperature and snow product in providing reliable data for robust model calibration and validation. The MOD10A1 is used as an aid in determining the temperature thresholds of LIP. It is noted that the MOD10A1 is not always of good quality in continuously retrieving the ice-covered ratio of a lake due to accidental misclassification and effects of cloud cover. To alleviate the uncertainties induced by the optical RS data, we use open water ratio (see Methods) instead of directly using ice-covered ratio of a lake in determining the temperature threshold, and the medians of the temperature thresholds are finally accepted for robust lake ice phenology identification. The reconstructed LIP dataset is validated by comparing against the published LIP data as described below. Validation of the reconstructed LIP dataset There is little in-situ observation of lake ice phenology in the TP available for a complete one-by-one validation of our reconstructed LIP dataset. We herein validate our dataset in two different ways. Firstly, we validate the effectiveness of our approach by comparing the reconstructed LIP against in-situ observations of two lakes in the TP 46 , 47 and one lake in Canda 48 . The comparison is shown in Supplementary Fig. S1 suggesting that the estimated ice phenology overall is comparable with the in-situ observations, where the root mean squared error (RMSE) of the estimated FUE in Nam Co is around 7 days, RMSE of the estimated BUE in Lesser Slave Lake is near 9 days and RMSE of the estimated in Qinghai Lake is around 12 days. It should be noted that the bias could be due to both the uncertainties in the model and in the climate input to the model as well. We further compare the reconstructed LIP against two existing datasets developed based on remotely sensed data, which were published by Cai et al . 32 and Qiu et al . 23 respectively. The dataset published by Cai et al . 32 is retrieved from MOD10A1 and Landsat data, providing mean lake ice phenology of 44 TP lakes for the period 2000–2017. Figure 4 shows the comparison of the mean annual LIP from our reconstructed dataset against that from Cai et al . for the 44 lakes. It can be seen that the LIP dataset produced by our approach is in high agreement with that solely based on MOD10A1, where the R 2 for all the four phenology indices (i.e., FUS, FUE, BUS and BUE) are higher than 0.96 and the root mean squared errors (i.e., RMSE) are all within 5 days. Fig. 4 Comparison of long-term mean lake ice phenology derived in this research against that from Cai et al . 32 , where FUS, FUE, BUS, BUE represents starting date of freezing-up, ending date of freezing-up, starting date of breaking-up, and ending date of breaking-up respectively. The unit of FUS, FUE, BUS and BUE is day of year with the number larger than 365 indicates dates in the next calendar year. Full size image The reconstructed LIP time series is also compared against the dataset published by Qiu et al . 23 , which was derived from microwave brightness temperature measured by AMSR-E, AMSR2 and MWRI. The dataset from Qiu et al . provides LIP of 51 lakes in High Asia for the period from 2002 to 2016, among which there are 35 lakes identical to the lakes in our dataset. For the 35 lakes in the period 2002–2015 (sample size = 35 × 14), the statistically significant correlations are found between the two datasets for all the six phenology indices with R 2 all above 0.75 (Fig. 5 ). The discrepancy between the two datasets is found to be the lowest in FUS (RMSE = 8.6 days) and to be the highest in BUS (RMSE = 15 days). The relative higher discrepancy of the two datasets in both BUS and FUE results at noticeable bias of the CID, which by definition equals to number of days between FUE and BUS. The uncertainties of both our reconstructed LIP dataset and those built on satellite-observations however need to be further quantified in the future particularly where in-situ observations becoming available. Fig. 5 Comparison of lake ice phenology derived in this research against that from Qiu et al . 23 , where IceD and CID represents ice-covered duration and fully ice-covering duration respectively with unit day. The meaning and unit of FUS, FUE, BUS and BUE are the same to that in Fig. 4 . Full size image Spatial distribution of LIP across the TP Figure 6 shows the long-term mean spatial distribution of the LIP indices based on the reconstructed dataset. Overall, the results suggest that the phenology is close related to latitude (Fig. S2 in Supplementary). For lake in higher latitude, it tends to freeze-up earlier but break-up later hence with longer ice-covered duration. Overall, from south to north, FUS ranges from October 15 th to January 15 th while FUE varies from November 9 th to January 21 th . The gradients of FUS and FUE with respect to latitude are around 6.7 days/degree and 5.7 days/degree respectively (Fig. S2 in Supplementary). For ice-breaking dates, from south to north, BUS ranges between 27 th January and 22 th June, while BUE varies between 2 nd March and 5 th July. The latitude gradients of BUS and BUE are about 11.8 days/degree and 8.5 days/degree respectively. From south to north, ice-covered duration ranges from 15 days to 215 days (CID) or from 85 to 255 days (IceD). The latitude gradients of CID and IceD are 17.5 days/degree and 15.2 days/degree respectively. Lake ice phenology is also affected by altitude (Fig. S3 in Supplementary). Statistically significant correlations are found between LIP and altitude particularly for lakes locating between 4500 to 5000 m (account for 70% of lakes in this dataset). Overall, FUS and FUE are in negative correlations with altitude, while BUS, BUE, IceD and CID are in positive correlations with altitude. It suggests that for lakes with higher altitude, earlier the FUS and FUE, later BUS and BUE and longer IceD and CID can all be expected. The altitude gradient of FUS and FUE is 9.4 days/hm and 5.7 days/hm respectively, while it is around −13 days/hm and −11.7 days/hm for BUS and BUE respectively. The altitude gradient of CID and IceD is estimated to be 18.9 days/hm and 21.4 days/hm (Fig. S3 in Supplementary). Fig. 6 Spatial distribution of lake ice phenology over the TP. The meaning of FUS, FUE, BUS, BUE, CID and IceD are the same to that in Fig. 4 and Fig. 5 . Full size image Code availability Codes in developing the lake ice phenology dataset are freely available. Codes of the modified air2water model for lake surface water temperature simulation are available at: . Codes for lake ice phenology identification are written using Matlab and are available at: .

Lakes on the Tibetan Plateau show a trend of later freezing-up, earlier breaking-up and thus shorten ice-covered duration since the late 1970s, according to a study by researchers from the International Research Center of Big Data for Sustainable Development Goals (CBAS) and the Aerospace Information Research Institute (AIR) of the Chinese Academy of Sciences (CAS). These findings explicitly reflect the changes in thermodynamics of lakes under a warmer climate and implicate substantial associated alteration in biogeochemical processes in this region. The study was published in Scientific Data on Dec. 2. The study was based on the lake ice phenology dataset reconstructed from remote sensing and modeling for 132 lakes over the Tibetan Plateau. Based on the reconstructed dataset, the research team also reported that for lakes from the southern to northern Tibetan Plateau, the completely ice-covered duration (CID) ranges from 15 days to 215 days. The latitude gradients of CID are 17.5 days/degree. For lakes located between 4,500 to 5,000 m, the altitude gradient of CID is estimated to be 18.9 days/hm. Ice phenology (i.e., the timing of ice-on and ice-off) records are rarely available for lakes over the Tibetan Plateau due to challenges and cost of ground-based observations. The research team developed a framework by combining the strengths of satellite-based observation and numerical modeling in reconstructing complete, consistent, and continuous time series for 132 lakes across the Tibetan Plateau for the period 1978 to 2016. The produced dataset can help to understand the response of lake thermodynamics to climate change and the subsequent hydrological and ecological impacts.

10.1038/s41597-022-01863-9

Medicine

Research suggests a 15-minute 'daily mile' could enhance health of the world's children

Ross A. Chesham et al. The Daily Mile makes primary school children more active, less sedentary and improves their fitness and body composition: a quasi-experimental pilot study, BMC Medicine (2018). DOI: 10.1186/s12916-018-1049-z Journal information: BMC Medicine

http://dx.doi.org/10.1186/s12916-018-1049-z

https://medicalxpress.com/news/2018-05-minute-daily-mile-health-world.html

Abstract Background The Daily Mile is a physical activity programme made popular by a school in Stirling, Scotland. It is promoted by the Scottish Government and is growing in popularity nationally and internationally. The aim is that each day, during class time, pupils run or walk outside for 15 min (~1 mile) at a self-selected pace. It is anecdotally reported to have a number of physiological benefits including increased physical activity, reduced sedentary behaviour, increased fitness and improved body composition. This study aimed to investigate these reports. Methods We conducted a quasi-experimental repeated measures pilot study in two primary schools in the Stirling Council area: one school with, and one without, intention to introduce the Daily Mile. Pupils at the control school followed their usual curriculum. Of the 504 children attending the schools, 391 children in primary classes 1–7 (age 4–12 years) at the baseline assessment took part. The follow-up assessment was in the same academic year. Outcomes were accelerometer-assessed average daily moderate to vigorous intensity physical activity (MVPA) and average daily sedentary behaviour, 20-m shuttle run fitness test performance and adiposity assessed by the sum of skinfolds at four sites. Valid data at both time points were collected for 118, 118, 357 and 327 children, respectively, for each outcome. Results After correction for age and gender, significant improvements were observed in the intervention school relative to the control school for MVPA, sedentary time, fitness and body composition. For MVPA, a relative increase of 9.1 min per day (95% confidence interval or 95%CI 5.1–13.2 min, standardised mean difference SMD = 0.407, p = 0.027) was observed. For sedentary time, there was a relative decrease of 18.2 min per day (10.7–25.7 min, SMD = 0.437, p = 0.017). For the shuttle run, there was a relative increase of 39.1 m (21.9–56.3, SMD = 0.236, p = 0.037). For the skinfolds, there was a relative decrease of 1.4 mm (0.8–2.0 mm, SMD = 0.246, p = 0.036). Similar results were obtained when a correction for socioeconomic groupings was included. Conclusions The findings show that in primary school children, the Daily Mile intervention is effective at increasing levels of MVPA, reducing sedentary time, increasing physical fitness and improving body composition. These findings have relevance for teachers, policymakers, public health practitioners, and health researchers. Peer Review reports Background The Daily Mile is a physical activity intervention developed at St Ninian’s Primary School (Stirling, Scotland) in 2012 [ 1 ]. The initial aim was to improve the fitness of children; although, there have since been many additional benefits anecdotally reported by children, teachers and parents [ 1 ]. These include improved physical activity, sedentary time, physical fitness, body composition, sleep, diet, concentration, well-being and obesity levels. However, no objectively measured scientific evidence has yet been gathered on the validity of these reports. The successful implementation of the Daily Mile at many schools, its continued maintenance and its increasing popularity seem to be a result of the simplicity of the activity, the autonomy given to classroom teachers over when they do it during the school day and the pupil-determined pace. Since its development, the Daily Mile has been rolled out across the country by the Scottish government [ 2 ]. It is estimated by Education Scotland that, from local authorities who responded to their query, ~ 50% of primary schools in Scotland are already doing the Daily Mile with a further 18% planning to do the Daily Mile soon (personal communication). It has also become popular in the rest of the UK with interest from the UK government [ 3 , 4 ]. Additionally, it has been introduced in the Netherlands, Belgium and parts of the USA with interest from many other countries [ 5 ], despite the lack of rigorous evidence on the efficacy of taking part. Globally, physical activity levels are low [ 6 ]. Furthermore, low physical activity levels in childhood are predictive of low physical activity levels in adulthood [ 7 ]. The World Health Organisation (WHO) considers policies and interventions to increase physical activity levels to be important in all age groups and that any potential harm of increasing physical activity is outweighed by the associated benefits [ 8 , 9 ]. In a recent accelerometer study on children (2–11 years old) from eight European countries, the proportion of children achieving 60+ minutes per day (recommended for 5–18 year olds [ 10 ]) of moderate to vigorous physical activity (MVPA) ranged from 9.5% to 34.1% in boys and from 2.0% to 14.7% in girls [ 11 ]. MVPA levels generally decline with age in both genders as sedentary time increases [ 12 ]. MVPA levels are also believed to be influenced by socioeconomic status; however, systematic reviews are unclear on the consistency of this relationship and more evidence needs to be gathered [ 13 ]. School-based physical activity interventions like the Daily Mile are appealing because they include whole classes, therefore they reach many children regardless of socioeconomic status, physical activity level or fitness level. They also break up sedentary time as they occur during lessons and have the potential to reach much larger proportions of the population than opt-in groups like sports clubs. Therefore, understanding the impact of the Daily Mile on physical activity and sedentary behaviour levels is of key importance. Overweight and obesity rates are of pandemic proportions and considered to be a key target for the WHO [ 8 ]. In Scotland, 30% of children (29% of boys and 32% of girls) aged 7–11 years were overweight or obese in 2015 [ 14 ], a figure similar to that in England [ 15 ]. At the same time, there is evidence of a decline in the performance of children in the 20-m shuttle run (an indicator of physical endurance fitness) [ 16 ]. Low fitness and low levels of physical activity in adults and children are associated with a number of risk factors for non-communicable diseases and adverse health outcomes including obesity, cardiovascular disease, diabetes, some cancers, low mood and poor cognitive function [ 17 , 18 , 19 , 20 , 21 ]. Additionally, both overweight and low fitness are also known to be related to lower socioeconomic status [ 22 , 23 ]. Studies into the impact of the Daily Mile on fitness and body composition are important for understanding its potential to improve public health, including health inequalities, and for developing future public health policies. The Daily Mile at least has the potential to impact on key areas of global public health. Whilst published evidence shows a positive health relationship between physical activity, particularly MVPA, and these outcomes [ 24 ], it is unknown whether 15 min of exercise, particularly with no expectation of intensity, will provide such benefits. Given the associated loss of academic classroom time (up to 75 min per week), it is paramount that evidence on the impact of the Daily Mile on each anecdotally reported benefit is gathered to ensure that government policies are appropriate. Furthermore, any associated benefits in physiological health are likely to be small [ 25 ]; thus, it is essential to use gold standard measurement techniques. Therefore, the aims of this study are to assess the anecdotally reported physiological benefits of participation in the Daily Mile. Specifically, using a repeated measures design and gold standard measurement techniques, we will assess whether the introduction of the Daily Mile into a primary school setting leads to increased MVPA, reduced sedentary time, improved fitness and improved body composition. Methods Study design and ethics “Using the Daily Mile to turn the WHEEL” (Well-being, Health, Exercise, Enjoyment and Learning) is a school-based quasi-experimental study designed to assess the anecdotally reported benefits of taking part in the Daily Mile. Ethical approval was obtained from the University of Stirling, School of Sport Research Ethics Committee (reference number 760). Approval was also obtained from the Director of Children, Young People and Education at Stirling Council. Eligibility and recruitment State primary schools in the Stirling Council areas were eligible for inclusion. Recruiting from only one council area reduced any potential variance in the delivery of education that may have impacted on outcome measurements. Two local primary schools were identified and approached: one that was not doing the Daily Mile but was intending to introduce it (the intervention school) and one that was not doing the Daily Mile and did not intend to introduce it (the control school). Both schools had expressed a desire to introduce the Daily Mile although the control school felt that it would not be possible due to the layout of its playground. Both schools had a range of levels of deprivation, although the majority of pupils were from higher socioeconomic quintiles (Fig. 1 ). Participants were children in all years (age 4–12 years) at the time of recruitment. Fig. 1 Trial profile. N is the number of school pupil participants. The percentage of female participants for each measurement is shown in parentheses. Individual boxes show the totals for each school at baseline, or at follow-up, or the totals for participants with valid measurements in both rounds. Not all pupils with follow-up measurements had been measured at the baseline (or vice versa), since different pupils were absent at the main and follow-up assessments, some refused to complete certain measurements at each time point and some pupils left or moved between schools. BMI body mass index, SIMD Scottish Index of Multiple Deprivation Full size image Once the schools had agreed to participate in the study, parents and guardians of the children were sent a letter and information sheet about the study with an opt-in consent form. For children in primary classes 4–7 (age 7–12 years), an additional child consent form was included. Information sessions were held in both schools to allow parents to ask questions and see the study equipment. They were also given the opportunity to contact the research team by email or phone to discuss the study. Information about being able to withdraw from the study at any stage was given in the information sheet, consent form and verbally. All of the children were also asked to confirm verbally that they were happy to take part on each day of testing. Intervention The Daily Mile is a school-based physical activity intervention made popular by a primary school in the Stirling Council area, Scotland [ 1 ]. It involves children going outside, at a time of the classroom teacher’s choosing, for ~ 15 min of exercise at a pace self-selected by each individual child. This is done during normal classroom time and is in addition to time spent in physical education or scheduled breaks. Typically, it involves laps of a football pitch or playground area. Children often talk as they go and perform a mixture of walking and running. Those who run the whole time will complete ~1 mile in 15 min. Children wear their normal school clothes; most wear their normal school shoes and jackets are only worn in cold or wet weather. It is completed on most days regardless of weather conditions. A leaflet produced by the originator school was given to the school implementing the Daily Mile. No additional instructions for initiation of the Daily Mile were given by the research team. Participant involvement The outcome measures selected for the study were chosen based on anecdotal reports of the influence of taking part in the Daily Mile. This information had been gathered from the children, their parents and their teachers by the originator school. The research question, study design and specific outcome measures were developed in part during meetings with the head teacher at the originator school, who was familiar with this information. After publication, the results from this study will be disseminated to the schools involved as copies of the manuscript and infographics, and during question and answer sessions in the schools. Outcome measures The primary physiology-related outcome measurements for the WHEEL project were accelerometer-assessed MVPA and sedentary time, fitness assessed using a 20-m shuttle run, and body composition assessed using skinfolds. Additional, cognitive- and well-being-related outcome measurements will be reported elsewhere. Participant assessments Baseline assessments (before the intervention) were carried out in October 2015 for the intervention school and March 2016 for the control school. Outcome assessments were completed in May 2016 for the intervention school and June 2016 for the control school. Identical protocols and procedures were used at both assessments. They were undertaken by trained fieldworkers. The fieldworkers worked in groups such that at least one member of each group was disclosure-checked under the Protecting Vulnerable Groups scheme [ 26 ]. Children completed tasks in a random order and were given stickers for completing each measurement type. Testing sessions lasted between 1 and 2 h and were carried out over 2 weeks, depending on class size and school timetable. ActiGraph accelerometers were used to assess physical activity and sedentary time. Five models of accelerometer were used: ActiGraph wGT3X-BT, wGT3X+, GT3X+, GT3X and GT1M. The GT1M and GT3X monitors are comparable when classifying total time spent in specific intensity categories [ 27 ]. There is also strong agreement between the GT1M, GT3X and GT3X+ accelerometers, making it acceptable to use them in the same study [ 28 ]. The primary difference between the wGT3X-BT and GT3X+ accelerometers is the ability of the wGT3X-BT to communicate wirelessly and all measurements should function identically (personal communication from ActiGraph). Accelerometers were worn around the waist on the left side with the same orientation to standardise the position. Children were asked to wear the belt for eight consecutive days during waking hours (except when bathing or swimming). A poster and instruction sheet with visual and written prompts were provided to each child to remind them of how and when to wear the accelerometer. Upon collection by the research team, the data were downloaded with the ActiLife 6 software (ActiGraph LLC, USA). A valid measurement required at least 10 h wear for 3 days [ 29 ]. A 60-s epoch was used and non-wear time was defined as strings of consecutive zeros lasting 60 min or more [ 30 ]. The accelerometer output is in counts per minute (cpm). Evenson cut points [ 31 ] were used to define time spent being sedentary (≤100 cpm) and time spent in MVPA (≥2296 cpm). The accelerometer data were corrected for wear time in addition to gender and age in days on the day of testing in the main analysis. The maximal multistage 20-m shuttle run test [ 32 ] was assessed using the Multistage Fitness Test CD (Sports Coach UK, UK) using the standard procedure. Cones marked shuttle boundaries and lanes. Between four (younger) and eight (older) children completed the test at the same time depending on their age group. Instructions and a demonstration of the test were provided prior to each group participating and additional verbal instructions were provided during the test as necessary. The test began at 8.5 km·h − 1 and after each minute increased by 0.5 km·h − 1 . When a child was unable to reach the 20-m line prior to the bleep twice in a row, they were asked to stop and their level and shuttle score was recorded. Shuttle run tests were performed outside on tarmac. Age-corrected \dot{\mathrm{V}}{\mathrm{O}}_2\max \dot{\mathrm{V}}{\mathrm{O}}_2\max scores were created according to the method of [ 32 ] for comparison with other studies (Additional file 1 : Tables S1–S3). However, this corrects for age in years and therefore, it was not used for the main analysis. Instead, shuttle distance corrected for gender and age in days on day of testing was used to give an improved resolution to the correction in the main analysis. All skinfold measurements were completed with the child behind a privacy screen with at least two disclosure-checked fieldworkers present. Fieldworkers taking skinfold measurements were trained by an International Society for the Advancement of Kinanthropometry (ISAK) Level 3 instructor prior to involvement in the study and all measurements were taken according to standard ISAK procedures [ 33 ]. Triceps, biceps, iliac crest and subscapular skinfolds were measured using Harpenden skinfold callipers (Baty International, UK). All measurements were taken from the right-hand side with the child standing. Where appropriate, they stood on an anthropometric box. Skinfold measurements were summed prior to analysis to give a sum of skinfolds (in millimetres). Skinfold data were corrected for gender and age in days on day of testing in the main analysis. Height was measured, to the nearest 1 mm, without shoes using the Leicester Height Measure (Seca, UK) according to standard ISAK procedures [ 33 ]. Body weight was measured without shoes in light clothing to the nearest 0.1 kg using electronic Sensa 804 scales (Seca, UK). Height (m) and weight (kg) were used to calculate the body mass index (BMI = weight/height 2 ). BMI z scores relative for age were calculated using UK 1990 reference data in the LMS Growth add-in for Microsoft Excel [ 34 ]. Healthy weight was defined as BMI z score < 1.04, overweight as BMI z score of 1.04–1.63 and obesity as BMI z score ≥ 1.64. All anthropometric measurements were taken twice and the average taken for the analyses. Where there was a substantial difference between the two measurements, a third measurement was taken and the median value was used for the analyses. Additional Information Schools provided the date of birth and postcodes for all consented pupils. The date of birth allowed the analyses to be corrected for age on day of testing. The postcode allowed the assignment of the Scottish Index of Multiple Deprivation (SIMD) [ 35 ]. SIMD combines data from seven different domains of deprivation into a single score: income, employment, health, education, access to services, crime and housing. However, it should be noted that this gives a postcode-specific deprivation score that may not reflect that of an individual household. Statistical analyses Descriptive statistics, Pearson χ 2 and odds ratios were calculated in Excel 2013. Baseline group frequency comparisons were performed using a Pearson χ 2 test. Baseline group mean comparisons were performed using Student’s t test (uncorrected) or general linear model ANOVA (corrected for age at time of testing, gender and age*gender). The main analyses were performed in SPSS Statistics (version 21.0.0.1). General linear model regression analyses with repeated measures were used to investigate the effect of doing the Daily Mile. Analyses of all outcome measures included an adjustment for the common confounders: age at time of testing, gender and age*gender. This controls for the effects of age and gender and any different effects of age in the two genders as well as any differences in length of time in the study. Analyses for MVPA and sedentary time were additionally corrected for accelerometer wear time as a covariate. Analyses were conducted first without and then with a correction for SIMD. Standardised mean differences (SMDs), or effect sizes, are calculated as a change measured in the intervention school relative to the control school as a proportion of the pooled standard deviation of the change. An SMD of 0.2–0.5 is considered to be small, an SMD of 0.5–0.8 is considered to be medium and an SMD of 0.8 or above is considered to be large [ 36 ]. No correction for multiple testing was made since all four primary outcome measures were anecdotally reported to be influenced by the Daily Mile. Results Figure 1 shows the study profile. Overall 77.6% of eligible pupils consented to the study (77.3% in the intervention school and 78.1% in the control school). A total of 371 pupils (247 and 124 in the intervention and control schools respectively) provided at least one measurement in both rounds of testing. All consented children, irrespective of whether they had all measurements or not, are included in Fig. 1 . For each outcome measurement, the information in Tables 1 and 2 is based only on children who had measurements at both time points. Table 1 Baseline characteristics of participants by study group and gender Full size table Table 2 Baseline characteristics of participants by study group and SIMD Full size table Consent for a complete set of measurements was given in the majority of cases; although, on the day of testing, some pupils did not wish to give verbal consent for skinfold measurements, had worn inappropriate clothing for some skinfold measurements or were unable to complete individual tests due to pre-existing minor injuries or ailments. In total, at both baseline and follow-up, between 84% and 94% of originally consented participants had data on fitness and body composition outcomes and the proportions were similar at both baseline and follow-up, in the intervention and control schools and in males and females (Fig. 1 ). Three pupils refused to wear accelerometers. The proportion of pupils with valid accelerometer data was lower than other measurements although the proportion of males and females was similar. Only 32% of participants in the intervention school and 67% in the control school had valid accelerometer data at baseline whilst at follow-up 65% and 64% respectively had valid accelerometer data (Fig. 1 ). This was partially because of the requirement that wearers had at least 10 hours of valid wear time on at least 3 days [ 31 ]. However, due to a desire to start doing the Daily Mile as soon as possible in the intervention school and a limited number of accelerometers (117), it was possible to collect true baseline data (i.e. prior to beginning the Daily Mile) from only a portion of the intervention school participants. These participants were selected at random based on the availability of pupils for the other physiological tests. The baseline characteristics of the participants are given in Table 1 . Age was significantly higher in the control school at baseline due to the difference in the time of baseline measurements meaning that the pupils were slightly further through the academic year at the control school (Table 1 ). The percentage meeting the physical activity guidelines of 60+ min MVPA per day and the sum of skinfolds were both higher in the control school. However, after correction for age, gender and age*gender, only sedentary time differed significantly between the schools at baseline, suggesting that the correct confounders were accounted for in the analysis. Additionally, the intra-school differences between genders and year groups were similar (Additional file 1 : Table S1). SIMD scores were similar across schools (χ 2 = 1.299, p = 0.254). For quintiles 4–5 (least deprived), the scores were 71% versus 65%, respectively, for the intervention school and control school. For quintiles 1–3 (most deprived), the scores were 29% versus 35%. These reflect the lower deprivation than the average across Scotland [ 35 ], since ~20% would be expected in each quintile. However, this excess of children from less deprived areas reduces the likelihood of observing an impact of the Daily Mile, rather than creating any potential artefacts, as the children are more likely to be fitter and less likely to be overweight or obese [ 22 , 23 ] at baseline. Nonetheless, children from areas with lower socioeconomic scores (quintiles 1–3) had similar minutes of MVPA and sedentary time compared to those with higher socioeconomic scores (quintiles 4–5). They were also equally likely to meet the physical activity guidelines. However, they had lower shuttle distance, and a higher sum of skinfolds and rates of overweight and obesity (Table 2 ). A full breakdown of baseline characteristics by socioeconomic group, school and gender is given in Additional file 1 : Table S2. In the main analysis, after adjustment for the common confounders of age, gender and age*gender, significant improvements were observed in the intervention school relative to the control school for MVPA (+9.1 min), sedentary time (−18.2 min), fitness (+39.1 m) and body composition (−1.4 mm; Table 3 ). These relationships persisted after including a correction for SIMD (Table 3 ). Baseline values and change in BMI z score and age-corrected \dot{\mathrm{V}}{\mathrm{O}}_2\max \dot{\mathrm{V}}{\mathrm{O}}_2\max scores are given in Additional file 1 : Table S1–S3 only for comparison with other studies. Table 3 Effect of introducing the Daily Mile on outcomes assessed immediately after the end of the intervention period Full size table A change in MVPA predicted a change in sedentary behaviour ( r = −0.559, n = 118, p < 0.001; Fig. 2a ) but did not predict changes in other primary outcome variables. A change in shuttle distance predicted a change in sum of skinfolds ( r = −0.203, n = 317, p < 0.001; Fig. 2b ) but did not predict changes in other primary outcome variables. These correlations were not significantly altered by the inclusion of SIMD ( r = −0.564 and −0.212, respectively) although a change in MVPA additionally predicted a change in shuttle distance ( r = 0.187, n = 115, p = 0.046). However, the relationship between a change in sum of skinfolds and a change in shuttle distance differed by school (interaction p = 0.043). The relationship was stronger for the intervention school than the control school ( r = −0.245 and −0.046, respectively). The relationship between a change in MVPA and a change in sedentary behaviour was not different by school (interaction p = 0.896). However, it should be noted that the calculations for MVPA and sedentary time are linked by the finite number of minutes in a day and a change in one is likely to result in a change in the other. Fig. 2 Relationship between change in ( a ) MVPA and sedentary behaviour and ( b ) shuttle distance and sum of skinfolds. Both graphs are drawn after correction for gender, age in days and gender*age in days. MVPA moderate to vigorous intensity physical activity Full size image Discussion In this primary-school-based quasi-experimental pilot study, which investigated the effects of taking part in the Daily Mile, we found evidence of a positive effect on our four primary outcomes—accelerometer-assessed time spent in MVPA, accelerometer-assessed time spent in sedentary behaviour, physical fitness and body composition—after correcting for the common confounders of age and gender with or without socioeconomic grouping. Comparisons with other studies Whilst no other studies have investigated the effect of taking part in the Daily Mile, some have investigated the effect of increasing physical activity throughout the school day or introducing short physical activity breaks into the school day itself [ 37 , 38 ]. On the whole, they have had mixed results, with some finding alterations in MVPA and others not. This is likely in part due to the different methods used to assess these behaviours, in part due to the different interventions involved and in part due to the different accelerometer cut-points that can be found in the literature. Additionally, some studies use self-reported physical activity measures, which although easier to administer on a large scale, can lead to differing estimates in comparison to accelerometry [ 39 ]. Undoubtedly, the age and demographic of the children also has an influence and an intervention that works in one setting may not work in another. Similarly, some studies have found changes in body composition or fitness whilst others have not [ 40 , 41 , 42 ]. The observed effect of the Daily Mile on fitness in the current study may be a result of the type of intervention activity involved (i.e. running) being similar to the fitness test. However, few studies have taken detailed physiological measurements and often assess BMI only. Changes in BMI are observed with some physical activity interventions but mostly in high-BMI groups. The decrease in the sum of skinfolds observed in this study without a concomitant change in BMI z score is likely due to the higher resolution of skinfolds and its utility in assessing body fatness without the confounding effect of muscle mass. Meaning of study findings Scottish government figures suggest that 73% of children in Scotland (77% of boys and 69% of girls) meet the physical activity guidelines [ 14 ]. However, this figure is based on self-reported questionnaire results rather than accelerometer assessment and is likely to contain bias [ 43 , 44 ]. Estimates by accelerometer of the percentage of children meeting the physical activity guidelines vary across Europe, from as low as 2% to as high as 63% [ 11 , 45 ]. The children in this study fall within this range and are likely typical of Scottish and European primary school children [ 46 , 47 ]. Regardless of how many meet the minimum recommended guidelines (at least 60 min per day), higher levels of MVPA are generally considered to be better. This study shows that introducing the Daily Mile into a primary school setting does increase the MVPA of children by 9.1 min (SMD = 0.407). Although the Daily Mile is a 15-min physical activity intervention, an increase of ~9 min is consistent with the pattern of running interspersed with periods of walking and chatting that is observed in children taking part (personal observations). Although the SMD would be considered to be small according to Cohen [ 36 ], small effects on a prevalent behaviour, such as physical inactivity, may have a high impact at the level of population health [ 48 ]. In addition, a change of this magnitude is close to the 10 min increase in MVPA previously associated with meaningful reductions in cardiometabolic risk in children and adolescents [ 49 ]. Sedentary time is less well studied than physical activity. Nonetheless, the available evidence suggests that the children in this study are typical of European children [ 11 , 45 ]. In some studies, sedentary time appears to be a predictor of chronic disease independent of physical activity levels [ 50 ]. Two aspects of sedentary behaviour appear to be key to this: total sedentary time and prolonged blocks of sedentary time. The Daily Mile is potentially able to address both these issues although the present analysis only investigates total sedentary time. Although children at the intervention school were less sedentary at the baseline (after correction for the common confounders), this would make it harder to observe a reduction in sedentary behaviour rather than easier. Despite this, this study shows an ~18 min reduction (SMD = 0.437) in average daily sedentary time with the introduction of the Daily Mile. Again, this is consistent with a target of 15 min of physical activity since the children will at least be up from their chairs for a slightly longer period. However, if done correctly, the Daily Mile also breaks up the sedentary time, as it should happen in the middle of lessons, so that the children are likely to be sitting before and after their Daily Mile. As for MVPA, the SMD would be considered to be small but may well have significant impacts on population health due to mass participation. Additionally, the data also show a strong correlation between increasing MVPA and reducing sedentary time. This suggests that children are not compensating for the increase in MVPA during the Daily Mile by sitting more at other times of the day: they are replacing sedentary time with MVPA. However, note that the calculations for MVPA and sedentary time are linked by the finite number of minutes in a day and may be more appropriately analysed in future studies using a compositional data analysis. The children in the IDEFICS study [ 51 ] have median values of age-corrected \dot{\mathrm{V}}{\mathrm{O}}_2\max \dot{\mathrm{V}}{\mathrm{O}}_2\max scores between 46.7 and 48.1 ml·kg − 1 ·min − 1 for boys and between 45.4 and 47.4 ml·kg − 1 ·min − 1 for girls between the ages of 6 and 9 years. Relatively, the children in the current study could be considered to have high aerobic fitness (see Additional file 1 : Tables S1 and S2 for age-corrected \dot{\mathrm{V}}{\mathrm{O}}_2\max \dot{\mathrm{V}}{\mathrm{O}}_2\max scores). This high baseline fitness would make it less likely that a change in fitness could be observed after a small increase in physical activity. Nonetheless, an improvement in fitness, as measured by shuttle distance (39.1 m, SMD = 0.236), was observed with the introduction of the Daily Mile. \dot{\mathrm{V}}{\mathrm{O}}_2\max \dot{\mathrm{V}}{\mathrm{O}}_2\max is linked with cardiovascular health and all-cause mortality [ 52 ]. Although the SMD would be considered to be small, it may have a significant impact on a population scale. The CARDIA study in young adults suggests that having a \dot{\mathrm{V}}{\mathrm{O}}_2\max \dot{\mathrm{V}}{\mathrm{O}}_2\max of 3.5 ml·kg − 1 ·min − 1 (approximately 1 metabolic equivalent) higher gives a reduction in all-cause mortality of ~15% [ 53 ]. Whilst we only see a relative increase of ~0.35 ml·kg − 1 ·min − 1 (Additional file 1 : Table S3) with the Daily Mile, this is still predictive of an ~1.5% reduction in all-cause mortality risk. Note that the conversion from shuttle distance to \dot{\mathrm{V}}{\mathrm{O}}_2\max \dot{\mathrm{V}}{\mathrm{O}}_2\max includes age in years and has, therefore, relatively lower resolution. It has also been suggested that having a higher cardiorespiratory fitness at a younger age confers the greatest survival benefit [ 52 ]. Furthermore, those with lower starting values appear likely to benefit to a greater extent [ 54 ]. This suggests that there are potentially useful health benefits associated with taking part in the Daily Mile. The children at both schools in this study had lower rates of overweight and obesity than are typical of Scottish children. The Scottish Health Survey reports overweight and obesity rates in 7–11 year olds as 30% (29% for boys and 32% for girls) [ 14 ]. Again, this makes it less likely that a change in adiposity could be observed after a small increase in physical activity. Still, a reduction in adiposity as measured by skinfold (1.4 mm, SMD = 0.246) was observed with the introduction of the Daily Mile. Again, although the SMD would be considered to be small, at the population level it may have significant impacts on levels of adiposity. It is also possible that the impact of the Daily Mile on body composition may be larger still in children with higher rates of overweight and obesity. This intervention may be a useful component within measures designed to help tackle the obesity pandemic [ 8 ]. The strong correlation between those who reduced their skinfolds the most and those who gained the most fitness may indicate a common cause. Given that the pace each child completes the Daily Mile at is self-determined, it is possible that the children who gained the most benefits took a particular approach to the Daily Mile. An insight into this may come when we interview the children taking part in the Daily Mile about their experiences. Evidence linking socioeconomic status to MVPA and sedentary behaviour is unclear [ 13 ]. This is in part due to the use of different methods of capturing these outcome measures but also due to different ways of assessing socioeconomic status in different countries. Nonetheless, these outcome measures do appear to associate with specific aspects of socioeconomic status in some studies. However, clear differences between higher and lower socioeconomic groupings could be seen in the current study for fitness and body composition: children from postcodes with higher deprivation had lower levels of fitness, higher sums of skinfolds and higher rates of overweight and obesity. This is consistent with the widely recognised health inequality gap [ 55 ]. However, no differences were seen between the socioeconomic groupings in response to the introduction of the Daily Mile, suggesting that it may be beneficial to all groups regardless of background. Note that this study was not intended to investigate this, and larger more powerful studies are needed to investigate this aspect of the Daily Mile. A summary of the study and its implications can be found in Box 1. Strengths and limitations of study This study is the first to investigate the widely publicised and adopted Daily Mile physical activity intervention. The intervention appears to be increasingly popular and has now been maintained in the originator school for more than five years. Thus, it is undoubtedly feasible to deliver and has been adopted locally in many areas. What was unknown was the efficacy for the anecdotally reported physiological benefits of taking part in the Daily Mile. Consent rates were high (>77% in both schools) as were the number of children successfully assessed at both time points for most outcome measures. We assessed MVPA and sedentary time using the gold standard accelerometer technique, we assessed fitness using the bleep test (which has been validated in this age group) and we assessed body composition using labour-intensive skinfold assessments rather than the more straightforward but lower resolution BMI. We acknowledge that there was a difference in sedentary time between the schools at baseline and the socioeconomic groupings were not reflective of the whole of Scotland, which are limitations of our study. However, these differences would be predicted to make any effects of the Daily Mile harder to observe, not easier. Changes were observed despite these differences. It would have been preferable to assess both the intervention and control schools at the same time of year to avoid any seasonal impact on physical activity. However, we believe that October and March should be similar enough to allow comparison [ 56 ]. Additionally, it would have been better to have had both schools involved in the study for the same length of time, although, correcting for age and gender should account for this difference. It is also possible that differences in the health and well-being policies within the schools contributed to differences in the results. However, the schools were selected to be from the same local authority and to be of similar socioeconomic make-up to minimise potential differences. As predicted, changes in outcome variables had effect sizes at the smaller end of the distribution (0.2–0.5). However, given the involvement of whole classes, small effects could have an important impact on population health. To gain further confidence in the results, this study should be replicated in a larger number of schools. Furthermore, no monitoring of adherence, or level of adherence, to the intervention was carried out, although the results suggest adherence was sufficient. Unanswered questions and future research Additional anecdotally reported benefits to the Daily Mile (cognition, behaviour and well-being) are currently being investigated [ 57 , 58 ]. It is essential that the current studies are replicated in a larger number of schools and countries to ensure that the findings are both robust and repeatable in different educational contexts. Future studies should include diet and sleep quality, which we are not yet investigating, to explore the potential mechanisms of impact. More attention should be given to when the Daily Mile is being done during the school day and whether it is breaking up sedentary time. Additionally, future studies should investigate whether MVPA and sedentary behaviour are changing on weekdays and/or weekend days. In 2015, the Scottish government launched the Scottish Attainment Challenge with the aim of achieving equity in educational outcomes for all Scottish children [ 59 ]. Whilst the current study found no difference in the response to the Daily Mile by socioeconomic grouping, both schools were heavily weighted towards less deprived catchment areas. Furthermore, the current study was not powered to detect such a complex interaction. The Daily Mile is a free, simple intervention that can be rolled out to schools regardless of socioeconomic status. It is necessary to conduct carefully designed studies to understand the impact of the Daily Mile in different socioeconomic settings and to understand whether it can have any impact on the attainment gap. The sample of children participating in this study included a number with challenging behaviours including autism spectrum disorders. Nonetheless, they took part in the Daily Mile and our investigations. Understanding the impact of the Daily Mile on children with differing learning needs should also be a future priority. This study shows the value of introducing the Daily Mile into schools. Whilst the Daily Mile has been introduced as policy across Scotland, many schools do not have appropriate outdoor facilities to allow their children to take part. One of the challenges for policymakers and other stakeholders is to consider how to introduce the Daily Mile or alternative interventions that have been shown to increase MVPA and fitness into such schools or how to adapt those schools to allow the introduction of appropriate interventions. Conclusions In conclusion, introducing the Daily Mile to the primary school day appears to be an effective intervention for increasing MVPA and reducing sedentary time and it has measurable impacts on key aspects of metabolic health: body composition and physical fitness. This study provides the first assessment of the Daily Mile and it will allow the development of evidence-based policy around introducing the Daily Mile to more schools. Box 1: What this study adds Why was this study done? Low physical activity, high sedentary behaviour, declining fitness levels and high levels of overweight and obesity are global problems that have been targeted by the World Health Organisation. The Daily Mile is an increasingly popular school-based physical activity intervention, backed by the Scottish government, which is anecdotally reported to lead to increased physical activity, reduced sedentary time, improved fitness and improved body composition. Pupils run or walk laps of the playground at a self-selected pace for 15 min during normal classroom time. It is increasingly popular throughout the UK, in parts of Europe and some schools in the USA. However, these reported benefits remain anecdotal and need to be quantitatively and objectively assessed to ensure that the loss of academic classroom time is providing the reported alternative benefits. What did the researchers do and find? Two schools in the Stirling Council area, Scotland, were recruited: one with intention to start the Daily Mile, the other without. Researchers assessed the physical activity and sedentary behaviour of children using accelerometers, their fitness using the bleep test and their body composition using skinfolds. This was done in both schools before and after the intervention school introduced the Daily Mile. This quasi-experimental pilot study found that, after correcting for age, gender and socioeconomic grouping, taking part in the Daily Mile did lead to an improvement in physical activity, sedentary behaviour, fitness and body composition of children in the intervention school relative to the control school. What do these findings mean? This suggests that the Daily Mile is a worthwhile intervention to introduce in schools and that it should be considered for inclusion in government policy. This study can underpin the policy already introduced by the Scottish government and the development of future policy in other parts of the UK and abroad. Show more Abbreviations 95%CI: 95% confidence interval BMI: Body mass index cpm: Counts per minute ISAK: International Society for the Advancement of Kinanthropometry MVPA: Moderate to vigorous intensity physical activity PE: Physical education SIMD: Scottish Index of Multiple Deprivation SMD: Standardised mean difference WHEEL: Well-being, Health, Exercise, Enjoyment and Learning WHO: World Health Organisation

Policymakers should consider introducing The Daily Mile to improve the health and fitness of schoolchildren around the world, according to new research led by the Universities of Stirling and Edinburgh. The first study of the popular Daily Mile initiative—which involves children taking a 15-minute break from class to do physical activity—has confirmed it improves fitness, body composition and activity levels in participants. The findings indicate The Daily Mile can help combat global problems such as low physical activity, high sedentary behaviour, declining fitness levels and high levels of obesity. The study was jointly led by Dr. Colin Moran and Dr. Naomi Brooks, of the University of Stirling's Faculty of Health Sciences and Sport, and Dr. Josie Booth, of the University of Edinburgh's Moray House School of Education. It also involved a number of other experts from Stirling and the University of the Highlands and Islands. Dr. Moran said: "Our research observed positive changes in children who participated in The Daily Mile intervention, compared to our control school where the scheme was not introduced. "It suggests that The Daily Mile is a worthwhile intervention to introduce in schools and that it should be considered for inclusion in government policy, both at home and abroad." The Daily Mile was founded in February 2012 by Elaine Wyllie, the then headteacher of St Ninians Primary School in Stirling, to improve the fitness of her pupils. Children are encouraged to run, jog or walk around their school grounds during a 15-minute break from class, which is in addition to normal intervals and physical education lessons. Following the scheme's success, the Scottish Government has outlined its desire for Scotland to become the first Daily Mile nation, with around half of the country's primary schools now implementing the approach. There has been interest from the UK Government and the scheme has attracted the attention of other countries, with the Netherlands, Belgium and parts of the USA among those to have already adopted the approach. The research team conducted their research at two primary schools within the Stirling Council area, with 391 pupils, aged between four and 12, participating. Each child underwent an initial assessment and then a follow-up later in the academic year. Between times, one school implemented The Daily Mile, while pupils at the other—known as the control school—followed their usual curriculum. Children wore accelerometers to record their average daily minutes of moderate to vigorous intensity physical activity (MVPA) and average daily sedentary behaviour. They also had skinfold measurements taken to check body fat, and were assessed on their performance at a multistage fitness test (known as a bleep test or shuttle run), where they ran between cones 20 metres apart between bleeps. After correcting for age and gender, the team witnessed significant improvements in the intervention school, relative to the control school. Revealing the findings of the research during a special event at St Ninians, Dr. Brooks explained: "We observed a relative increase of 9.1 minutes per day in terms of MPVA and a relative decrease of 18.2 minutes per day in sedentary time. Children at the intervention school covered, on average, 39.1 metres more during the shuttle run, while their body composition improved too." There were similar results when the data was adjusted to additionally account for socioeconomic circumstances. Dr. Booth said: "Schools can help support pupils to be more active by taking part in The Daily Mile. The benefits of an active lifestyle are wide reaching and important for education as well as for health." The Scottish Government's Public Health and Sport Minister, Aileen Campbell, said: "Scotland's Daily Mile initiative is catching the imagination of the UK and beyond, and this research is even more validation of the impact it can have on people's lives. "By taking small steps and jogging, walking or running for 15 minutes, people can make huge changes to their health and wellbeing. We want Scotland to be the first 'Daily Mile Nation', with nurseries, colleges, universities and workplaces joining over 800 primary schools and regularly taking part. "There is great momentum behind this initiative, and I hope this research encourages even more schools and workplaces to sign up and reap the benefits of becoming more active more often." Elaine Wyllie, of The Daily Mile Foundation, said: "I founded The Daily Mile as I became concerned about the lack of physical fitness displayed by pupils and wanted to find a solution. The Daily Mile started with my simple belief that it would help children lead more active lives and to encourage the development of healthy habits in their futures. "With my pupils I saw that 15 minutes of daily activity rapidly improved pupils' fitness, health and concentration in the classroom. "I am delighted that this new research underlines what I found and I look forward to the day when every school does The Daily Mile."Susan McGill, convenor of Stirling Council's Children and Young People Committee said: "Improving children's health and wellbeing is a key priority for all schools and nurseries in Stirling. "We are so pleased to know that an initiative started in Stirling has now been proven to have such a positive impact on the lives of so many young people." The paper, The Daily Mile makes primary school children more active, less sedentary and improves their fitness and body composition: a quasi-experimental pilot study, is published in BMC Medicine.

10.1186/s12916-018-1049-z

Medicine

Gene mutation enhances cognitive flexibility in mice, study suggests

Jia–Hua Hu et al, Activity-dependent isomerization of Kv4.2 by Pin1 regulates cognitive flexibility, Nature Communications (2020). DOI: 10.1038/s41467-020-15390-x Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-020-15390-x

https://medicalxpress.com/news/2020-03-gene-mutation-cognitive-flexibility-mice.html

Abstract Voltage-gated K + channels function in macromolecular complexes with accessory subunits to regulate brain function. Here, we describe a peptidyl-prolyl cis - trans isomerase NIMA-interacting 1 (Pin1)-dependent mechanism that regulates the association of the A-type K + channel subunit Kv4.2 with its auxiliary subunit dipeptidyl peptidase 6 (DPP6), and thereby modulates neuronal excitability and cognitive flexibility. We show that activity-induced Kv4.2 phosphorylation triggers Pin1 binding to, and isomerization of, Kv4.2 at the pThr 607 -Pro motif, leading to the dissociation of the Kv4.2-DPP6 complex. We generated a novel mouse line harboring a knock-in Thr607 to Ala (Kv4.2TA) mutation that abolished dynamic Pin1 binding to Kv4.2. CA1 pyramidal neurons of the hippocampus from these mice exhibited altered Kv4.2-DPP6 interaction, increased A-type K + current, and reduced neuronal excitability. Behaviorally, Kv4.2TA mice displayed normal initial learning but improved reversal learning in both Morris water maze and lever press paradigms. These findings reveal a Pin1-mediated mechanism regulating reversal learning and provide potential targets for the treatment of neuropsychiatric disorders characterized by cognitive inflexibility. Introduction Rapidly activating and inactivating somatodendritic voltage-gated K + (Kv) A-type currents regulate action potential (AP) repolarization and repetitive firing and prevent backpropagation into the dendrites of hippocampal pyramidal neurons 1 , 2 . Kv4.2, a member of the Shal-type family, is the prominent A-type voltage-gated potassium channel expressed in hippocampal CA1 pyramidal neuron dendrites 1 . Kv4.2’s role in controlling of dendritic excitability impacts neuronal plasticity and contributes to learning and memory 3 , 4 , 5 . Kv4.2 activity remodels synaptic NMDA receptors by regulating the relative synaptic NR2B/NR2A subunit composition ratio at hippocampal synapses 6 . Ablation of Kv4.2 in mice abolishes the gradual reduction in GluN2B/GluN2A subunit ratio during post-natal development and results in a higher proportion of silent synapses in adulthood 7 . Aberrant Kv4.2 activity is also implicated in Autism Spectrum Disorder (ASD) 8 , temporal lobe epilepsy 9 , 10 , 11 , and Fragile X syndrome 12 , 13 . Considerable evidence suggests that Kv4.2 channels function in macromolecular protein complexes with accessory subunits, including the K + channel interacting proteins (KChIP1–4) and dipeptidyl peptidases 6 and 10 (DPP6 and DPP10) 14 . DPP6 is a type II transmembrane protein that increases Kv4.2 membrane expression and single channel conductance and accelerates the inactivation and recovery from inactivation of Kv4 subunit-containing channels 15 , 16 . In CA1 hippocampal pyramidal neurons, Kv4.2-mediated currents increase in density from the soma to distal dendrites 1 . However, this gradient is abolished in DPP6 KO mice 17 . In addition to its roles in modulating multiple aspects of Kv4.2 function, DPP6 appears to regulate hippocampal synaptic development independently of Kv4.2 (ref. 18 ). Recent studies have identified DPP6 and DPP10 as genes associated with autism 19 , amyotrophic lateral sclerosis 20 , 21 and neurodegeneration 22 . Thus, the regulation of the Kv4.2-DPP6 complex may not only affect Kv4.2 channel activity but also influence Kv4.2-independent functions of DPP6. However, little is known about how the stability or composition of this complex is regulated. In the present study, we report a Pin1-dependent mechanism that regulates the composition of the Kv4.2-DPP6 complex, neuronal excitability and cognitive flexibility. Pin1 is a prolyl isomerase that selectively binds to and isomerizes phospho-Ser/Thr-Pro (pSer/Thr-Pro) bonds 23 . pSer/Thr-Pro motifs in certain proteins can exist in two sterically distinct cis and trans conformations and Pin1 specifically accelerates the cis / trans conversion to regulate post-phosphorylation signaling 23 . Mis-regulation of Pin1 plays an important role in a growing number of pathological conditions including Alzheimer disease (AD), where it may protect against age-dependent neurodegeneration 24 , 25 , 26 , 27 . We identified Pin1 as a Kv4.2 binding partner via a TAP-MS pulldown assay. Subsequent biochemical studies revealed that Pin1-Kv4.2 binding is direct and via the canonical Pin1 binding motif. Stimuli including seizure induction and exposure to enriched, novel environments increased Kv4.2 phosphorylation at the Pin1 binding site T607 by p38 MAPK in the mouse cortex and hippocampus. Using biochemical and electrophysiological techniques, we showed that Pin1 activity is required for the dissociation of the Kv4.2-DPP6 complex and this action alters neuronal excitability. To confirm these observations, we generated a mouse line containing a Kv4.2 T607A (Kv4.2TA) mutation that abolished the phosphorylation and subsequent isomerization of an important C-terminal Pin1 motif. These mutant mice phenocopied those treated with pharmacological inhibitors of Pin1, which suggests a Pin1-dependent mechanism of Kv4.2 regulation. Intriguingly, Kv4.2TA mice exhibited normal initial learning but improved reversal learning in multiple behavioral tasks, introducing Pin1 isomerase regulation of Kv4.2 as a novel mechanism impacting cognitive flexibility. Results Pin1 binds to Kv4.2 at T607 Kv4.2 accessory subunits were identified by yeast two-hybrid screens and immunopurification over a decade ago 28 , 29 . Whether there are other Kv4.2 binding proteins that modulate Kv4.2 function is unknown. Here we took advantage of recently-developed Tandem Affinity Purification (TAP) combined with mass spectrometry (MS) techniques to identify Kv4.2 binding proteins in neurons and HEK-293T cells. We purified complexes of lentivirally expressed TAP-tagged Kv4.2 in cultured hippocampal neurons (Supplementary Fig. 1a ). MS analysis showed interaction with the well-established Kv4.2 accessory subunits DPP6/10 and KChIP1-4, verifying the validity of our Kv4.2 TAP-MS screen (Supplementary Fig. 1b ). This result is similar to the proteomic analyses of Kv4.2 complex in mouse brain using Kv4.2 antibody pulldown 30 . Using the same TAP technique to purify exogenously-expressed TAP-tagged Kv4.2 from HEK-293T cells, we identified Pin1 as a Kv4.2 binding partner (Supplementary Fig. 1c-f ). As shown in the MS list, Kv4.2 has many intracellular binding partners when expressed in HEK-293T cells. However, the majority of the binding partners are protein synthesis and degradation machinery proteins (Supplementary Fig. 1c, d ). This binding was confirmed by the co-immunoprecipitation (co-IP) of endogenous Pin1 with Kv4.2 in mouse brain lysates (Fig. 1a , uncropped images of all western blots are provided in the Supplementary Information file), and immunostaining of cultured hippocampal neurons revealed that Pin1 colocalized with Kv4.2 (Fig. 1b ). Since Pin1 substrate binding requires phosphorylation, we showed that Kv4.2 binding to Pin1 is significantly reduced when it’s dephosphorylated by Lambda protein phosphatase (Supplementary Fig. 2a ). To examine if Kv4.2 and Pin1 binding occurs via the canonical Pin1 binding interface, we employed the Pin1 WW domain mutant (W34A) and the PPIase domain mutant (R68A, R69A). When co-expressed with Kv4.2 in HEK-293T cells, both Pin1(W34A) and Pin1(R68A, R69A) mutants exhibited significantly reduced binding to Kv4.2 (Fig. 1c ). Thus, the Kv4.2-Pin1 interaction appears to be direct and involves conventional Pin1 binding domains. Fig. 1: Pin1 binds to Kv4.2 at pT607 and elicits structural rearrangements in Kv4.2. a Pin1 co-immunoprecipitated with Kv4.2 in mouse brain lysates. Forebrain lysates from WT and Kv4.2 KO were immunoprecipitated with mouse (ms) or rabbit (rb) anti-Kv4.2 antibodies. Both total lysates and immunoprecipitates were blotted with anti-Kv4.2 or Pin1 antibodies. Data from three independent experiments. b Cultured hippocampal neurons (DIV 10) were immunostained with anti-Pin1 along with anti-Kv4.2. Pin1 co-localized with Kv4.2, indicated with arrows. Scale bars: 20 μm top panels, 5 μm bottom. Data from four coverslips in two independent experiments. c , Pin1 mutants reduced Pin1-Kv4.2 binding. Myc-Kv4.2 was co-transfected alongside HA-Pin1 with or without WW (W34A) or PPIase domain (R68, R69A) point mutants into HEK-293T cells. Kv4.2 was immunoprecipitated from detergent lysates with anti-Myc antibody. Samples were analyzed by western blotting with anti-HA and anti-Myc antibodies. n = 3 each group. d Alignment of Kv4.2 C-terminal sequences from various species. The putative Pin1 binding site is conserved. Bold residues show preferred Pin1 binding context. e Pin1 selectively binds to the phosoho-T607-containing Kv4.2 peptide. Synthetic Kv4.2-peptides were conjugated to Affi-Gel 15 Sepharose beads and incubated with lysate from HA-Pin1 transfected HEK-293T cells. n = 4 each group. f Kv4.2 T607A mutation significantly reduced Pin1 binding. HA-Pin1 and Myc-Kv4.2 mutants were co-transfected into HEK-293T cells. Pin1 co-immunoprecipitation with Kv4.2 was assayed. Kv4.2 T607 is required for Pin1 binding. n = 3 each group. g Molecular modeling of Kv4.2 phospho-peptide binding to Pin1. h Highlight of Kv4.2 pT602 peptide binding to the Pin1 WW domain. i Highlight of Kv4.2 pT607 peptide binding to the Pin1 PPIase domain. j Dose-dependent proteolysis of Kv4.2 by subtilisin. Asterisk, non-specific bands; arrowhead, 47kD band; arrow, 33kD band. Data repeated in two independent experiments. k Pin1 blocked Kv4.2 subtilisin digestion while Pin1C113S (an isomerase dead mutant) did not. Pin1 block was lost when Kv4.2 was dephosphorylated by Lambda protein phosphatase (PP). Quantification of the 47kD degradation fragment. n = 4 each group. Data was repeated in four independent experiments. Data are presented as mean ± SEM, * p < 0.05, *** p < 0.001, Paired t -tests. Full size image There are three S/T-P sites at the C-terminus of Kv4.2 that can be phosphorylated by extracellular signal–regulated kinases (ERKs) 31 . These phosphorylated S/T-P motifs might serve as putative Pin1 binding sites. Pin1 preferentially targets to a pSer/Thr-Pro motif that is surrounded by multiple upstream hydrophobic residues such as isoleucine, valine, tyrosine and/or phenylalanine, and a downstream arginine or lysine residue 23 , 32 , 33 . There are three isoleucine prior to the two T-P motifs and an arginine downstream (Fig. 1d ), providing a better Pin1 binding context. Sequence alignment revealed that these S/T-P sites are conserved from zebrafish to human (Fig. 1d ), suggesting a conserved function of Kv4.2 across the species. To identify the Pin1-binding site(s) in Kv4.2, we synthesized non-phospho- and phospho-Kv4.2 peptides containing T602, T607 or S616 and conjugated them with Affi-Gel 15 Sepharose beads. Peptide pulldown assays revealed that Pin1 binds weakly to the T602 phosphorylated peptide but strongly to the T607 phosphorylated peptide (Fig. 1e ). Interestingly, Pin1 showed even stronger binding to the peptide with dual phosphorylation of T602 and T607 (Fig. 1e ). Pin1 did not bind to the S616 phosphorylated peptide (Fig. 1e ). Furthermore, co-IP studies in HEK-293T cell lysates using Kv4.2 mutants with abolished phosphorylation sites showed that T602A or S616A mutants did not affect Kv4.2-Pin1 binding while T607A or T602A/T607A mutants dramatically reduced the binding (Fig. 1f ), which is consistent with the peptide pulldown assay (Fig. 1e ). These data support the idea that Pin1 directly binds to Kv4.2 at T602 and T607 sites, where the latter site is involved in a greater degree of binding. Previous studies have shown that the PPIase domain of Pin1 is able to bind to the pS/T-P motif in addition to the WW domain 34 , 35 . pS/T-P motifs that have an additional P residue in the +1 position, pS/T-P-P, seem to be targeted by the WW domain but not the PPIase domain of Pin1 (ref. 36 ). Therefore, a substrate with multiple phosphate binding sites could allow for the simultaneous binding of multiple Pin1 domains 36 , 37 . Kv4.2-Pin1 binding modeled by an interplay of “manual” manipulation with the UCSF Chimera software showed the first pT-P-P binding to the Pin1 WW domain and the second pT-P binding to the Pin1 PPIase domain (Fig. 1g–i ). The Shal-type family contains three members: Kv4.1, Kv4.2, and Kv4.3. They are all expressed in the hippocampus but with different expression patterns 38 . We examined the possibility of Pin1 binding to other Shal-type family members. Human sequence alignment revealed that these S/T-P sites are conserved among the members (Supplementary Fig. 2b ). The Pin1 binding context is also conserved except Kv4.1 lacking an arginine after the two T-P sites (Supplementary Fig. 2b ). Co-IP showed that Pin1 binds to all the Kv4 members but not a non-Shal-type family member, Kv3.4, that does not contain the S/T-P motif at its C-terminal (Supplementary Fig. 2c ). Pin1 binds to Kv4.2 and Kv4.3 better than Kv4.1 (Supplementary Fig. 2c ), which is consistent with the lack of arginine in Kv4.1 (Supplementary Fig. 2b ). Thus, Pin1 likely regulates all Kv4 members. Pin1 elicits structural rearrangements in Kv4.2 To verify whether the prolyl-isomerase activity of Pin1 induces conformational changes in Kv4.2, we performed a partial proteolysis assay on purified Kv4.2. This assay relies on the observation that Pin1-dependent structural changes impair proteolysis by subtilisin serine endopeptidase that is sensitive to substrate structure 39 , 40 , 41 . Myc-Kv4.2 was purified from HEK-293T cells that were transfected with myc-Kv4.2 construct and subjected to subtilisin digestion. Myc-Kv4.2 was dose-dependently degraded by subtilisin (Fig. 1j ). It was mainly degraded into a 47kD fragment and a 33kD fragment (Fig. 1j ). Incubation with GST-Pin1 prior to subtilisin significantly blocked the 47kD fragment degradation compared to GST control and GST-Pin1C113S, an isomerase dead mutant (Fig. 1k ). Furthermore, the blockage effect of GST-Pin1 was abolished by Lambda protein phosphatase treatment before GST-Pin1 incubation (Fig. 1k ). These data indicate that Pin1 is recruited by Kv4.2 in a phosphorylation-dependent manner and it promotes structural rearrangements. Kv4.2 T607 phosphorylation and Pin1 binding is dynamic In order to further study the role of Kv4.2 phosphorylation at T602 and T607, we characterized phospho-T602 and phospho-T607-specific antibodies using site-specific mutations of Kv4.2 (Supplementary Fig. 3a, b ) and Lambda protein phosphatase treatment (Supplementary Fig. 3c ). Phosphorylation of Kv4.2 at T602 and T607 sites were detected in mouse brain lysates by western blot (Fig. 2a–c ). To investigate the dynamic regulation of Kv4.2 phosphorylation, we exposed mice to a novel, enriched environment (EE) which has previously been shown to downregulate dendritic Kv4.2 function 42 . EE exposure induced T607 phosphorylation but not T602 phosphorylation in the mouse hippocampus (Fig. 2a ). EE exposure induced similar changes in the cortex (Supplementary Fig. 4a ). Fig. 2: Enriched novel environment exposure and seizure induce Kv4.2 phosphorylation at a Pin1 binding site. a Enriched novel environment (EE, 1 h) induces phosphorylation of Kv4.2 at Thr607 but not Thr602 in mouse hippocampus. n = 5 in each group. T -test, * p < 0.05. b Kainic acid-induced seizure (25 mg/kg, i.p., 15 min) induces phosphorylation of Kv4.2 at Thr607 but not Thr602 in mouse hippocampus. n = 4 in each group. T -test, *** p < 0.001. c PTZ-induced seizure (50 mg/kg, i.p., 15 min) induces phosphorylation of Kv4.2 at Thr607 and Thr602 in mouse hippocampus. n = 4 in each group. * p < 0.05, T -test, ** p < 0.01. d PTZ-induced seizure increases Pin1 binding to Kv4.2. GST or GST-Pin1-linked beads were incubated with brain lysates from mice subjected to saline or PTZ administration. n = 5 in each group. T -test, ** p < 0.01. e Mouse brain lysates from WT mice w or w/o PTZ administration (50 mg/kg, i.p., 15 min) were incubated with excess anti-Kv4.2, anti-Kv4.2-pT602 or anti-Kv4.2-pT607 antibodies. Immunoprecipitation (IP) samples were blotted with Kv4.2 antibody. In WT mouse brains, pT607 Kv4.2 is almost half as abundant as pT602. However, PTZ administration increased the amount of pT607 until it reached a similar level as pT602. n = 4 for each group. T -test, ** p < 0.01. f Mouse brain lysates from WT mice were incubated with excess anti-Kv4.2-pT602 or normal IgG antibodies. Immunoprecipitation (IP) samples were blotted with anti-Kv4.2 pT607 antibody. pT602 and pT607 dual phosphorylation was observed in mouse brain. Data was repeated in two independent experiments. g Mouse brain lysates from WT mice w or w/o PTZ administration (50 mg/kg, i.p., 15 min) were incubated with excess anti-Kv4.2-pT607 antibody. IP samples were blotted with anti-Kv4.2 pT602 antibody. PTZ-induced seizure increases the dual phosphorylation of T602 and T607 in mouse brain. Data was repeated in two independent experiments. Data are presented as mean ± SEM. Full size image Temporal lobe epilepsy has also been shown to decrease Kv4.2 availability 9 , 11 . Here we found seizure induced by kainic acid (KA) increased T607 phosphorylation but not T602 phosphorylation in the mouse hippocampus (Fig. 2b ). Seizure induced by pentylenetetrazole (PTZ) showed increased T607 phosphorylation and T602 phosphorylation in the mouse hippocampus (Fig. 2c ). It also increases Kv4.2 T607 phosphorylation but not T602 phosphorylation in the mouse cortex (Supplementary Fig. 4b ). These data suggest that Kv4.2 T607 phosphorylation is dynamically regulated in the mouse brain. As Pin1 only binds to phosphorylated substrates, we hypothesized that the induction of Kv4.2 T607 phosphorylation would increase Pin1-Kv4.2 association. Accordingly, GST-Pin1 pulldown experiments revealed that PTZ-induced seizures significantly enhanced Kv4.2 pulldown by Pin1 (Fig. 2d ). We next investigated the prevalence of phosphorylated T602 and T607 in the mouse brain. Total, phospho-T602, and phospho-T607 Kv4.2 was immunoprecipitated by saturating specific antibodies and quantified by western blot. We found 10.20 ± 0.62% of Kv4.2 was phospho-T602 and 5.10 ± 0.71% was phospho-T607 in the mouse forebrain (Fig. 2e ). Phospho-T607 increased to 10.78 ± 1.30% while T602 was un-altered (11.18 ± 0.41%) with seizure induction by PTZ (Fig. 2e ). We also detected phospho-T607 Kv4.2 when brain samples were immunoprecipitated with phospho-T602 antibody (Fig. 2f ). Furthermore, increased T602 phosphorylation was detected when immunoprecipitation was performed with the phospho-T607 antibody after PTZ administration (Fig. 2g ). These data suggest that Kv4.2 is dually phosphorylated at sites T602 and T607 in the mouse brain and that their phosphorylation is regulated by neuronal activity. P38 phosphorylates Kv4.2 at T607 Kv4.2 T607 has previously been reported to be phosphorylated by ERK in vitro 31 . We verified this finding with co-expression assays in HEK-293T cells. Kv4.2 phosphorylation at T602 and T607 was increased when Kv4.2 was co-expressed with Erk1 or MEKDD (a constitutively active MEK mutant) (Supplementary Fig. 5a ). However, these increases were small, which led us to consider other proline-directed kinases that could phosphorylate these two sites. We examined the individual effects of CDK5/p35, GSK3β and p38α on Kv4.2 phosphorylation. Among these proline-directed kinases, p38α had the most robust effect on Kv4.2 phosphorylation (Fig. 3a ). Moreover, a p38α point mutant with abolished kinase activity largely blocked Kv4.2 phosphorylation when both constructs were expressed in HEK-293T cells (Fig. 3a ). Exogenous p38α co-immunoprecipitated with Kv4.2 in HEK-293T cell lysates (Fig. 3b ) which supports the notion that Kv4.2 is a substrate of p38α. Furthermore, EE exposure phospho-activated p38 in mouse hippocampus and cortex as reported by western blot using a phospho-p38 antibody (Fig. 3c , Supplementary Fig. 5b ). Seizure induced by KA or PTZ also activated p38 in the hippocampus (Fig. 3d, e ). These data are consistent with the effects of EE, KA, and PTZ on the induction of Kv4.2 T607 phosphorylation (Fig. 2a–c ). Interestingly, the p38 inhibitor SB203580 blocked the induction of Kv4.2 phosphorylation by PTZ-induced seizure in the mouse hippocampus (Fig. 3f ), while PTZ-induced Kv4.2 phosphorylation is only partly reduced by the MEK inhibitor SL327 (Fig. 3g ). These findings suggest that p38 is the primary kinase responsible for the dynamic phosphorylation of Kv4.2 at the T607 site in mouse hippocampus. Fig. 3: P38 phosphorylates Kv4.2 at a C-terminal Pin1 binding site. a P38 phosphorylates Kv4.2 at Thr602 and Thr607. Kv4.2 and p38α constructs (agf: p38 kinase dead mutant) were co-transfected into HEK-293T cells. Lysates were analyzed by western blotting with anti-pT602 and anti-pT607 specific antibodies. n = 15 for ctl, 9 for p38 and p38 agf. T -test, *** p < 0.001. b p38 binds to Kv4.2. Kv4.2 and p38α were co-transfected into HEK-293T cells. Detergent lysates were incubated with anti-Myc antibody and analyzed by western blotting with anti-Flag and anti-Myc antibodies. Data was repeated in two independent experiments. c , Enriched novel environment activates p38 in mouse hippocampus. n = 6 in each group. T -test, * p < 0.05. d KA-induced seizure activates p38 in mouse hippocampus. n = 8 in each group. T -test, ***p < 0.001. e PTZ-induced seizure activates p38 in mouse hippocampus. n = 6 for ctl and 5 for PTZ. T -test, *** p < 0.001. f SB203580, a potent p38 inhibitor (20 mg/kg, i.p., 15 min) blocked PTZ-induced phosphorylation of Kv4.2 T607 in mouse hippocampus. n = 4 in each group. T -test, ** p < 0.01. g SL327, a selective MEK inhibitor (30 mg/kg, i.p., 15 min) did not block PTZ-induced phosphorylation of Kv4.2 T607 in mouse hippocampus. n = 7 in each group. * p < 0.05, ** p < 0.01. t -test. Data are presented as mean ± SEM. Full size image P38-Pin1-Kv4.2 pathway regulates Kv4.2-DPP6 complex formation Both the biophysical properties and surface expression of Kv4.2 are regulated by its auxiliary subunit DPP6 (refs. 29 , 43 ). We wondered if the Kv4.2-DPP6 complex is regulated by Kv4.2 phosphorylation and Pin1 activity. As PTZ-induced seizure enhanced Kv4.2 phosphorylation at T607 by p38 (Figs. 2, 3 ), we sought to determine if this seizure model also alters Kv4.2-DPP6 binding. From co-IP, we found that PTZ-induced seizure reduced Kv4.2-DPP6 binding in the mouse brain (Fig. 4a, b ). This Kv4.2-DPP6 complex dissociation was blocked by the p38 inhibitor SB203580, but not the MEK inhibitor SL327 (Fig. 4a ), suggesting that p38 is required for the dissociation of the Kv4.2-DPP6 complex. Furthermore, the PTZ-induced Kv4.2-DPP6 complex dissociation was blocked by the Pin1 inhibitor Juglone (Fig. 4b ). In cultured mouse neurons, synaptic stimulation with α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA, 50 µM) for 15 min resulted in decreased Kv4.2-DPP6 binding, which was opposed by the expression of PinC113S, an isomerase dead mutant (Fig. 4c ). These data suggest that Pin1 activity is required for the dissociation of the Kv4.2-DPP6 complex in response to neuronal activity. Fig. 4: P38-Pin1 pathway regulates composition of the Kv4.2-DPP6 complex. a P38 inhibitor SB203580 blocked PTZ-induced Kv4.2-DPP6 dissociation while MEK inhibitor SL327 did not. Mouse forebrain lysates with or w/o SB203580 (20 mg/kg, i.p., 20 min) or SL327 (30 mg/kg, i.p., 20 min) or PTZ administration (60 mg/kg, i.p., 20 min) were immunoprecipitated with anti-Kv4.2 antibody. PTZ-injected mice showed decreased Kv4.2-DPP6 binding, blocked by preinjection of SB203580 but not SL327. n = 5 for each group. b Pin1 inhibitor juglone blocked PTZ-induced Kv4.2-DPP6 dissociation. Forebrain lysates with or w/o juglone (15 mg/kg, i.p., 15 min) or PTZ administration (60 mg/kg, i.p., 20 min) were immunoprecipitated with an anti-Kv4.2 antibody. PTZ-injected mice showed decreased Kv4.2-DPP6 binding while juglone-preinjected mice exhibited normal Kv4.2-DPP6 binding. n = 4 for ctl, 5 for PTZ and Juglone/PTZ. c Pin1C113S mutant blocked AMPA-induced Kv4.2-DPP6. Cultured cortical neurons infected with GFP or Pin1C113S lentivirus were treated with 50uM AMPA for 15 min and processed for immunoprecipitation with anti-Kv4.2 antibody. AMPA treatment reduced Kv4.2-DPP6 binding in GFP but not in Pin1C113S infected neurons. n = 6 for each group. d Seizure-induced Pin1-Kv4.2 association is abolished in Kv4.2TA mice. GST or GST-Pin1-linked beads were incubated with brain lysates from WT and Kv4.2TA mice with or without PTZ administration (60 mg/kg, i.p., 15 min). n = 4 WT/Ctl, WT/PTZ and Kv4.2TA/Ctl, n = 3 for Kv4.2TA/PTZ. e Seizure-induced Pin1-Kv4.2 association is abolished in Kv4.2TA mice. Forebrain lysates from WT and Kv4.2TA mice were immunoprecipitated with rabbit anti-Kv4.2 antibody, with or without PTZ administration (60 mg/kg, i.p., 15 min). Pin1-Kv4.2 association is induced by PTZ in WT but abolished in Kv4.2TA mice. n = 4 each group. f PTZ-induced Kv4.2-DPP6 dissociation is abolished in Kv4.2TA mice. Forebrain lysates from WT and Kv4.2TA mice with or w/o PTZ administration (60 mg/kg, i.p., 20 min) were immunoprecipitated with anti-Kv4.2 antibody. PTZ treatment decreased Kv4.2-DPP6 binding in WT but not in Kv4.2TA mice. n = 3 for each group. Data are presented as mean ± SEM. Paired T -test, ** p < 0.01, ** p < 0.01 vs ctl, ## p < 0.01 Kv4.2TA vs WT, *** p < 0.001. Full size image To further study Pin1’s role in regulating Kv4.2 channel complexes, we created a Kv4.2 T607A mutant knock-in mouse (Kv4.2TA) where Thr607 was mutated to Ala using CRISPR-Cas9 techniques to prevent its phosphorylation and subsequent Pin1 binding (Supplementary Fig. 6a ). The mice were identified by PCR followed by sequencing (Supplementary Fig. 6b ). Kv4.2TA mice were born with expected Mendelian ratios, with no differences in mortality rate or weight of heterozygous or homozygous Kv4.2TA mice compared to WT littermates. There were no significant differences in the total protein expression of Kv4.2 in the hippocampus between WT and Kv4.2TA mice (Supplementary Fig. 6c ). Kv4.2TA mice were also verified to have abolished Kv4.2 T607 phosphorylation by western blot (Supplementary Fig. 6c ). Furthermore, the structure of the hippocampus appears normal by Nissl staining (data not shown). Additionally, the general distribution of Kv4.2 in apical dendrites appears to be unaltered in the hippocampus of Kv4.2TA mice relative to WT as detected by immuno-labelling (Supplementary Fig. 6d ). To determine if Pin1 binding to Kv4.2 was impaired, we measured Kv4.2 pulldown by GST-Pin1 in Kv4.2TA and WT mouse forebrains with or without PTZ administration. Kv4.2 pulldown was reduced in Kv4.2TA mice compared to that of WT littermates (Fig. 4d ) in basal conditions. Interestingly, PTZ-induced seizure did not increase Pin1 binding to Kv4.2 in Kv4.2TA mice as it did in WT littermates (Fig. 4d ). We also performed the Kv4.2 and Pin1 Co-IP experiment under the same condition, and the result is consistent with the GST-Pin1 pulldown (Fig. 4e ). These data indicate that Pin1-Kv4.2 binding is dynamically regulated in WT mice but abolished in Kv4.2TA mice. We next examined if the regulation of Kv4.2-DPP6 binding is altered in Kv4.2TA mice. Total DPP6 expression is normal in Kv4.2TA mice (Fig. 4f ). However, Kv4.2-DPP6 dissociation by PTZ-induced seizure was abolished in Kv4.2TA mice (Fig. 4f ). This data supports the notion that both Kv4.2 phosphorylation at T607 and Pin1 activity regulate Kv4.2-DPP6 complex formation. Pin1 activity and phosphorylation of Kv4.2 at T607 regulate neuronal excitability Substantial evidence supports a role for Kv4.2-containing A-type K + channels and their associated auxiliary subunits in the regulation of the intrinsic excitability of CA1 pyramidal neurons. Along with other voltage-gated ion channels localized to the somatodendritic compartment of pyramidal cells, Kv4.2 contributes to the firing mode of the cell by regulating back-propagating action potential amplitude and the after-hyperpolarization of individual spikes in a train 1 , 44 , 45 . Thus, Pin1 regulation of the Kv4.2 channel complex could impact the excitability of hippocampal pyramidal neurons. To test whether Pin1 isomerization of Kv4.2 affects excitability, whole-cell somatic current-clamp recordings were performed in CA1 pyramidal neurons in adult mouse acute slices. We first asked if Pin1 regulates membrane excitability in WT mice. We utilized a Pin1 inhibitor, PiB, that has been shown to block the catalytic activity of Pin1 (ref. 46 ). Recordings were performed in the presence PiB (4 µM) following pre-incubation of slices with PiB included in the recovery solution. PiB significantly reduced the AP firing frequency of CA1 pyramidal cells compared to vehicle (0.1% DMSO) at each current step (Fig. 5a, b ). Notably, PiB application resulted in a characteristic irregularity of spiking during current injections with prolonged intermittent pauses (Fig. 5a, b ). To assess whether the PiB-induced reduction in excitability is mediated through a Kv4-dependent mechanism, we co-applied the Kv4-specific blocker, AmmTX3 (250 nM) 47 along with PiB in the extracellular bath. Indeed, bath application of AmmTX3 reversed the suppressive effects of PiB alone (Fig. 5a, b ), indicating that the effect of PiB on neuronal excitability is mediated by Kv4 channels. Additional electrical properties, including the resting membrane potential (RMP), membrane capacitance, and the shape of individual APs were unchanged between vehicle, PiB, and AmmTX3 treatments although PiB application did slightly reduce input resistance and increase rheobase (Table 1 ). Fig. 5: P38-Pin1-Kv4.2 pathway regulates neuronal excitability. a , b Pin1 inhibitor PiB (4 µM) reduces pyramidal cell excitability in WT mouse hippocampal brain slices. a Current steps at 100 pA and 200 pA result in reduced firing frequency with PiB application (teal) ( n = 13) relative to vehicle (green) ( n = 14). This reduction is rescued by co-application of AmmTX3 (250 nM) (blue) ( n = 7). Scale: 30 mV / 250 ms. b , PiB significantly reduces AP firing frequency relative to vehicle in response to 100, 150, and 200 pA somatic current injections. This reduction is rescued by AmmTX3 application. Two-way ANOVA, * p < 0.05, ** p < 0.01. c , d Mutation of Kv4.2 T607 Pin1 binding site phenocopies pharmacological inhibition of Pin1 in WT. c Pyramidal cells from Kv4.2TA slices display reduced firing frequency relative to WT over a range of increasing current injections. Scale: 30 mV/250 ms. d AP firing frequency is significantly reduced after 150 and 200 pA somatic current injections in Kv4.2TA ( n = 22) relative to WT ( n = 20), two-way ANOVA, ** p < 0.01, ***p < 0.001. e , f Pin1 inhibition with PiB has no effect in Kv4.2TA mice. e Pyramidal cells from Kv4.2TA slices display similar AP firing patterns when treated with vehicle ( n = 14) and PiB ( n = 14). Scale: 30 mV/250 ms. f No significant changes are observed in Kv4.2TA pyramidal cell AP firing frequency with PiB exposure, Two-Way ANOVA, p > 0.05. g Brain slicing and recovery activates p38 and increases pT607 of Kv4.2. Adult mouse brains were either sliced as for electrophysiological recordings or dissected as for biochemical assays. Same brain regions were used. n = 4 for ctl and 3 for slicing. T -test. * p < 0.05, *** p < 0.001. Data are presented as mean ± SEM. Full size image Table 1 Neuronal excitability and subthreshold membrane properties in WT and Kv4.2TA hippocampal CA1 pyramidal neurons with and without pharmacological treatment. Full size table We next assessed whether the Pin1-Kv4-dependent reduction in excitability through pharmacological manipulation was replicated by mutation of the Pin1 binding site T607 within Kv4.2. To test this, we performed whole-cell current-clamp recordings in hippocampal slices from WT and Kv4.2TA mice in regular ACSF. We found that the input/output curves of firing frequency displayed a rightward shift in Kv4.2TA cells relative to WT (Fig. 5c, d ). At peak current injection (+200 pA), the average firing frequency in Kv4.2TA pyramidal cells was nearly half of that in WT cells (Fig. 5c, d ). As with pharmacological blockade of Pin1 in WT, we noted an irregular AP spiking pattern in Kv4.2TA cells (Fig. 5c, d ). We also found that the peak fast after hyperpolarization (fAHP) was significantly increased in Kv4.2TA (Table 1 ), which coincided with an overall, significant increase in the inter-spike interval (Table 1 ). Additional properties including RMP, membrane capacitance and AP shape were unchanged between the two mouse lines (Table 1 ). These data suggest a role for Kv4.2 phosphorylation at T607 in the regulation of neuronal excitability. Furthermore, we found that PiB exposure to Kv4.2TA slices did not significantly affect excitability in hippocampal pyramidal neurons (Fig. 5e, f ), contrary to its effect in WT cells (Fig. 5a, b ) and also note that reduced excitability of Kv4.2TA neurons was consistent in each experimental condition (Fig. 5 ). Therefore, pharmacological blockade of Pin1 did not augment any reduction in excitability induced by genetic manipulation of the Pin1 binding site, suggesting an important role for this specific Pin1-Kv4.2 interaction in its regulation of neuronal excitability. The basal level of Kv4.2-DPP6 protein complex seems unaltered in Kv4.2TA mice compared to WT littermates in biochemistry experiments (Fig. 4f ) whereas we found reduced neuronal excitability in Kv4.2TA mice (Fig. 5c, d ). We hypothesized that the slicing and recovery process in recording experiments activates p38 and triggers Pin1-dependent changes. To examine this, we measured p38 phosphorylation in sliced brain in comparison with un-sliced brain. The results showed that slicing and recovery did not alter the expression of p38 protein (Ctl: 100 ± 6.23%; slicing: 93.30 ± 2.80%, p = 0.4262) and ERK (Ctl: 100 ± 3.85%; slicing: 100.36 ± 3.31%, p = 0.9492), but largely activates p38 (over 10 fold) and increases Kv4.2 phosphorylation at T607 (Fig. 5g ). These data suggest that the slicing and recovery process before recording activated the p38-Pin1-Kv4.2 pathway, leading to the excitability changes in the Kv4.2TA mice. Pin1 activity and phosphorylation of Kv4.2 at T607 regulate A-current The reduced excitability observed in Kv4.2TA neurons and in response to pharmacological blockade of Pin1 in WT is suggestive of enhanced I A in these cells. Studies of DPP6’s effect on Kv4.2 have revealed that DPP6 increases macro I A amplitude and accelerates recovery from inactivation 17 , 48 . Since the Kv4.2-DPP6 complex is mis-regulated in Kv4.2TA mice (Fig. 4f ), we anticipated disruption of Pin1-Kv4.2 interaction would alter I A . To test this, we performed voltage-clamp recordings from outside-out patches pulled from CA1 pyramidal somata. As in analysis of firing properties, we first measured I A in WT slices exposed to Pin1 blocker, PiB (4 µM). We found that PiB exposure significantly increased I A density relative to vehicle (.1% DMSO) (Fig. 6a, b ). Additionally, as we identified p38 and MEK-mediated phosphorylation at the Pin1 binding site on Kv4.2, we tested their effect in facilitating Pin1 regulation of I A . We identified that pharmacological blockade of p38 (SB230580) also significantly increased I A density relative to vehicle while MEK inhibition (PD98059) displayed no significant effect (Fig. 6a, b ). Further, consistent with the observed effects on firing suggestive of enhanced I A in Kv4.2TA mice, isolation of I A revealed a significant increase in current density in patches pulled from Kv4.2TA cells relative to WT (Fig. 6d, e ). Additionally, while changes in macro current inactivation, rise time and voltage-dependence of activation and inactivation were indistinguishable between the lines (Fig. 6c , Supplementary Tab. 1 ), a leftward shift in the normalized recovery from inactivation curve was identified in Kv4.2TA cells (Fig. 6h, i ). Single exponentials fitted to the normalized recovery curves yielded a statistically significant reduction in the time constant of I A recovery in Kv4.2TA cells, suggesting these channels recover more quickly from inactivation relative to WT (Fig. 6i ). Importantly, pharmacological blockade of Pin1 had no effect on I A in Kv4.2TA mice (Supplementary Fig. 7 ). Taken together, we show that block of p38 kinase and Pin1 results in enhanced I A density in the soma of CA1 pyramidal cells. The T607A mutation occludes the effects of pharmacolocal Pin1 bockade on neuronal excitability and I A , supporting the notion that Kv4.2 phosphorylation at T607 and Pin1 isomerization of the Kv4.2 pT607-P bond regulate the intrinsic excitability of CA1 pyramidal neurons through the modulation of both Kv4.2 channel availability and recovery from inactivation kinetics. Furthermore, these data provide evidence that blocking Pin1-Kv4.2 interaction may increase the proportion of Kv4.2 channels in complex with DPP6. Fig. 6: P38-Pin1-Kv4.2 pathway regulates A-current. a – c I A recorded from outside-out somatic patches from CA1 pyramidal cells from WT mice. a , Trace of transient I A. inactivating I A was isolated by subtracting total K + measured from a step to +40 mV from a −120 mV pre-pulse from a subsequent step to +40 mv from −30 mV. Scale: 10 pA/100 ms. b I A density in outside-out patches is significantly increased with PiB ( n = 12) or SB203580 ( n = 12) but not PD98059 ( n = 13), ordinary one-way ANOVA (two-tailed), * p < 0.05, ** p < 0.01. c No significant difference in decay kinetics was observed among the drug treatments and vehicle control, Kruskal–Wallis test (two-tailed), p > 0.05 . d Trace of transient I A in WT and Kv4.2TA mice. e I A density in outside-out patches is significantly increased in Kv4.2TA mice ( n = 15) relative to WT ( n = 14), two-tailed unpaired T -test, * p < 0.05. f No significant difference in decay kinetics was observed between Kv4.2TA and WT, two-tailed Mann–Whitney test, p > 0.05. g – i I A recovery from inactivation. g Sample traces of I A recovery from inactivation in Kv4.2TA and WT. Scale: 20 pA/200 ms. h Normalized recovery curves from Kv4.2TA ( n = 11) shows faster recovery relative to WT ( n = 9). i Single-exponentials fitted to normalized recovery curves yielded significantly reduced tau in Kv4.2TA relative to WT, unpaired two-tailed T -test. ** p < 0.01. Data are presented as mean ± SEM. Full size image Kv4.2TA mice demonstrate enhanced cognitive flexibility Kv4.2 KO mice have shown impairment in learning and memory 3 . In light of the physiological deficit and the accompanying biochemical changes, we sought to determine whether the disruption of Kv4.2 phosphorylation and Pin1 binding might alter any cognitive functions. In an open field test, Kv4.2TA mice showed normal locomotion, not significantly different from WT littermates (Supplementary Fig. 8a, b ). In addition, Kv4.2TA mice displayed similar center vs perimeter time as WT littermates (Supplementary Fig. 8a, b ), suggesting that their anxiety level was normal, too. We then employed the Morris water maze task to test hippocampal-dependent spatial memory. Both WT and Kv4.2TA mice showed similar performance in the training sessions (no main effect of genotype or genotype × session interaction, Fig. 7a ) as well as in a probe trial (Fig. 7b, c ). We then tested reversal learning by moving the hidden platform to the opposite quadrant of the pool. Kv4.2TA mice learned the new target location faster than WT littermates (effect of genotype: F 1,27 = 11.92, p = 0.0018 session: F 3,81 = 33.58, p = 1E-6; genotype × session: F 3,81 = 3.71, p = 0.015; Fig. 7d ). In addition, they spent more time in the new target quadrant and less time in old target quadrant during the reversal probe trial (Fig. 7e, f ). This difference suggests that, although Kv4.2 T607 phosphorylation deletion did not affect the acquisition of spatial memory itself, it led to enhanced behavioral flexibility when the location of the platform was changed during the reversal learning. To investigate whether other forms of behavioral flexibility were also affected, we performed an operant reversal test 49 . In this task, Kv4.2TA mice showed normal acquisition of lever pressing behavior and no significant difference in reaching the learning criteria on a fixed ratio (FR1) schedule (Fig. 7g ). In the following 5 random ratio (RR2) sessions, Kv4.2TA mice and WT littermates received similar numbers of rewards. However, when the reward lever was switched, Kv4.2TA mice exhibited faster reversal learning than WT littermates (Fig. 7h, i ), as in the water maze (Fig. 7d ). Kv4.2TA mice decreased inactive lever pressing and increased active lever pressing more rapidly than WT controls (effect of genotype: F 1,19 = 5.017, p = 0.037; session: F 4,76 = 4.110, p = 0.0045; Fig. 7h ). Kv4.2TA mice reached the high level of active lever press while WT littermates barely started the reversal learning on the first day (Fig. 7h ). On the second day of reversal learning, Kv4.2TA mice retained the high active lever pressing activity while WT littermates started reversal learning and caught up Kv4.2TA mice (effect of genotype x session: F 4,76 = 4.514, p = 0.0025; Fig. 7i ). These data show that disruption of Kv4.2 phosphorylation at T607 site and Pin1 binding/isomerization contributes to an enhanced rate of reversal learning suggesting improved cognitive flexibility. Fig. 7: Improved reversal learning in Kv4.2TA relative to WT mice. a–c Kv4.2TA mice showed normal learning of the initial platform location in the Morris water maze task. Average escape latencies of four trials per round for WT ( n = 16 mice) and Kv4.2TA mice ( n = 14 mice) over a training period of 3 days (two rounds each day). b Swimming time heat maps during the probe trial of the WT and Kv4.2TA mice at day 5. c Time spent in each quadrant during the probe trial. d – f Kv4.2TA mice showed improved reversal learning in the Morris water maze. d Average escape latencies of four trials per round for WT and Kv4.2TA mice over a reversal training period of 2 days (two rounds each day). e Swimming time heat maps during the reversal probe trial of the WT and Kv4.2TA mice at day 8. f Time spent in each quadrant during the reversal probe trial. Two-way ANOVA, * p < 0.05. g – i Kv4.2TA mice displayed improved lever press reversal learning. g Number of trainings to reach criteria in initial learning. h Number of active lever press in the 1st day of reversal learning. n = 10 for WT, n = 11 for Kv4.2TA. Two-way ANOVA, * p < 0.05. i Number of active lever press in the 2nd day of reversal learning. n = 10 for WT, n = 11 for Kv4.2TA. Two-way ANOVA, * p < 0.05. Data are presented as mean ± SEM. Full size image Discussion The present study describes a Pin1 isomerase-dependent mechanism that regulates the composition of the Kv4.2-DPP6 complex, neuronal excitability, and cognitive flexibility (Fig. 8 ). This mechanism occurs in a subset of neurons that are activated by neuronal activity or other stimulations. Pin1 was identified as a Kv4.2 binding partner by a TAP-MS assay in HEK-293T cells. As Pin1 is a cell proliferation regulator, examination of its substrates thus far has mainly focused on cell cycle proteins that play a pivotal role in cancer 50 . Increasingly, studies have shown that Pin1 isomerizes proteins in the brain such as APP 25 , Tau 51 , mGluR5 (ref. 52 ), PSD-95 (ref. 41 ), and CRMP2A 53 . We provide here the first report of a voltage-gated channel, Kv4.2, that is directly modified by Pin1. The effects of this modification were found to be important for neuronal and cognitive function. Pin1 is a peptidyl-prolyl cis - trans isomerase that catalyzes the isomerization of peptidyl-prolyl peptide bonds. Pin1 differs from other isomerases as it is, so far, the only known prolyl isomerase that specifically catalyzes isomerization of certain Ser/Thr-Pro bonds upon their phosphorylation 54 . Isomerization of Ser/Thr-Pro motifs is especially important because kinases and phosphatases specifically recognize the cis or trans conformation of the prolyl peptide bond of their substrates 55 and phosphorylation further slows down the isomerization rate of proline 56 . Pin1 enhances the cis / trans conformational changes by reducing the free energy barrier, resulting in a markedly increased conversion rate up to 100- to 1000-fold 57 . The fast switch provides the correct conformation and precise timing for further activation and could be critical for modulating channel function in response to transient neuronal activity. Loss of phosphorylation essentially locks the channel into one confirmation. As described here, we show Pin1 acts as a molecular switch that mediates the activity-dependent regulation of a channel complex, thereby affecting neuronal excitability. Fig. 8: Working model of Pin1-dependent Kv4.2-DPP6 complex remodeling that underlies neuronal excitability and cognitive inflexibility. a In WT mice, stimulations such as seizure and exposure to a novel environment trigger the phosphorylation of Kv4.2 at T607, which allows Pin1 binding to pT602 and pT607 which subsequently isomerizes the pT607-P bond. This process changes the conformation of Kv4.2, which dissociates the Kv4.2-DPP6 complex and increases neuronal excitability and cognitive inflexibility. b In Kv4.2 TA mice, the 607 site is no longer phosphorylatable so that Pin1’s effect on Kv4.2 is abolished. The Kv4.2-DPP6 complex is stable, neuronal excitability is reduced, and cognitive flexibility is improved. Full size image By using the TAP technique to purify exogenously-expressed TAP-tagged Kv4.2 from HEK-293T cells, many intracellular proteins were identified in addition to Pin1 (Supplementary Fig. 1c, d ). The majority of the binding partners are protein synthesis and degradation machinery proteins, such as ribosomal proteins, eukaryotic initiation factors, proteasome subunits and ubiquitin-specific proteases (Supplementary Fig. 1c, d ). This is reasonable since the exogenously expressed protein underwent active translation and degradation. Kv4.2-Pin1 binding is direct and requires critical amino acids that can bind to other substrates in the Pin1 WW (W34) and PPIase (R68, R69) domains. The Pin1 binding motif in Kv4.2 involves two adjacent pT-P motifs (TPPVTTP) which is similar to that of mGluR5 (TPPSPF) 52 . They even share the same pattern of phospho-regulation, i.e., the first T-P phosphorylation is not altered by stimulation while the second S/T-P phosphorylation is dynamic 52 . Interestingly, the first phosphorylation site of both proteins contains the T-P-P motif that is a better fit for binding the Pin1 WW domain while the second S/T-P motif binds to the catalytic domain 36 . This could be a common mechanism of how Pin1 regulates of dually phosphorylated proteins. In mGluR5, there is a second Pin1 binding motif in its C-terminal. The Kv4.2 mutant experiment (Fig. 2c ) showed there is about 40% Pin1 binding left when the T602 and T607 sites were mutated, suggesting there may exist another Pin1 binding site. However, dynamic Pin1 binding to Kv4.2 is dependent on T607 site (Fig. 5d, e ). Although ERK can phosphorylate the three proline-directed sites (T602, T607, and S616) in vitro 31 , we found that p38 is a better proline-directed kinase for the T607 site of Kv4.2. Extensive and intensive studies highlighted the role of p38 in the stress responses, such as osmotic shock, UV irradiation, and inflammatory cytokines 58 . We have found exposure to an enriched novel environment and seizure induction by PTZ or KA activate p38 and increase Kv4.2 phosphorylation at T607 in mice (Figs. 2, 3 ). Importantly, we also found that p38-Pin1-Kv4.2 pathway regulates Kv4.2-DPP6 complex (Fig. 4 ) and neuronal excitability (Fig. 5 ). The mechanism how Pin1-elicited Kv4.2 conformation change leads to Kv4.2-DPP6 disassociation is interesting and needs to be elucidated in follow up studies. The p38 MAPK pathway is possible target for the treatment a number of neurodegenerative diseases, such as AD 59 . Thus, this Kv4.2 phosphorylation-Pin1 mechanism could be applied to treat pathological conditions and neurodegeneration diseases 22 . As Kv4.2 containing channels are the primary carriers of the subthreshold, transient A-current, their impact on membrane excitability in rodent hippocampal pyramidal cells is well-documented 2 , 60 . We confirmed the significant contribution of T607 phosphorylation in mediating this influence as reduced excitability was observed in CA1 pyramidal cells of Kv4.2TA mice. This was further bolstered by our finding that the Pin1 blocker, PiB, decreased neuronal excitability in WT but not Kv4.2TA neurons, implying a floor effect in the mutant cells where the engagement of Pin1 and Kv4.2 is already maximally inhibited. Further, we show this reduction in excitability can be traced to alterations in I A . Pharmacological and genetic disruption of the p38-Pin1-Kv4.2 cascade resulted in enhanced I A density in CA1 pyramidal somata. It is well established that modulation of Kv4.2 surface expression and/or kinetics/voltage-dependent properties, through the alteration of auxiliary subunits, impacts intrinsic excitability 43 . Although we identified a remarkable similarity in the firing properties of neurons from WT mice treated with Pin1 inhibitors and Kv4.2TA mice without pharmacological intervention, we did not observe significant alterations in subthreshold excitability in these mice relative to WT. This indicates the possibility that additional ion channels impacting sub-threshold membrane properties in CA1 pyramidal cells, such as Kv4.1 and Kv4.3 (Supplementary Fig. 2b, c ), may be regulated by Pin1 as these interactions would also be impaired by broad Pin1 inhibition. Our biochemistry data showed that Kv4.2-DPP6 dissociation is impaired in Kv4.2TA mice, indicating that the Kv4.2-DPP6 complex is more stable without phosphorylation at T607. In heterologous expression systems, association of DPP6 in the tripartite Kv4.2-KChIP-DPP6 complex leads to increased current density, faster recovery from inactivation, and more rapid inactivation 15 , 16 . Our voltage-clamp recordings support this notion as we identified increased I A density in Kv4.2TA mice compared to WT littermates and a non-significant trend toward faster macro current decay, which was also observed with p38 inhibition. Interestingly, the recovery from inactivation kinetics in Kv4.2TA mice displayed a shift to faster recovery, consistent with more channels in complex with DPP6, further supporting our hypothesis. Moreover, that the Kv4-specific blocker AmmTX3 (250 nM) occluded the effects of PiB on WT mice, suggests that the PiB-induced reduction in excitability is mediated by Kv4 channels that are associated with DPP6, since the high-affinity blockade of Kv4 channels by AmmTX3 depends on the presence of DPP6 (ref. 47 ). It is intriguing that constitutive knockout of DPP6 does not result in significant alterations in somatic I A (ref. 17 ); however, evidence is suggestive of homeostatic compensation in the soma of DPP6 KO mice, which preserves relative excitability 17 . It is likely this compensation is absent in Kv4.2TA mice given our findings that firing properties are also significantly altered. The basal level of Kv4.2-DPP6 protein complex is not altered in Kv4.2TA mice compared to WT littermates in biochemistry experiments (Fig. 4f ). However, we saw reduced neuronal excitability (Fig. 5c, d ) and increased I A (Fig. 6d, e ) in Kv4.2TA mice compared to WT littermates. This difference likely results from technical differences between biochemical and electrophysiological experiments. To determine this, we measured p38 phosphorylation in sliced brain in comparison with un-sliced brain. The result showed that slicing and recovery largely activates p38, and Kv4.2 phosphorylation at T607 is also increased after slicing (Fig. 5g ). These data suggested that slicing and recovery process before recording has already activated p38-Pin1-Kv4.2 pathway, and the data is consistent with our hypothesis. Taken together, our data demonstrate that Pin1 regulates the composition of the Kv4.2-DPP6 complex and neuronal excitability. These changes may then impart additional, so-far undetermined, downstream effects in the neuron. Cognitive flexibility is the ability to appropriately adjust one’s behavior according to a changing environment. Greater cognitive flexibility is associated with favorable outcomes throughout the lifespan. Here we showed that reduced neuronal excitability unexpectedly left initial learning and memory intact and improved reversal learning in Kv4.2TA mice. Cognitive flexibility has previously been associated with both NMDAR- and mGluR-dependent long term depression 61 , 62 , 63 , 64 . Further research is required to attribute a cellular function to the enhancement in reversal learning observed in Kv4.2TA mice. Cognitive inflexibility is observed in various psychiatric disorders such as autism spectrum disorder (ASD) 65 , schizophrenia 66 , suicidal ideation 67 , and anxiety and mood disorders 68 . Considering that both Kv4.2 and DPP6 are implicated in such psychiatric disorders 8 , 22 , 69 , the stability of the Kv4.2-DPP6 complex might be a common factor of pathophysiology. It will be interesting to examine if the T607A mutation can rescue cognitive inflexibility in mouse models of psychiatric or neurodegenerative disorders. Taken together, our results reveal that disrupting the activity-dependent isomerization of Kv4.2 by Pin1 stabilizes the Kv4.2-DPP6 complex and improves cognitive flexibility. Stabilization of the Kv4.2-DPP6 complex might represent a promising strategy for enhancing adaptive cognitive behavior and correcting maladaptive cognitive deficits in a number of neuropsychiatric conditions. Methods Expression constructs The human Myc-DDK-Kv4.2 construct was purchased from Origene (RC215266). All of the other expression constructs were made by PCR. Internal deletions and point mutations were generated using either the QuikChange Site–Directed Mutagenesis Kit (Stratagene) or the megaprimer method. PCR products were cloned into expression vectors pGEX 4T2 (Pharmacia) and pRK5 (Genentech), with Myc, flag or HA tags as we reported previously 52 . All constructs were verified by sequencing. Chemicals All chemicals were purchased: KA (Sigma, K0250), PTZ (Sigma, P6500), SB203580 (Tocris, 1202), PD 98059 (Tocris, 1213), SL327 (Tocris, 1969), S-AMPA (Tocris, 0254), Juglone (Millipore, 420120), PiB (Sigma, B7688), AmmTX3 (Alomone, 305). For injections, KA and PTZ were dissolved in saline; SB203580, SL327 and Juglone were dissolved in DMSO and 10% Tween 80. Antibodies Mouse anti-Kv4.2 (NeuroMab, 75-016) was used at 1:2000 for western blot, 1:200 for immunostaining, Rabbit anti-Kv4.2 (Sigma, P0233) was used at 1:2000 for western blot, rabbit anti-Kv4.2 (Sigma, HPA029068) was used at 1:200 for staining, pT602 (Santa Cruz, SC-16983-R) was used at 1:1000 for western blot, pT607 (Santa Cruz, SC-22254-R) was used at 1:500 for western blot, Pin1 (Santa Cruz, SC-46660) was used at 1:100 for staining, 1:1000 for western blot, Pin1 (Millipore, 07-091) was used at 1:3000 for western blot, p38 (Cell Signaling, 9212 s) was used at 1:1000 for western blot, p-p38 (Cell Signaling, 4511 s) at 1:1000 for western blot, DPP6 (Abcam, 41811) was used at 1:2000 for western blot, Myc (Millipore, 05-419) was used at 1:10000 for western blot, HA(Santa Cruz, SC-805) was used at 1:1000 for western blot, Actin (Sigma, A-1978) was used at 1:10000 for western blot; Alexa Fluor 488 goat anti-mouse (Invitrogen, A-11029) was used at 1:500; Alexa Fluor 488 goat anti-rabbit (Invitrogen, A-11034) was used at 1:500; Alexa Fluor 555 goat anti-mouse (Invitrogen, A-21424) was used at 1:500; Alexa Fluor 555 goat anti-rabbit (Invitrogen, A-21429) was used at 1:500; Alexa Fluor 680 goat anti-mouse (Invitrogen, A-21057) was used at 1:10000; Alexa Fluor 680 goat anti-rabbit (Invitrogen, A-21076) was used at 1:10000; IRDye 800CW goat anti-mouse (Licor, 926-32210) was used at 1:10000, IRDye 800CW goat anti-rabbit (Licor, 926-32211) was used at 1:10000. Mouse models Kv4.2TA mice were generated using CRISPR-Cas9 techniques. Briefly, CRISPR sgRNA (CCTGTCGTCGCCTTCTGGGG) was made by in vitro transcription using ThermoFisher’s sgRNA synthesis service. B6D2F1 mice (JAX Stock No. 100006) were used as embryo donors for this study. For the injection step, sgRNA (20 µg / ml), Cas9 mRNA (100 µg/ml, Trilink Biotechnologies) and corresponding single strand oligos with mutations (100 µg/ml) were injected into the cytoplasm of fertilized eggs, which then were cultured overnight in M16 medium. Those embryos that reached the two-cell stage of development were implanted into the oviducts of pseudo-pregnant surrogate mothers (CD-1, Charles River). Mice born to these foster mothers were genotyped by PCR amplification followed by DNA sequencing to identify mice with correct mutations. Mice were group housed in plastic mouse cages with free access to standard rodent chow and water. The colony room was maintained at 22 ± 2 °C with a 12 hr: 12 h light: dark cycle. Kv4.2TA mice were backcrossed at least three generations onto C57/Bl6J mice. All animal procedures were performed in accordance with guidelines approved by the National Institute of Child Health and Human Development Animal Care and Use Committee and in accordance with NIH guidelines. Cell culture and transfection HEK-293T cells used in biochemistry experiments were obtained from Dr. Paul Worley’s lab 70 . HEK-293T cells were cultured in DMEM medium containing 10% FBS. Transfections were performed with X-tremeGENE 9 according to the manufacturer’s specifications. Cells were harvested about 40 h after transfection. Neuronal culture Mouse hippocampal neuron cultures from embryonic day 18 (E18) pups were prepared as reported previously 71 . In all, 1 × 10 6 neurons were added to each well of a six–well plate (Corning) with cover slips coated with poly–L–lysine. Growth medium consisted of Neurobasal medium (Invitrogen) supplemented with 5% FBS (Hyclone), 2% B27, 1% Glutamine (Invitrogen), 100 U/mL penicillin, and 100 U / mL streptomycin (Invitrogen). Neurons were fed twice per week with glia-conditioned growth medium. DIV14-17 neurons were used for biochemistry experiments. Rat hippocampal neuron cultures from embryonic day 18 (E18) pups were prepared similarly as above. DIV14-17 neurons were used for the immunostaining experiment. Tandem affinity purification-mass spectrometry (TAP-MS) assay Kv4.2 was subcloned into the TAP tag vector that was obtained from Agilent (pCTAP, #240102). TAP-tagged Kv4.2 was then subcloned into the lentivirus vector (Dr. Paul Worley’s lab) to generate TAP-Kv4.2-IRES-GFP Lentivirus. TAP-Kv4.2-IRES-GFP Lentivirus and IRES-GFP control Lentivirus were generated using a standard protocol 72 . Rar hippocampal neuron cultures from embryonic day 18 (E18) pups were prepared as described above. Neurons were infected by TAP-Kv4.2-IRES-GFP Lentivirus or IRES-GFP control Lentivirus on the day of the culture and harvested at DIV14. TAP-Kv4.2 was purified using the TAP purification kit from Agilent (#240107) with some modifications. The samples were run on 10% SDS-PAGE gel (Novex/Invitrogen). The gels that contains the protein samples were excised, separated into high molecular and low molecular weight samples, and sent to the Taplin Mass Spectrometry Facility at Harvard University for in-gel digestion using trypsin and mass spectrometric analysis. Peptide pulldown The following peptides were synthesized: Non-Phospho for T602 and T607: KAIISIPTPPVTTPEGDDR; pT602: KAIISIP-pT-PPVTTPEGDDR; pT607: KAIISIPTPPVT-pT-PEGDDR; pT602, pT607: KAIISIP-pT-PPVT-pT-PEGDDR; Non-Phospho for S616: KEGDDRPESPEYSGG; pS616: KEGDDRPE-pS-PEYSGG. The peptides were conjugated to Affi-Gel 15 (Bio-Rad) according to the manufacturer’s instructions. Peptide-linked Affi-Gel was incubated for 3 h at 4 °C with Pin1 protein that was expressed in HEK-293T cells, and then washed clear for western blot analysis. Co-immunoprecipitation and immunoprecipitation assays Mouse brain tissues or HEK-293T cells were used in co–immunoprecipitation assays as previous reported in ref. 70 . For Kv4.2-Pin1 co-IP experiment, co-IP buffer (1 X PBS, pH 7.4, with 0.8% Triton X–100, phosSTOP and Complete™ EDTA–Free protease inhibitors) were added (1:20 for brain tissues and 400 µl for a 6-well of HEK-293T cells), and the samples were sonicated. After centrifugation, the supernatant was mixed with 2–3 µg of Kv4.2 (NeuroMab, 75-016, or Sigma, P0233) or myc (Millipore, 05-419) antibodies for 3–4 h at 4 °C. Next, 40 µl of protein G magnetic beads (Bio-Rad, 161-4023) was added for an additional 2 h or overnight. The protein beads were washed three times with co-IP buffer. The protein samples were eluted with SDS loading buffer and analyzed by gel electrophoresis and western blotting. For Kv4.2-DPP6 co-IP experiment, co-IP buffer (1 X PBS, pH 7.4, with 1% Triton X–100, 0.2% Chaps, phosSTOP and Complete™ EDTA–Free protease inhibitors) were added (1:20 for brain tissues), and the samples were sonicated. After centrifugation, the supernatant was mixed with 3–4 µg of Kv4.2 antibody (NeuroMab, 75-016) for 3 h at 4 °C. Next, 50 µl of 1:1 protein G–Sepharose slurry (GE Healthcare, 17-0886-02) was added for an additional 3 h. The protein beads were washed three times with IP buffer. The protein samples were eluted with SDS loading buffer and analyzed by gel electrophoresis and western blotting. For Kv4.2 phosphorylation detection, brain tissues sonicated in co-IP buffer (1 X PBS, pH 7.4, with 1% Triton X–100, phosSTOP and Complete™ EDTA–Free protease inhibitors). After centrifugation, the supernatant was mixed with 3 µg of Kv4.2 (NeuroMab, 75-016) antibody for 3 h at 4 °C. Next, 50 µl of 1:1 protein G–Sepharose slurry (GE Healthcare, 17-0886-02) was added for an additional 3 h. The protein beads were washed three times with IP buffer. The protein samples were eluted with SDS loading buffer and analyzed by gel electrophoresis and western blotting. Molecular modeling Models of the bound phosphotyrosine-proline sequence motifs were developed from published structures. These models were primarily based on PDB: 2N10 (ref. 73 ) for the WW domain, and PDB: 2Q5A 74 for the Catalytic domain. Modeling was done by an interplay of “manual” manipulation with UCSF Chimera 75 and energy minimization (to prevent steric overlap and optimize salt-bridges and hydrogen bonds) with CHARMM 76 . Images were generated with UCSF Chimera. Subtilisin proteolysis Myc-Kv4.2 (co-transfected with p38) was expressed in HEK293T cells and purified with myc-magnetic beads (Thermo Scientific, #88842). Kv4.2 Phosphorylation at T602 and T607 after purification was measured by saturated phospho-specific and total Kv4.2 antibody pulldown combined with western blot. About 2/3 of purified Myc-Kv4.2 was phosphorylated at T607 and was used for proteolysis. Equal amount of Myc-Kv4.2 were then incubated with 100 ng of either GST, GST-Pin1, and GST-Pin1C113S in a buffer containing 50 mM HEPES, pH 7.5, 100 mM NaCl, 1 mM MgCl 2 , supplemented with phosphatase inhibitors. After 30 min incubation at room temperature, reaction mixtures were cooled on ice, and subtilisin (Sigma-Aldrich, P5380) was added for a further 1 min on ice. The reaction was stopped by the addition of boiling sample buffer, and the proteolytic fragments were resolved by 4–12% SDS-PAGE and visualized by western blot analysis. Western blot and quantification Protein samples were mixed with 4x LDS sample buffer (Invitrogen NP0007) and 10x sample reducing agent (Invitrogen NP0007) to a final concentration of 1×. Samples were loaded on 4–12% Bis-Tris gradient gel (Invitrogen 12-well, NP0322; 15-well, NP0323). The proteins were transferred to Immobilon-FL PVDF membrane (EMD Millipore, IPFL00010). The membrane was blocked with Odyssey blocking buffer (Li-COR, 927-40000) for 1 h at room temperature, followed by incubation with primary antibody in PBS overnight at 4 °C. The membrane was then washed with PBST (PBS, pH 7.4, and 0.1% Tween-20) three times and incubated with secondary antibody in PBS for another hour. After three washes with PBS, the membrane was scanned using an Odyssey imaging system (LI-COR) according to the manufacturer’s protocol. Quantification of western blots was carried out using the gel analysis function in ImageJ within the linear range of detection which is determined by using serial dilutions of a representative sample. Immunostaining Cultured hippocampal neurons (DIV10) were fixed with 4% PFA, and permeabilized with 0.2% Triton X-100 in PBS. Cells were then blocked with 10% horse serum at RT for 1 h and then incubated with mouse anti-Pin1 antibody (Santa Cruz, SC-46660, 1:100) and rabbit Kv4.2 antibody (Sigma, HPA029068, 1:200) at 4°C overnight. After washing, cells were incubated with anti-mouse-555 and anti-rabbit-488 secondary antibodies at RT for 1 h. After washing, cells were then mounted on slides with anti-fade mounting medium containing 4′,6-diamidino-2-phenylindole (DAPI, Invitrogen, P36962 ) and imaged using a Zeiss 710 laser scanning confocal microscope equipped with a ×63 objective. Acute hippocampal slice preparation For all electrophysiological recordings, adult male (5–7 weeks) mice were used. Mice were anesthetized in isoflurane and decapitated. Brains were removed and washed with ice-cold sucrose cutting solution. The sucrose solution was made up of the following (in mM): 60 NaCl, 3 KCl, 28 NaHCO 3 , 1.25 NaH 2 PO 4 , 5 Glucose, 0.5 CaCl 2 , 7 MgCl 2 . Brain hemispheres were dissected and mounted following a 45° cut of the dorsal cerebral hemisphere(s). Modified transverse slices (300 μm) were made by a Leica VT1200S vibrating microtome in ice-cold sucrose that was continuously bubbled with carbogen (95% O 2 /5% CO 2 ). Slices were recovered at 32 °C in sucrose solution for 30 min at which time the solution temperature was slowly lowered to room temperature where it remained for the remainder of the recording day. Whole-cell current-clamp recordings Following a 1-hour recovery in sucrose cutting solution, hippocampal slices were transferred to a recording chamber submerged in artificial cerebral spinal fluid (ACSF) with the temperature maintained at 33 °C (±1 °C). The ACSF contained the following (in mM): 125 NaCl, 2.5 KCl, 25 NaHCO 3 , 1.25 NaH 2 PO 4 , 25 Glucose, 2 CaCl 2 , 1 MgCl 2 (pH 7.4). The recording chamber was continuously perfused with carbogen-bubbled ACSF at a rate of 2–3 mL/min. Somatic whole-cell patch-clamp recordings were performed on identified somata of hippocampal CA1 pyramidal neurons. Pyramidal neurons were identified using infrared Differential Interference Contrast (DIC) on an upright Leica Axioskop 2. Cells were patched with 3–6 MΩ borosilicate glass pipettes pulled from a Narishige vertical puller and filled with K + Gluconate-based intracellular solution consisting of the following (in mM): 20 KCl, 125 K-Gluconate, 1 EGTA, 4 NaCl, 4 Na 2 ATP, 0.3 NaGTP, 10 HEPES, 10 Phosphocreatine with pH adjusted with KOH and HCl to a final value of 7.25–7.30 and an osmolarity of 290–300 mOsm. Firing properties were measured from whole-cell recordings in the conditions described above unless otherwise noted (described in detail below). All data were recorded with a Multiclamp 700b amplifier (Molecular Devices) and a Digidata 1440 A digitizer. Signals were low-pass filtered at 5 kHz and digitized at 10 kHz using Clampex 10.7 software and were acquired in bridge balance mode to compensate series resistance. Liquid junction potential was not corrected for. Subthreshold membrane properties were measured after initial break-in in order to avoid alteration in equilibrium potentials and dialysis as a result of solution exchange. Whole-cell capacitance and series resistance were measured from Multiclamp 700B commander. A voltage step of −10 mV was initiated, and the decay tau of the whole-cell capacitive transient was used to calculate these parameters. Resting membrane potential was measured after switching to I = 0. Any recordings where series resistance exceeded 25 MΩ or resting membrane potential was greater (more depolarized) than −55 mV were discarded. Input resistance was calculated as the slope of the I–V curve in response to current steps from −50 to 50 pA in 50 pA steps (three steps in total). To evoke APs in patched CA1 pyramidal neurons, square 1 s current pulses were elicited in 50 pA steps with current injections ranging from −200 to +200 pA in 50 pA steps. In some cases, for instance in the measurement of rheobase, smaller step sizes of 20 pA were used to enhance the resolution of the average minimum current magnitude required to elicit APs. Three sweeps at each magnitude were elicited and the average response of the three sweeps was used for each cell. All measures of AP waveform were taken from the first spike in a train in response to a 200 pA injection, and inter-spike measurements, including inter-spike interval and after-hyperpolarization amplitude, were recorded between the first two spikes in a train elicited by a 200 pA square current injection (see associated figures). Outside-out somatic patch voltage-clamp recordings of A-current A-current was recorded in voltage-clamp mode from outside-out patches pulled from somata of identified pyramidal cells in hippocampal slices. Outside-out patch pipettes (3–6 MΩ) were filled with K + Gluconate-based intracellular solution consisting of the following (in mM): 20 KCl, 125 K-Gluconate, 1 EGTA, 4 NaCl, 4 Na 2 ATP, 0.3 NaGTP, 10 HEPES, 10 Phosphocreatine. All recordings were performed at room temperature (24–25 °C). Tetrodotoxin citrate (TTX) (500 nM), Gabazine (2 µM), and CNQX (2 µM) were added to the extracellular bath (ACSF) in order to block voltage-gated Na + channels and ligand-gated channels. A-current density was calculated using a standard subtraction protocol. In short, a pre-pulse step to −120 mV for 600 ms from a holding potential of -65 mV was initiated to relieve inactivation of A-type K + channels. Voltage was stepped to +40 mV for 500 ms to measure total outward K + current. A subsequent sweep consisted of a pre-pulse step to −30 mV to inactivate the transient current followed by a step to +40 mV to measure the non-inactivating current. Peak amplitude of the non-inactivating current was subtracted from total outward K + current amplitude offline to isolate I A . Leakage and capacitive currents were digitally subtracted. Peak amplitude was normalized to current density by dividing by patch capacitance, which was measured following formation of the outside-out patch configuration and was analyzed with custom written code in MATLAB version R2018a (MathWorks). Inactivation kinetics were measured by fitting a single exponential to each isolated I A current trace and were analyzed with Graphpad Prism. Traces are averages of 10-30 sweeps. Recovery from inactivation was calculated using a double voltage-step (−120 to +30 mV) protocol with increasing interstep intervals between steps. Interstep intervals were (in ms): 5, 10, 15, 20, 25, 50, 100, 200, 500, and 1000. Recovery from inactivation was calculated as the ratio of the peak amplitude of the second current trace relative to the point where the initial A-current was fully inactivated and was normalized to the peak amplitude of the initial A-current. Single exponentials were fitted to the non-linear fitted curves of normalized recovery from inactivation and the taus from each individual recovery curve were averaged and compared. Voltage-dependent activation was recorded using a voltage step protocol consisting of a 600 ms prestep to −120 mV, followed by a series of 13 steps (10 mV each from −80 to 40 mV). Leakage and capacitive currents were digitally subtracted. The same protocol but with a 600 ms prestep from -65 mV to -30 mV was used to record non-inactivating currents, which were offline subtracted from the overall K + currents to obtain inactivating A-type K + currents. Currents were then converted into conductances, normalized to peak conductance at 40 mV, plotted against the holding voltage and fitted with a Boltzman-function to obtain V 1/2-activation and k activation . Voltage-dependent inactivation was recorded using a voltage-step protocol consisting of a series of 14 presteps (each 600 ms and 10 mV, from −120 to 10 mV) followed by an activating voltage step to 20 mV. Leakage and capacitive currents were digitally subtracted. Peak currents were measured for each step and the I/V relationship was fitted with a Boltzmann function to obtain V 1/2-inactivation and k inactivation . All data were recorded with a Multiclamp 700b amplifier (Molecular Devices) and a Digidata 1440A digitizer, were digitized at 10 kHz and low-pass filtered at 2 kHz using a Bessel filter. Behavioral assays Open field task Novelty-induced locomotor activity was assessed in a novel open-field square arena (50 × 50 cm) constructed of white Plexiglas as previously described 77 . Mice were acclimated to the testing room for at least ~10 min and then placed in the arena and left to explore freely for 30 min. Sessions were performed once a day for two successive days. The distance traveled and time spent in different areas of the maze were measured. Results were analyzed with “ANY-maze” software (ANY-maze, Wood Dale, IL, USA). Data were compiled from two independent experimental cohorts. Male mice used in the first and second cohorts were ~8.5–10 and ~15–16 weeks old, respectively. Morris water maze The Morris water maze task was performed to evaluate hippocampus-dependent spatial navigation learning and memory 78 , 79 . The water maze consisted of a 120 cm circular pool (depth 50 cm), filled ~40 cm deep with 20–22 °C water containing a 10 cm wide square platform. External high contrast cues were placed on the interior of the pool above the water surface to aid with spatial navigation. Trials were video recorded and scored by ANY-maze software (ANY-maze, Wood Dale, IL, USA) for measures including latency to find the hidden platform, total distance traveled, and swim speed. The latency to the platform of the training trials was measured manually with a stopwatch. General mouse handling was performed as follows: Mice were acclimated to the testing room for at least ~1 h before testing. Each mouse was placed into the water maze facing the wall in one of four possible quadrant positions, which was pseudo-randomly varied by training session. Mice were given 60 s to find the platform and a ~15 s platform rest interval. If a mouse was unable to find the platform in the allocated time, it was gently guided to the platform and allowed to rest for ~15 s. Mice were then patted dry with a cloth and put back into a warm cage for ~15 s after each trail. For data analyses, a latency time of 60 s was ascribed to mice that failed to reach the platform without guidance. On Day 1 mice were trained in the visible platform version of the Morris water maze task to assess general swimming and visual ability. The platform was ~1 cm above the clear water surface with a red flag placed on the platform to increase its visibility. Each mouse underwent two visible platform training sessions and the location of the platform was varied between sessions. No significant difference in escape latency was apparent between genotypes. The water was made opaque with nontoxic white and red paint between Days 1 and 2. On Day 2 through Day 4 (sessions 1–6), mice were trained for the Hidden Platform protocol where the flag was removed from the platform and additional water was added to the pool to submerge the platform ~1 cm below the surface. Mice were given a total of 24 training trials (4 trials per session, two sessions per day for three successive days). On Day 5, the platform was removed, and mice underwent a 60 s probe trial to determine the amount of time spent exploring the target quadrant. On Day 6 and Day 7, mice were trained for reversal learning (2 sessions per day for 4 total sessions) where the Hidden Platform was moved to the opposite quadrant. On Day 8, the platform was removed, and the mice underwent a 60 s probe trial to determine the amount of time spent exploring the target quadrant. ~16–17-weeks-old male mice were used. Data were compiled from two independent experimental cohorts. Lever press Operant reversal learning was performed as previously described in ref. 49 . Mice were food restricted to 85–90% of their free-feeding weight over several days prior to testing and throughout the experiment. Mice were trained to lever press on an FR1 schedule in daily sessions with an endpoint of 30 rewards or 30 min. When they collected all of the rewards in the allotted time for 3 consecutive days, animals progressed to RR2 sessions lasting 25 min, during which they could earn unlimited rewards. After 5 training days, levers were reversed such that the reinforced lever became non-reinforced and vice versa. Mice were given daily 15-minute reversal learning sessions for 5 days. Statistical analysis Biochemistry and behavior data were analyzed by Origin 2018b by two–tailed Student’s t test and two–way ANOVA, respectively. Electrophysiology data were analyzed by GraphPad Prism 7 (7.0d). For all measures of I A in outside-out somatic patches the experimenter was blinded to the genotype. For measures of firing properties, the experimenter was aware of these conditions. Sample sizes were not predetermined with any statistical methods but were chosen based on numbers reported in similar publications in the field. All statistical tests were two-tailed. Specifically, for electrophysiological analysis of I A and AP shape in WT and Kv4.2TA mice and pharmacological analysis (PiB treatment) in Kv4.2TA, unpaired t-test (with Welch’s correction) was used (passed normality testing). For electrophysiological analysis of the pharmacological impact on I A in WT slices a One-way ANOVA (ordinary), or One-Way ANOVA on Ranks (Kruskal–Wallis) was used and were corrected for multiple comparisons with Dunnett’s test (ordinary) or Dunn’s test (Ranks) respectively. The use of parametric or non-parametric analysis was determined after testing for normal distribution in the data using the D’Agostino & Pearson normality test for all electrophysiological data (alpha Level = 0.05). Non-parametric statistics were used if the data failed normality testing. For all analysis of pharmacological treatment effects in WT, significance was probed relative to control (vehicle). All analysis of firing frequency in response to sequential current steps, a two-way ANOVA with Sidak’s post hoc test was used. All the data are presented as mean ± SEM. Reporting summary Further information on experimental design is available in the Nature Research Reporting Summary linked to this paper. Data availability PDB:2N10, PDB:2Q5A, and other data that support the findings of this study are available from the corresponding authors in reasonable request. The source data underlying all figures and tables are provided as a Source Data file.

Researchers at the National Institutes of Health have discovered in mice what they believe is the first known genetic mutation to improve cognitive flexibility—the ability to adapt to changing situations. The gene, KCND2, codes for a protein that regulates potassium channels, which control electrical signals that travel along neurons. The electrical signals stimulate chemical messengers that jump from neuron to neuron. The researchers were led by Dax Hoffman, Ph.D., chief of the Section on Neurophysiology at NIH's Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD). It appears in Nature Communications. The KCND2 protein, when modified by an enzyme, slows the generation of electrical impulses in neurons. The researchers found that altering a single base pair in the KCND2 gene enhanced the ability of the protein to dampen nerve impulses. Mice with this mutation performed better than mice without the mutation in a cognitive task. The task involved finding and swimming to a slightly submerged platform that had been moved to a new location. Mice with the mutation found the relocated platform much faster than their counterparts without the mutation. The researchers plan to investigate whether the mutation will affect neural networks in the animals' brains. They added that studying the gene and its protein may ultimately lead to insights on the nature of cognitive flexibility in people. It also may help improve understanding of epilepsy, schizophrenia, Fragile X syndrome, and autism spectrum disorder, which all have been associated with other mutations in KCND2.

10.1038/s41467-020-15390-x

Chemistry

Chemists uncover a means to control catalytic reactions

Kelvin Anggara et al, Bond selectivity in electron-induced reaction due to directed recoil on an anisotropic substrate, Nature Communications (2016). DOI: 10.1038/ncomms13690 Journal information: Nature Communications

http://dx.doi.org/10.1038/ncomms13690

https://phys.org/news/2016-12-chemists-uncover-catalytic-reactions.html

Abstract Bond-selective reaction is central to heterogeneous catalysis. In heterogeneous catalysis, selectivity is found to depend on the chemical nature and morphology of the substrate. Here, however, we show a high degree of bond selectivity dependent only on adsorbate bond alignment. The system studied is the electron-induced reaction of meta-diiodobenzene physisorbed on Cu(110). Of the adsorbate’s C-I bonds, C-I aligned ‘Along’ the copper row dissociates in 99.3% of the cases giving surface reaction, whereas C-I bond aligned ‘Across’ the rows dissociates in only 0.7% of the cases. A two-electronic-state molecular dynamics model attributes reaction to an initial transition to a repulsive state of an Along C-I, followed by directed recoil of C towards a Cu atom of the same row, forming C-Cu. A similar impulse on an Across C-I gives directed C that, moving across rows, does not encounter a Cu atom and hence exhibits markedly less reaction. Introduction Bond-selective chemistry due to differences in energy barriers continues to play a major part in the understanding of thermal heterogeneous catalysis 1 , 2 . A notable advance in recent years has been the employment of selective excitation of a molecule incident or present on a surface as a means to influence the reaction probability. For vibrationally and translationally excited gaseous beams incident at metal surfaces, reactions have been reported to be mode-specific 3 , 4 , 5 , bond-selective 6 , 7 and gas-phase alignment-dependent 8 . In the case of molecules adsorbed at surfaces, electrons from scanning tunnelling microscopy (STM) have been employed to excite vibrationally or electronically the adsorbate, resulting in molecular translation 9 , 10 , 11 , rotation 11 , 12 , 13 , isomerization 14 and desorption 9 , 15 , 16 . The breaking of a selected chemical bond in the adsorbate has also been achieved by varying the electron energy 16 , 17 , 18 , 19 or, for extended adsorbates, the excitation location 19 . Here, we show a high degree of bond selectivity due to a cause that, to our knowledge, has not previously been noted, namely the directed recoil of products formed as a result of electron-induced surface reaction. The approach is quite general since directed recoil of products along what was the prior bond direction has been widely reported 16 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 . When the directed recoil occurs at an anisotropic surface, we find bond-selective surface reaction occurs. We show that different alignments, even of a chemically identical bond, can result in a hundred-fold alteration in reaction probability, with corresponding bond selectivity. The example given is that of an electron from an STM tip impinging on a single meta-diiodobenzene (mDIB) molecule on Cu(110) giving two orders of magnitude greater probability of breaking a carbon-iodine (C-I) bond lying ‘Along’ (AL) a Cu row than the C-I that lies ‘Across’ (AC) the rows. Molecular dynamics (MD) calculations employing an approximate anionic potential, followed by impulsive recoil across a ground potential computed by density functional theory (DFT) reproduce the observed bond selectivity, explaining it as due to directed recoil across the anisotropic surface of the Cu. Results Experiment Experiments were performed on a Cu surface held at 4.6 K. Two physisorbed states of mDIB were observed: one termed ‘Row’ 28 and the other ‘Diagonal’, the latter being the principal subject of the present report. An STM image of the Diagonal adsorption state on Cu(110) is shown in Fig. 1a , with a theoretical simulation directly below. Figure 1: Bond-selective electron-induced reaction of physisorbed mDIB on Cu(110). STM images (EXPT) and simulations (TH) show the selective dissociation of the C-I bond. ( a ) An intact physisorbed mDIB. The white cross indicates the midpoint of mDIB in the STM image. The two C-I bond directions, ‘AL’ and ‘AC’, are given by blue lines and designated as C-I(AL) and C-I(AC). ( b ) A reactive outcome as a result of the breaking of C-I(AL) in the prior mDIB reagent. ( c ) A reactive outcome for the single observed case of mDIB breaking C-I(AC). In the pictured final states (PIC), the green dashed circle indicates the Cu atom to which the IPh was bound, and the blue dashed line the direction of the broken C-I bond. All experimental STM images shown here are 25 × 17 Å 2 in size, recorded at a sample bias of +0.1 V and a tunnelling current of 0.1 nA. Full size image From the simulation, the benzene ring is located at a short-bridge site with one C-I directed along the Cu row and the other at 126° from the Cu row (measured clockwise from [1 0]), leading to the two lobes that were observed, corresponding to the two C-I bonds with I atoms located atop Cu atoms, as evident in the simulated STM image. The bond lengths of both C-I bonds in the physisorbed state were computed as ∼ 2.1 Å, and the heat of adsorption as 1.0 eV. This Diagonal adsorption state was also predicted by a computational study performed by Panosetti and Hofer 29 . The midpoint between the two lobes in Fig. 1a , EXPT, is indicated by a white cross used as the origin of the spatial distribution of the products following electron-induced reaction; see Fig. 1a–c . Electron-induced reaction of physisorbed mDIB was initiated by tunnelling electrons from the STM tip placed over the centre of the intact physisorbed reagent at a sample bias ≥+1.0 V. The experimental findings are illustrated in Fig. 1 , with, at left, sample STM images for the initial and final states of the major reaction path breaking the C-I bond AL the Cu rows, and, at the right, the minor path breaking the C-I bond directed AC the Cu rows. For 139 reactive cases, the breaking of only one of the two C-I bonds was observed, namely, the AL bond of the physisorbed mDIB, in preference to the AC bond. The bond selectivity amounted to a 99.3% probability for C-I(AL), as against 0.7% for the alternate C-I(AC). A representative case of C-I(AL) bond breaking is shown in Fig. 1b . The reaction products are a chemisorbed I atom and a chemisorbed iodophenyl (IPh). A comparison between experiment ( Fig. 1b , EXPT) and theory ( Fig. 1b , TH and PIC) allowed us to identify the binding sites of both products with respect to the prior mDIB. The I atom was found at the closest four-fold hollow site adjacent to the initial I atom position in the reagent. The IPh product was bound to the atop Cu atom underneath the reagent mDIB (green dashed circle, Fig. 1b , PIC). The I atom and IPh were found to recoil in opposite directions AL the Cu row ( Supplementary Fig. 1a and Supplementary Note 1 ), as expected for the dissociation of C-I(AL) aligned along the row. The most probable recoil distance of the I atom measured from the white cross of Fig. 1 was 4.7 Å. Figure 1c and Supplementary Fig. 1b show the minor path consisting of 1 out of 139 cases. In this case, the products were found to recoil in the opposite direction AC the rows ( Supplementary Note 1 ), from which we conclude that the bond dissociated was C-I(AC). The observed strong preference for the breaking of C-I(AL) rather than C-I(AC) shows that the electron-induced reaction of the Diagonal physisorbed mDIB is markedly bond-selective. As the reaction of Row physisorbed mDIB did not exhibit any detectable bond selectivity between its symmetrical C-I bonds 28 , the bond-selective reaction of Diagonal is attributed to its asymmetric adsorption geometry. This reactive bond selectivity will be examined below by the MD theory. The number of electrons involved in triggering a reactive event was determined experimentally for the major path by measuring the average reaction rate as a function of the tunnelling current in the range of 0.6–18 nA, at a constant sample bias of +1.3 V. As shown in Fig. 2a , the reaction rate scaled linearly with the tunnelling current, evidencing a one-electron process. This linear relationship also excludes an electric field effect as a major cause of reaction 30 . Figure 2: Evidence of electron-induced process. ( a ) Rate versus current plot (in log scale) for the electron-induced reaction of physisorbed mDIB on Cu(110). The plot is shown for C-I(AL) reaction (major path) with a linear fit (red line) of the data. The slope of the linear fit was determined to be 1.1±0.2. The good quality of the fit was evidenced by the coefficient of determination ( R 2 ) of 0.927. Reaction rates were measured at +1.3 V. The error bar is the standard error from the exponential fitting of the data at each current (see Methods for details). ( b ) Projected-density-of-states (pDOS) calculation for the physisorbed mDIB. The pDOS shows the LUMO of mDIB to be 0.8 eV above the Fermi level. The pDOS of the mDIB molecule is given by the black line; the dashed and dotted lines give the pDOS of the I atoms and the C atoms next to the I atoms. The inset in panel b visualizes the LUMO of mDIB (isocontour=0.0005 e Å −3 ). The nodal planes between the C and I atoms show the LUMO to be of σ*(C-I) antibonding character. Full size image As in previous studies of the electron-induced reaction of aryl iodides by ∼ 1 eV electrons on Cu(110) 26 , 27 , 28 , 31 , 32 , we ascribe the reaction to adsorbate electronic excitation. The computed projected density of states is shown in Fig. 2b . It exhibits a nodal plane between carbon and iodine atoms for the lowest unoccupied molecular orbital (LUMO), indicative of a σ* antibonding character with respect to the C-I bond. This C-I antibonding orbital was computed to be 0.8 eV above the Fermi level. The yield for electrons of ∼ 1.3 eV was measured as ∼ 10 −9 reactive events per electron for the major path. The high single-electron energy of at least 1.0 eV required to give reaction argues against vibrational excitation as the source of induced reaction. Theory The observed bond selectivity can be understood in terms of the different recoil directions of the AL and AC fragments following an electron-induced repulsive impulse in the LUMO of one or the other C-I bond, leading to different reaction paths across the anisotropic Cu(110) surface. As in previous work 25 , 26 , 27 , 28 , 31 , 32 we used the ‘Impulsive Two-State’ (I2S) model to simulate the MD of the electron-induced reaction. In this two-electronic state model, the MD was first followed for the 192-atom system on an approximate anionic repulsive potential-energy surface (PES) obtained by the transfer of an electron to the valence shell of a halogen atom for a period of femtoseconds (a time t *). Thereafter, the atoms were returned with their accumulated momenta to the ab initio DFT ground PES for MD over a period of picoseconds needed to reach the reacted final states. In the present case, the added charge comprises one electron at the I atom of C-I(AL) to favour breaking of that bond, or alternatively one electron at the I atom of C-I(AC) to favour the alternate bond breaking. Figure 3a shows the dynamics for an electron added at C-I(AL), using the minimum t * of 20 fs required for reaction. Figure 3a shows that the impulse, with momenta carried over to the ground PES, stretched C-I(AL) causing it to break at ∼ 150 fs concurrently with the formation of the C-Cu (i.e., IPh-Cu) and I-Cu bonds. The distance versus time plot of Fig. 3a′ in the panel below shows bond extension of C-I(AL) from its initial separation of 2.1 to a 3.0 Å, accompanied by the formation of C-Cu and I-Cu bonds at their equilibrium separations (all equilibrium separations being indicated by horizontal dotted lines). Figure 3b shows that for the same t * the placing of one electron in the I atom of the alternate C-I(AC) bond did not break that C-I bond. This failure to break C-I(AC), as evident from the distance versus time plot of Fig. 3b′ , is connected with the larger separation between the C atom of C-I(AC) from its nearest Cu atom neighbour (C-Cu is 2.3 Å in the AL case, 3.0 Å for AC). We show below that this difference in C-Cu separation is significant for the extent of C-Cu binding energy. Figure 3: Computed trajectories for the bond-selective reaction of mDIB using the system’s 192 atoms. The trajectories, calculated using the I2S model with t *=20 fs, give the time evolution of the system to its final state. ( a ) Trajectory for one electron placed at the I atom of C-I(AL); the green circle indicates the nearest Cu atom to the C atom of C-I(AL). ( b ) Trajectory for one electron placed at the I atom of C-I(AC). The yellow highlights indicate that the system is in the anionic state. In ( a′ ) and ( b′ ) below, the time evolution of the C-I, I-Cu and C-Cu separations are shown as black, blue and green lines, respectively, for each trajectory, AL and AC. The C-Cu bond separation is measured from the Cu atom nearest to the carbon of the C-I bond: C-I(AL) in ( a′ ) and C-I(AC) in ( b′ ). The horizontal dotted lines give the equilibrium separations of the atomic pairs, as indicated. I2S, Impulsive Two-State. Full size image In Fig. 4 , we superimpose the MD taken from a full calculation of the motion of the 192 atoms onto a two-dimensional cut through the potential-energy hypersurface that gives the dependence of the potential energy on the two C-I separations. In drawing this PES we have kept the angle between the two C-I bonds fixed at its initial 120°. This constant angle is a good approximation to the PES governing the initial reaction dynamics, since the motion calculated for the full hypersurface shows the angle between the two C-I bonds decreasing by only 10° in the first ∼ 200 fs, during which C-I(AL) extends by 1.0 Å. On this restricted PES, the energy barrier to dissociate C-I(AL) is shown to be 0.7 eV for that coordinate, while the barrier to dissociate C-I(AC) is markedly higher for the alternate bond extension (the crest of the barrier along C-I(AC) is at 1.5 eV, lying beyond the range of Fig. 4 ). Figure 4: Trajectories using the I2S model superimposed on a restricted ground PES. The restricted cut through the ground potential-energy hypersurface was obtained by varying only the two C-I bond distances at a fixed bond angle of 120°. The trajectory from the I2S model was based on a full molecular dynamics calculation using all 192 atoms in the system. ( a ) The trajectory for one electron added to the I atom of the AL bond, C-I(AL). ( b ) The trajectory for one electron added to the I atom of the AC bond, C-I(AC). The energy values shown are relative to the initial state (IS). The yellow arrow in the trajectories encompass the 20 fs duration of the repulsion on the anionic PES; the red line indicates the subsequent ∼ 3 ps motion on the ground PES. I2S, Impulsive Two-State. Full size image We attribute the lower barrier along C-I(AL) to the greater C-Cu binding, and hence greater stabilization along this coordinate than along C-I(AC) (see Fig. 3a′ which show the C-Cu bond forming AL, but not AC). The extent of stabilization due to C-Cu binding can be estimated from the Bader version of the bond-length versus bond-order ( n ) relationship for atomic pairs 33 , which gives n =0.5 for C-Cu at the transition state (C-I(AL)=3.0 Å) in Fig. 4a as compared with n =0.3 for extension of the other coordinate, C-I(AC), to 3.0 Å. For a computed ∼ 3.0 eV bond energy of C-Cu, this leads to ∼ 0.6 eV greater stabilization due to C-Cu binding in the AL coordinate than the AC coordinate. Accordingly, the lower barrier for the C-I(AL) (0.7 eV) than C-I(AC) (>1.2 eV) in Fig. 4 can be ascribed to stronger C-Cu binding and hence greater stabilization AL rather than AC. This finding shows that the binding of the products to the surface, late in the course of reaction, can influence reaction barriers with resultant bond selectivity. The trajectories shown in Fig. 4 are taken from the motion on the full potential, with the two C-I bond separations plotted here on the ‘fixed bond-angle’ PES whose energy contours are shown. Bond selectivity is caused by the impulsive motion on the anionic PES, which impels the system on the ground PES to stretch C-I(AL) giving surface reaction over a 0.7 eV barrier along r 1 , but does not stretch C-I(AC) sufficiently to surmount the higher barrier along r 2 . The low barrier for C-I(AL), as indicated above, is due to the proximity of an underlying Cu atom along the row direction, stabilizing the system through C-Cu bond formation. It should be noted that the bond selectivity being reported is specific to the impulsive dynamics (see Fig. 4 ) being a consequence of the directed recoil of the atoms liberated in electron-induced reaction. The C-I bond breaking could also in principle be obtained thermally, although this was not studied here. We have, however, calculated, ab initio , the different energy barriers to reaction for the individual breaking of the two C-I bonds by C-I extension along the alternative minimum-energy paths of the ground PES. The computed barrier heights relevant to thermal reaction were found to be 0.3 eV for the previously favoured AL C-I and 0.2 eV for the previously disfavoured AC C-I. Thermal reaction is therefore predicted to exhibit a modest bond selectivity but in the opposite sense to that observed and modelled here for electron-induced reaction. Discussion The high degree of bond selectivity in electron-induced surface reaction reported here is linked to directed recoil in relation to the anisotropy of the substrate, resulting in stronger bonding to the substrate when one bond is broken than the other. We anticipate similar bond-selective reaction for electron-induced processes on other anisotropic surfaces, the cause being the greater proximity of a substrate atom to one bond in an adsorbate than to another. This proximity leads in the present instance to 100 × greater reactivity for the adsorbate bond more closely adjacent to a substrate atom, despite the fact that the bonds being broken in the reaction are chemically identical. The proximity, in turn, is determined by the geometry of the physisorbed adsorbate relative to the substrate atoms. Directed recoil together with substrate anisotropy can therefore be expected to lead to bond selectivity in other electron-induced surface reactions, for example, in chemical reactions at surfaces being subjected to generalized electron radiation, in an extension of conventional heterogeneous catalysis. Methods Experiment The experiments were conducted at 4.6 K, using an ultrahigh vacuum Omicron low-temperature STM with a base pressure of <3.0 × 10 −11 mbar. The Cu(110) surface was prepared by cycles of Ar + sputtering (0.6 keV) and annealing (800 K). mDIB (98%; Sigma-Aldrich) was outgassed by 3–4 cycles of freeze pump thaw before being dosed via a capillary tube directed at the Cu surface. During dosing the substrate temperature rose to 12.6 K. The surface was imaged by STM using the constant-current mode. The bias reported was the sample bias. The reactions of mDIB were induced by tunnelling electrons from the STM tip. This is briefly described as follows: (i) after imaging an intact physisorbed mDIB molecule, the STM tip was placed over the molecule (white cross in Fig. 1 ); (ii) the tip height was adjusted according to a predefined set of sample bias ( V set ) and tunnelling current ( I set ); (iii) the feedback loop was deactivated, and a surface bias ( V reaction ) was applied for up to 10 s while the tunnelling current ( I reaction ) was recorded as a function of time. The reaction was identified by a single discontinuity of the current at time t . Subsequent imaging of the same area confirmed the occurrence of the electron-induced reaction products. Their locations were obtained relative to the reagent position, using the WSxM software 34 . The reaction order with respect to the current was established by examining the reaction rate as a function of the current ( I reaction ). To determine the reaction rate ( R ), we have followed the procedure first described by Ho and co-workers 12 and modified it for two competing pathways 28 . The times ( t , see above) measured from repeated experiments at a particular I reaction were binned according to the Doane’s formula 35 ; the resulting histogram being normalized and fitted with a single-parameter exponential function (e − Rt ). The value of R obtained from the exponential fitting was plotted for each I reaction as a log–log graph with an error bar that represents the standard error from the exponential fitting. The slope of the linear fit between R and I reaction in the log–log graph gives the reaction order with respect to the current. In the linear fitting, each point was weighted by its standard error. Theory Plane-wave-based DFT calculations were performed using Vienna Ab Initio Simulation Package (VASP 5.2.11) 36 , 37 installed at the SciNet supercomputer 38 . The parameterization followed the earlier work 26 , 27 , 28 , 31 , 32 of this laboratory, using the projected augmented wave, generalized-gradient approximation 39 , 40 , Perdew–Bruke–Ernzerhof (PBE) functional 41 and the second version of Grimme’s semiempirical correction for dispersion force corrections (PBE-D2) 42 . Calculations of the geometry relaxation, MDs and barrier height were restricted to the Γ-point at the centre of the surface Brillouin zone, using a cutoff energy of up to 450 eV with the dipole correction along the surface normal direction and no spin polarization. The Cu(110) surface was represented by a (6x6) slab that consisted of 180 Cu atoms in five layers with a vacuum gap of 16 Å. In the geometry relaxation calculations, all atoms were allowed to relax, except the bottom two layers of the Cu-slab, until the residual force on each atom was <0.01 eV Å −1 . Given that PBE-D2 overestimates the physisorption energy 43 , the energy was corrected by RPBE-D2 in a denser k-mesh (3x3x1) from the structure relaxed using PBE-D2. The STM simulation used the Tersoff-Hamann approximation and was visualized using the Hive software 44 , 45 . The molecular structures and the electron-charge densities were visualized using the VESTA software 46 . The Impulsive Two-State model 25 , 26 , 27 , 28 , 31 , 32 implemented in the MD calculations was applied to simulate the electron-induced reaction. To introduce the C-I repulsion caused by an electron occupying the LUMO of the intact mDIB, we used the anionic pseudopotential method 47 , 48 , which has been developed and tested in a number of cases in this laboratory for electron-induced reaction at both metallic 25 , 26 , 27 , 28 , 31 , 32 and semiconducting surfaces 49 . In this method, an electron was placed in an I atom of the molecule by exciting the 4 d core electron of the I atom to its 5 p valence shell. The MD calculations were run as a microcanonical ensemble using a time step of 0.5 fs. The reaction trajectory was obtained by evolving the system on the anionic PES for a brief residence time of t *, and afterwards, on the ground PES for up to ∼ 3 ps. The resulting product distribution was compared with the experimental one. The Climbing-Image Nudged Elastic Band technique 50 was used to compute the minimum-energy path and to locate the transition states for mDIB dissociation on Cu(110) on the ab initio ground PES. The Climbing-Image Nudged Elastic Band calculations used the initial and final states observed in the electron-induced experiment. The number of images used, including the initial and final states, was 10 for C-I(AL) and 13 for C-I(AC). The calculations were conducted until the forces orthogonal from the band were <0.02 eV Å −1 . Data availability The authors declare that the data supporting the findings of this study are available from the corresponding author on request. Additional information How to cite this article: Anggara, K. et al . Bond selectivity in electron-induced reaction due to directed recoil on an anisotropic substrate. Nat. Commun. 7, 13690 doi: 10.1038/ncomms13690 (2016). Publisher’s note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Scientists at the University of Toronto have found a way to make catalysis - the use of catalysts to facilitate chemical reactions - more selective, breaking one chemical bond 100 times faster than another. The findings are described in a study published in Nature Communications. The team of researchers, led by Nobel Prize-winning chemist John Polanyi, employed a combination of experiment and theory to discover that the position of the molecule on the catalytic surface is a key factor in determining the rate at which particular bonds break. "We found that the microscopic positioning of the molecule relative to the catalytic surface below, rendered the catalyst highly bond-specific," says Polanyi, University Professor in the Department of Chemistry at U of T. "The closer the alignment of the bonds of the molecule to the rows of atoms in the catalyst, the more selective was the reaction." The scientists investigated a chemical reaction that involved breaking carbon-to-iodine bonds in the organic molecule iodobenzene, by means of metallic copper, a common catalyst. The reaction was initiated by an electron coming from the tip of a microscope, which attached itself to the iodobenzene. "We observed acceleration in the reactivity of the carbon-to-iodine bonds when those bonds were aligned along rows of copper atoms in the catalyst, as compared to the bonds aligned across the copper rows," says Kelvin Anggara, a PhD candidate in Polanyi's research group and a lead author of the study. "The copper surface acted more strongly on bonds that were nearby than on bonds that were further away," Anggara says. "We saw 100-fold differences in reactivity between bonds pointing in specific directions on the catalyst." The experiment could be explained by a mathematical model developed by the researchers over the past few years, which enabled them to produce a computer-generated movie of the motions of the atoms involved in the bond-breaking at the copper surface. It was the movie that revealed the reason why the copper catalyzed the bonds along its rows in preference to bonds across the rows. "The copper atoms along the rows were slightly closer together, by about the diameter of a single atom, than the atoms across the rows," says Anggara. "This closer spacing promoted the breaking of bonds lying along the rows." The method is rooted in the study of chemical reactions taking place at the surface of solid materials that has guided Polanyi and his colleagues for decades. Following his receipt of the Nobel Prize in chemistry in 1986 for observing the molecular motions in chemical reactions occurring in gases, Polanyi began studying the reactions of individual molecules lying on well-defined catalytic surfaces. Polanyi says scientists are only beginning to understand how catalysis operates, and that the shift towards green chemistry makes knowing as much as possible about catalysts and how they reduce waste caused by chemical reactions more important than ever before. "The challenge for the future will be to fabricate metal catalysts embodying atomic patterns that speed chemical reactions along pathways that lead to desired products," said Polanyi. "Recent advances in the construction of surfaces, atom-by-atom, lend themselves to the fabrication of such engineered-catalysts. We're now a bit closer to that, since we begin to understand what patterns of atoms make the best catalysts." The findings are reported in the study "Bond selectivity in electron-induced reaction due to directed recoil on an anisotropic substrate." Support for the research was provided by the Natural Sciences and Engineering Research Council of Canada, U of T NSERC General Research Fund and the Connaught International Scholarship for Doctoral Studies. Computations were performed on the SciNet HPC Consortium supercomputer at U of T.

10.1038/ncomms13690

Other

Three extinct squirrel-like species discovery supports earlier origin of mammals in late Triassic

Three new Jurassic euharamiyidan species reinforce early divergence of mammals, Nature, dx.doi.org/10.1038/nature13718 Journal information: Nature

http://dx.doi.org/10.1038/nature13718

https://phys.org/news/2014-09-extinct-squirrel-like-species-discovery-earlier.html

Abstract The phylogeny of Allotheria, including Multituberculata and Haramiyida, remains unsolved and has generated contentious views on the origin and earliest evolution of mammals. Here we report three new species of a new clade, Euharamiyida, based on six well-preserved fossils from the Jurassic period of China. These fossils reveal many craniodental and postcranial features of euharamiyidans and clarify several ambiguous structures that are currently the topic of debate. Our phylogenetic analyses recognize Euharamiyida as the sister group of Multituberculata, and place Allotheria within the Mammalia. The phylogeny suggests that allotherian mammals evolved from a Late Triassic (approximately 208 million years ago) Haramiyavia -like ancestor and diversified into euharamiyidans and multituberculates with a cosmopolitan distribution, implying homologous acquisition of many craniodental and postcranial features in the two groups. Our findings also favour a Late Triassic origin of mammals in Laurasia and two independent detachment events of the middle ear bones during mammalian evolution. Main Mammalia Linnaeus, 1758 Allotheria Marsh, 1880 Euharamiyida (new clade) Shenshou lui gen. et sp. nov. Bi, Wang, Guan, Sheng and Meng Etymology. Shen , from pinyin of the Chinese word, meaning deity, divinity or cleaver; shou , from pinyin of the Chinese word for creature, animal or beast; specific name after Lu Jianhua, the collector of the holotype. Holotype. A nearly complete skeleton from an adult individual (LDNHMF2001, Lande Museum of Natural History, Tangshan, Hebei Province, China) (Figs 1a and 2d, and Extended Data Figs 1a and 2). Three specimens are referred to as the paratypes (see Supplementary Information , section A). Paratypes. Three specimens are referred to as the paratypes (see section A of Supplementary Information ). Locality and horizon. The Tiaojishan Formation, Daxishan site of Linglongta, Jianchang County, Liaoning Province, China; the locality was dated as approximately 160 million years ago (within the Oxfordian) (see Supplementary Information , section C, for age constraint). Diagnosis. Medium-sized euharamiyidan with an estimated body mass of 300 g ( Supplementary Information , section D). Dental formula I 1 -C 0 -P 2 -M 2 /I 1 -C 0 -P 1 -M 2 (I, incisor; C, canine; P, premolar; M, molar; superscript, upper teeth; subscript, lower teeth). The only pair of upper incisors (I 2 ) are in contact so that a facet is present on the medial side of each tooth; I 2 with two cusps. The mesial upper premolar (P 3 ) small (not basined) and the ultimate premolar (P 4 ) not significantly larger than upper molars. Upper molars with two main cusps in the buccal row, separated by a low ridge; three cusps in the lingual one, of which the penultimate cusp (B2) is the largest. P 4 sub-molariform with a main mesiolingual cusp (a1) and a long basined heel with two rows of cusps; M 1 with three cusps in each cusp row, M 2 with four cusps in the lingual and three in the labial row (Fig. 2d; see Supplementary Information , section A, for differential comparisons). The terminology we use to designate cusps in allotherian teeth is as follows: for upper teeth, the buccal (labial or lateral) cusp row is A, lingual row is B, cusps are numbered from the distal end; lower teeth, the lingual cusp row is a, the buccal row is b (lower case), numbering starts from the mesial (anterior) end. Eleutherodontidae Kermack et al. , 1998 Xianshou gen. nov. Wang, Meng, Bi, Guan and Sheng Etymology. Xian , from pinyin of the Chinese word meaning celestial being or immortal. Locality and horizon. Same as Shenshou . Diagnosis. Dental formula: I 2 -C 0 -P 2 -M 2 /I 1 -C 0 -P 1 -M 2 ; upper and lower molars ovoid in outline, with a shallow central basin. Differ from Sineleutherus in having three well-separated cusps of I 2 , non-molariform P 4 with a hypertrophic mesiolingual cusp (a1) and a weakly basined heel. Differ from Eleutherodon in the ovoid upper molars by absence of third row of cusps, and lack of cuspules and transverse ridges in the central basin. Differ from Shenshou and Arboroharamiya in having an extra-small I 1 , ovoid upper and lower molars, more distally positioned distobuccal cusp (A1) on P 4 and M 1 ; a hypertrophic mesiolingual cusp (a1) on lower molars (Fig. 2e, f; see Supplementary Information , section A, for differential comparisons, and Extended Data Figs 5 and 6 for additional figures). Xianshou linglong sp. nov. Wang, Meng, Bi, Guan and Sheng Etymology. linglong , from pinyin of the Chinese word, meaning ‘exquisite’, and also after the town name Linglongta where the specimen came from. Holotype. A skeleton preserved on a split slab of laminated siltstone (IVPP V16707A-B, Institute of Vertebrate Paleontology and Paleoanthropology, Chinese Academy of Sciences, Beijing, China; Figs 1b and 2e, and Extended Data Figs 1b and 5 ). Diagnosis. Estimated body mass of 83 g. I 2 with three cusps; I 3 extremely small and budlike. Upper molars with sharp cusps and stronger and sharper ridges (flutings) than those in Shenshou and X. songae ; M 1 with two main cusps at the mesial and distal end and connected by a low and sharp ridge in each cusp row; an enlarged and more distally extended distobuccal cusp (A1) on P 4 and M 1 ; P 4 enlarged with a hypertrophic mesiolingual cusp (a1) and a small talonid heel; cusp a1 of lower molar procumbent, projecting mesially beyond the crown. Xianshou songae sp. nov. Meng, Guan, Wang, Bi and Sheng Etymology. The specific name is after Rufeng Song, the collector of the holotype specimen. Holotype. A skeleton preserved partial skull, mandible, and most of the postcranial skeleton (BMNHC-PM003253, Beijing Natural History Museum, China) (Figs 1c and 2f, and Extended Data Fig. 6 ). Diagnosis. A small euharamiyidan with an estimated body mass of 40 g. Differ from X. linglong by having significantly smaller body size and in having P 4 more transversely oval, cusp A1 on P 4 and M 1 proportionally smaller and less distally extended. The lingual row of M 1 bears three cusps, of which the middle (B2) is the largest; lower molars anteroposteriorly short with a vertical hypertrophic mesiolingual cusp (a1) and low buccal cusps. Craniodental features The general skull shape of the new euharamiyidans is therian-like with a broad basicranial region and a tapered rostrum in dorsal and ventral views ( Extended Data Figs 3 and 4 ) but multituberculate-like in lateral view (Fig. 2a, b). A small septomaxilla is probably present ( Extended Data Fig. 2a ). The zygomatic arch is slender but fully developed, with the anterior root lateral to P 4 . As reflected by the skull shape, the cranial cavity is more inflated than those of eutriconodontans. The glenoid fossa is anterolateral to the bulging promontorium of the petrosal and anteroposteriorly oriented, and lacks the postglenoid process ( Extended Data Figs 4b and 5b ). The dentary is similar to that of Arboroharamiya 1 and multituberculates in having a distinct diastema between the incisor and premolar, a small coronoid process, a masseteric fossa extending anteriorly to the level of P 4 and a low-positioned mandibular condyle that is orientated ventrodorsally. A vestigial coronoid is present on the medial side of the jaw ( Extended Data Fig. 2b ), posteroventral to M 2 , similar to that of Arboroharamiya and the Jurassic multituberculate Kuehneodon 2 . A reduced coronoid, usually indicated by a scar on the dentary, is also present in some more advanced mammals, such as Amphitherium 3 and Zhangheotherium 4 . This is in contrast to sizable coronoids that cover the anterior portion of the postdentary unit in primitive mammaliaforms, such as Morganucodon 5 and Haldanodon 6 . Unlike multituberculates but similar to Arboroharamiya , the dentary has a small angular process that inflects medially. Dentitions are preserved in situ in the six specimens ( Figs 1 and 2 , and Extended Data Figs 1 , 2 , 3 , 4 , 5 , 6 , 7 ), confirming the tooth identification and orientation of Arboroharamiya 1 as well as reinterpretation 1 of some previously known euharamiyidans 7 , 8 , 9 , 10 . The cusp arrangement and morphology from Haramiyavia , Thomasia , Arboroharamiya and the new species reported here show consistently larger row A cusps (buccal) than row B cusps in upper molars, and larger row a cusps (lingual) than row b cusps in lower molars. In the lower cheek teeth, the mesiolingual cusp (a1) is the largest, whereas the distobuccal one (A1) is the largest in the upper teeth. The lingual lower cusp row bites into the central valley of the upper molars as indicated by wear facets in the upper and lower teeth ( Extended Data Fig. 7a–c ). These dental features and occlusal patterns should also be applicable to those known only from isolated teeth, such as Eleutherodon and Sineleutherus. It is also noteworthy that similar tooth cusp morphology and occlusal patterns are present in the upper and lower second molars of some Jurassic multituberculates 11 , in which the mesiolingual cusps are the tallest among cusps in the lower molar. The molars with two rows of multiple cusps and their wear patterns, coupled with the morphologies of the lower jaw and the glenoid fossa, convincingly indicate palinal jaw movement during mastication 12 . Figure 1: The holotypes of three euharamiyidan species. a , Holotype (LDNHMF2001) of Shenshou lui . b , Holotype (IVPP V16707A) of Xianshou linglong . c , Holotype specimen (BMNHC-PM003253) of Xianshou songae . See Extended Data Fig. 1 for interpretations of the skeletal features. PowerPoint slide Full size image Figure 2: Teeth, skull and mandibles of euharamiyidans. a , Reconstruction of the skull and jaws of Xianshou linglong (the small I 1 is on the medial side of the large I 2 ). b , The labial view of the lower jaw of Xianshou linglong . c , Lingual view of the lower jaw of Xianshou linglong . Part of the ascending ramus and coronoid bone have not been preserved and are reconstructed based on Shenshou lui . d , Right I 2 , M 1,2 , and left P 4 –M 2 of Shenshou lui in occlusal view. e , Left I 1 , I 2 , P 4 ,M 1 and P 4 –M 2 of Xianshou linglong in occlusal view. Dashed lines represent the alveolus of P 3 . f , Right P 3 –M 2 , and P 4 and M 1 of Xianshou songae in occlusal view. See text for description of the skull and mandible. Dental formula are I 1? -C 0 -P 2 -M 2 /I 1 -C 0 -P 2 -M 2 (the upper incisor is not preserved in this species). Shenshou and Xianshou have a multicusped I 2 and a greatly enlarged I 2 that is fully covered with enamel. There is no upper or lower canine. There are one lower and two upper premolars, and two upper and lower molars. Small enamel ridges (flutings) vary in size and density on cheek tooth cusps and basins. Upper premolars are broadly basined with main cusps located peripherally. Lower premolar has an enlarged a1, which is hypertrophic in Xianshou and Arboroharamiya 1 , but there is no serration. Cusp A1 (distobuccal) of the upper premolar and molars is the largest and extends distally. Cusp a1 (mesiolingual) of the lower molars is the largest cusp, extending mesially. The lingual cusps of lower molars bear wear on their lingual sides, indicating that they bit into the central valley of the upper molars in mastication. Molars have a fusiform (spindle-shaped) basin that is closed mesially by a ridge in the upper molars and distally in the lower molars. The ridge can be erased owing to wear so that the central valley becomes confluent in Shenshou . Scale bars: a – c , 10 mm; d – f , 2 mm. Cusp terminology follows ref. 12 . See photographs of teeth in Extended Data Figs 2 , 4b, c , 5 , 6 and 7a–c . PowerPoint slide Full size image Postcranial skeleton Although the vertebral column varies considerably in extant mammals, particularly those from the southern continents 13 , the general or primitive cervical and thoracolumbar counts for therian mammals are conservative 14 . The axial skeleton of the new euharamiyidans is therian-like in possessing 7 cervical and 19 thoracolumbar vertebrae (13 thoracic and 6 lumbar; Fig. 3 ). In contrast to the docodontan Castorocauda 15 , the eutriconodont Liaoconodon and Yanoconodon 16 , 17 , and the symmetrodont Akidolestes 18 , which all have lumbar ribs, euharamiyidans (including Arboroharamiya ) have true lumbar vertebrae that lack ribs but have obliquely oriented facets of the prezygapophysis, elaboration of the transverse process, and the laminapophysis (a mammalian neomorph that split into the anterior metapophysis and posterior anapophysis) 19 ( Fig. 3b ). These features define a distinct thoracolumbar boundary ( Fig. 3a ) and, along with fixation of cervical vertebrae 20 and well-developed ribcage, are indicative of development of the diaphragm, a unique feature that allows mammals to progressively increase ventilation in adapting to fast movement 19 , 20 . Figure 3: Skeletal features of euharamiyidans. a , Reconstruction of the skeleton of Shenshou lui , based in part on Xianshou linglong and Xianshou songae . b , The vertebral column of Shenshou lui . c , The anterior section of the rib cage in Shenshou lui . d , The right scapular of Shenshou lui . e , The distal end of the humerus in Shenshou lui . f , Reconstruction of the hand of Shenshou lui , based on LDNHMF2001 and WGMV-001. g , Reconstruction of the foot of euharamiyidans, based on Shenshou lui (JZT-D061) and Xianshou songae . The cervical ribs, present in monotremes and many Mesozoic mammals 4 , 17 , 20 , 24 , are absent. Similar to therians, such as Eomaia 32 and Sinodelphys 27 , the first ten thoracic ribs are connected to the ossified manubrium and sterna via costal cartilages and the last three ribs float freely. The capitulum of each rib articulates between two thoracic centra at a ventral position. The last thoracic rib (13th) is identified as the diaphragmatic vertebra because the articular facets of its zygapophyses with the first lumbar are more vertical than horizontal, as in many extant mammals 49 . The anticlinal vertebrate is located at lumbar four. The tail consists of 18–22 caudal vertebrae, with the proximal ones bearing expanded transverse processes as in extant prehensile therians 50 . The scapular blade is roughly rectangular and has a large and deep infraspinous fossa but lacks a supraspinous fossa. The clavicle is strap-shaped and slightly curved anteriorly. Its rough proximal end indicates a flexible contact with the depression on the clover-leaf-shaped interclavicle. The distal end of the humerus has a bulbous radial condyle, a relatively smaller spherical ulna condyle, and a prominent entepicondyle. The femur has a hemispherical head with an extensive articular surface but a short neck, the greater trochanter is triangular and does not extend above the head. The distal condyles are small with a shallow patellar groove. The ankles are comparable to those of multituberculates, such as Kryptobaatar 24 , in that the calcaneus has a short and mediolaterally compressed tuber calcanei and are partly superposed by the astragalus. The cuboid is positioned obliquely and has little contact with the calcaneus so that metatarsal V is separated from the calcaneus. The entocuneiform is elongated, and its joint with metatarsal I is offset anteriorly from the joint of the intermediate cuneiform and metatarsal II, similar to that of Asioryctes and Eomaia 32 . Metapodials and proximal phalanges possess a well-developed palmar groove for the digital flexor muscle tendon. Metatarsals I and V are short and robust compared to others. The distal ends of proximal and intermediate phalanges are well trochleated. Terminal phalanges are compressed laterally and curved strongly with a sharp tip; each has a massive digital flexor tubercle on the ventral side and a small dorsally extended ridge. as, astragalus; ce, centrale; cm, calcaneum; ct, capitate; cu, cuboid; ec, ectocuneiform; en, entocuneiform; hm, hamate; L, lumbar; ltp, lumbar transverse process; lu, lunate; me, mesocuneiform; T, thoracic; td, trapezoid; tm, trapezium; sp, scaphoid; pi, pisiform; tq, triquetrum; ts, tarsal spur. The 10 mm scale applies to panel a and the 5 mm scale bar applies to all other postcranial elements. See photographs of skeletons and postcranial structures in Extended Data Figs 1 , 3 , 4a , 5a and 7d–g . PowerPoint slide Full size image The scapula ( Fig. 3d and Extended Data Fig. 7g ) is primitively similar to that of Megazostrodon , Haldanodon and monotremes 21 , 22 in that it lacks the supraspinous fossa, but differs from those taxa in that it has a reduced coracoid that is fused to the scapula and forms a small part of the glenoid fossa that faces ventrally. The clavicle ( Fig. 3c ) is similar to that of multituberculates 23 . As in multituberculates the humerus has a slight shaft torsion (approximately 15°) 23 , in contrast to a strong proximodistal torsion in cynodonts and other early mammals 4 , 24 . The distal end of the humerus is highly comparable to that of multituberculates 23 , 24 ( Fig. 3e and Extended Data Fig. 7e ). In the wrist, the hamate is hypertrophied and the scaphoid and triquetrum are enlarged, as in some extant arboreal marsupials 25 , 26 and the scansorial Sinodelphys 27 . Metacarpal V is offset from the hamate and the pisiform is sizable, similar to those of Zhangheotherium 4 ( Fig. 3g and Extended Data Fig. 8a ). The epipubic bone, a plesiomorphic feature common in several groups of Mesozoic mammals, including eutherians 28 , is absent. The pelvis is similar to that of therians in being shallow and differs from the deep pelvis of multituberculates and monotremes in having a reduced pubis and an ischium with a straight dorsal margin and an ischial tuberosity that is slender and less extended dorsoposteriorly 24 . The ilium is shorter than that of multituberculates. The femur has a short neck, similar to eutriconodonts 16 , 29 , 30 , symmetrodonts 4 , 18 , 31 and more primitive forms 21 , 22 , but different from multituberculates, in which the femoral head has a long neck and the greater trochanter projects beyond the head 24 . There is no parafibula, differing from Eomaia and Jeholodens 30 , 32 . The proximal end of the tibia is roughly symmetrical, contrasting to that of multituberculates 33 . The bony extratarsal spur is distinctive and displays different relative sizes in different species reported here. The spur or its os calcaris exists in monotremes 34 and is common in Mesozoic mammals and their close relatives 15 , 35 . As in Arboroharamiya and some extant arboreal didelphids and cheirogaleid primates 1 , the manus and pes are characterized by relatively short metapodials and long phalanges ( Fig. 3f, g and Extended Data Fig. 8 ). The limb features are collectively indicative of scansorial and/or arboreal adaptation 1 , 26 , 32 . Phylogenetic relationship A major unsolved problem in mammalian taxonomy and phylogeny concerns Allotheria (Multituberculata and Haramiyida) 9 , 12 , 33 , 36 , 37 , 38 , 39 , which affects how we view the early evolution of mammals. Most recent studies present contrasting hypotheses that either place allotherians in mammals 1 , indicating an explosive model for the origin of mammals in the Late Triassic (approximately 208 million years ago) and a long-fuse model for the origin of therians, or separate them from multituberculates and place them outside mammals 40 , suggesting an explosive model for the origin of the Mammalia in the Middle Jurassic epoch. In light of the new data reported here, we are able to revise characters used in previous studies 1 , 11 , 40 , discuss existing problems relating to Megaconus ( Supplementary Information , section E), the postdentary trough in ‘haramiyidans’ (now demonstrated to be a paraphyletic group) ( Supplementary Information , section F) and Hadrocodium ( Supplementary Information , section G) and conduct phylogenetic analyses including the new species ( Supplementary Information , sections H–J). The result shown in Fig. 4 (see also Extended Data Fig. 9 ) is consistent with most recent phylogenetic analyses 1 , 11 , 41 , 42 , which suggest that the ‘haramiyidans’ are related to multituberculates to form Allotheria within the Mammalia. This supports a Late Triassic origin of mammals 1 . This hypothesis gains support from some Late Triassic ‘symmetrodontans’, such as Kuehneotherium and Woutersia , that are considered taxonomically to be members of the trechnotherian mammals 33 , 43 , 44 . Our analyses also recognize a new clade, here named Euharamiyida, that pairs with Multituberculata; this is consistent with the view that previously discovered ‘haramiyidans’ seem to form a paraphyletic group, from which multituberculates were derived 12 , 36 , 39 . Primitive species traditionally placed in ‘haramiyidans’, such as Haramiyavia and Thomasia , form the stem members of allotherians. This resultant topology remains the same whether Megaconus was included ( Extended Data Fig. 10a ) or both Megaconus and Hadrocodium were excluded ( Extended Data Fig. 10b ) (see discussions in Supplementary Information , sections E and G). Figure 4: Phylogeny of mammals with focus on Allotheria. This simplified cladogram is based on the consensus tree ( Extended Data Fig. 9 ) computed from a data matrix with 113 taxa and 495 characters (see Supplementary Information ). The dashed line of ‘symmetrodontans’ indicates occurrences of taxa that existed in the Late Triassic, such as Kuehneotherium and Woutersia 33 , 43 , but are not included in our phylogenetic analyses because of their fragmentary preservation. Tree nodes represent the following clades: (1) Mammaliaformes; (2) Mammalia; (3) unnamed clade consisting of Eutriconodonta, Allotheria and Trechnotheria 33 ; (4) Allotheria; (5) Trechnotheria; and (6) Cladotheria. Mammaliaformes, Crown Mammalia, Allotheria, Multituberculata and Euharamiyida are also marked as nested colour blocks from the most to the least inclusive group. PowerPoint slide Full size image Character evolution of early mammals With the discoveries of the new euharamiyidans, it becomes increasingly evident that the cranial and postcranial features of euharamiyidans and multituberculates are similar to each other and to other mammals. However, the fundamental obstacle in interpreting their mammalian affinity remains the fact that the tooth pattern consists of two main rows of multiple cusps that are capable of longitudinal (palinal) chewing function in allotherians 12 , 39 , 45 . If allotherians were placed outside mammals, it is equally difficult to derive the allotherian tooth pattern from other mammaliaformes, such as tritylodontids. Our phylogenetic analyses ( Fig. 4 ) suggest that the primitive allotherian tooth pattern, as represented by Haramiyavia , was probably derived by developing an extra cusp row, or rows, from a triconodont-like tooth pattern or even from a tooth pattern with an initially reversed triangular cusp arrangement. It was noted that in lateral view the teeth of Haramiyavia are more similar to those of Sinoconodon and Morganucodon than to those of multituberculates 46 . We note, however, that the tooth pattern of Haramiyavia is also similar to that of Woutersia , which co-existed with Theroteinus , another Late Triassic haramiyidan 36 , 43 . The tooth morphology and occlusion of euharamiyidans indicate that, if the allotherian tooth pattern was derived from a triconodont tooth pattern, the secondary cusp row has to be added on the buccal side in the lower teeth. In the conventional view, however, development of extra cusps on the lingual cingula is common, but buccal cingula are rare on lower molars 45 , although exceptions exist, such as Hallautherium 47 . Nonetheless, the orientation of an isolated tooth in early mammals is not always certain, as demonstrated in the case of eleutherodontids 1 (this study). There is no convincing evidence to rule out the possibility that additional cusps could be added on the buccal side of the tooth in early mammals. Better material with teeth in situ from each taxon of interest, such as Woutersia , is needed to test this hypothesis. Interpretations of character evolution in early mammals depends on their phylogeny. If ‘haramiyidans’ were separated from multituberculates and placed outside mammals, while multituberculates fell within mammals 40 , then numerous similar craniodental and postcranial features, particularly the molar pattern with two cusp rows and bilateral occlusion, must have evolved independently in ‘haramiyidans’ and multituberculates during different periods of time. In addition, detachment of the postdentary bones from the dentary would have evolved at least four times independently in ‘haramiyidans’, multitubuculates, monotremes and therians. However, our phylogeny ( Fig. 4 and Extended Data Figs 9 and 10 ) indicates that Euharamiyida and Multituberculata were probably derived from a Haramiyavia- like common ancestor at a minimum oldest age (according to current fossil records; future finds may reveal an earlier ancestor) in the Late Triassic and diversified thereafter during the Jurassic epoch, with known euharamiyidans adapting to a scansorial and/or arboreal lifestyle which may explain their rare fossil record. In contrast to interpreting numerous parallelisms in ‘haramiyidans’ and multituberculates 40 , our hypothesis favours homologous acquisition of many similar craniodental and postcranial features in euharamiyidans and multituberculates, such as reduction of teeth, enlargement of the lower incisors, possessing only two molars in each side of the upper and lower jaws, and a palinal chewing motion. Moreover, euharamiyidans are similar to multituberculates in lacking the postdentary trough and Meckelian groove, indicating the presence of the definitive mammalian middle ear 3 . If the reinterpretation is correct—that the dentary of Haramiyavia has only the Meckelian groove (see Supplementary Information , section F)—then the clade containing Eutricondonta, Allotheria and Trechnotheria 33 ( Fig. 4 and Extended Data Fig. 9 ) would have evolved from a common ancestor that had a transitional mammalian middle ear 16 . This clade and the geological and geographic occurrences of its earliest known members are consistent with accumulating evidence from Gondwana landmasses that shows a cosmopolitan distribution of members in the clade 48 and suggest a Laurasian origin of mammals. Finally, by reinterpretating Hadrocodium as having postdentary bones (see Supplementary Information , section G), our phylogeny suggests that detachment of the postdentary bones evolved twice independently during the early evolution of mammals, once in the clade leading to monotremes and once towards the clade containing Eutricondonta, Allotheria and Trechnotheria.

Paleontologists have described three new small squirrel-like species that place a poorly understood Mesozoic group of animals firmly in the mammal family tree. The study, led by scientists at the American Museum of Natural History and the Chinese Academy of Sciences, supports the idea that mammals—an extremely diverse group that includes egg-laying monotremes such as the platypus, marsupials such as the opossum, and placentals like humans and whales—originated at least 208 million years ago in the late Triassic, much earlier than some previous research suggests. The study is published today in the journal Nature. "For decades, scientists have been debating whether the extinct group, called Haramiyida, belongs within or outside of Mammalia," said co-author Jin Meng, a curator in the Museum's Division of Paleontology. "Previously, everything we knew about these animals was based on fragmented jaws and isolated teeth. But the new specimens we discovered are extremely well preserved. And based on these fossils, we now have a good idea of what these animals really looked like, which confirms that they are, indeed, mammals." The three new species—Shenshou lui, Xianshou linglong, and Xianshou songae—are described from six nearly complete 160-million-year-old fossils found in China. The animals, which researchers have placed in a new group, or clade, called Euharamiyida, likely looked similar to small squirrels. They weighed between 1 and 10 ounces and had tails and feet that indicate that they were tree dwellers. "They were good climbers and probably spent more time than squirrels in trees," Meng said. "Their hands and feet were adapted for holding branches, but not good for running on the ground." The members of Euharamiyida likely ate insects, nuts, and fruit with their "strange" teeth, which have many cusps, or raised points, on the crowns. Mammals are thought to evolve from a common ancestor that had three cusps; human molars can have up to five. But the newly discovered species had two parallel rows of cusps on each molar, with up to seven cusps on each side. How this complex tooth pattern evolved in relation to those of other mammals has puzzled scientist for many decades. This reconstruction shows arboreal mammals in a Jurassic forest. The three animals on the left side represent the three new species of euharamiyidan mammals. The other two represent a gliding species and another euharamiyidan, respectively, that were reported earlier. Credit: Zhao Chuang Despite unusual tooth patterning, the overall morphology, or physical characteristics, seen in the new haramiyidan fossils is mammalian. For example, the specimens show evidence of a typical mammalian middle ear, the area just inside the eardrum that turns vibrations in the air into ripples in the ear's fluids. The middle ears of mammals are unique in that they have three bones, as evidenced in the new fossils. However, the placement of the new species within Mammalia poses another issue: Based on the age of the Euharamiyida species and their kin, the divergence of mammals from reptiles had to have happened much earlier than some research has estimated. Instead of originating in the middle Jurassic (between 176 and 161 million years ago), mammals likely first appeared in the late Triassic (between 235 and 201 million years ago). This finding corresponds with some studies that used DNA data. The holotype specimen of Senshou lui, which represents a new species of euharamiyidan mammal. It is a nearly complete skeleton that indicates a gracile body with a tail and long fingers that were adapted for an arboreal life in Jurassic forests. Credit: ©AMNH/J. Meng "What we're showing here is very convincing that these animals are mammals, and that we need to turn back the clock for mammal divergence," Meng said. "But even more importantly, these new fossils present a new suite of characters that might help us tell many more stories about ancient mammals."

dx.doi.org/10.1038/nature13718

Medicine

New cancer immunotherapy approach turns human cells into tiny anti-tumor drug factories

Gonzalo Almanza et al. Extracellular vesicles produced in B cells deliver tumor suppressor miR-335 to breast cancer cells disrupting oncogenic programming in vitro and in vivo, Scientific Reports (2018). DOI: 10.1038/s41598-018-35968-2 Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-018-35968-2

https://medicalxpress.com/news/2018-12-cancer-immunotherapy-approach-human-cells.html

Abstract The successful implementation of miRNA (miR) therapies in humans will ultimately rely on the use of vehicles with improved cellular delivery capability. Here we tested a new system that leverages extracellular vesicles (EVs) laden with a tumor suppressor miRNA (miR-335) produced in B cells by plasmid DNA induction (iEVs). We demonstrate that iEVs-335 efficiently and durably restored the endogenous miR-335 pool in human triple negative breast cancer cells, downregulated the expression of the miR-335 target gene SOX4 transcription factor, and markedly inhibited tumor growth in vivo . Remarkably, iEVs-335 mediated transcriptional effects that persisted in tumors after 60 days post orthotopic implantation. Genome-wide RNASeq analysis of cancer cells treated in vitro with iEVs-335 showed the regulation of a discrete number of genes only, without broad transcriptome perturbations. This new technology may be ideally suited for therapies aimed to restore tumor suppressor miRNAs in cancer cells, disrupting the oncogenic program established after escape from miRNA control. Introduction Micro-RNAs (miRNAs) are evolutionarily conserved 20–30 nucleotides that represent a large family of gene expression regulators through their ability to prevent translation of specific mRNA into protein 1 , 2 . Individual miRNAs may repress up to hundreds of transcripts 3 and can regulate diverse processes including cell growth, metabolism, immunity, inflammation, and cancer. miRNA mutations or mis-expression exist in human cancers suggesting that miRNAs can function either as tumor suppressors or oncogenes (oncomiRs) 4 , 5 . Consequently, selective miRNA restoration or oncomiR suppression represent new avenues to cancer therapy. miR-335 is implicated in the growth and metastasis of the triple negative breast cancer cell line MDA-MB-231 derivative 4175 (LM2) cell line 6 . Clinically, triple negative breast cancer patients whose primary tumors have low miR-335 expression have a shorter median time to metastatic relapse 6 . Reportedly, miR-335 inhibits tumor re-initiation but is then silenced by genetic and epigenetic mechanisms 7 . One of the targets of miR-335 is SOX4, a transcription factor involved in embryonic development and cell fate determination 8 , 9 , 10 and in epithelial to mesenchymal transition (EMT) 11 . SOX4 expression is elevated in various tumors, including lymphoma, colorectal, cervical, lung, pancreatic, and breast cancer (Human Protein Atlas portal: ). The deregulated expression of this developmental factor has been correlated with increased cancer cell proliferation, cell survival, inhibition of apoptosis, and induction of EMT 12 . Experiments in mice with conditional deletion of SOX4 in stratified epithelia showed resistance to chemical carcinogenesis leading to delayed onset and tumor size reduction 13 . Recently, we demonstrated that B cells can be reprogrammed for the enforced biogenesis and synchronous release of short noncoding (snc)RNAs 14 . sncRNAs were packaged and enriched as cargo in extracellular vesicles (EVs) induced in B cells (iEVs), with an estimate content of 3.6 copy number/iEV 15 . Here, we demonstrate that iEVs programmed to contain miR-335 cargo deliver and durably restore miR-335 to LM2 cells, modulate target mRNA expression in vitro and in vivo , and greatly reduce the growth of orthotopic LM2 tumors in immune deficient NSG mice. Interestingly, regulation was confined to a discrete number of genes, without broad transcriptome perturbations. Results A plasmid expressing miR-335 doublets in B cells At the outset, we reasoned that restoring miR-335 content in LM2 cells would be best achieved by transfecting B cells with a plasmid engineered with two miR-335 precursor stem loops 15 . The general approach and the generation of iEVs in B cells by transfection with plasmid DNA are shown in Fig. 1A . We engineered pCMVmir carrying two pre-miR-335 stem loops in tandem with a nucleotide linker (Fig. 1B ). Transfection experiments were performed in the murine myeloma cell line J558L to determine the efficiency of iEVs-335 from cells transfected with a pCMVmir coding for one or two pre-miR-335 stem loops, respectively. miR-335 relative quantification (RQ) in isolated iEVs-335 produced by J558L cells transfected with a single pre-miR-335 stem loop plasmid was modestly increased over control EVs from untransfected J558L cells. In contrast, miR-335 abundance in iEVs-335 produced by J558L cells carrying a pre-miR-335 doublet showed a nearly ~250 fold over that of cells transfected with the singlet (Fig. 1C ). Therefore, all subsequent experiments were performed using a pCMVmir carrying two pre-miR-335 stem loops. The iEVs were characterized as having an average size of 100 nm (Supplementary Fig. 1A ) and expressing CD63 and CD81 (Supplementary Fig. 1B). Negative staining by electron microscopy shows iEVs to be circular structures of ~ 100 nm diameter with a homogeneous cavity (Fig. 1D ). Because of these features iEVs have exosome like characteristics 16 . Figure 1 Experimental model and cartoon of dual miR-335 plasmid. ( A ) Schematic of experimental model involving the transfection of murine J558L B cells with pCMVmir.335, the production of induced extracellular vesicles (iEVs), and subsequent treatment on LM2 cells. ( B ) Schematic diagram of tandem hsa-mir335 stem (pre-miR) loops with an intervening spacer sequence. ( C ) Differential abundance of miR-335 in iEVs from J558L cells transfected with pCMVmir.335 containing a single or dual pre-miR-335 sequence, respectively. 10 6 J558L cells were transfected with 1 µg of plasmid DNA, and the supernatant was collected 48 hrs later. iEVs were isolated by precipitation, counted and analyzed (10 6 ) by RT-qPCR amplification using RT-specific primers for miR-335 and SnoRNA202 as a control. RQ (Relative Quantity). Results refer to the mean ± SD of a representative experiment out of three independent experiments. ( D ) Negative staining electron micrographs of iEVs-335. Magnification: Inset (6800x), Main frame 9300x. Full size image Effects of iEVs containing miR-335 on LM2 cells in vitro To establish the minimum threshold for effective miR-335 restoration in target LM2 cells, we quantified miR-335 content in LM2 cells incubated in vitro for 48 hrs with iEVs-335 over a range of iEVs:LM2 cell ratios (4 × 10 2 –10 4 iEVs: LM2 cell). The miR-335 copy number increased in a dose dependent manner, with a > 4 fold increase over untreated LM2 cells at the 4 × 10 3 dose (Fig. 2A ). Next, we measured the effect on two miR-335 targets, SOX4 and tenascin C ( TNC ) 6 . Restoration of miR-335 expression in LM2 cells was associated with a dose dependent reduction in SOX4 mRNA expression (Fig. 2B ). TNC expression reduction was less pronounced but also persisted. Two control mRNAs, CTNNB1 ( β-catenin) and hTERT , which are constitutively expressed in cancer cells, were unaffected, suggesting that the effect on SOX4 mediated by iEVs-335 was specific (Fig. 2C ). Collectively, we concluded that iEVs internalized into LM2 cells release their miR-335 cargo and effectively modulate their target mRNAs, particularly SOX4 . Treatment with iEVs-335 did not affect LM2 cell viability. LM2 cells were incubated with 4 × 10 4 iEVs-335:LM2 cells for 48 hours and subsequently cultured under standard culture conditions for an additional 8 days. Cell viability measured by 7-AAD staining did not change in a noticeable way relative to untreated and sham EVs-treated cells (Fig. 2D ), suggesting that neither the mere contact/internalization of iEVs nor the cargo content had per se an immediate effect on cell survival. Figure 2 Effects of iEVs-335 on LM2 cells. ( A ) Titration of iEVs-335 input/cell and restoration of endogenous miR-335 content in LM2 cells following co-culture for 48 hrs. Results are expressed as miR-335 copy number/LM2 cell, and refer to the mean ± SD of a representative experiment out of three independent experiments. ( B ) Target modulation of SOX4 and TNC in LM2 cells treated with increasing numbers of iEVs/cell. ( C ) Modulation of bystander genes CTNNB1 (β-catenin) and human (h) TERT . Samples were pre-amplified and then subject to RT-qPCR amplification using RT-specific primers. RQ (Relative Quantity). Results refer to the mean ± SD of a representative experiment out of three independent experiments. ( D ) Viability of LM2 cells treated with iEVs-335 (4 × 10 4 /cell) as compared with untreated (Unt) LM2 cells or LM2 cells treated with sham EVs (4 × 10 4 /cell), by 7-AAD exclusion staining by flow cytometry. Results refer to the mean ± SD of a representative experiment out of three independent experiments. Objects (the syringe, Petri dish and mouse) in panel A are from Openclipart.org ( ). Full size image Suppression of orthotopic tumor in vivo The ability of iEVs-335 to control LM2 tumorigenicity was tested in an orthotopic model by LM2 implantation in the mammary fat pad. Briefly, LM2 cells were incubated with 4 × 10 4 fold iEVs-335, or control EVs per LM2 cells, for 48 hrs to allow for their uptake/internalization, and the intracellular release of miR-335. NSG mice were then injected in the mammary fat pad with 4 × 10 5 LM2 cells. Mice were imaged on day 45 and 60, at which point they were sacrificed. Figure 3A outlines the experimental design. On day 45, 4 out of 6 control mice implanted with LM2 cells only, and 5 out of 5 mice implanted with LM2 cells pretreated with sham EVs, had tumors by bioluminescence (not shown). Only 4 out of 9 mice implanted with LM2 cells pretreated with iEVs-335 had tumor. On day 60, all control mice including those implanted with untreated LM2 cells alone and those implanted with LM2 cells pretreated with sham EVs, had large tumors. Upon macroscopic examination all mice had peritoneal invasion and in few instances bone or lymphatic invasion. Among the iEVs-335 group, 4 out of 9 mice had a tumor by in vivo imaging but the tumors were considerably smaller than those in mice implanted with LM2 cells treated with sham EVs (Fig. 3B ). Local invasion was found in 1 out of 4 tumor-bearing mice only. The average tumor size (mm 3 ) was 1,682 ± 250 in the 6 mice given LM2 cells alone and 1,896 ± 479 in sham EVs-treated LM2 cells, respectively. In contrast, the average size (mm) of the four tumors pretreated with iEVs-335 was 7.2 ± 9.8 (Fig. 3C ). Likewise, the average weight (g) was 1.3 ± 0.6 in the 6 mice given LM2 cells alone and 2.3 ± 1.2 for tumors from sham EVs treated LM2 cells. The average weight (gr) of tumors treated with iEVs-335 was 0.16 ± 0.18 (Fig. 3D ). Thus, pretreatment of LM2 cells with iEVs-335 dramatically impeded tumor growth in vivo . Figure 3 iEVs-335 treatment impedes orthotopic tumor growth in immune compromised mice. ( A ) Schematic representation of the experimental design. LM2 cells were treated by co-culture with iEVs-335 for 48 hrs prior to injection in the mammary fat pad of 10–12 week old NSG mice. Mice were given LM2 cells as one of three groups: untreated (N = 6), pretreated with sham EVs (N = 5), and pretreated with iEVs-335 (N = 9). ( B ) Day 60 bioluminescence images representative of orthotopic tumors formed by LM2 cells treated with either sham EVs (left) or iEVs-335 (right). At sacrifice, volume (mm 3 ) ( C ) and weight (g) ( D ) measured for all tumors in the three groups specified in ( A ). ( E – L ) RT-qPCR values (RQ) of endogenous miR-335 content ( E ), control miRNA Let-7a ( F ), SOX4 ( G ), TNC (H), CTNNB1 (I), and h TERT (L), in explanted tumors born out of untreated LM2 cells, LM2 cells pretreated with sham EVs, and LM2 cells pretreated with iEVs-335 (N = 4). A Grubb’s test was performed to exclude one tumor volume data from the LM2 untreated condition, which was a significant outlier from all other values (Z = 1.9395). Full size image Next, we measured the endogenous levels of miR-335 in the explanted tumors to see if differential tumor growth amongst groups was associated with higher miR-335 levels. The endogenous (RQ) miR-335 value was 1.0 ± 0.06 for the 6 control mice and 0.8 ± 0.02 for tumors pretreated with sham EVs, whereas it was 4.7 ± 0.7 in the four small tumors born out of LM2 cells pretreated with iEVs-335 (Fig. 3E ). Endogenous levels of Let-7a , a miRNA predicted to be unaffected by iEVs treatment, did not vary between treatment conditions (Fig. 3F ). To assess that higher miR-335 levels were functionally relevant, we quantified the mRNA levels of predicted endogenous targets miR-335, SOX4 and TNC , and found them to be considerably reduced compared to those in tumors from all control mice (Fig. 3G–H ). Oddly, tumors from sham EVs treated LM2 cells had higher TNC expression compared to untreated LM2 cells. No variation was noted in the mRNA levels of two control genes, CTNNB1 and hTERT (Fig. 3I–L ). Collectively, the data show effective restoration of the endogenous miR-335 content by iEVs-335 with surprisingly long-lasting effects on target mRNAs regulation in treated cancer cells. Durability of miR-335 restoration in LM2 cells Tumor growth suppression was associated with a high content of miR-335 and a concomitant reduction of target mRNAs 60 days after iEVs treatment. Because the average miRNA half-life is estimated to be ~5 days 17 , it became important to investigate the longevity of miR-335 restoration/target mRNA regulation by iEVs-335 in LM2 cells. To this end, cultured LM2 cells were treated for 48 hrs as follows: iEVs-335, scrambled miRNA iEVs, and sham EVs, respectively. At the end of a two-day treatment, cells were thoroughly washed and cultured in complete medium for an additional 8 days. Treatment with iEVs-335 resulted in a marked increase in miR-335 content with a peak on day 4. This did not occur in LM2 cells treated with either scrambled miRNA iEVs or sham EVs. Correspondingly, mRNA levels of SOX4 remained depressed through day 8 (Fig. 4A,B ). Significantly, in subsequent in vitro experiments, we found higher miR-335 levels through day 50 despite no further manipulations of the cells (Supplementary Fig. 2 ). To further explore this phenomenon, we performed two experiments. First, we looked at the potential transfer of precursor molecules from pCMVmir(335) transfected iEV-producing J558L cells by probing total RNA with primers specifically designed to anneal a region upstream of the multi-cloning site of pCVMmir(335). The result of this experiment shows no amplification in the iEVs (Supplementary Fig. 3 ), ruling out the potential carryover of precursor miR-335 molecules from J558L cells to iEVs. Second, we sought to determine whether mature miR-335 in recipient LM2 cells was due the exogenous miR-335 provided by iEVs or was instead contributed by a de novo synthesis of miR-335 induced by iEV treatment. To this end, LM2 cells were first treated with aurintricarboxylic acid (ATA), an inhibitor of de novo synthesis of miRNA 18 , for 24 or 48 hours. LM2 cells were then washed three times and cultured in fresh cDMEM for 4 or 8 days, with or without the addition of iEVs miR-335. The experimental design is depicted in Supplementary Fig. 4A . ATA inhibited the endogenous production of miR-335 at both 24 and 48 hr time points (Supplementary Fig. 4B,C - left panels). On day 4 and 8 post-ATA treatment, the endogenous production of miR-335 did not increase in ATA-treated LM2 cells that had not been subsequently treated with iEVs miR-335 (Supplementary Fig. 4B,C - right panels). In contrast, LM2 cells treated with iEVs miR-335 expressed mature miR-335 (Supplementary Fig. 4B,C - right panels). This argues strongly against the possibility that mature miR-335 in LM2 cells treated with iEVs may be due to the activation of the biogenesis of endogenous miR-335. Figure 4 ( A ) Expression of miR-335 and SOX4 in LM2 cells treated with iEVs-335, scramble iEVs or sham EVs for 2 days (D2 on) and subsequently on day 4 and 8 following removal of iEVs (D4 off and D8 off). ( B ) Whole transcriptome analysis of iEVs-335 treatment of LM2 cells. A heatmap showing relative expression of SOX4 and target genes in LM2 cells 4 days post-treatment (D4 off) with iEVs-335, iEVs-scrambled, sham EVs and untreated. Log2 gene expression values were converted to z-scores. Color in the heatmap indicates the mean z-score across replicates from the same condition. Genes that are significantly differentially expressed in iEV-335 treated LM2 cells relative to the pooled control conditions are indicated with*. ( C ) Relative quantification (RQ) of MMP1 , FSCN1 and TGFβ in LM2 cells treated with iEVs-335 4 days after removal of treatment. ( D ) TGFβ signaling via SMAD genes colored according to log 2 fold change in gene expression between miR-335 treated and control conditions. Significantly up- or down-regulated genes in iEVs-335 treated LM2 cells relative to control conditions are indicated by red text. Full size image Collectively, these findings suggest that the uptake of iEVs with a miR-335 payload leads to restoration of the endogenous miR-335 pool beyond the average half-life of miRNA. Since our findings do not substantiate de novo biogenesis of endogenous miR-335, the observed long-lasting effects may then be due to a combination of slow intracellular release and protection from degradation. RNASeq analysis of LM2 cells treated with iEVs-335 RNASeq interrogation was performed to assess effects on the transcriptome in LM2 cells after treatment with iEVs-335. Transcriptome profiles were compared from three replicates each of iEVs-335, iEVs scrambled, sham EVs and untreated LM2 cells. We performed unsupervised clustering of whole-transcriptome profiles and observed very high correlation across treatment conditions, with the highest correlation observed among samples treated with iEVs-335, suggesting small but consistent differences in gene expression relative to the control conditions (Supplementary Fig. 5 ). As predicted, SOX4 expression was down-regulated in LM2 cells treated with iEVs-335 relative to pooled control conditions (untreated, iEVs-scrambled and sham EVs) (Fig. 4B ; and Supplementary Fig. 6 ). However, SOX4 target genes as a whole were not significantly differentially expressed across conditions although two of the genes, MMP1 (Matrix metalloproteinase-1) and FSCN1 (Fascin Actin-Binding Protein 1), were significantly down-regulated in response to treatment with iEVs-335 (Fig. 4B , Supplementary Table 1 ). The down-regulation was confirmed by RT-qPCR (Fig. 4C ). Aside from SOX4 , only two other miR-335 target genes were significantly down regulated, FOXA2 and DKK1 (Supplementary Fig. 6 , Supplementary Table 2 ). Exploratory analysis of differentially expressed (DE) genes revealed overall 14% of expressed protein coding genes (1739 out of 12134) were affected by iEVs-335 treatment at a 5% false discovery rate, and only 5.7% (696) had a fold change of at least 1.5. DE genes were predominantly down-regulated by iEVs-335 treatment (Supplementary Fig. 7 ; Supplementary Table 3 ) and enrichment analysis implicated a small number of pathways associated with cellular signaling (TGFβ, TNFα and mTOR), energy metabolism (cholesterol homeostasis and hypoxia), and differentiation (myogenesis) (Supplementary Table 4 ). We also performed weighted gene co-expression network analysis 19 to identify networks of genes with highly correlated expression. Although WGCNA analysis detected two clusters of co-expressed genes, these clusters were only weakly correlated with miR-335 treatment (rho ~0.15; p < 0.65). The TGFβ pathway has previously been implicated in breast cancer stem cell maintenance and EMT transition 20 via transcriptional regulation of genes including transcription factor SOX4 , and transcriptional repressors ID1 21 and SNAI1 22 . All three of these genes were significantly down-regulated in iEVs-335 treated LM2 cells (Fig. 4D ), as was TGFβ1 itself. Thus, treatment with mir-335 may also indirectly down-regulate production of SOX4 via a reduction in TGFβ signaling. The combined down-regulation of TGFβ1 and downstream targets SOX4 , ID1 and SNAI1 suggests that the reduced ability of miR-335 treated LM2 cells to generate tumors in mice may be in part due to loss of TGFβ signaling. Discussion Here we show that iEVs generated in B cells and enriched in miR-335 can be used to restore miR-335 endogenous content in triple negative cancer cells that have a reduced content, enabling the regulation of SOX4, a transcription factor that is overexpressed in cancer cells. The internalization of iEVs-335 in LM2 cells was effective at down-regulating SOX4 mRNA in vitro and in vivo . We verified that these effects were not restricted to LM2 cells as other breast cancer cell lines, ovarian and pancreatic cell lines, were also similarly affected (Supplementary Fig. 8 ). Surprisingly, tumors born out of LM2 cells treated ex vivo with iEVs-335 showed a profound SOX4 downregulation sixty days after orthotopic implantation, suggesting that SOX4 negative regulation by a single treatment iEVs-335 is long-lasting. The mechanism for the durable effects of iEVs treatment is presently unknown. The fact that the viability of LM2 cells treated with iEVS-335 was consistently preserved up to day 8 in culture argues against a direct apoptotic or toxic effect. Rather, it suggests that reconstitution of miR-335 in target cells triggered durable gene regulation, such as that of genes involved in tumor growth and local invasion. Figure 4A showed a selective downregulation of SOX4 during the first week after treatment. RNASeq analysis on day 4 post-treatment (Fig. 4B ) revealed a selective perturbation of the SOX4 transcriptional network with a significant downregulation of two genes MMP1 and FSCN1 (Supplementary Table 1 ) even though other genes were also affected without reaching, however, the level of significance. MMP1, a collagenase that cleaves collagen Type I, II, III, VII and X, is overexpressed in a variety of cancers types including breast cancer and in circulating cancer cells 23 . Fascin1, a member of the cytoskeleton protein family, has been reported to regulate invasion of breast and pancreatic cancer cells by increasing cell motility 24 , 25 . Interestingly, neither the RNASeq analysis not a RT-qPCR could reveal a statistically relevant perturbation of TMEM2 , a transmembrane protein gene reported to be a SOX4 target (Supplementary Fig. 9 ) 26 . We also observed down-regulation of other hallmark cancer genes 27 in the miR-335 treated condition relative to control (Supplementary Table 5 ), in particular genes associated with cell cycle, apoptosis and phosphorylation (kinase signaling) hallmarks (p < 0.05). Furthermore, we generated a Volcano plot that illustrates that there is a bias for hallmark associated genes (in red) to have significant decreased expression in association with miR-335 treatment (Supplementary Fig. 10 ). Consistent with the idea of the activation of a regulatory cascade is a report showing that miR-335 may exert its tumor-controlling effects via upregulation of BRCA1 mRNA 28 . RNA sequencing analysis suggested that overall, treatment with iEVS-335 had a modest effect on gene expression. While expression profiles across samples were highly correlated (Supplementary Fig. 5 ), samples treated with iEVs-335 showed even higher correlation, consistent with a small number of differences in gene expression between treated and control samples. Only three miR-335 target genes were significantly down-regulated ( SOX4 , FOXA2 and DKK1 ) and downstream of SOX4 , only FSCN1 and MMP1 showed a significant decrease in expression. When pathway level enrichment for differentially expressed genes was assessed, only a small number of biological pathways were detected. These pathways included TGFβ, TNFα and mTOR signaling, pathways that regulate cell growth, proliferation, differentiation, apoptosis and epithelial to mesenchymal transition. Thus, a small perturbation in gene expression by miR-335 treatment was associated with specific down-regulation of key pathways required for tumorigenesis. Remarkably, a one-time internalization of iEVs-335 proved to be efficient to suppress and in some instances block, orthotopic tumor formation, confirming the tumor-suppressor function of miR-335. A possible interpretation is that once internalized into LM2 cells, iEVs-335 degrade slowly, releasing their payload over time, leading to selected gene perturbations that, as demonstrated herein, relate to genes of the primary mRNA target transcriptional network, as well as other genes with participatory role such as TGFβ, which drives malignant progression, invasiveness and dissemination 29 . In addition to their slow release and intracellular degradation, iEVs proved to be superior to equimolar concentrations of soluble miR-335 mimics in their ability to downregulate the SOX4 mRNA target (Supplementary Fig. 11 ). Expression of miR-335 is reduced in various cancer types in humans besides triple negative breast cancer 30 , 31 , 32 , 33 , 34 , 35 , and low miR-335 expression levels have been associated with reduced recurrence-free survival, representing an independent indicator of poor overall survival 7 , 32 . As demonstrated here, iEVs-335 could lend themselves as a new form of therapeutic intervention in cancers in which genomic interrogation documents a decrease of tumor suppressor miR-335 and/or an increase of SOX4 . Optimization of iEVs as vehicles of miRNA therapies in general will further require the addition of simple, cost-effective modalities for precision tissue targeting in vivo upon systemic administration to enhance therapeutic efficacy while reducing off-target effects. Material and Methods Mice 8–10 week old NOD scid gamma (NSG) mice were purchased from The Jackson Laboratories. Cell lines and chemicals MDA MB 231–4175 (LM2) cells are human TNBC cells derivative of MDA-MB 231 cells stably transduced with a lentivirus expressing a triple-fusion reporter (abbreviated “TGL”) encoding herpes simplex virus thymidine kinase 1, green florescence protein and firefly luciferase 36 . LM2 cells were kindly obtained from the Cell Repository of the Memorial Sloan-Kettering Cancer Center (New York, NY). MDA-MB 231, SKBr3, SKOV3, and PANC1 cells were purchased from the American Type Cell Collection. Aurintricarboxylic acid (ATA) was purchased from Sigma-Aldrich. Plasmid Constructs A dual miRNA construct containing miR-335-miR-335 was synthesized with unique SgfI/XhoI ends by Integrated DNA Technologies (IDT, Coralville, IA). A single miRNA (miR-335) scramble (uuguauuauuuuuaacauaugaaugaauua) construct was similarly synthesized with unique SgfI/XhoI ends. Constructs were cloned into the pCMVmir (Origene, Rockville, MD) expression vector by digesting with SgfI and XhoI, and subsequent ligation of the insert into the pCMVmir vector. The ligation mixture was transformed into TOP10 competent cells (Life Technologies, Carlsbad CA). Transformed cells were plated, and clones were selected and grown overnight at 37 °C. DNA was extracted with Promega Wizard Plus SV Minipreps DNA Purification System (Promega, Madison WI). The resulting plasmid was termed pCMV dual mir335. The insert was verified via sequencing. The plasmid was stored at −20 °C until transfection. Single miRNA construct containing miR-335 was generated through excision from the dual miRNA construct by digestion and ligation using unique restriction sites (SgfI-MluI or NotI-XhoI) within the minigene to yield pCMV miR-335. The correctness of each plasmid construct was verified by sequencing. Cell Culture and Transfection J558L mouse B cell myeloma cells were grown in suspension in cRPMI with 10% fetal bovine serum (FBS). Cells were grown to 80% confluence. 2 × 10 6 cells were transfected with 1 μg of pCMVmiR plasmid utilizing the Lonza VACA-1003 transfection kit V in a Nuclefector 2b device (Lonza, Walkersville, MD). Cells were allowed to recover in a T25 flask upright at 37 °C with 5% CO 2 for 48 hrs. In experiments in which sncRNA copy number was determined, transfected J558L cells were cultured in EXO-FBS-50A-1 exosome-depleted FBS (Exo-FBS, Systems Biosciences, Mountain View, CA). EV Isolation and enumeration Forty-eight hours post-transfection 1 mL of culture supernatant was collected and incubated with 0.5 mL of Total Exosome Isolation solution (Life Technologies, Carlsbad, CA) for 1 hr at room temperature. The EV-containing mixture was spun at 16,000 RPM at 4 °C for 1 hr. The EV pellet was resuspended in 100 µL of PBS at room temperature and stored in 1.5 mL Eppendorf tubes at −20 °C until use. EVs isolated from untransfected or sham transfected (electroporated only) J558L cells served as a control. The number of vesicles recovered was determined by Nanoparticle Tracking Analysis (NTA) on a NanoSight LM-10HS equipped with a 405 nm laser (NanoSight, Wiltshire, UK) calibrated with polystyrene latex microbeads at 100 nm and 200 nm prior to analysis. Resuspended vesicles were diluted 1:100-1:300 with PBS to yield 20–100 objects per frame. iEVs were manually injected into the sample chamber at room temperature. Each sample was measured in triplicate at camera setting 14 with acquisition time of 30 seconds and detection threshold setting of 7. At least 200 completed tracks were analyzed per video. The NTA analytical software version 2.3 was used for capturing and analyzing the data. Negative staining electron microscopy A morphological analysis of iEVs was performed by standard negative stain method. Briefly, Formvar-carbon-coated copper grids (100 mesh, Electron Microscopy Sciences, Hatfield, PA) were placed on 20 μl drops of each sample solution displayed on a Parafilm sheet. After allowing material to adhere to the grids for 10 minutes, grids were washed x3 by rising through 200 μl drops of milli-Q water before being left for 1 min on 2% (wt/vol) uranyl acetate (Ladd Research Industries, Williston, VT). Excess solution was removed with Whatman 3MM blotting paper, and grids were left to dry for a few minutes before viewing. Grids were examined using a Tecnai Spirit G2 BioTwin transmission electron microscope operating at 80 kV. Images were recorded using an Eagle 4 K digital camera. Western blot analysis 1 × 10 6 EVs for each condition shown in Supplementary Fig. 1B were processed for Western blot analysis as follows. iEVs were counted by NTA and lysed at 70 °C for 5 minutes in 5 × loading dye buffer in a total volume of 20 µL. 15 µL samples were loaded into a 4–20% mini protean TGX gel (Bio-Rad). Precision Plus Protein Standard was used as a marker for weight (Bio-Rad). Gel was ran at 100 volts for 75 minutes and then transferred to 0.2 µm PVDF membrane using a Trans Blot Turbo device (Bio-Rad). Wester blotting protocol from Cell Signal Inc. was followed for antibody detection. Primary antibodies rat anti-mouse CD63 monoclonal antibody (Cat#564222, BD Biosciences) and hamster anti-mouse CD81 monoclonal antibody (Cat#559519) were both used at 1:1,000 dilution. Horseradish peroxidase-labeled rabbit antibodies to rat/hamster IgG were used at 1:2,500 dilution. Bound antibodies were revealed with Clarity Western ECL substrate (Bio-Rad). Blot was exposed for 10 minutes with Blue Devil autoradiography film (Genesee Scientific). RNA Extraction and Copy Number Determination 1 × 10 6 transfected or untransfected J558L cells, and 1 mL of culture supernatant, were collected for RNA extraction using ZYGEM RNAtissue Plus System (Zygem, Hamilton, NZ) according to the manufacturer’s protocol. RNA from cell supernatant (200 µL) was extracted with the Qiagen miRNeasy Serum/Plasma kit following the manufacturer’s protocol. EVs extraction was performed using the ZYGEM RNAtissue Plus System. cDNA was generated from intracellular and iEV miRNA with Taqman small RNA assays. Input RNA was normalized to 100 ng/sample for intracellular and exosome RNA, and to 25 ng/sample for extracellular miRNA. Taqman MicroRNA Reverse Transcription Kit was utilized for all samples per manufacturer’s instructions. Cycling conditions for qPCR were: 40 cycles, 96 °C denature 30 secs, 60 °C anneal/extension 30 secs. Results are expressed as RQ (Relative quantity of sample) that was calculated using the formula: Relative Quantity target = E target (Cq (control) − Cq (treatment)). Abbreviations: E = Efficiency of primer set; C q (control) = Average C q for the control or untreated sample; C q (treatment) = Average C q for treated sample; Target = The gene of interest or reference gene. Copy number was determined in samples normalized at 100 ng cDNA/reaction run concomitantly with a standard curve constructed with known amounts (100–0.01 ng) of miR-335 cDNA and an endogenous control standard curve constructed using known amounts (100–0.01 ng) of snoRNA202 cDNA (Applied Biosystems snoRNA202 – assay No. 001232 - specific reverse transcription primers). Samples were run in duplicate. Relative expression was determined by the Ct value of test samples vs. the endogenous control. Once the amount (ng) of specific target was determined, the copy number present in each reaction was calculated using the following formula: (ng × 6.0223 × 10 23 )/(number of nucleotides × 1.0 × 10 9 × 650) as indicated in . Copy number/EV determination was calculated as follows: [Total copy number/No. EVs sample]. Treatment of LM2 cells with iEVs and in vivo studies LM2 cells were plated at 1 × 10 6 and treated with iEVs at 4 × 10 4 iEVs:LM2 cell ratio for 48 hrs. After treatment cells were washed x3, and resuspended in PBS until implanted (4 × 10 5 ) into the right mammary fat pad in 50 μl. Mice were monitored for tumor take by palpation. When tumors became palpable, tumor size was determined through two-dimensional caliper measurements every three days. On day 30 and prior to sacrifice on day 60 mice received 6 mg of D-luciferin in PBS i.p., allowed to rest for 6 minutes, and imaged in a Xenogen IVIS system. At sacrifice tumors were resected, weighed and measured by caliper. Tumor volume was calculated using the ellipsoid formula: V = ½ (H × W 2 ). All animal work was approved by the UCSD Institutional Animal Use and Care Committee. RNA sequencing Triplicate RNA samples from LM2 cells corresponding to four different conditions as shown in Fig. 4A,B were sequenced to a targeted 25 million reads per sample. Sailfish v0.7.4 was used to estimate transcript abundance as transcripts per million (TPM) from single end 75 bp sequencing reads. Transcript levels were log2 transformed, and any genes with mean TPM <1.0 were excluded from further analysis. Samples were clustered by transcriptome profile using agglomerative hierarchical clustering with Euclidean distance and single linkage, and Pearson correlation was calculated between pairs of samples. Differential expression was determined by comparison of the iEVs-335 treated condition to the pooled control conditions. Individual genes were assessed for differential expression using a t -test and multiple testing correction was performed using the Benjamini-Hochberg method. Twenty-six SOX4 target genes identified by anti-SOX4 anti-body were obtained from Lee et al . 26 . Overall enrichment of the 26 gene set for differentially expressed genes was evaluated using Fisher’s exact test. Mir-335 target genes were obtained from mirTarBase 37 , and only targets supported by strong experimental evidence were retained (Supplementary Table 2 ). We preformed exploratory gene set enrichment analysis (GSEA) 38 using the stand-alone software (v3.0) to identify other biological activities perturbed by miR-335. GSEA was run with default parameters on three MSigDB collections: c2 (curated gene sets), c6 (oncogenic signature) and h (hallmark gene sets). A false discovery rate of 0.05 was used as the cutoff to consider a gene set as enriched for differentially expressed genes. No gene sets in the c2 or c6 groups met the significance criteria. WGCNA was run on whole transcriptome profiles with default parameters using the WGCNA R package v1.61 19 . Precursor miR-335 analysis 1 × 10 5 LM2 cells (treated with iEV mir335) and J558L transfected with pCMVmir335 were extracted with Qiagen DNA blood mini kit. 1 × 10 6 miR-335-laden iEVs were also processed with a Qiagen DNA blood mini kit. 50 ng of nucleic acid or volumetric equivalent were used in each 20 µL reaction and allow to anneal with primers pCMV ins chk F 5′- TTGTAATACGACTCACTATAGG −3′ and pCMV ins chk R 5′-GGATCTGTTCAGGAAACAGC-3′. Thermocycling parameters were as follows: 96 °C 10 min, 30 cycles of 96 °C denaturing, 53.7 °C annealing and 72 °C extension.

Cancer immunotherapy—efforts to better arm a patient's own immune system to attack tumors—has shown great potential for treating some cancers. Yet immunotherapy doesn't work for everyone, and some types of treatment can cause serious side effects. In a new approach, researchers at University of California San Diego School of Medicine are turning B cells, best known for producing antibodies, into factories that assemble and secrete vesicles or sacs containing microRNAs. Once internalized by cancer cells, these small pieces of genetic material dampen a gene that spurs tumor growth. In mice, breast tumors treated with this approach were fewer and significantly smaller than in untreated tumors. The study is published in the December 4 issue of Scientific Reports. "Once further developed, we envision this method could be used in situations where other forms of immunotherapy don't work," said senior author Maurizio Zanetti, MD, professor of medicine at UC San Diego School of Medicine and head of the Laboratory of Immunology at UC San Diego Moores Cancer Center. "The advantages are that this type of treatment is localized, meaning potentially fewer side effects. It's long-lasting, so a patient might not need frequent injections or infusions. And it would likely work against a number of different tumor types, including breast cancer, ovarian cancer, gastric cancer, pancreatic cancer and hepatocellular carcinoma." MicroRNAs don't encode proteins. Instead, microRNAs bind messenger RNAs that do encode proteins, inhibiting their translation or hastening their degradation. Normal cells use microRNAs to help fine-tune which genes are dialed up or down at different times. MicroRNAs tend to be less active in cancer cells, which can allow growth-related proteins to run wild. In this study, Zanetti and team used miR-335, a microRNA that specifically dampens SOX4, a transcription factor that promotes tumor growth. They added a miR-335 precursor to B cells in the lab. Once inside, through a naturally occurring process, the cells convert the precursor into mature, active miR-335 and package it into vesicles, small, membrane-coated sacs that bud off from the cell. Each B cell can produce 100,000 miR-335-containing vesicles per day—enough to treat 10 cancer cells. To test this new system, the researchers treated human breast cancer cells with miR-335-containing vesicles or sham vesicles in the lab. Then they transplanted the cancer cells to mice. After 60 days, 100 percent (5/5) of the mice with mock-treated cancer cells had large tumors. In contrast, 44 percent (4/9) of the mice with miR-335 vesicle-treated cancer cells had tumors. On average, the tumors in the treated mice were more than 260 times smaller than those in the mock-treated mice (7.2 vs. 1,896 mm3). And the treatment was long-lasting—miR-335 levels were still elevated in the treated mice 60 days after the vesicles and cancer cells were transplanted. "We were surprised to find that even small changes in cancer cell gene expression after miR-335 treatment were associated with specific down-regulation of molecules key to tumor growth," said study co-author Hannah Carter, Ph.D., assistant professor of medicine at UC San Diego School of Medicine. Other research groups and pharmaceutical companies are using tumor suppressor microRNAs therapeutically. What's new here, said researchers, is the method for producing and delivering them. According to Zanetti, this therapy could be developed in two ways. First, by first harvesting vesicles from B cells in a lab, then administering only the vesicles, as they did here, or second, by administering the B cells themselves. He says the challenge now will be to develop ways to ensure the B cells or vesicles get as close to a tumor as possible. This would be easier in some types of cancer, where the tumor is readily accessible by injection. But many cancers are difficult to access. Zanetti and colleagues are currently working to improve the delivery system, maximize efficiency and diminish side effects. "Ideally, in the future we could test patients to see if they carry a deficiency in miR-335 and have an overabundance of SOX4," Zanetti said. "Then we'd treat only those patients, cases where we know the treatment would most likely work. That's what we call personalized, or precision, medicine. We could also apply this technique to other microRNAs with other targets in cancer cells and in other cell types that surround and enable tumors."

10.1038/s41598-018-35968-2

Medicine

Genetic study provides most comprehensive map of risk to date of breast cancer risk

Fine-mapping of 150 breast cancer risk regions identifies 191 likely target genes, Nature Genetics (2020). DOI: 10.1038/s41588-019-0537-1 Journal information: Nature Genetics

http://dx.doi.org/10.1038/s41588-019-0537-1

https://medicalxpress.com/news/2020-01-genetic-comprehensive-date-breast-cancer.html

Abstract Genome-wide association studies have identified breast cancer risk variants in over 150 genomic regions, but the mechanisms underlying risk remain largely unknown. These regions were explored by combining association analysis with in silico genomic feature annotations. We defined 205 independent risk-associated signals with the set of credible causal variants in each one. In parallel, we used a Bayesian approach (PAINTOR) that combines genetic association, linkage disequilibrium and enriched genomic features to determine variants with high posterior probabilities of being causal. Potentially causal variants were significantly over-represented in active gene regulatory regions and transcription factor binding sites. We applied our INQUSIT pipeline for prioritizing genes as targets of those potentially causal variants, using gene expression (expression quantitative trait loci), chromatin interaction and functional annotations. Known cancer drivers, transcription factors and genes in the developmental, apoptosis, immune system and DNA integrity checkpoint gene ontology pathways were over-represented among the highest-confidence target genes. Main Genome-wide association studies (GWASs) have identified genetic variants associated with breast cancer risk in more than 150 genomic regions 1 , 2 . However, the variants and genes driving these associations are mostly unknown, with fewer than 20 regions studied in detail 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 . Here, we aimed to fine-map all known breast cancer susceptibility regions using dense genotype data on >217,000 subjects participating in the Breast Cancer Association Consortium (BCAC) and the Consortium of Investigators of Modifiers of BRCA1/2 (CIMBA). All samples were genotyped using the OncoArray 1 , 2 , 21 or the iCOGS chip 22 , 23 . Stepwise multinomial logistic regression was used to identify independent association signals in each region and to define credible causal variants (CCVs) within each signal. We found genomic features significantly overlapping the CCVs. We then used a Bayesian approach, integrating genomic features and genetic associations, to refine the set of likely causal variants and calculate their posterior probabilities. Finally, we integrated genetic and in silico epigenetic expression and chromatin conformation data to infer the likely target genes of each signal. Results Most breast cancer genomic regions contain multiple independent risk-associated signals We included 109,900 cases of breast cancer and 88,937 controls, all of European ancestry, from 75 studies in the BCAC. Genotypes (directly observed or imputed) were available for 639,118 single nucleotide polymorphisms (SNPs), deletions/insertions and copy number variants (CNVs) with a minor allele frequency (MAF) ≥ 0.1% within 152 previously defined, risk-associated regions (Supplementary Table 1 and Fig. 1 ). Multivariate logistic regression confirmed associations for 150 out of 152 regions at a significance threshold of P < 10 −4 (Supplementary Table 2a ). To determine the number of independent risk signals within each region, we applied stepwise multinomial logistic regression, deriving the association of each variant, conditional on the more significant ones, in order of statistical significance. Finally, we defined CCVs in each signal as variants with conditional P values within two orders of magnitude of the index variant 24 . We classified the evidence for each independent signal, and its CCVs, as either strong (conditional P < 10 −6 ) or moderate (10 −6 < conditional P < 10 −4 ). Fig. 1: Flowchart summarizing the study design. Logistic regression summary statistics were used to select the final set of variants to run stepwise multinomial regression. These results were meta-analyzed with CIMBA to provide the final set of strong independent signals and their CCVs. Through case-only analysis, we identified significant differences in effect sizes between ER-positive and ER-negative breast cancer and used this to classify the phenotype for each independent signal. With these strong CCVs, we ran the bio-features enrichment analysis, which identified the features to be included in the PAINTOR models, together with the OncoArray logistic regression summary statistics and the OncoArray linkage disequilibrium. Both multinomial regression CCVs and PAINTOR high-posterior-probability (PP) variants were analyzed with INQUISIT to determine high-confidence target genes. Finally, we used the set of high-confidence target genes to identify enriched pathways. iCOGS and OncoArray Cox regression was conditional on the index variants from BCAC strong signals. Full size image From the 150 genomic regions, we identified 352 independent risk signals containing 13,367 CCVs, 7,394 of which were within the 196 strong-evidence signals across 129 regions (Fig. 2a,b ). The number of signals per region ranged from 1 to 11, with 79 (53%) containing multiple signals. We noted a wide range of CCVs per signal, but in 42 signals there was only a single CCV: for these signals, the simplest hypothesis is that the CCV was causal (Fig. 2c,d and Table 1 ). Furthermore, within signals with few CCVs (<10), the mean scaled combined annotation-dependent depletion score was higher than in signals with more CCVs (13.1 versus 6.7 for CCVs in exons; P t -test = 2.7 × 10 −4 ), suggesting that these are more likely to be functional. Fig. 2: Determining independent risk signals and CCVs. a , Number of independent signals per region, identified through multinomial stepwise logistic regression. b , Signal classification as strong- or moderate-confidence signals. c , d , Number of CCVs per signal in strong- ( c ) and moderate-confidence signals ( d ), identified through multinomial stepwise logistic regression. e , Subtype classification of strong signals into ER positive, ER negative and signals equally associated with both phenotypes (ER neutral) from the BCAC analysis. f , Subtype classification from the meta-analysis of BCAC and CIMBA. Numbers in brackets show the numbers of CCVs from the meta-analysis of BCAC and CIMBA. g , Number of variants at different posterior probability thresholds. In total, 15 variants reached a posterior probability of ≥80% by at least one of the three models (ER all, ER positive or ER negative). Full size image Table 1 Signals with single CCVs and variants with a posterior probability of >80% Full size table The majority of breast tumors express the estrogen receptor (ER positive), but ~20% do not (ER negative); these two tumor types have distinct biological and clinical characteristics 25 . Using a case-only analysis for the 196 strong-evidence signals, we found 66 signals (34%; containing 1,238 CCVs) where the lead variant conferred a greater relative risk of developing ER-positive tumors (false discovery rate (FDR) = 5%), and 29 (15%; 646 CCVs) where the lead variant conferred a greater risk of ER-negative cancer tumors (FDR = 5%) (Supplementary Table 2b and Fig. 2e ). The remaining 101 signals (51%; 5,510 CCVs) showed no difference by ER status (referred to as ER neutral). Patients with BRCA1 mutations are more likely to develop ER-negative tumors 26 . Hence, to increase our power to identify ER-negative signals, we performed a fixed-effects meta-analysis, combining association results from BRCA1 mutation carriers in CIMBA with the BCAC ER-negative association results. This meta-analysis identified ten additional signals (seven ER-negative and three ER-neutral), making 206 strong-evidence signals (17% ER negative) containing 7,652 CCVs in total (Fig. 2f ). More than one-quarter of the CCVs (2,277) were accounted for by one signal, resulting from strong linkage disequilibrium with a CNV. The remaining analyses focused on the other 205 strong signals across 128 regions (Supplementary Table 2c ). The proportion of the familial relative risk (FRR) of breast cancer explained by all 206 strong signals was 20.6%, compared with 17.6% when only the lead SNP for each region was considered. The proportion of the FRR explained increased by a further 3% (to 23.6%) when all 352 signals were considered (Supplementary Table 2d ). CCVs are over-represented in active gene regulatory regions and transcription factor binding sites (TFBSs) We constructed a database of mapped genomic features in seven primary cells derived from normal breast and 19 breast cell lines using publicly available data, resulting in 811 annotation tracks in total. These ranged from general features (such as whether a variant was in an exon or in open chromatin) to more specific features (such as cell-specific transcription factor binding or histone marks (determined through chromatin immunoprecipitation followed by sequencing (ChIP-Seq) experiments) in breast-derived cells or cell lines). Using logistic regression, we examined the overlap of these genomic features with the positions of 5,117 CCVs in the 195 strong-evidence BCAC signals versus the positions of 622,903 variants excluded as credible candidates in the same regions (Supplementary Fig. 1a and Supplementary Table 3 ). We found significant enrichment of CCVs (FDR = 5%) in four genomic features (open chromatin, actively transcribed genes, gene regulatory regions and binding sites), as described below. Fig. 3: Overlap of CCVs with gene regulatory regions, gene bodies and TFBSs. a , Breast cancer CCVs overlap with chromatin states and broad breast cell epigenetic marks. HMEC, human mammary epithelial cells. Open chrom, open chromatin; heterochrom, heterochromatin b , c , Breast cancer CCVs ( b ) and autoimmune CCVs ( c ) overlap with breast cell epigenetic marks. vHMEC, variant HMEC. Luminal pr, luminal progenitor. d , e , Breast cancer CCVs ( d ) autoimmune CCVs ( e ) and overlap with autoimmune-related epigenetic marks. In a , b and d , the column ‘strong CCVs’ represents analysis with all CCVs at strong signals, while the remaining columns represent analysis of CCVs at strong signals stratified by phenotype. Logistic regression robust variance estimation for clustered observations was used, and Wald test Χ 2 P values were estimated using 67,136 ER-positive and 17,506 ER-negative cases, together with 88,937 controls. Non-significant P values are shown in dark gray. Significance was defined as an FDR of 5%, which corresponds to the following P value thresholds: P = 1.66 × 10 −2 (strong signals); P = 2.42 × 10 −2 (ER positive); P = 3.02 × 10 −3 (ER negative); and P = 1.76 × 10 −3 (ER neutral). f – h , Significant ER-positive ( f ), ER-negative ( g ) and ER-neutral CCVs ( h ) overlap with TFBSs. TFBSs found significant for ER-positive CCVs are highlighted in red ( x axis labels). Full size image Open chromatin As shown in Fig. 3a , DNase I hypersensitive sites sequencing and formaldehyde-assisted isolation of regulatory elements sequencing showed significant enrichment of CCVs in open chromatin in ER-positive breast cancer cell lines and normal breast. Conversely, we found depletion of CCVs within heterochromatin (determined by the H3K9me3 mark in normal breast, and by chromatin state in ER-positive cells 27 ). Actively transcribed genes Significant enrichment of CCVs was also found in actively transcribed genes in normal breast and ER-positive cell lines (as defined by H3K36me3 or H3K79me2 histone marks; Fig. 3a ). Enrichment was larger for ER-neutral CCVs than for those affecting either ER-positive or ER-negative tumors. Gene regulatory regions CCVs overlapped distal gene regulatory elements in ER-positive breast cancer cells lines (defined by H3K4me1 or H3K27ac marks; Fig. 3b ). This was confirmed using the Encyclopedia of DNA Elements (ENCODE) definition of active enhancers in MCF-7 cells (enhancer-like regions defined by combining DNase and H3K27ac marks), as well as the definition of refs. 27 , 28 (Supplementary Table 3 ). Under these more stringent definitions, enrichment among ER-positive CCVs was significantly larger than ER-negative or ER-neutral CCVs. Data from ref. 27 showed that 73% of active enhancer regions overlapped by ER-positive CCVs in ER-positive cells (MCF-7) are inactive in the normal human mammary epithelial (HMEC) breast cell line; thus, these enhancers appear to be MCF-7 specific. We also detected significant enrichment of CCVs in active promoters in ER-positive cells (defined by H3K4me3 marks in T-47D), although the evidence for this effect was weaker than for distal regulatory elements (defined by H3K27ac marks in MCF-7; Fig. 3b ). Only ER-positive CCVs were significantly enriched in T-47D active promoters. Conversely, CCVs were depleted among repressed gene regulatory elements (defined by H3K27me3 marks) in normal breast (Fig. 3b ). As a control, we performed similar analyses with autoimmune disease CCVs 29 ( Methods ) and relevant B and T cells (Fig. 3b–e ). The strongest evidence of enrichment of breast cancer CCVs was found at regulatory regions active in ER-positive cells (Fig. 3b ), whereas enrichment of autoimmune CCVs was in regulatory regions active in B and T cells (Fig. 3e ). We also compared the enrichment of our CCVs in enhancer-like and promoter-like regions (defined by ENCODE; Supplementary Fig. 1b ). The strongest evidence of enrichment of ER-positive CCVs in enhancer-like regions was found in MCF-7 cells—the only ER-positive cell line in ENCODE (Supplementary Fig. 1b ). These results highlight both the tissue specificity and disease specificity of these histone-marked gene regulatory regions. Binding sites We observed significant enrichment of CCVs in the binding sites for 40 TFBSs determined by ChIP-Seq (Fig. 3f–h ). The majority of the experiments were performed in ER-positive cell lines (90 TFBSs; 20 with data in ER-negative cell lines, 76 with data in ER-positive cell lines and 16 with data in normal breast). These TFBSs overlap each other and histone marks of active regulatory regions (Supplementary Fig. 2 ). Enrichment in five TFBSs ( ESR1 , FOXA1 , GATA3 , TCF7L2 and E2F1 ) has been reported previously 2 , 30 . All 40 TFBSs were significantly enriched in ER-positive CCVs (Fig. 3f ), seven were also enriched in ER-negative CCVs and nine were enriched in ER-neutral CCVs (Fig. 3g–h ). ESR1 , FOXA1 , GATA3 and EP300 TFBSs were enriched in all CCV ER subtypes. However, the enrichment for ESR1 , FOXA1 and GATA3 was stronger for ER-positive CCVs than for ER-negative or ER-neutral CCVs. CCVs significantly overlap consensus transcription factor binding motifs We investigated whether CCVs were also enriched within consensus transcription factor binding motifs by conducting a motif search within active regulatory regions (ER-positive CCVs at H3K4me1 marks in MCF-7). We identified 30 motifs from eight transcription factor families, with enrichment in ER-positive CCVs (FDR = 10%; Supplementary Table 4a ) and a further five motifs depleted among ER-positive CCVs. To assess whether the motifs appeared more frequently than by chance at active regulatory regions overlapped by our ER-positive CCVs, we compared motif presence in a set of randomized control sequences ( Methods ). Thirteen of 30 motifs were more frequent at active regulatory regions with ER-positive CCV enrichment; these included seven homeodomain motifs and two forkhead factors (Supplementary Table 4b ). When we looked at the change in predicted binding affinity, 57 ER-positive signals (86%) included at least one CCV predicted to modify the binding affinity of the enriched TFBSs (at least twofold; Supplementary Table 4c ). Forty-eight ER-positive signals (73%) had at least one CCV predicted to modify the binding affinity greater than tenfold. This analysis validates previous reports of breast cancer causal variants that alter the DNA binding affinity for FOXA1 (refs. 3 , 30 ). Bayesian fine-mapping incorporating functional annotations and linkage disequilibrium As an alternative statistical approach for inferring likely causal variants, we applied PAINTOR 31 to the same 128 regions (Fig. 1 ). In brief, PAINTOR integrates genetic association results, linkage disequilibrium structure and enriched genomic features in an empirical Bayes framework and derives the posterior probability of each variant being causal, conditional on available data. To eliminate artefacts due to differences in genotyping and imputation across platforms, we restricted PAINTOR analyses to cases and controls typed using OncoArray (61% of the total). We identified seven variants with a high posterior probability (HPP ≥ 80%) of being causal for overall breast cancer, and ten for the ER-positive subtype (Table 1 ); two of these had a HPP > 80% for both ER-positive and overall breast cancer. These 15 HPP variants (HPPVs; ≥80%) were distributed across 13 regions. We also identified an additional 35 variants in 25 regions with HPP (≥50 and <80%) for ER-positive, ER-negative or overall breast cancer (Fig. 2g ). Consistent with the CCV analysis, we found evidence that most regions contained multiple HPPVs; the sum of posterior probabilities across all variants in a region (an estimate of the number of distinct causal variants in the region) was >2.0 for 84 out of 86 regions analyzed for overall breast cancer, with a maximum of 16.1 and a mean of 6.4. For ER-positive cancer, 46 out of 47 regions had total posterior probability of >2.0 (maximum: 18.3; mean: 6.5). For ER-negative cancer, 17 out of 23 regions had a total posterior probability of >2.0 (maximum: 9.1; mean: 3.2). Although for many regions we were not able to identify HPPVs, we were able to reduce the proportion of variants needed to account for 80% of the total posterior probability in a region to <5% for 65 regions for overall breast cancer, 43 regions for ER-positive breast cancer and 18 regions for ER-negative breast cancer (Supplementary Fig. 3a–c ). PAINTOR analyses were also able to reduce the set of likely causal variants in many cases. After summation of the posterior probabilities for CCVs in each of the overall breast cancer signals, 39 out of 100 strong-evidence signals had a total posterior probability of >1.0. The number of CCVs in these signals ranged from 1 to 375 (median: 24), but the number of variants needed to capture 95% of the total posterior probability in each signal ranged from 1 to 115 (median: 12), representing an average reduction of 43% in the number of variants needed to capture the signal. PAINTOR and CCV analyses were generally consistent, yet complementary. Only 3.3% of variants outside of the set of strong-signal CCVs for overall breast cancer had a posterior probability of >1%, and only 48 (0.013%) of these had a posterior probability of >30% (Supplementary Fig. 3d ). At ER-positive and ER-negative signals, respectively, 3.1 and 1.6% of the non-CCVs at strong signals had a posterior probability of >1%, and 40 (0.019%) and 3 (0.003%) of these had a posterior probability of >30% (Supplementary Fig. 3e–f ). For the non-CCVs at strong-evidence signals with a posterior probability of >30%, the relatively HPP may be driven by the addition of functional annotation. Indeed, the incorporation of functional annotations more than doubled the posterior probability for 64 out of 88 variants when compared with a PAINTOR model with no functional annotations. CCVs co-localize with variants controlling local gene expression We used four breast-specific expression quantitative trait loci (eQTL) datasets to identify a credible set of variants associated with differences in gene expression (expression variants): tumor tissue from the Nurses’ Health Study (NHS) 32 and The Cancer Genome Atlas (TCGA) 33 ; and normal breast tissue from the NHS and the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) 34 . We then examined the overlap of expression variants (for each gene, expression variants were defined as those variants that had a P value within two orders of magnitude of the variant most significantly associated with that gene’s expression) with CCVs ( Methods ). There was significant overlap of CCVs with expression variants from both the NHS normal and breast cancer tissue studies (normal breast: odds ratio (OR) = 2.70; P = 1.7 × 10 −5 ; tumor tissue: OR = 2.34; P = 2.6 × 10 −4 ; Supplementary Table 3 ). ER-neutral CCVs overlapped with expression variants in normal tissue more frequently than ER-positive and ER-negative CCVs (OR ER neutral = 3.51; P = 1.3 × 10 −5 ). Cancer risk CCVs overlapped credible expression variants in 128 out of 205 signals (62%) in at least one of the datasets (Supplementary Table 5a,b ). Sixteen additional variants with a posterior probability of ≥30%, not included among the CCVs, also overlapped with a credible expression variant (Supplementary Table 5a,b ). Transcription factors and known somatic breast cancer drivers are over-represented among prioritized target genes We assumed that causal variants function by affecting the behavior of a local target gene. However, it is challenging to define target genes or to determine how they may be affected by the causal variant. Few potentially causal variants directly affect protein coding: we observed 67 out of 5,375 CCVs and 19 out of 137 HPPVs (≥30%) in protein-coding regions. Of these, 33 (0.61%) were predicted to create a missense change, one a frameshift and another a stop gain, while 30 were synonymous (0.59%; Supplementary Table 5c ). In total, 499 CCVs at 94 signals, and four additional HPPVs (≥30%), are predicted to create new splice sites or activate cryptic splice sites in 126 genes (Supplementary Table 5d ). These results are consistent with previous observations that the majority of common susceptibility variants are regulatory. We applied an updated version of our pipeline integrated expression quantitative trait and in s ilico prediction of GWAS targets (INQUISIT) 2 to prioritize potential target genes from 5,375 CCVs in strong signals and all 138 HPPVs (≥30%; Supplementary Table 2c ). The pipeline predicted 1,204 target genes from 124 out of 128 genomic regions examined. As a validation, we examined the overlap between INQUISIT predictions and 278 established breast cancer driver genes 35 , 36 , 37 , 38 , 39 . Cancer driver genes were over-represented among high-confidence (level 1) targets, with a fivefold increase over expected levels from CCVs and a 15-fold increase from HPPVs ( P = 1 × 10 −6 ; Supplementary Fig. 4a ). Notably, 13 cancer driver genes ( ATAD2 , CASP8 , CCND1 , CHEK2 , ESR1 , FGFR2 , GATA3 , MAP3K1 , MYC , SETBP1 , TBX3 , XBP1 and ZFP36L1 ) were predicted from the HPPVs derived from PAINTOR. Cancer driver gene status was consequently included as an additional weighting factor in the INQUISIT pipeline. Transcription factor genes 40 were also enriched among high-confidence targets predicted from both CCVs (twofold; P = 4.6 × 10 −4 ) and HPPVs (2.5-fold; P = 1.8 × 10 −2 ; Supplementary Fig. 4a ). In total, INQUISIT identified 191 target genes supported by strong evidence (Supplementary Table 6 ). Significantly more genes were targeted by multiple independent signals ( n = 165) than expected by chance ( P = 4.3 × 10 −8 ; Supplementary Fig. 4b and Fig. 4 ). Six high-confidence predictions came only from HPPVs, although three of these ( IGFBP5 , POMGNT1 and WDYHV1 ) had been predicted at lower confidence from CCVs. Target genes included 20 that were prioritized via potential coding/splicing changes (Supplementary Table 7 ), ten via promoter variants (Supplementary Table 8 ) and 180 via distal regulatory variants (Supplementary Table 9 ). We illustrate the genes prioritized via multiple lines of evidence in Fig. 4 . Fig. 4: Predicted target genes are enriched in known breast cancer driver genes and transcription factors. Target genes ( n = 69) that fulfill at least one of the following criteria: (1) is targeted by more than one independent signal; (2) is a known driver gene; (3) is a known transcription factor gene; (4) its binding sites (as determined by ChIP-Seq (ChIP-Seq BS)) are significantly overlapped by CCVs; or (5) its consensus (transcription factor) motif is significantly overlapped by CCVs. Asterisks indicate genes with published functional follow-up. Full size image Three examples of INQUISIT using genomic features to predict target genes On the basis of capture Hi-C and chromatin interaction analysis by paired-end tag sequencing (ChIA-PET) data, NRIP1 is a predicted target of intergenic CCVs and HPPVs at chr21q21 (Supplementary Fig. 5a ). Multiple target genes were predicted at chr22q12, including the driver genes CHEK2 and XBP1 (Supplementary Fig. 5b ). A third example at chr12q24.31 is a more complicated scenario with two level 1 targets: RPLP0 (ref. 41 ) and a modulator of mammary progenitor cell expansion, MSI1 (ref. 42 ) (Supplementary Fig. 5c ). Target gene pathways include DNA integrity checkpoint, apoptosis and developmental processes and the immune system We performed pathway analysis to identify common processes using INQUSIT high-confidence target protein-coding genes (Fig. 5a ) and identified 488 Gene Ontology terms and 307 pathways at an FDR of 5% (Supplementary Table 10 ). These were grouped into 98 themes by common ancestor Gene Ontology terms, pathways or transcription factor classes (Fig. 5b ). We found that 23% (14/60) of the ER-positive target genes were classified within developmental process pathways (including mammary development), 18% were classified in immune system pathways and a further 17% were classified in nuclear receptor pathways. Of the genes targeted by ER-neutral signals, 21% (18/87) were classified in developmental process pathways, 19% were classified in immune system pathways and a further 18% were classified in apoptotic process pathways. The top themes of genes targeted by ER-negative signals were DNA integrity checkpoint processes and the immune system, each of which contained 19% of genes (7/37), and apoptotic processes (16%). Fig. 5: Predicted target genes by phenotype and significantly enriched pathways. a , Venn diagram showing the associated phenotype (ER positive, ER negative or ER neutral) for the level 1 target genes, predicted by the CCVs and HPPVs. Asterisks denote ER-positive or ER-negative target genes also targeted by ER-neutral signals. b , Heat map showing clustering of pathway themes over-represented by INQUISIT level 1 target genes. Colors represent the relative number of genes per phenotype within enriched pathways, grouped by common themes (ER positive, ER negative, ER neutral or all phenotypes together (strong)). cAMP, cyclic adenosine monophosphate; CARM1, coactivator associated arginine methyltransferase 1; cGMP, cyclic guanosine monophosphate; EGFR, epidermal growth factor receptor; FGFR, fibroblast growth factor receptor; GATA, GATA transcription factors; MAPK, mitogen activated protein kinase; MET, MET proto-oncogene receptor tyrosine kinase; NOTCH, notch protein; PTEN, phosphatase and tensin homolog; PTK6, protein tyrosine kinase 6; RAS, RAS protein; ROBO, roundabout receptors; ROS, reactive oxygen species; TGFBR, transforming growth factor beta receptor; WNT, WNT proteins. Full size image Novel pathways revealed by this study include tumor necrosis factor-related apoptosis-inducing ligand (TRAIL) signaling, the AP-2 transcription factors pathway, and regulation of IκB kinase/nuclear factor-κB (NF-κB) signaling. Of note, the latter of these is specifically over-represented among ER-negative target genes. We also found significant over-representation of additional carcinogenesis-linked pathways, including cyclic adenosine monophosphate, NOTCH, phosphoinositide 3-kinase, RAS and WNT/β-catenin, and of receptor tyrosine kinase signaling, including fibroblast growth factor receptor, epidermal growth factor receptor and transforming growth factor-β receptor 43 , 44 , 45 , 46 , 47 . Finally, our target genes are also significantly over-represented in DNA damage checkpoint and DNA repair pathways, as well as programmed cell death pathways, such as apoptotic processes, regulated necrosis and death receptor signaling-related pathways. Discussion We have performed multiple complementary analyses on 150 breast cancer-associated regions originally found by GWASs, and identified 362 independent risk signals, 205 of these with high confidence ( P < 10 −6 ). The inclusion of these new variants increases the explained proportion of familial risk by 6% compared with that explained by the lead signals alone. We observed that most regions contain multiple independent signals, with the greatest number (nine) in the region surrounding ESR1 and its co-regulated genes, and on 2q35, where IGFBP5 appears to be a key target. We used two complementary approaches to identify likely causal variants within each region: a Bayesian approach, PAINTOR (which integrated genetic associations, linkage disequilibrium and informative genomic features, providing complementary evidence) and a more traditional, multinomial regression approach. PAINTOR supported most associations found by multinomial regression and also identified additional variants. Specifically, the Bayesian method highlighted 15 variants that are highly likely to be causal (HPP ≥ 80%). From these approaches, we identified a single variant, likely to be causal, at each of 34 signals (Table 1 ). Of these, only rs16991615 ( MCM8 ; NP_115874.3:p.E341K) and rs7153397 ( CCDC88C ; NM_001080414.2:c.5058 + 1342G>A; a cryptic splice-donor site) were predicted to affect protein-coding sequences. However, in other signals, we also identified four coding changes previously recognized as deleterious: the stop gain rs11571833 ( BRCA2 ; NP_000050.2:p.K3326*) 48 ; two CHEK2 coding variants (the frameshift rs555607708 (refs. 49 , 50 ) and a missense variant, rs17879961 (refs. 51 , 52 )); and a splicing variant (rs10069690 in TERT , which results in the truncated protein INS1b 19 , decreased telomerase activity, telomere shortening and increased DNA damage response 53 ). Having identified potential causal variants within each signal, we aimed to uncover their functions at the DNA level, as well as trying to predict their target gene(s). Across all 150 regions, a notable feature is that many likely causal variants implicated in ER-positive cancer risk lie in gene regulatory regions marked as open and active in ER-positive breast cells, but not in other cell types. Moreover, a significant proportion of potential causal variants overlap the binding sites for transcription factor proteins ( n = 40 from ChIP-Seq) and co-regulators ( n = 64 with the addition of computationally derived motifs). Furthermore, nine proteins also appear in the list of high-confidence target genes; hence, the following genes and their products have been implicated by two different approaches: CREBBP , EP300 , ESR1 , FOXI1 , GATA3 , MEF2B , MYC , NRIP1 and TCF7L2 . Most proteins encoded by these genes already have established roles in estrogen signaling. CREBBP , EP300 , ESR1 , GATA3 and MYC are also known cancer driver genes that are frequently somatically mutated in breast tumors. In contrast with ER-positive signals, we identified fewer genomic features enriched in ER-negative signals. This may reflect the common molecular mechanisms underlying their development, but the power of this study was limited, despite including as many patients with ER-negative tumors as possible from the BCAC and CIMBA consortia. Less than 20% of genomic signals confer a greater risk of ER-negative cancer and there are few publicly available ChIP-Seq data on ER-negative breast cancer cell lines. The heterogeneity of ER-negative tumors also may have limited our power. Nevertheless, we have identified 35 target genes for ER-negative likely causal variants. Some of these already had functional evidence supporting their role: including CASP8 (ref. 54 ) and MDM4 (ref. 55 ). However, most targets currently have no reported function in ER-negative breast cancer development. Finally, we examined the Gene Ontology pathways in which target genes most often lie. Of note, 14% (25/180) of all high-confidence target genes and 19% of ER-negative target predictions are in immune system pathways. Among the significantly enriched pathways were T cell activation, interleukin signaling, Toll-like receptor cascades and I-κB kinase/NF-κB signaling, as well as processes leading to activation and perpetuation of the innate immune system. The link between immunity, inflammation and tumorigenesis has been studied extensively 56 , although not primarily in the context of susceptibility. Five ER-negative high-confidence target genes ( ALK , CASP8 , CFLAR , ESR1 and TNFSF10 ) lie in the IκB kinase/NF-κB signaling pathway. Interestingly, ER-negative cells have high levels of NF-kB activity compared with ER-positive cells 57 . A recent expression–methylation analysis on breast cancer tumor tissue also identified clusters of genes correlated with DNA methylation levels: one enriched in ER signaling genes and a second in immune pathway genes 58 . These analyses provide strong evidence for more than 200 independent breast cancer risk signals, identify the plausible cancer variants and define likely target genes for the majority of these. However, notwithstanding the enrichment of certain pathways and transcription factors, the biological basis underlying most of these signals remains poorly understood. Our analyses provide a rational basis for such future studies into the biology underlying breast cancer susceptibility. Methods Study samples Epidemiological data for European women were obtained from 75 breast cancer case–control studies participating in the BCAC (cases: 40,285 iCOGS and 69,615 OncoArray; cases with ER status available: 29,561 iCOGS and 55,081 OncoArray; controls: 38,058 iCOGS and 50,879 OncoArray). Details of the participating studies, genotype calling and quality control are given in refs. 2 , 22 , 23 , respectively. Epidemiological data for BRCA1 mutation carriers were obtained from 60 studies providing data to the CIMBA (affected: 1,591 iCOGS and 7,772 OncoArray; unaffected: 1,665 iCOGS and 7,780 OncoArray). This dataset has been described in detail previously 1 , 59 , 60 . All studies provided samples of European ancestry. Any non-European samples were excluded from the analyses. Variant selection and genotyping Similar approaches were used to select variants for inclusion on the iCOGS and OncoArray, and these are described in detail elsewhere 2 , 21 . Both arrays included a dense coverage of variants across known susceptibility regions (at the time of their design), with sparser coverage of the rest of the genome. Twenty-one known susceptibility regions were selected for dense genotyping using iCOGS and 73 regions were selected for OncoArray. These regions were 1-megabase (Mb) intervals centered on the published lead GWAS hit (combined into larger intervals where these overlapped). For iCOGS, all known variants from the March 2010 release of the 1000 Genomes Project with a MAF > 0.02 in Europeans were identified, and all those correlated with the published GWAS variants at r 2 > 0.1 ( r 2 , Pearson’s squared correlation coefficient), together with a set of variants designed to tag all remaining variants at r 2 > 0.9, were selected to be included in the array ( ). For OncoArray, all designable variants correlated with the known hits at r 2 > 0.6, plus all variants from lists of potentially functional variants on RegulomeDB and a set of variants designed to tag all of the remaining variants at r 2 > 0.9, were selected. In total, across the 152 regions considered here, 26,978 iCOGS- and 58,339 OncoArray-genotyped variants passed the quality control criteria. We imputed genotypes for all of the remaining variants by using IMPUTE2 (ref. 61 ) and the October 2014 release of the 1000 Genomes Project as a reference. Imputation was conducted independently in the iCOGS and OncoArray subsets. To improve accuracy at low-frequency variants, we used the standard IMPUTE2 MCMC algorithm for follow-up imputation, which includes no pre-phasing of the genotypes and increased both the buffer regions and the number of haplotypes to use as templates (a more detailed description of the parameters used can be found in ref. 21 ). We thus genotyped or successfully imputed 639,118 variants (all with an imputation info score ≥ 0.3 and a MAF ≥ 0.001 in both iCOGS and OncoArray datasets). Imputation summaries and coverage for each of the analyzed regions stratified by allele frequency can be found in Supplementary Table 1b . BCAC statistical analyses Per-allele odds ratios and standard errors were estimated for each variant using logistic regression. We ran this analysis separately for iCOGS and OncoArray, and for overall, ER-positive and ER-negative breast cancer. The association between each variant and breast cancer risk was adjusted by study (iCOGS) or country (OncoArray), and eight (iCOGS) or ten (OncoArray) ancestry-informative principal components. The statistical significance for each variant was derived using a Wald test. Defining appropriate significance thresholds for association signals To establish an appropriate significance threshold for independent signals, all variants evaluated in the meta-analysis were included in logistic forward selection regression analyses for overall breast cancer risk in iCOGS, run independently for each region. We evaluated five P value thresholds for inclusion: <1 × 10 −4 , <1 × 10 −5 , <1 × 10 −6 , <1 × 10 −7 and <1 × 10 −8 . The most parsimonious iCOGS models were tested in OncoArray, and the FDR at the 1% level for each threshold was estimated using the Benjamini–Hochberg procedure. At a 1% FDR threshold, 72% of associations, significant at P < 10 −4 , were replicated on iCOGS, and 94% of associations, significant at P < 10 −6 , were replicated on OncoArray. Based on these results, two categories were defined: strong-evidence signals (conditional P < 10 −6 in the final model) and moderate-evidence signals (conditional P < 10 −4 and P ≥ 10 −6 in the final model). Identification of independent signals To identify independent signals, we ran multinomial stepwise regression analyses, separately in iCOGS and OncoArray, for all variants displaying evidence of association ( n variants = 202,749). We selected two sets of well-imputed variants (imputation info score ≥ 0.3 in both iCOGS and OncoArray): (1) common and low-frequency variants (MAF ≥ 0.01) with a logistic regression P value inclusion threshold of ≤0.05 in either the iCOGS or OncoArray datasets for at least one of the three phenotypes (overall, ER positive and ER negative breast cancer); and (2) rarer variants (MAF ≥ 0.001 and <0.01), with a logistic regression inclusion P value of ≤ 0.0001. The same parameters used for adjustment in logistic regression were used in the multinomial regression analysis (R function multinom), which simultaneously estimated per-allele odds ratios for ER-positive and ER-negative breast cancer. The multinomial regression estimates were combined using a fixed-effects meta-analysis weighted by the inverse variance. Variants with the lowest conditional P value from the meta-analysis of both European cohorts at each step were included in the multinomial regression model. However, if the new variant to be included in the model caused collinearity problems due to high correlation with an already selected variant, or showed high heterogeneity ( P < 10 −4 ) between iCOGS and OncoArray after being conditioned by the variant(s) in the model, we dropped the new variant and repeated this process. At 105 of 152 evaluated regions, the main signal showed genome-wide significance, while 44 were marginally significant (9.89 × 10 −5 ≥ P > 5 × 10 −8 ). For two regions, there were no variants significant at P < 10 −4 (chr14:104712261–105712261; rs10623258; multinomial regression P = 2.32 × 10 −4 ; chr19:10923703–11923703; rs322144; multinomial regression P = 3.90 × 10 −3 ). Four main differences in the datasets used here and in the previous paper may account for this: (1) our previous paper 2 included data from 11 additional GWASs (14,910 cases and 17,588 controls) that have not been included in the present analysis in order to minimize differences in array coverage, and because ER status data were substantially incomplete and individual-level data were not available for all GWASs; (2) the present analysis was based on estimating separate risks for ER-positive and ER-negative disease, whereas in our previous paper the outcome was overall breast cancer risk; ER status was available for only 73% of the iCOGS and 79% of the OncoArray breast cancer cases; (3) for the set of samples genotyped with both arrays, ref. 2 used the iCOGS genotypes, while the present study included OncoArray genotypes to maximize the number of samples genotyped with a larger coverage; and (4) the imputation procedure was modified (in particular using one-step imputation without pre-phasing) to improve the imputation accuracy of less frequent variants. We used a forward stepwise approach to define the number of independent signals within each associated genomic region. First, we identified the index variant of the main signal in the region, and then ran multinomial logistic regression for all of the other variants, adjusted by the index variant, to identify additional variants that remained independently significant within the model. We repeated this process, adjusting for identified index variants, until no more additional variants could be added. In this way, we found from 1–11 independent signals within the 150 regions that containing a genome-wide significant main signal. Selection of a set of CCVs For each independently associated signal, we first defined CCVs likely to drive its association as those variants with P values within two orders of magnitude of the most significant variant for that signal, after adjusting for the index variant of other signals within that region (as identified in the forward stepwise regression above; Supplementary Fig. 6a ) 24 . For each region, we then attempted to obtain the best-fitting model by successively fitting models in which the index variant for each signal was replaced by other CCVs for that signal, adjusting for the index variants for the other signals (Supplementary Fig. 6b ). Where a model with a higher chi-squared value was obtained, the index variant was replaced by the CCV in the best model (Supplementary Fig. 6c,d ). This process was repeated until the model (that is, the set of index variants) did not change further (Supplementary Fig. 6g ). This procedure was performed first for the set of strong signals (that is, considering models including only the strong signals). Once a final model had been obtained for the strong signals, the index variants for the strong signals were considered fixed and the process was repeated for all signals, this time allowing the index variants for the weak signals (but not the strong signals) to vary. Using this procedure, we could define the best model for 140 out of 150 regions, but for ten regions this approach did not converge (chr4:175328036–176346426, chr5:55531884–56587883, chr6:151418856–152937016, chr8:75730301–76917937, chr10:80341148–81387721, chr10:122593901–123849324, chr12:115336522–116336522, chr14:36632769–37635752, chr16:3606788–4606788 and chr22:38068833–39859355). For these ten regions, we defined the best model, from among all possible combinations of credible variants, as that with the largest chi-squared value. Finally, we redefined the set of CCVs for each signal using the conditional P values, after adjusting for the revised set of index variants. Again, for the strong signals, we conditioned on the index variants for the other strong signals, while for the weak signals we conditioned on the index variants for all of the other signals. Case-only analysis Differences in the effect size between ER-positive and ER-negative disease for each index-independent variant were assessed using a case-only analysis. We performed logistic regression with ER status as the dependent variable and the lead variant at each strong signal in the fine-mapping region as the independent variables. We used FDR (5%) to adjust for multiple testing. OncoArray-only stepwise analysis To evaluate whether the lower coverage in iCOGS could affect the identification of independent signals, we ran stepwise multinomial regression using only the OncoArray dataset. We identified 249 independent signals. Ninety-two signals, in 67 fine-mapping regions, achieved a genome-wide significance level (conditional P < 5 × 10 −8 ). Of these, 205 signals were also identified in the meta-analysis with iCOGS. Nine independent variants across ten regions were not evaluated in the combined analysis due to their low imputation information score in iCOGS. Of these nine signals, two signals would be classified as main primary signals: rs114709821 at region chr1:145144984–146144984 (OncoArray imputation information score = 0.72); and rs540848673 at region chr1:149406413–150420734 (OncoArray imputation information score = 0.33). Given the low number of additional signals identified in the OncoArray dataset alone, all analyses were based on the combined iCOGS/OncoArray dataset. CIMBA statistical analysis CIMBA provided data from 60 retrospective cohort studies consisting of 9,445 unaffected and 9,363 affected female BRCA1 mutation carriers of European ancestry. Unconditional (that is, single-variant) analyses were performed using a score test based on the retrospective likelihood of observing the genotype conditional on the disease phenotype 62 , 63 . Conditional analyses, where more than one variant is analyzed simultaneously, cannot be performed in this score test framework. Therefore, conditional analyses were performed by Cox regression, allowing for adjustment of the conditionally independent variants identified by the BCAC/DRIVE analyses. All models were stratified by country and birth cohort, and adjusted for relatedness (unconditional models used kinship-adjusted standard errors based on the estimated kinship matrix; conditional models used cluster robust standard errors based on phenotypic family data). Data from the iCOGS array and OncoArray were analyzed separately and combined to give an overall BRCA1 association by fixed-effects meta-analysis. Variants were excluded from further analyses if they exhibited evidence of heterogeneity (heterogeneity P < 1 × 10 −4 ) between iCOGS and OncoArray, had a MAF < 0.005, were poorly imputed (imputation information score < 0.3) or were imputed to iCOGS only (that is, they must have been imputed to OncoArray or iCOGS and OncoArray). Meta-analysis of ER-negative cases in BCAC with BRCA1 mutation carriers from CIMBA BRCA1 mutation carrier association results were combined with the BCAC multinomial regression ER-negative association results in a fixed-effects meta-analysis. Variants considered for analysis must have passed all previous quality control steps and have had MAF ≥ 0.005. All meta-analyses were performed using the METAL software 64 . Instances where spurious associations might occur were investigated by assessing the linkage disequilibrium between a possible spurious association and the conditionally independent variants. High linkage disequilibrium between a variant and a conditionally independent variant within its region causes model instability through collinearity, and convergence of the model likelihood maximization may not be reliable. Where the association appeared to be driven by collinearity, the signals were excluded. Heritability estimation To estimate the frailty-scale heritability due to all of the fine-mapping signals, we used the formula: $$h^2 = 2\left( {\gamma ^{\prime T} R\gamma ^\prime - \tau ^{\prime T} I\tau ^\prime } \right)$$ Here, \(\gamma ^\prime = \gamma \sqrt {{\mathbf{p}}\left( {1 - {\mathbf{p}}} \right)}\) and \(\tau ^{\prime} = \tau \sqrt {{\mathbf{p}}\left( {1 - {\mathbf{p}}} \right)}\) , where p is a vector of allele frequencies, γ are the estimated per-allele odds ratios, τ are the corresponding standard errors and R is the correlation matrix of genotype frequencies. To adjust for the overestimation resulting from only including signals passing a given significance threshold, we adapted the approach of ref. 65 , based on maximizing the likelihood, conditional on the test statistic passing the relevant threshold. Since our analyses were based on estimating ER-negative and ER-positive odds ratios simultaneously, the method needed to be adapted to maximize a conditional bivariate normal likelihood. Following ref. 65 , we then estimated mean square error estimates based on a weighted mean of the maximum likelihood estimates and the naïve estimates, which were shown to be unbiased in the 1-degree of freedom case. The estimated effect sizes for overall breast cancer were computed as a weighted mean of the ER-negative and ER-positive estimates, based on the proportions of each subtype in the whole study (weights: 0.21 and 0.79). The results were then expressed in terms of the proportion of the FRR to first-degree relatives of affected women, using the formula h 2 /(2log[ λ ]), where the FRR λ was assumed to be 2 (ref. 2 ). eQTL analysis Total RNA was extracted from normal breast tissue in formalin-fixed paraffin-embedded breast cancer tissue blocks from 264 NHS participants 32 . Transcript expression levels were measured using the Glue Grant Human Transcriptome Array version 3.0 at the Molecular Biology Core Facilities, Dana-Farber Cancer Institute. Gene expression was normalized and summarized into log 2 values using RMA (Affymetrix Power Tools version 1.18.012). Quality control was performed using GlueQC and arrayQualityMetrics version 3.24.014. Genome-wide data on variants were generated using the Illumina HumanHap550 BeadChip as part of the Cancer Genetic Markers of Susceptibility initiative 66 . Imputation to the 1000KGP Phase 3 version 5 ALL reference panel was performed using MACH to pre-phase measured genotypes, and minimac to impute. Expression analyses were performed using data from the TCGA and METABRIC projects 34 , 38 . The TCGA eQTL analysis was based on 458 breast tumors that had matched gene expression, copy number and methylation profiles, together with the corresponding germline genotypes available. All 458 individuals were of European ancestry, as ascertained using the genotype data and the Local Ancestry in Admixed Populations (LAMP) software package (LAMP estimate cut-off > 95% European) 67 . Germline genotypes were imputed into the 1000 Genomes Project reference panel (October 2014 release) using IMPUTE version 2 (refs. 68 , 69 ). Gene expression had been measured on the Illumina HiSeq 2000 RNA sequencing (RNA-Seq) platform (gene-level RSEM normalized counts 70 ), copy number estimates were derived from Affymetrix SNP 6.0 (somatic copy number alteration minus germline copy number variation called using the GISTIC2 algorithm 71 ), and methylation beta values were measured on the Illumina Infinium HumanMethylation450. Expression QTL analysis focused on all variants within each of the 152 genomic intervals that had been subjected to fine-mapping for their association with breast cancer susceptibility. Each of these variants was evaluated for its association with the expression of every gene within 2 Mb that had been profiled for each of the three data types. The effects of tumor copy number and methylation on gene expression were first regressed out using a method described previously 72 . eQTL analysis was performed by linear regression, with residual gene expression as the outcome, germline SNP genotype dosage as the covariate of interest, and ESR1 expression and age as additional covariates, using the R package Matrix eQTL 73 . The METABRIC eQTL analysis was based on 138 normal breast tissue samples resected from patients with breast cancer of European ancestry. Germline genotyping for the METABRIC study was also done on the Affymetrix SNP 6.0 array, and gene expression in the METABRIC study was measured using the Illumina HT12 microarray platform (probe-level estimates). No adjustment was implemented for somatic copy number and methylation status since we were evaluating eQTLs in normal breast tissue. All other steps were identical to the TCGA eQTL analysis described above. Genomic features enrichment We explored the overlap of CCVs and excluded variants with 90 transcription factors, ten histone marks and DNase hypersensitivity sites in 15 breast cell lines and eight normal human breast tissues. We analyzed data from the Encyclopedia of DNA Elements (ENCODE) Project 74 , 75 , Roadmap Epigenomics Projects 76 , the International Human Epigenome Consortium 27 , 77 , Pellacani et al. 78 , TCGA 33 , METABRIC 34 , the ReMap database (we included 241 transcription factor annotations from ReMap (from a total of 2,825), which showed at least 2% overlap for any of the phenotype SNP sets) 79 and other data obtained through the National Center for Biotechnology Information Gene Expression Omnibus. Promoters were defined following the procedure defined in ref. 78 (that is, ±2 kilobases (kb) from a gene transcription start site) using an updated version of the RefSeq genes (refGene version updated 11 April 2017) 80 . Transcribed regions were defined using the same version of RefSeq genes. lncRNA annotation was obtained from GENCODE (version 19) 81 To include eQTL results in the enrichment analysis we: (1) identified all of the genes for which summary statistics were available; (2) defined the most significant eQTL variant for each gene (index eQTL variant; P value threshold ≤ 5 × 10 −4 ); and (3) classified variants with P values within two orders of magnitude of the index expression variant as the credible set of eQTL variants (that is, the best candidates to drive expression of the gene). Variants within at least one eQTL credible set were defined as expression variants. We evaluated the overlap between eQTL credible sets and CCVs (risk variants credible set). We evaluated the enrichment of CCVs for genomic features using logistic regression, with CCV (versus non-CCV variants) being the outcome. To adjust for the correlation among variants in the same fine-mapping region, we used robust variance estimation for clustered observations (R function multiwaycov). The associated variants at an FDR of 5% were included in a stepwise forward logistic regression procedure to select the most parsimonious model. A likelihood ratio test was used to compare multinomial logistic regression models with and without equality effect constraints to evaluate whether there was heterogeneity among the effect sizes for ER positive, ER negative or signals equally associated with both phenotypes (ER neutral). To validate the disease specificity of the regulatory regions identified through this analysis, we followed the same approach for the autoimmune-related CCVs from ref. 29 ( n = 4,192). Variants excluded as candidate causal variants, and within 500 kb upstream and downstream of the index variant for each signal, were classified as excluded variants ( n = 1,686,484). We then tested the enrichment for both the breast cancer and autoimmune CCVs with breast and T and B cell enhancers. We also evaluated the overlap of our CCVs with ENCODE enhancer-like and promoter-like regions for 111 tissues, primary cells, immortalized cell lines and in vitro-differentiated cells. Of these, 73 had available data for both enhancer- and promoter-like regions. Transcription binding site motif analysis We conducted a search to find motif occurrences for transcription factors at active regulatory regions significantly enriched in CCVs. For this, we used two publicly available databases: Factorbook 82 and JASPAR 2016 (ref. 83 ). For the search using Factorbook, we included the motifs for the transcription factors discovered in the cell lines where significant enrichment was found in our genomic features analysis. We also searched for all of the available motifs for Homo sapiens in the JASPAR database (JASPAR CORE 2016; TFBSTools 84 ). Using the USCS sequence (BSgenome.Hsapiens.USCS.hg19) as a reference, we created fasta sequences with the reference and alternative alleles for all of the variants included in our analysis plus 20 base pairs flanking each variant. We used FIMO (version 4.11.2; Grant et al. 85 ) to scan all of the fasta sequences, searching for the JASPAR and Factorbook motifs to identify any overlap of any of the alleles for each of the variants (setting the P value threshold to 10 −3 ). We subsequently determined whether our CCVs were more frequency overlapping a particular transcription factor binding motif when compared with the excluded variants. We ran these analyses for all of the strong signals, but also strong signals stratified by ER status. Also, we subset this analysis to the variants located at regulatory regions in an ER-positive cell line (MCF-7 marked by H3K4me1; ENCODE identification: ENCFF674BKS) and evaluated whether the ER-positive CCVs overlapped any of the motifs more frequently than the excluded variants. We also evaluated the change in total binding affinity caused by the ER-positive CCV alternative allele for all but one (2:217955891:T:<CN0>:0) of the ER-positive CCVs (MatrixRider 86 ). Subsequently, we evaluated whether the MCF-7 regions demarked by H3K4me1 (ENCODE identification: ENCFF674BKS) and overlapped by ER-positive CCVs were enriched in known TFBS motifs. First, we subset the ENCODE bed file ENCFF674BKS to identify MCF-7 H3K4me1 peaks overlapped by the ER-positive CCVs ( n = 107), as well as peaks only overlapped by excluded variants ( n = 11,099), using BEDTools 87 . We created fasta format sequences using genomic coordinate data from the intersected bed files. To create a control sequence set, we used the script included with the MEME Suite (fasta-shuffle-letters) to create ten shuffled copies of each sequence overlapped by ER-positive CCVs ( n = 1,070). We then used AME 88 to interrogate whether the 107 MCF-7 H3K4me1 genomic regions overlapped by ER-positive CCVs were enriched in known TFBS consensus motifs when compared with the shuffled control sequences, or with the MCF-7 H3K4me1 genomic regions overlapped only by excluded variants. We used the command line version of AME (version 4.12.0), selecting as a scoring method the total number of positions in the sequence whose motif score P value was <10 −3 , and using a one-tailed Fisher’s exact test as the association test. PAINTOR analysis To further refine the set of CCVs, we performed empirical Bayes fine-mapping using PAINTOR to integrate marginal genetic association summary statistics, linkage disequilibrium patterns and biological features 31 , 89 . PAINTOR derives jointly the posterior probability for causality of all variants along the respective contribution of genomic features, in order to maximize the log-likelihood of the data across all regions. PAINTOR does not assume a fixed number of causal variants in each region, although it implicitly penalizes non-parsimonious causal models. We applied PAINTOR separately to association results for overall breast cancer (in 85 regions determined to have at least one ER-neutral association or ER-positive and ER-negative association), ER-positive breast cancer (in 48 regions determined to have at least one ER-positive-specific association) and ER-negative breast cancer (in 22 regions determined to have at least one ER-negative-specific association). To avoid artefacts due to mismatches between the linkage disequilibrium in study samples and the linkage disequilibrium matrix supplied to PAINTOR, we used association logistic regression summary statistics from OncoArray data only, and estimated the linkage disequilibrium structure in the OncoArray sample. For each endpoint, we fit four models with increasing numbers of genomic features selected from the stepwise enrichment analyses described above: model 0 (with no genomic features; assumes each variant is equally likely to be causal a priori); model 1 (with those genomic features selected with the stopping rule P < 0.001); model 2 (with those genomic features selected with the stopping rule P < 0.01); and model 3 (with those genomic features selected with the stopping rule P < 0.05). We used the Bayesian information criterion (BIC) to choose the best-fitting model for each outcome. As PAINTOR estimates the marginal log-likelihood of the observed Z scores using Gibbs sampling, we used a shrunk mean BIC across multiple Gibbs chains to account for the stochasticity in the log-likelihood estimates. We ran PAINTOR four times to generate four independent Gibbs chains, and estimated the BIC difference between model i and model j as \(\Delta _{ij} = \left( {\frac{{100}}{{V + 100}}} \right)\left( {{\rm{BIC}}_i - {\rm{BIC}}_j} \right)\) . This assumes an n (0,100) prior on the difference, or roughly a 16% chance that model i would be decisively better than model j (that is |BIC i − BIC j | > 10). We then proceeded to choose the best-fitting model in a stepwise fashion: starting with a model with no annotations, we selected a model with more annotations in favor of a model with fewer if the larger model was a considerably better fit (that is, Δ ij > 2). Model 1 was the best fit according to this process for overall and ER-positive breast cancer, while model 0 was the best fit for ER-negative breast cancer. Differences between the PAINTOR and CCV outputs may be due to several factors. By considering functional enrichment and joint linkage disequilibrium among all SNPs, PAINTOR may refine the set of likely causal variants. Rather than imposing a hard threshold, PAINTOR allows for a gradient of evidence supporting causality; in addition, the two sets of calculations are based on different summary statistics: CCV analyses used both iCOGS and OncoArray genotypes, while PAINTOR used only OncoArray data (Fig. 1 and Methods ). Variant annotation Variant genome coordinates were converted to assembly GRCh38 with liftOver and uploaded to Variant Effect Predictor 90 to determine their effect on genes, transcripts and protein sequence. The commercial software Alamut Batch version 1.6 was also used to annotate coding and splicing variants. PolyPhen-2 (ref. 91 ), SIFT 92 and MAPP 93 were used to predict the consequences of missense coding variants. MaxEntScan 94 , Splice-Site Finder and Human Splicing Finder 95 were used to predict splicing effects. INQUISIT analysis Logic underlying INQUISIT predictions Briefly, genes were considered to be potential targets of candidate causal variants through effects on: (1) distal gene regulation; (2) proximal regulation; or (3) a gene’s coding sequence. We intersected CCV positions with multiple sources of genomic information, including chromatin interactions from capture Hi-C experiments performed in a panel of six breast cell lines 96 , ChIA-PET 97 and Hi-C 98 . We used computational enhancer–promoter correlations (PreSTIGE 99 , IM-PET 100 , FANTOM5 (ref. 101 ) and super-enhancers 28 ), results for breast tissue-specific expression variants from multiple independent studies (TCGA, METABRIC and NHS; Methods ), allele-specific imbalance in gene expression 102 , transcription factor and histone modification ChIP-Seq from the ENCODE and Roadmap Epigenomics Projects, together with the genomic features found to be significantly enriched as described above, gene expression RNA-Seq from several breast cancer lines and normal samples, and topologically associated domain boundaries from T-47D cells (ENCODE 103 ; Methods). To assess the impact of intragenic variants, we evaluated their potential to alter splicing using Alamut Batch to identify new and cryptic donors and acceptors, and several tools to predict the effects of coding sequence changes (see ‘Variant annotation’ section). Variants potentially affecting post-translational modifications were downloaded from the ‘A Website Exhibits SNP On Modification Event’ database ( ) 104 . The output from each tool was converted to a binary measure to indicate deleterious or tolerated predictions. Scoring hierarchy Each target gene prediction category (distal, promoter or coding) was scored according to different criteria. Genes predicted to be distally regulated targets of CCVs were awarded points based on physical links (for example, CHi-C), computational prediction methods, allele-specific expression or expression variant associations. All CCVs and HPPVs were considered as potentially involved in distal regulation. Intersection of a putative distal enhancer with genomic features found to be significantly enriched (see ‘Genomic features enrichment’ for details) were further upweighted. Multiple independent interactions were awarded an additional point. CCVs and HPPVs in gene proximal regulatory regions were intersected with histone ChIP-Seq peaks characteristic of promoters and assigned to the overlapping transcription start sites (defined as −1.0 kb to +0.1 kb). Further points were awarded to such genes if there was evidence of expression variant association or allele-specific expression, while a lack of expression resulted in down-weighting as potential targets. Potential coding changes, including missense, nonsense and predicted splicing alterations, resulted in the addition of one point to the encoded gene for each type of change, while lack of expression reduced the score. We added an additional point for predicted target genes that were also breast cancer drivers. For each category, scores ranged from 0 to 7 (distal), 0–3 (promoter) or 0 to 2 (coding). We converted these scores into ‘confidence levels’: level 1 (highest confidence; distal score > 4, promoter score ≥ 3 and coding score > 1); level 2 (1 ≤ distal score ≤ 4, promoter score = 1 or 2 and coding score = 1); and level 3 (0 < distal score < 1, 0 < promoter score < 1 and 0 < coding < 1). For genes with multiple scores (for example, those predicted as targets from multiple independent risk signals or predicted to be impacted in several categories), we recorded the highest score. Driver and transcription factor gene enrichment analysis was carried out using INQUISIT scores before adding a point for driver gene status. Modifications to the pipeline since original publication 2 included: Topologically associated domain boundary definitions from ENCODE T-47D Hi-C analysis. Previously, we used regions from ref. 98 . eQTL (addition of NHS normal and tumor samples). Allele-specific imbalance using TCGA and Genotype-Tissue Expression RNA-Seq data 102 . Capture Hi-C data from six breast cell lines 96 . Additional bio-features derived from global enrichment in this study. Variants affecting sites of post-translational modification 104 . Multi-signal targets To test whether more genes were targeted by multiple signals than would be expected by chance, we modeled the number of signals per gene by negative binomial regression (R function glm.nb; package MASS) and Poisson regression (R function glm; package stats) with ChIA-PET interactions as a covariate, and adjusted by fine-mapping region. Likelihood ratio tests were used to compare goodness of fit. Rootograms were created using the R function rootogram (package vcd). Pathway analysis The pathway gene set database dated 1 September 2018 was used 105 ( ). This database contains pathways from Reactome 106 , the NCI Pathway Interaction Database 107 , Gene Ontology 108 , HumanCyc 109 , MSigdb 110 , NetPath 111 and Panther 112 . All duplicated pathways, defined in two or more databases, were included. To provide more biologically meaningful results, only pathways that contained ≤200 genes were used. We interrogated the pathway annotation sets with the list of high-confidence (level 1) INQUISIT genes. The significance of over-representation of the INQUISIT genes within each pathway was assessed with a hypergeometric test using the R function phyper as follows: $$P\left( {x|n,m,N} \right) = 1 - \mathop {\sum }\limits_{i = 0}^{x - 1} \frac{{\left( {\begin{array}{*{20}{c}} m \\ i \end{array}} \right)\left( {\begin{array}{*{20}{c}} {N - m} \\ {n - i} \end{array}} \right)}}{{\left( {\begin{array}{*{20}{c}} N \\ n \end{array}} \right)}}$$ where x is the number of level 1 genes that overlap with any of the genes in the pathway, n is the number of genes in the pathway, m is the number of level 1 genes that overlap with any of the genes in the pathway dataset ( m strong GO = 145; m ER-positive GO = 50; m ER-negative GO = 27; m ER-neutral GO = 73; m strong pathways = 121; m ER-positive pathways = 38; m ER-negative pathways = 21; m ER-neutral pathways = 68) and N is the number of genes in the pathway dataset ( N genes GO = 14,252; N genes pathways = 10,915). We only included pathways that overlapped with at least two level 1 genes. We used the Benjamini–Hochberg FDR 113 at the 5% level. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The credible set of causal variants (determined by either multinomial stepwise regression or PAINTOR) is provided in Supplementary Table 2c . Further information and requests for resources should be directed to M.K.B. ([email protected]).

A major international study of the genetics of breast cancer has identified more than 350 DNA 'errors' that increase an individual's risk of developing the disease. The scientists involved say these errors may influence as many as 190 genes. The results, published today in the journal Nature Genetics, provide the most comprehensive map of breast cancer risk variants to date. The researchers involved, from over 450 departments and institutions worldwide, say the findings will help provide the most detailed picture yet of how differences in our DNA put some women at greater risk than others of developing the disease. The majority of the DNA is identical between individuals, but there are some differences, known as genetic variants, and these changes can have a profound effect, increasing an individual's susceptibility to disease. Our DNA—the blueprint for the human body—contains between 20,000-25,000 genes. Many of these code for proteins, the building blocks that make up the human body. Genetic variants can be located within genes, altering the protein. However, most of genetic variants are located outside genes, sometimes regulating the function of genes, turning their 'volume' up or down or even off. Finding which gene is targeted by these variants is not straightforward. Most diseases are complex, polygenetic diseases—in other words, no single genetic variant or gene causes the disease, but rather the combination of a number of them act together to increase the likelihood that an individual will develop a particular disease. Breast cancer is one such disease. Previous genome-wide association studies (GWAS), which involve comparing the DNA of patients against that of healthy controls, have found around 150 regions of the genome that clearly affect breast cancer risk. Within these regions, researchers know there are one or more genetic changes that affect the risk of developing cancer, but rarely are they able to pinpoint the specific variants or genes involved. Fine-mapping studies, such as this one, allow scientists to narrow down which variants contributing to the disease, how they might work and predict which are the genes involved. "We know from previous studies that variants across our DNA contribute towards breast cancer risk, but only rarely have scientists have been able to identify exactly which genes are involved," said Dr. Laura Fachal from the Wellcome Sanger Institute. "We need this information as it gives us a better clue to what is driving the disease and hence how we might treat or even prevent it." In this new study, researchers from hundreds of institutions worldwide collaborated to compare the DNA of 110,000 breast cancer patients against that of some 90,000 healthy controls. By looking in much closer detail than was previously possibly, they identified 352 risk variants. It is not yet clear exactly how many genes these target, but the researchers have identified 191 genes with reasonable confidence; less than one in five of these had been previously recognised. "This incredible haul of newly-discovered breast cancer genes provides us with many more genes to work on, most of which have not been studied before," said Dr. Alison Dunning from the University of Cambridge. "It will help us build up a much more detailed picture of how breast cancer arises and develops. But the sheer number of genes now known to play a role emphasises how complex the disease is." Of the newly-discovered genetic variants, a third predispose women towards developing hormone-responsive breast cancer, the type of disease found in four out of five breast cancer patients, which respond to hormonal treatments such as tamoxifen. 15% of the genetic variants predispose women to the rarer type, estrogen-receptor-negative breast cancer. The remaining genetic variants play a role in both types of breast cancer. In the majority of cases, the genetic change affected gene expression—in other words, how active a particular gene was and how much of a particular protein it created—rather than altering the type of protein itself. For instance, nine different variants regulate the same gene, the Estrogen Receptor (ESR1) gene. Many other variants affect places in the DNA where the Estrogen Receptor protein binds, and, in turn regulates other genes. This highlights the importance of the ESR1 gene and its protein product, the Estrogen Receptor, in breast cancer development. While each genetic variant only increases the risk of developing breast cancer by a very small amount, the researchers say that added together, these will allow them to 'fine tune' genetic testing and give women a much clearer picture of their genetic risk. This will then allow doctors and clinicians to provide advice on the best strategy for reducing their risk and preventing onset of the disease. Professor Doug Easton, also from the University of Cambridge, said: "Our work would not have been possible without the help of the 200,000 volunteers who allowed us to study their DNA. It is also testament to the work of hundreds of researchers from all over the world who collaborated on this study."

10.1038/s41588-019-0537-1

Nano

Research team develops new compact and energy-efficient nanoscale microwave oscillators

The paper, "Ultralow-current-density and bias-field-free spin-transfer nano-oscillator," has been published online in the journal Scientific Reports (Nature Publishing Group). www.nature.com/srep/2013/13031 … /full/srep01426.html Journal information: Scientific Reports

http://www.nature.com/srep/2013/130312/srep01426/full/srep01426.html

https://phys.org/news/2013-03-team-compact-energy-efficient-nanoscale-microwave.html

Abstract The spin-transfer nano-oscillator (STNO) offers the possibility of using the transfer of spin angular momentum via spin-polarized currents to generate microwave signals. However, at present STNO microwave emission mainly relies on both large drive currents and external magnetic fields. These issues hinder the implementation of STNOs for practical applications in terms of power dissipation and size. Here, we report microwave measurements on STNOs built with MgO-based magnetic tunnel junctions having a planar polarizer and a perpendicular free layer, where microwave emission with large output power, excited at ultralow current densities and in the absence of any bias magnetic fields is observed. The measured critical current density is over one order of magnitude smaller than previously reported. These results suggest the possibility of improved integration of STNOs with complementary metal-oxide-semiconductor technology and could represent a new route for the development of the next-generation of on-chip oscillators. Introduction Oscillators are basic components of many communication, navigation and measurement systems. They provide a periodic output that can be used to generate electromagnetic energy for radiation, for high speed digital systems, clocking functions and to enable frequency up and down conversion if used as a local oscillator in communication circuits and for many other applications. There is a strong interest, driven by cost and performance, to develop improved microwave oscillators for on-chip integration and the spin-transfer nano-oscillator (STNO) is a promising candidate for this due to its combination of characteristics such as frequency tunability, nanoscale size, broad range of working temperature and relatively easy integration with complementary metal-oxide-semiconductor (CMOS) technology 1 , 2 , 3 , 4 , 5 , 6 , 7 . The STNO is not only one of the more complex non-linear dynamical systems (of great interest from fundamental point of view), but it is also important being one of the smallest auto-oscillators known in nature (hence of great interest from technological point of view). The remaining challenges that need to be addressed are the simultaneous optimization of a variety of properties, including high frequency operation, frequency tunability, narrow spectral linewidth, large output power and operation in both absence of external magnetic fields and with low applied current densities. Here, we demonstrate that the interfacial perpendicular anisotropy (IPA) between the ferromagnetic electrodes and the tunnel barrier of magnetic tunnel junctions (MTJs) employing the material combination of CoFeB/MgO is a key ingredient to improve the dynamic properties of STNOs 8 . In general, IPA can be used for the design of high efficiency spintronic devices. It permits the realization of MTJs for storage applications with high thermal stability and low switching current 8 , 9 and STNOs with very large emitted power ( i.e. > 0.95 μW delivered to a matched load) 10 . In previous studies, IPA was used to orient the magnetization of both the free and polarizer layers out of the film plane 8 , or to reduce the out-of-plane demagnetizing field ( H d ) while maintaining the orientation of both of the two magnetizations in the film plane 10 . In this work, we designed STNOs with an asymmetric MTJ consisting of an in-plane polarizer and a perpendicular free layer achieved by controlling the trade-off between IPA and H d . This orthogonal magnetic configuration is different from the one in our previous study 10 and is more efficient in order to excite large angle free layer precession (large output power) in absence of a bias magnetic field. The main objective of our experiment is to solve the key technological issues related to the integration of STNOs with CMOS technology: elimination of the need for an external magnetic bias field and reduction of drive current densities. The former goal will allow for eliminating the current lines necessary to create the magnetic field for the STNO biasing, while the latter is necessary to reduce the size of the current driving transistors (small currents allow for smaller transistors, hence making the overall oscillator smaller. For high-current oscillators, the actual size of the oscillator would be determined by the transistor size, rather than by the MTJ) and to control power dissipation. Results Spin-transfer torque oscillator The samples studied have a core magnetic stack consisting of a synthetic antiferromagnet (SAF) Co 70 Fe 30 /Ru/Co 40 Fe 40 B 20 layer and a Co 20 Fe 60 B 20 free layer (FL) separated by a MgO insulator as shown in Fig. 1a . We introduce a Cartesian coordinate system where the x-axis is the direction of the polarizer (+x), the y-axis and the z-axis are the hard in-plane and the out-of-plane axes respectively. The SAF layer is designed to have an in-plane easy axis 9 , 10 serving as a polarizer, while the Fe-rich Co 20 Fe 60 B 20 FL with a thickness of t = 1.60 ~ 1.62 nm was chosen to achieve the proper IPA, which favors the out-of-plane (perpendicular) magnetic configuration in the free layer (see Methods). Electron-beam lithography and ion milling were used to define and etch the MTJs resulting in pillar-shaped devices with nominal dimensions of 150 nm × 70 nm. The samples are different from the 50–70 nm diameter point contact spin valves with out-of-plane magnetized CoNi free layers studied in refs. 11 , 12 . A d.c. bias current is injected into the sample through a bias Tee as shown in the measurement setup in our previous work 10 , where we define the positive current I as electrons flowing from the polarizer to the free layer. A time-varying voltage produced by the oscillations of the magnetization via the tunneling magnetoresistance (TMR) effect is recorded using a 9 kHz–26.5 GHz spectrum analyzer. The measurements were carried out at room temperature. We show detailed data from a single device with free layer thickness t = 1.60 nm, however similar results have been achieved for more than 5 devices for each free layer thickness. Figure 1 Sample structure and properties. (a) Schematic of the sample layer structure consisting of an in-plane magnetized fixed (polarizer) layer and an out-of-plane magnetized free layer. (b) Resistance as a function of in-plane magnetic field ( H ∥ ) and perpendicular magnetic field ( H ⊥ ) for sample 1 ( t = 1.60 nm), inset in (b) is the resistance as a function of H ⊥ , the black (red) arrow denotes the magnetization direction of the reference (free) layer. (c) Resistance-Current curve at zero applied magnetic field, AP (P) denotes the antiparallel (parallel) configurations between the free and fixed layers. (d) Microwave spectra as a function of d.c. current bias I at zero applied magnetic field, the curves are offset by approximately 20 nW GHz −1 along the vertical axis for clarity. Inset: full width at half maximum (FWHM, or linewidth) (triangles) and f 0 (circles) of the STNO sample 1 as a function of I . Full size image Magneto-resistance transport properties Figure 1b shows the resistance of a typical nanopillar device as a function of the external in-plane field ( H || ) at a sub-critical bias current of I = 10 μA, revealing a TMR ratio of 79%. As H || increases from −700 Oe to +1100 Oe, the resistance increases gradually as the magnetization of the free layer changes gradually from parallel to antiparallel relative to the polarizer. One source of asymmetry of the field scan is the dipolar coupling between the polarizer and the free layer, which (at low bias currents and zero external field) induces the free layer magnetization to tilt at a small angle away from z-axis ( Fig. 1a ). The additional switching near H || = −700 Oe and +1100 Oe corresponds to reorienting of the polarizer magnetization. The resistance curve scan as function of the out-of-plane field (inset of Fig. 1b ) clearly indicates the perpendicular free layer. Spin-transfer torque dynamics The devices exhibit dynamical behavior in a range of external fields (see Supplementary Fig. S1 online), but in the rest of the paper we will focus on the data measured at zero bias magnetic field, being the most important result of this experiment. Fig. 1c shows the resistance versus applied current I , an intermediate resistance (IR) state is seen in the range of I = −0.34 to + 0.65 mA, while the high resistance state (AP) for I < −0.34 mA and the low resistance state (P) for I > 0.65 mA are observed. The large-power microwave emission is only observed in the IR state and for negative currents, i.e. B1 regime. We focused on frequency-domain measurements to establish the existence and study the characteristics (oscillation frequency f 0 , linewidth Δ f 0 and output power) of these STNOs. Fig. 1d displays typical microwave power density spectra obtained for different values of I for sample 1 ( t = 1.60 nm). For all of the current values (| I| ≤ 0.3 mA), a single peak (with Lorentzian line shape) is observed in the GHz-range that is characterized by the excitation of an out-of-plane oscillation mode as indicated by the micromagnetic simulations (see Methods). The oscillation frequency exhibits a red shift, consistent with previous studies in point contact geometries with in-plane polarizer and out of plane FL 11 , with a slope of 1.75 GHz mA −1 and a minimum linewidth of 28 MHz (inset Fig. 1d ). The estimated threshold current I c is ~ −45 μA (see Supplementary Note 2 online) 13 , which corresponds to a current density of J c = −5.4 × 10 5 A/cm 2 , this value can be further reduced by controlling the thickness of the free layer t , e.g. at t = 1.62 nm J c = −1.2 × 10 5 A/cm 2 ( I c = −10 μA), which is to the best of our knowledge the smallest value measured to date. The observation of the ultralow J c value can be explained as follows. In the presence of the IPA the J c is approximately proportional to the effective demagnetizing field 14 ( , where is the IPA in the FL, which increases with decreasing the FL thickness 8 , 9 ). In our case, the strong IPA in the FL significantly counteracts the effect of H d , resulting in the ultralow J c and is responsible for the strong dependence of J c on the FL thickness. As can be observed, the effect of the FL thickness on IPA has a non-trivial influence on the dynamical properties of the STNOs (for example a larger thickness corresponds to a smaller critical current for ) and a complete systematic study will be presented elsewhere. Fig. 2 displays a comparison between the oscillation power as a function of the current for the two thicknesses t = 1.60 and 1.62 nm. A complete set of data for t = 1.62 is included in the Supplementary Note 3 online (the dynamical properties of the STNO for t = 1.60 and 1.62 nm are summarized in the last two rows of Table 1 ). The maximum oscillation power is larger at t = 1.60 nm with the best result achieved for a current I = −0.3 mA where the measured output power is 18 nW (> 60 nW delivered to a matched load, see Supplementary Note 4 online) with an oscillation frequency and a linewidth of 850 MHz and 73 MHz, respectively. In the absence of external fields, an additional property is the reduction of total power dissipation, at the best working condition for t = 1.60 nm the dissipated power is of the order of 50–70 μW which is a very small value for oscillators at sub-micrometer size. Hence, these results open the possibility to design nanoscale oscillators with improved dynamic performance and lower power dissipation. Table 1 Comparison among the performance parameters (external field, critical current density, oscillation power, oscillation frequency and the minimum linewidth) of different STNO solutions and the ones reported in this work at zero field and for two thicknesses t (1.60 nm and 1.62 nm). Here P mes is the measured maximum power integrated from the oscillation peak, P max is the maximum power delivered to a matched load and the hyphen (−) indicates cases where no data are available Full size table Figure 2 Power of STNO samples. Integrated power of fundamental signals at zero applied magnetic field as a function of I . Olive circles are for 1.60 nm sample and blue squares for 1.62 nm sample. Full size image Comparison with simulations We performed micromagnetic simulations considering the same experimental framework to identify the origin of the steady-state magnetization oscillation. The simulations have been performed by solving the Landau-Lifshitz-Gilbert-Slonczewski equation (See Methods: Micromagnetic simulations). As in the experimental data, the magnetization precession is observed at negative currents and it is related to the excitation of a mode with an out-of-plane oscillation axis and a trajectory which is initially circular and expands (i.e. the output power increases) as the current increases (compare trajectories at −82 and −164 μA in Fig. 3 (left)). At large currents the oscillation axis moves towards the x-y plane (trajectory at −288 μA) and for values of I < −0.34 mA the AP-state is obtained as in the experimental data. The magnetization dynamics are characterized by the excitation of a spatially quasi-uniform mode (see also Supplementary Video online). Fig. 3 (right) shows two examples of snapshots at I = −0.3 mA, the arrows indicate the in-plane component of the magnetization while the colors are related to the m x (blue negative, red positive). At high currents ( I < −0.3 mA), the simulations show that for some range of time the dynamics is switched off (see Supplementary Fig. S4 online). We believe that the presence of this behavior could be the origin of the non-uniform dynamics (experimental low frequency tail) measured in the high current regime (see Supplementary Fig. S3 online). We observed that this behavior is more evident in the range of thickness where the IPA is comparable with the out-of-plane demagnetizing field. The simulations also predict a decrease in the oscillation frequency as a function of I at a rate of ~1.8 GHz mA −1 , this value is consistent with the experimentally observed rate of ~1.75 GHz mA −1 ( Fig. 4 ). Figure 3 Micromagnetic simulations for sample 1. Left. Trajectories of the average magnetization vector on the unit sphere as computed from micromagnetic simulations ( I = −82, −164 and −288 μA). Right: example of two snapshots of the spatial distribution of the magnetization indicating the uniform dynamics (the color means the x-component of the magnetization blue negative, red positive). Full size image Figure 4 Dependence of microwave frequencies on current for sample 1. The blue circles show experimental data as a function of current bias at zero applied magnetic field. The black squares show the results from micromagnetic simulations. Full size image Discussion Microwave emission with large output power (in the presence of external magnetic fields and large drive current densities) has been already measured in STNOs with IPA 10 , 15 . However, to the best of our knowledge this work is the first experimental demonstration of large oscillation power without bias field and with ultralow current density J c below 6 × 10 5 A/cm 2 . For microwave generation in the absence of external magnetic fields, various other solutions have been recently proposed. For example, the idea based on a wave-like angular dependence of the spin torque 16 , the incorporation of a perpendicular polarizer into spin-valve structures 17 , magnetic vortex oscillations 6 , 18 or a tilted free layer 19 , but the resulting microwave power in those STNOs is smaller (<1 nW) and the excitation current is higher (>10 6 A/cm 2 ) 11 , 16 , 17 , 18 . When compared with the best results obtained from previous field-free STNOs, the data in this work show microwave emission with output power at least one order of magnitude larger and critical current densities one order of magnitude smaller. Table 1 compares the dynamical properties of STNO solution achieved without magnetic field. Importantly, the oscillation frequency is of the same order as that obtained from other types of STNOs. Finally, the STNOs in this work exhibit a tunability of ~1.75 GHz mA −1 that is substantially larger than the ones measured in spin-transfer driven vortex oscillations (0.03 GHz mA −1 in Ref. 6 and 0.08 GHz mA −1 in Ref. 18 ). This, however, comes at the cost of the larger linewidth compared to those obtained in magnetic vortex self-oscillations 6 , 18 . The origin of the linewidth is related primarily to thermal fluctuations and to the coupling between the oscillator phase and power as in other STNOs 13 . One possible remedy is the use of these STNOs as a basis of an array of phase locked STNOs, in that scenario it is expected that a significant decrease in the linewidth and an increase in the output power to over the tens of μW 20 , 21 could be achieved. A different solution would be the application of a low frequency current modulation as demonstrated in Ref. 22 . The possibilities opened by these results will eliminate some of the key issues related to on-chip integration of STNOs with CMOS technology. This direct integration and the reduced power consumption may potentially open applications in portable electronic devices and wireless modules such as embedded communications and power efficient local clock signal generation in digital systems. Our findings also provide a key ingredient in the development of ultralow–critical-current and zero-field spin-wave sources in magnonic logic devices 23 . Methods Sample preparation The magnetic stacks were deposited by sputtering in a Singulus TIMARIS PVD system and contain a layer structure: PtMn (15)/Co 70 Fe 30 (2.3)/Ru (0.85)/Co 40 Fe 40 B 20 (2.4)/MgO (0.8)/Co 20 Fe 60 B 20 (1.5 ~ 2.0) (the thicknesses are in nm). Details of the growth and fabrication process are similar to that in our previous work 9 . Note that the IPA gradually increases as the Co 20 Fe 60 B 20 FL thickness ( t ) decreases 8 , 9 . The direction of the FL magnetic moments is mainly determined by the competition between the IPA ( ) and the H d . For a thicker FL, e.g. t = 1.70 nm 10 , it has been verified that the magnetization is in-plane and it requires an external magnetic field to produce microwave signals. By contrast, the thinner t < 1.60 nm results in a larger and consequently a larger critical current density. We find that, the microwave oscillations with no external field exist in a thickness range from 1.55 to 1.65 nm and that in particular, for the FL thickness between 1.60 and 1.62 nm, better microwave performance in terms of power and drive current density is achieved. Micromagnetic simulation We numerically solve the Landau-Lifshitz-Gilbert-Slonczewski equation which also includes the field-like torque term T OP 24 , 25 . T OP is considered dependent on the square of the bias voltage 26 up to a maximum value of 25% of the in-plane torque computed for a current density | J | = 6.0 × 10 6 A/cm 2 . The total torque, including also the in-plane component T IP is: where g is the gyromagnetic splitting factor, is the gyromagnetic ratio, is the Bohr magneton, q(V) is the voltage dependent parameter for the perpendicular torque, is the spatial dependent current density, V is the voltage (computed from Fig. 1c in the main text), t is the thickness of the free layer and e is the electron charge 27 . The effective field takes into account the standard micromagnetic contributions (exchange, self-magnetostatic) and the magnetic coupling between free and polarizer layer as well as the Oersted field due to I . The presence of the IPA has been modeled as an additional contribution to the effective field. The parameters used for the CoFeB have been: saturation magnetization M s = 1.1 × 10 6 A/m, IPA constant k u = 7.4 × 10 5 J/m 3 , exchange constant A = 2.0 × 10 −11 J/m, damping parameter α = 0.01. The polarization function has been computed by Slonczewski 28 , 29 , being not dependent on the bias voltage, where m and m p are the normalized magnetizations of the free and polarizer layer, respectively. To fit the critical current density, we use for the spin-polarization the value 0.57 30 . Micromagnetic simulations provide useful insight into the nature of the excited mode in the STNO with IPA. They indicate the excitation of an out-of plane mode characterized by a quasi-uniform spatial distribution of the magnetization.

By using improved magnetic materials, based on the control of interface properties of ultra-thin magnetic films, researchers from the Suzhou Institute of Nano-tech and Nano-bionics, Chinese Academy of Sciences (SINANO), the University of California at Los Angeles(UCLA), and the University of Messinahave made major experimental improvements to develop a more compact, more energy-efficient generation of a mobile communication device known as spin transfer nano-oscillator (STNO). STNOs use the spin of electrons to create steady microwave oscillations needed for various applications in mobile communications, unlike current silicon-based oscillators which use their charge. The SINANO team's improved oscillator has great potential to be used in future portable electronic devices and wireless modules, systems on a chip, and for power-efficient local clock signal generation in digital systems. The STNOs are composed of two distinct magnetic layers. One layer has a fixed magnetic polar direction, while the other layer's magnetic direction can be manipulated to gyrate by passing an electric current through it. This allows the structure to produce very precise oscillating microwaves. The STNO's key advantage over existing technologies is that it can combine large tunability and low energy with nanoscale size, as well as broad working temperature ranges. Yet while STNOs are potentially superior in many respects to existing microwave oscillator technologies, their microwave signals mainly rely on both large drive currents and the application of external magnetic fields, which hinders the implementation of STNOs for practical applications in terms of power dissipation and size. By using magnetic layers with perpendicular magnetic anisotropy– similar to those used in spin-transfer torque memory – the SINANO team demonstrated large microwave signals at ultralow current densities (<5.4×105A/cm2) and in the absence of any bias magnetic fields. This eliminates the need to move large numbers of electrons through wires, and also eliminates the need for permanent magnets or conducting coils to provide the bias magnetic field, thus significantly saving both energy and space. The results are microwave oscillators that generate much less heat due to their lower current, making them more energy-efficient. "Previously, there had been no demonstration of a spin-transfer oscillator with sufficiently high output power, low drive current density, and simultaneously without the need for an external magnetic field, hence preventing practical applications," said the lead researcher ZENG Zhongming, SINANO professor at the SINANO Nanofabrication Facility. "We have realized all these requirements in a single device." "The ability to excite microwave signals at ultralow current density and in zero magnetic field is exciting in nano-magnetism. This work presents a new route for the development of the next-generation of on-chip oscillators." said co-author G. Finocchio, who is an assistant professor at the University of Messina, Italy. "Ultra-low-power spintronic devices have the potential to transform the electronics industry, with the most immediate example being in the area of nonvolatile magnetic memory (MRAM). This work shows that similar materials and devices may also bring nanoscale spintronic oscillators one step closer to reality," said Pedram Khalili, a research associate and program manager at UCLA and co-author of the paper. "Thesedevices can be integrated with standard CMOS logic manufacturing processes, enabling a wide range of products from standalone memory and microwave components to systems on a chip."

www.nature.com/srep/2013/13031 … /full/srep01426.html

Other

Six degrees of separation: Why it is a small world after all

Nicholas Jarman et al, Self-organisation of small-world networks by adaptive rewiring in response to graph diffusion, Scientific Reports (2017). DOI: 10.1038/s41598-017-12589-9 Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-017-12589-9

https://phys.org/news/2017-10-degrees-small-world.html

Abstract Complex networks emerging in natural and human-made systems tend to assume small-world structure. Is there a common mechanism underlying their self-organisation? Our computational simulations show that network diffusion (traffic flow or information transfer) steers network evolution towards emergence of complex network structures. The emergence is effectuated through adaptive rewiring: progressive adaptation of structure to use, creating short-cuts where network diffusion is intensive while annihilating underused connections. With adaptive rewiring as the engine of universal small-worldness, overall diffusion rate tunes the systems’ adaptation, biasing local or global connectivity patterns. Whereas the former leads to modularity, the latter provides a preferential attachment regime. As the latter sets in, the resulting small-world structures undergo a critical shift from modular (decentralised) to centralised ones. At the transition point, network structure is hierarchical, balancing modularity and centrality - a characteristic feature found in, for instance, the human brain. Introduction Complex network structures emerge in protein 1 and ecological networks 2 , social networks 3 , the mammalian brain 4 , 5 , 6 , and the World Wide Web 7 . All these self-organising systems tend to assume small–world network (SWN) structure. SWNs may represent an optimum in that they uniquely combine the advantageous properties of clustering and connectedness that characterise, respectively, regular and random networks 8 . Optimality would explain the ubiquity of SWN structure; it does not inform us, however, whether the processes leading to it have anything in common. Here we will consider whether a single mechanism exists that has SWN structure as a universal outcome of self-organisation. In the classic Watts and Strogatz algorithm 9 , a SWN is obtained by randomly rewiring a certain proportion of edges of an initially regular network. Thereby the network largely maintains the regular clustering, while the rewiring creates shortcuts that enhance the networks connectedness. As it shows how these properties are reconciled in a very basic manner, the Watts-Strogatz rewiring algorithm has a justifiable claim to universality. However, the rewiring compromises existing order than to rather develop over time and maintain an adaptive process. Therefore the algorithm is not easily fitted to self-organising systems. In self-organising systems, we propose, network structure adapts to use - the way pedestrians define walkways in parks. Accordingly, we consider the effect of adaptive rewiring: creating shortcuts where network diffusion (traffic flow or information transfer) is intensive while annihilating underused connections. This study generalises previous work on adaptive rewiring 10 , 11 , 12 , 13 , 14 . While these studies have shown that SWN robustly emerge through rewiring according to the ongoing dynamics on the network, the claim to universality has been frustrated by need to explicitly specify the dynamics. Here we take a more general approach and replace explicit dynamics with an abstract representation of network diffusion. Heat kernels 15 capture network-specific interaction between vertices and as such they are, for the purpose of this article, a generic model of network diffusion. We study how initially random networks evolve into complex structures in response to adaptive rewiring. Rewiring is performed in adaptation to network diffusion, as represented by the heat kernel. We systematically consider different proportions of adaptive and random rewirings. In contrast with the random rewirings in the Watts-Strogatz algorithm, here, they have the function of perturbing possible equilibrium network states, akin to the Boltzmann machine 16 . In this sense, the perturbed system can be regarded as an open system according to the criteria of thermodynamics. In adaptive networks, changes to the structure generally occur at a slower rate than the network dynamics. Here, the proportion of these two rates is expressed by what we call the diffusion rate (the elapsed forward time in the network diffusion process before changes in the network structure). Low diffusion rates bias adaptive rewiring to local connectivity structures; high diffusion rates to global structures. In the latter case adaptive rewiring approaches a process of preferential attachment 17 , 18 . We will show that with progressive adaptive rewiring, SWNs always emerge from initially random networks for all nonzero diffusion rates and for almost any proportion of adaptive rewirings. Depending on diffusion rate, modular or centralised SWN structures emerge. Moreover, at the critical point of phase transition, there exists a network structure in which the two opposing properties of modularity and centrality are balanced. This characteristic is observed, for instance, in the human brain 19 , 20 , 21 . We call such a structure hierarchical. In sum, adaptation to network diffusion represents a universal mechanism for the self–organisation of a family of SWNs, including modular, centralised, and hierarchical ones. Results For each pair ( τ , p ), where τ is a heat kernel parameter, and p is the rewiring probability, 100 independently evolved networks are obtained by our adaptive rewiring algorithm (see Methods). Resulting networks are described according to measures of small-world structure, modularity, and centrality. Where appropriate, the average of such measures is taken. Small-World Structure The small-worldness index S provides a canonical measure of the degree to which a network is small-world. Here, we take a slightly modified version, in which the normalised clustering coefficient ( C ) is multiplied by the normalised global efficiency ( E ) 8 , such that S = ( C / C r ) × ( E / E r ), where C r and E r are measures of C and E for an equivalent Erdös-Rényi (ER) random network 22 . In doing so, S is also defined on disconnected networks. For random networks, S ≈ 1 and so the greater the (positive) deviation of S from one, the greater the degree of small-worldness. For comparison, we include the average small-worldness values for networks constructed by the Watts-Strogatz algorithm (100 independently constructed networks for each p = 0, 1/500, 2/500, …, 1). In Fig. 1a we observe the average small-worldness index S as a function of random rewiring probability p . Unless stated otherwise, we consider networks arising from adaptive rewiring for random rewiring probability p ∈ {0, 1/30, …, 29/30, 1}. A striking result is that SWN emergence is observed for all sample values of τ nonzero, no matter how small or large. Moreover, for all nonzero τ that have been explored a greater maximum small-worldness is achieved than with the Watts-Strogatz algorithm. Figure 1 ( a ) Depicts the small-world index S as a function of decreasing random rewiring probability p ∈ {0, 1/30, …, 29/30, 1}: Coloured lines indicate values of heat kernel parameter \(\tau \in \mathrm{\{0,}\,\varepsilon ,\,\mathrm{1,}\,\mathrm{8,}\,\delta \}\) , black line indicates the Watts-Strogatz algorithm with random rewiring probability \(p\in \mathrm{\{0,}\,\mathrm{1/500,}\,\ldots ,\,\mathrm{499/500,}\,\mathrm{1\}}\) . ( b ) Depicts the average modularity Q as a function of decreasing random rewiring probability \(p\in \mathrm{\{0,}\,\mathrm{1/30,}\,\ldots ,\,\mathrm{29/30,}\,\mathrm{1\}}\) : Coloured lines indicate values of heat kernel parameter \(\tau \in \mathrm{\{0,}\,\varepsilon ,\,\mathrm{1,}\,\mathrm{8,}\,\delta \}\) . ( c,d ) Single trial. Example modular SWN. Adjacency matrices mapped to an n -by- n grid where rows (and columns) represent vertices and white indicates the existence of an edge. Rows and columns of adjacency matrices have been permuted to visualise the modules, in accordance with 28 . ( c ) \((\tau ,\,p)\,=\,(\varepsilon ,\,\mathrm{0.1)}\) ; ( d ) \((\tau ,\,p)\,=\,\mathrm{(1,}\,\mathrm{0.3)}\) . Full size image The degree of network adaptation to network diffusion, 1 − p , for which maximum small-worldness is obtained depends on the rate of diffusion τ . The values of τ are taken from the set {0, 10 −15 , 1, 8, 10 15 }. For the sake of convenience, we denote 10 −15 = ε , 10 15 = δ . For τ = 0, adaptive rewiring is a random process; emergent networks thus reflect those of the initial ER ones. For τ ∈ {8, δ }, the heat kernel reflects the degree distribution, and thus adaptive rewiring approaches a process of preferential attachment (see Methods). For maximum small-worldness, adaptive rewiring in response to local diffusion, for τ ∈ { ε , 1}, requires small p , i.e. small degree of random rewiring, while more global diffusion - preferential attachment -, for τ ∈ {8, δ }, requires larger p . Interestingly, in Supplementary Fig. S1 , for τ ∈ {8, δ } and p large, the network achieves an even greater efficiency than an equivalent random ER network, i.e. it is more well integrated. Modular Structure The modularity index Q is an optimised statistic of network partitioning into non-overlapping communities. The value Q is calculated as the proportion of intra-modular connections minus the expected proportion of inter-modular connections for an equivalent ER random network under the same community structure 23 , 24 . In Fig. 1b we observe the average modularity index Q as a function of random rewiring probability p . We observe that modularity can be controlled by choice of pair ( τ , p ). This is discussed in further detail in the section Critical Network Structure . For τ ∈ { ε , 1} networks emerge with near-maximal degrees of modularity as p → 0. On the other hand, for τ ∈ {8, δ } and over all sampled p ∈ [0, 1] emergent networks posses no community structure, i.e. modularity is essentially switched off. In fact, we see a lesser degree of modularity than in an equivalent random ER network. In Fig. 1c,d we present the adjacency matrices, permuted to visualise the modules, from randomly sampled networks resulting from single independent trials of the algorithm with pairs ( τ , p ), where p = p ( τ ) is chosen dependent on τ so that S is at maximum. In both Fig. 1c,d where ( τ , p ) = ( ε , 0.1) and ( τ , p ) = (1, 0.3), respectively, emergent modules are relatively uniform with a dense intra-connectivity and sparse inter-connectivity. Centralised Structure Properties of centrality are characterised using the measures of PageRank, the degree, assortativity, and maximised coreness statistic. The latter three network measures are evaluated for pairs ( τ , p ) where p is chosen dependent on τ such that S is at maximum. The PageRank centrality vector, a variant of eigenvector centrality, is defined as the stationary distribution achieved by instantiating a Markov chain on a network 25 , 26 , i.e., the probability distribution that a random walker is located at a given vertex. PageRank centrality takes into account global communication patterns, mediated by longer path lengths and patterns of convergence and divergence, whereas some of the more common centrality measures, such as closeness and betweenness centrality, do not 27 . We denote as π the (normalised) maximum element of the PageRank vector. We take equal initial PageRank probability, and take the damping factor (the probability of transitioning to an adjacent vertex) as 0.85 28 , i.e., the probability of random vertex hopping is 0.15. Since the PageRank vector sums to one, then its mean value is 1/ n . For convenience we normalise π by this mean value. In Fig. 2a we observe the average PageRank value π as a function of random rewiring probability p . As with modularity, we see that centrality can be controlled depending on the choice of τ . For τ ∈ {8, δ } emergent network structures exhibit values of π considerably (positively) far from that of the ER networks, indicating large deviations of the maximum component from the mean of the PageRank vector. Therefore, there exists at least one vertex having a greatly increased likelihood of being traversed in a random Markov chain than all others. On the other hand, for τ ∈ { ε , 1} and over all sampled p ∈ [0, 1], emergent networks posses no such central vertices, i.e. centrality is tuned off. In fact, we see a lesser degree of centrality than in an equivalent random ER network. Figure 2 ( a ) Depicts the average π - maximum element of PageRank vector normalised by its mean - as a function of decreasing random rewiring probability \(p\in \mathrm{\{0,}\,\mathrm{1/30,}\,\ldots ,\,\mathrm{29/30,}\,\mathrm{1\}}\) : Coloured lines indicate values of heat kernel parameter \(\tau \in \mathrm{\{0,}\,\varepsilon ,\,\mathrm{1,}\,\mathrm{8,}\,\delta \}\) . ( b ) Depicts the bar-plot in which the height of individual bars is the average number of vertices having degree d v , where \({d}_{v}\le 20\) . Inset bar-plot for vertex degrees d v , where 20 ≤ d v ≤ 70. Probability density function (PDF) curves fitted to d v : truncated normal PDF for \(\tau \in \mathrm{\{0,}\,\varepsilon ,\,\mathrm{1\}}\) and truncated and normalised lognormal PDF for \(\tau \in \mathrm{\{8,}\,\delta \}\) . Coloured bars (and curves) indicate values of heat kernel parameter \(\tau \) ; for each, \(p\) is chosen dependent on \(\tau \) such that S is at maximum. ( c,d ) Single trial. Example centralised SWN. Adjacency matrices mapped to an \(n\) -by- \(n\) grid where rows (and columns) represent vertices and white indicates the existence of an edge. Rows and columns of adjacency matrices have been permuted to visualise the modules, in accordance with 28 . ( c ) Depicts \((\tau ,\,p)\,=\,\mathrm{(8,\; 0.5667)}\) ; ( d ) depicts \((\tau ,\,p)=(\delta \mathrm{,\; 0.5667)}\) . Full size image In accordance with Fig. 1b , depending on the values of the pair ( τ , p ), emergent networks are either modular or centralised. The phase transition of network structure is discussed in the section Critical Network Structure . In Fig. 2b we present a bar-plot of networks’ degree distribution . The height of individual bars is the average number of vertices, over 100 independently evolved networks, having degree d v where d v = 1, …, 70. The degree distribution for τ ∈ { ε , 1} fits a truncated normal function, while for τ ∈ {8, δ } it fits a truncated log-normal function. Moreover, for τ ∈ {8, δ } vertices emerge having remarkably high degrees (=70). In Fig. 2c,d we present the adjacency matrices, permuted to visualise the modules, from randomly sampled networks resulting from single independent trials of the algorithm with pairs ( τ , p ), where p = p ( τ ) is chosen dependent on τ so that the values of S are at maximum. In both Fig. 2c where ( τ , p ) = (8, 0.5667), and Fig. 2d where ( τ , p ) = ( δ , 0.5667), we observe a small subset of hub vertices connecting to many peripheral vertices. The assortativity coefficient a describes the “assortative mixing” of vertex degrees, i.e. the preference for high-degree vertices to attach to other high-degree vertices 29 . In Table 1 row a , for τ ∈ {8, δ } and p = p ( τ ) such that S is at maximum, a strong negative correlation indicates that vertices of a high degree typically connect to vertices of a low degree. On the other hand, for τ ∈ { ε , 1} an approximately zero correlation indicates no preference of connections between vertices of varying degrees. Table 1 Column wise τ , p = p ( τ ) such that S is at maximum. Row wise: a assortativity coefficient; c maximised core-periphery statistic. Values presented are averages over trials. Full size table The maximised coreness statistic c measures the extent to which a network may be well-partitioned into two non-overlapping groups of vertices, a core and a periphery group 23 , 30 . In Table 1 row c , for τ ∈ {8, δ } and p = p ( τ ) such that S is at maximum, values close to one indicate that the network may be well-partitioned into non-overlapping groups of core and peripheral vertices. On the other hand, for τ ∈ { ε , 1} values close to zero indicate no such core-periphery partition. In sum, we note that for τ ∈ {8, δ }, and p = p ( τ ) such that S is at maximum, networks emerge as centralised, with a strong core, and that those core vertices connect to a high number of peripheral vertices. On the other hand, for τ ∈ { ε , 1}, and all sampled p ∈ [0, 1], networks exhibit none of these properties. Critical Network Structure We consider the transition between modularity and centrality, and show that at the phase transition of network structure, the two seemingly opposing properties are reconciled. Properties of modularity are characterised by Q while properties of centrality are characterised by π . In Fig. 3a,b , for parameters τ ∈ {4.50, 4.55, …, 5.45, 5.50}, and p ∈ {0, 1/30, …, 29/30, 1}, we present Q (Fig. 3a ), and π (Fig. 3b ), averaged over 100 independently evolved networks. In the domain ( τ , p ) there is a broad region of high modularity where both τ and p are low, and a broad region of high centrality in the remainder. Where the domain of modularity ends, the domain of centrality begins; the system exhibits a critical transition from modular (decentralised) to centralised structure as a function of the pair ( τ , p ). The phase transition region between the two is relatively sharp with respect to both τ and p . Figure 3 ( a,b) In the plane \(\tau \in \mathrm{\{4.50,}\,\mathrm{4.55,}\,\ldots ,\,\mathrm{5.45,}\,\mathrm{5.50\}}\) along the horizontal axis and random rewiring probability \(p\in \mathrm{\{0,}\,\mathrm{1/30,}\,\ldots ,\,\mathrm{29/30,}\,\mathrm{1\}}\) along the vertical axis. ( a) Depicts the modularity index \(Q\) ; ( b) Depicts π , the maximum element of the PageRank vector normalised by its mean value. ( c) For \(\tau \,=\,5\) , along the horizontal axis random rewiring probability \(p\in \mathrm{\{0.400,}\,\mathrm{0.402,}\,\ldots ,\,\mathrm{0.598,}\,\mathrm{0.600\}}\) . Along the vertical axis are \(Q\) the modularity index (left, blue), and \(\pi \) the maximum element of PageRank vector normalised by its mean value (right, red). ( d ) Single trial. Example critical SWN. Adjacency matrix mapped to an \(n\) -by- \(n\) grid where rows (and columns) represent vertices and white indicates the existence of an edge. Rows and columns of adjacency matrices have been permuted to visualise the modules, in accordance with 28 . Pair \((\tau ,p\mathrm{)\; =\; (5,\; 0.522)}\) . Full size image In Fig. 3c , we fix τ = 5 and take Q and π as functions of p ∈ {0.400, 0.402, …, 0.598, 0.600}, averaged over 100 independently evolved networks. It is clear that modularity and centrality are opposing, however, at the boundary of modularity and centrality, where they intersect at around p = 0.52, there is a small domain of p for which networks are a blend of both modular and central structure: each of Q and π are considerably large, indicating the presence of both network structures. Furthermore, the value of small-worldness for pair ( τ , p ) = (5, 0.522) is S = 5.32, indicating the network is also strongly small-world. In Fig. 3d we present the adjacency matrix, permuted to visualise the modules, resulting from one trial of the algorithm with ( τ , p ) = (5, 0.522). We observe a competition between modular and centralised structure; the simultaneous existence of densely connected communities (decentralised) and a core of high degree vertices connecting to many low degree peripheral vertices (centralised). In Supplementary Fig. S2a–d we present four additional randomly sampled networks resulting from single independent trials of the algorithm with ( τ , p ) = (5, 0.522). These additional figures support the notion that centrality and modularity are opposing, that at the point of phase transition they are reconciled, and that this is critical, i.e. they are competitive. The adjacency matrices exhibit some degree of both centrality and modularity: emergent networks may appear as more centralised (Fig. S2a–b ), or more modular (Fig. S2c ), or a blend of the two (Fig. S2d ). Conclusion and Discussion Small-world structure offers optimal efficiency in network communication 8 ; and has been shown to facilitate synchronisation in a range of oscillator networks 31 , 32 , 33 . Here, we studied whether a simple generic mechanism could be responsible for their emergence. We proposed a mechanism of network self-organisation that relies on ongoing network diffusion; over time, the network is rewired adaptively, rendering it conform to the patterns of network diffusion. With some probability p , the adaptive rewiring process is perturbed by random rewiring. Small-world structure emerged for almost any proportion of random rewiring, moreover, networks reached higher degrees of SWN structure than those in the Watts and Strogatz algorithm 9 . Patterns of network diffusion may be biased by local or global connectivity structures using the diffusion rate τ . For all (positive) nonzero diffusion rates SWN structure emerges; for small τ , SWNs are modular (decentralised), whereas for large τ , SWNs are centralised. For the latter, adaptive rewiring approaches a process of preferential attachment. Modularity versus centrality constitutes an important dimension in the characterisation of networks in the human brain, where they play a role in terms of (structural and functional) segregation and integration 34 , 35 . For intermediate values of τ and p there is a critical transition point at which network structures emerge that blend modularity and centrality. We may call these “hierarchical” 36 . Such networks are desirable for natural information processing systems like the human brain, in which a core of centralised components represents a global workspace and the decentralised modules represent autonomous client systems 19 , 20 , 21 . The criticality of these architectures renders them all but robust. At the level of the neuro-anatomy of the brain, it would probably involve dynamic maintenance to keep these architectures at the critical point. As a property of functional architecture, the criticality would render cognition extremely flexible, enabling rapid switching between centralised and modular processes 37 . Methods Here we will provide a formal definition of network diffusion, an algorithm for adaptive rewiring, and a description of a set of computational simulations to demonstrate the role of adaptive rewiring in the generation of small-world networks. The MATLAB code of the algorithm is included in the Supplementary materials. Notation In what follows we consider graphs that are undirected. A graph is an ordered pair G = ( V , E ) where V is the set of vertices and E is a subset of V × V called the edges. If X is a finite set, then | X | denotes its cardinality. The total number of vertices and edges in the graph are | V | = n and | E | = m , respectively. Two vertices u , v ∈ V are called adjacent if ( u , v ) ∈ E . For an n 1 × n 2 matrix B , B ij corresponds to the entry in the i -th row and j -th column, where \(i,j\,\in \,{{\mathbb{N}}}_{0}\) , and \(i\le {n}_{1}\) , \(j\le {n}_{2}\) . The adjacency matrix of a simple graph \(G\) is a square \(n\times n\) matrix \(A\) with entries \({a}_{uv}\,=\,1\,{\rm{if}}\,(u,v)\in E\) , and \({a}_{uv}\,=\,0\) otherwise. For undirected graphs \(A\) is symmetric. It is typically the case that \({a}_{uu}\,=\,0\) , i.e. no self–loops. The degree \({d}_{v}\) of a vertex \(v\) is the number of vertices adjacent to vertex \(v\) : \({d}_{v}={\sum }_{u\in V,u\ne v}{a}_{vu}\) . The matrix \(D\) is the diagonal matrix of degrees having entries \({D}_{uv}={d}_{u}\) if \(u=v\) and 0 otherwise. For a given set of \(n\) vertices V the complete graph is denoted as K n and its edge set is denoted as \({E}^{{K}_{n}}\) . The compliment of an edge set E , denoted as \({E}^{c}\) , is \({E}^{c}={E}^{{K}_{n}}\backslash E\) . Network Diffusion The Laplacian matrix of the graph \(G\) is \(L=D-A\) . The normalised Laplacian matrix, \( {\mathcal L} \) , is regarded as more appropriate for dealing with irregular graphs, $$ {\mathcal L} ={D}^{-\mathrm{1/2}}L{D}^{-\mathrm{1/2}}=I-{D}^{-\mathrm{1/2}}A{D}^{-\mathrm{1/2}}$$ (1) with the convention that \({D}_{uu}^{-1}\,=\,0\) for \({d}_{u}\,=\,0\) . All eigenvalues of \( {\mathcal L} \) are real (since \( {\mathcal L} \) is symmetric real) and confined to the interval \(\mathrm{[0,}\,\mathrm{2]}\) , in accordance with Gershgorin circle theorem 38 , and relate well to other graph invariants, such as random walks (or Markov chains), in a way that the eigenvalues of the Laplacian matrix \(L\) and adjacency matrices often fail to do 39 . Let \({\lambda }_{i}\) denote the eigenvalues of \( {\mathcal L} \) with eigenvectors \({v}_{i}\) , and \({\omega }_{i}\) the eigenvalues of the corresponding Markov process \(M\) with eigenvectors \({u}_{i}\) . Then, \({\lambda }_{i}\mathrm{=1}-{\omega }_{i}\) and \({v}_{i}={D}^{\mathrm{1/2}}{u}_{i}\) . Whereas \( {\mathcal L} \) incorporates information of the local connectivity of vertices, the introduction of a graph kernel provides a global connectivity metric. Physical processes such as diffusion suggest a natural way of constructing a kernel from such local information 15 . Network diffusion is formally represented by the Exponential Heat Kernel of the graph (cf. Theorem 10.11 in 39 ). Definition 1 ( Exponential heat kernel ) Let \( {\mathcal L} \) be the normalised Laplacian matrix for an \(n\times n\) real symmetric matrix and \(t\ge 0\) . The exponential heat kernel of \( {\mathcal L} \) , denoted by \(h(t)\) , is the symmetric and positive definite \(n\times n\) matrix , $$h(t)={e}^{-t {\mathcal L} }=\sum _{k\mathrm{=0}}^{\infty }\frac{{(-t)}^{k}}{k!}{ {\mathcal L} }^{k}\mathrm{.}$$ (2) In particular \(h\mathrm{(0)}\,=\,I\) , the identity matrix . The matrix exponential is a weighted sum of walks 40 . Coefficients \(\frac{{(-t)}^{k}}{k!}\) in Equation ( 2 ) allow for biasing of path lengths in the construction of \(h(t)\) , where for small \(t\) shorter paths carry greater weight and longer paths carry lesser weight, and vice versa . In our simulations we use the parameter \(\tau =t\) . The matrix \(h(t)\) as \(t\to \infty \) can be expressed by the leading eigenvector associated with the zero eigenvalue of \( {\mathcal L} \) . Since \( {\mathcal L} \) is real and symmetric, there exists an orthonormal matrix \(Q\) such that \( {\mathcal L} =Q{\rm{\Lambda }}{Q}^{-1}\) where \(Q\) is the matrix of eigenvectors and \({\rm{\Lambda }}\) is the diagonal matrix of eigenvalues. It is easily shown that substitution of this eigendecomposition into the Taylor expansion yields \(h(t)=Q{e}^{-t{\rm{\Lambda }}}{Q}^{-1}\) . Let \({{\rm{\Lambda }}}_{ii}={\lambda }_{i}\) and order the eigenvalues such that \(0\,=\,{\lambda }_{0}\le {\lambda }_{1}\le \cdots \le {\lambda }_{n-1}\) . If \(G\) is connected, then \( {\mathcal L} \) has one simple zero eigenvalue \({\lambda }_{0}\) . Then, the first column of \(Q\) contains the leading eigenvector, denoted as \(q\) , associated with the zero eigenvalue \({\lambda }_{0}\) . Then, \(q=\frac{{D}^{1/2}{\bf{\text{1}}}}{\sqrt{{{\bf{\text{1}}}}^{{\rm{T}}}D{\bf{\text{1}}}}}\) for \({\bf{\text{1}}}\) the \(n\) -vector of ones, and \({lim}_{t\to \infty }h(t)=q{q}^{{\rm T}}\) . If \(G\) is regular - all vertices have equal degree \(d={d}_{v}\) for all \(v\in V\) - then, \( {\mathcal L} =\frac{1}{d}L\) , and \(q\in {\rm{s}}{\rm{p}}{\rm{a}}{\rm{n}}\,(1,\ldots ,1)\) , hence \({lim}_{t\to {\rm{\infty }}}h(t)=\frac{1}{n}{{\bf{\text{11}}}}^{{\rm{T}}}\) . However, if \(G\) is irregular, then \(q\in {\rm{s}}{\rm{p}}{\rm{a}}{\rm{n}}\,(\sqrt{{d}_{1}},\ldots ,\sqrt{{d}_{n}})\) , thus \({lim}_{t\to {\rm{\infty }}}h(t)=\frac{1}{{\bf{\text{1}}}D{{\bf{\text{1}}}}^{{\rm{T}}}}{D}^{1/2}{{\bf{\text{1}}}}^{{\rm{T}}}{\bf{\text{1}}}{D}^{1/2}\) . The use of \( {\mathcal L} \) over \(L\) in construction of the heat kernel becomes apparent for \(G\) irregular. Assuming \(G\) is irregular, then the off-diagonal entries of \(h(t)\) as \(t\to \infty \) are proportional to the square root of the vertex degrees. Thus, for \(t\) taken arbitrarily large, irregularities in \( {\mathcal L} \) also appear in \(h(t)\) , i.e. information of network structure is still contained in \(h(t)\) . This property does not hold if we were to replace \( {\mathcal L} \) with \(L\) in the construction of \(h(t)\) . Indeed, denote the heat kernel constructed using \(L\) as \({h}_{L}(t)\) , then for \(G\) irregular, \({lim}_{t\to {\rm{\infty }}}{h}_{L}(t)=\frac{1}{n}{{\bf{\text{11}}}}^{{\rm{T}}}\) . Note also, that for \(\alpha \, > \,0\) where \(\alpha \) may be taken arbitrarily small, \(h(\alpha )\ne I\) , i.e. off-diagonal entries of \(h(\alpha )\) are nonzero, and hence \(h(\alpha )\) contains information of network structure. This property holds for both the use of \( {\mathcal L} \) and \(L\) in construction of the heat kernel. Adaptive rewiring algorithm Consider an undirected graph with number of vertices \(n\) and number of edges \(m\) . For convenience we take \(m=\lfloor 2\rho n(n-\mathrm{1)}\rceil \) , the nearest integer, where \(\rho =\frac{log(n)}{n}\) (natural logarithm), i.e. twice the critical connection density for which a random Erdös–Rényi (ER) graph is connected with probability one 41 , 42 . We consider self–organisation starting from a random network. The network is progressively rewired, with probability \(p\) at random and with probability \(1-p\) according to the current network diffusion. The process can be described in algorithmic form: Step 0: Generate an undirected random graph \(G\) of the Erdös–Rényi type. Begin with the graph \(G=(V,E)\) such that \(|V|=n\) and \(|E|\,=\,0\) . Select uniformly at random the pair \(u,v\) from the set \(\{u,v\in V|u\ne v,(u,v)\in {E}^{c}\}\) and add the (undirected) edge \((u,v)\) to the edge set \(E\) . Repeat until \(|E|=m\) . Step 1: Select a vertex \(v\) uniformly at random from all nonzero degree vertices \(v\in \{u\in V|{d}_{u}\ne 0\,{\rm{and}}\,{d}_{u}\ne n-\mathrm{1\}}\) . Step 2: Delete the edge \((v,{u}_{1})\) and add the edge \((v,{u}_{2})\) where vertices \({u}_{1}\) and \({u}_{2}\) are selected by the following criteria: With probability \(p\) go to 2i, otherwise go to 2ii, i. Vertices \({u}_{1}\) and \({u}_{2}\) are selected uniformly at random from the sets \({u}_{1}\in \{u\in V|(v,u)\in E\}\) and \({u}_{2}\in \{u\in V|(v,u)\in {E}^{c}\}\) . ii. For adjacency matrix \(A\) (of graph \(G\) ), calculate the heat kernel \(h(t)\) for \(t=\tau \) , where \(\tau \) is a chosen parameter. Vertices \({u}_{1}\) and \({u}_{2}\) are chosen such that, for all \(u\in V\) and \(u\ne v\) , $${u}_{1}:{h}_{v{u}_{1}}(t)\le {h}_{vu}(t)\,{\rm{for}}\,{\rm{all}}\,(v,u)\in E$$ $${u}_{2}:{h}_{v{u}_{2}}(t)\ge {h}_{vu}(t)\,{\rm{for}}\,{\rm{all}}\,(v,u)\in {E}^{c}\mathrm{.}$$ where \({h}_{uv}(t)\) is the \(u,v\) entry in matrix \(h(t)\) . In case of ties \({u}_{1},{u}_{2}\) are chosen arbitrarily. Step 3: Repeat from Step 1 until k edge rewirings have been made. All simulations were performed using MATLAB R2014. In Step 3 we take \(k\,=\,4m\) ; simulations without upper limit on k show sufficient convergence after only m rewirings. We simulate 100 independent trials for each pair \((\tau ,p)\) . In analysing the networks generated by the algorithm all measures used are provided by the Brain Connectivity Toolbox 28 . Note that for \(\tau \gg 1\) the heat kernel approaches the matrix \(\frac{1}{{\bf{\text{1}}}D{{\bf{\text{1}}}}^{{\rm{T}}}}{D}^{1/2}{{\bf{\text{1}}}}^{{\rm{T}}}{\bf{\text{1}}}{D}^{1/2}\) and so rewiring biases toward high degree vertices, hence, adaptive rewiring approaches a process of preferential attachment 17 , 18 .

It's a small world after all - and now science has explained why. A study conducted by the University of Leicester and KU Leuven, Belgium, examined how small worlds emerge spontaneously in all kinds of networks, including neuronal and social networks, giving rise to the well-known phenomenon of "six degrees of separation". Many systems show complex structures, of which a distinctive feature is small-world network organization. They arise in society as well as ecological and protein networks, the networks of the mammalian brain, and even human-built networks such as the Boston subway and the World Wide Web. The researchers set out to examine whether this is a coincidence that such structures are so wide-spread or is there a common mechanism driving their emergence? A study recently published in Scientific Reports by an international team of academics from the University of Leicester and KU Leuven showed that these remarkable structures are reached and maintained by the network diffusion, i.e. the traffic flow or information transfer occurring on the network. The research presents a solution to the long-standing question of why the vast majority of networks around us (WWW, brain, roads, power grid infrastructure) might have a peculiar yet common structure: small-world topology. The study showed that these structures emerge naturally in systems in which the information flow is accounted for in their evolution. Nicholas Jarman, who recently completed his PhD degree at the Department of Mathematics, and is first author of the study, said: "Algorithms that lead to small-world networks have been known in scientific community for many decades. The Watts-Strogatz algorithm is a good example. The Watts-Strogatz algorithm, however, was never meant to address the problem of how small-world structure emerges through self-organisation. The algorithm just modifies a network that is already highly organised." Professor Cees van Leeuwen, who led the research at KU Leuven said: "The network diffusion steers network evolution towards emergence of complex network structures. The emergence is effectuated through adaptive rewiring: progressive adaptation of structure to use, creating short-cuts where network diffusion is intensive while annihilating underused connections. The product of diffusion and adaptive rewiring is universally a small-world structure. The overall diffusion rate controls the system's adaptation, biasing local or global connectivity patterns, the latter providing a preferential attachment regime to adaptive rewiring. The resulting small-world structures shift accordingly between decentralised (modular) and centralised ones. At their critical transition, network structure is hierarchical, balancing modularity and centrality - a characteristic feature found in, for instance, the human brain." Dr Ivan Tyukin from the University of Leicester added: "The fact that diffusion over network graph plays crucial role in keeping the system at a somewhat homeostatic equilibrium is particularly interesting. Here we were able to show that it is the diffusion process, however small or big gives rise to small-world network configurations that remain in this peculiar state over long intervals of time. At least as long as we were able to monitor the network development and continuous evolution". Alexander Gorban, Professor in Applied Mathematics, University of Leicester commented: "Small-world networks, in which most nodes are not neighbours of one another, but most nodes can be reached from every other node by a small number of steps, were described in mathematics and discovered in nature and human society long ago, in the middle of the previous century. The question, how these networks are developing by nature and society remained not completely solved despite of many efforts applied during last twenty years. The work of N. Jarman with co-authors discovers a new and realistic mechanism of emergence of such networks. The answer to the old question became much clearer! I am glad that the University of Leicester is a part of this exciting research."

10.1038/s41598-017-12589-9

Medicine

How a mutated gene wreaks havoc on white matter

Jeremy M. Baskin et al. The leukodystrophy protein FAM126A (hyccin) regulates PtdIns(4)P synthesis at the plasma membrane, Nature Cell Biology (2015). DOI: 10.1038/ncb3271 Journal information: Nature Cell Biology

http://dx.doi.org/10.1038/ncb3271

https://medicalxpress.com/news/2015-11-mutated-gene-wreaks-havoc-white.html

Abstract Genetic defects in myelin formation and maintenance cause leukodystrophies, a group of white matter diseases whose mechanistic underpinnings are poorly understood 1 , 2 . Hypomyelination and congenital cataract (HCC), one of these disorders, is caused by mutations in FAM126A , a gene of unknown function 3 . We show that FAM126A, also known as hyccin, regulates the synthesis of phosphatidylinositol 4-phosphate (PtdIns(4)P), a determinant of plasma membrane identity 4 , 5 , 6 . HCC patient fibroblasts exhibit reduced PtdIns(4)P levels. FAM126A is an intrinsic component of the plasma membrane phosphatidylinositol 4-kinase complex that comprises PI4KIIIα and its adaptors TTC7 and EFR3 (refs 5 , 7 ). A FAM126A–TTC7 co-crystal structure reveals an all-α-helical heterodimer with a large protein–protein interface and a conserved surface that may mediate binding to PI4KIIIα. Absence of FAM126A, the predominant FAM126 isoform in oligodendrocytes, destabilizes the PI4KIIIα complex in mouse brain and patient fibroblasts. We propose that HCC pathogenesis involves defects in PtdIns(4)P production in oligodendrocytes, whose specialized function requires massive plasma membrane expansion and thus generation of PtdIns(4)P and downstream phosphoinositides 8 , 9 , 10 , 11 . Our results point to a role for FAM126A in supporting myelination, an important process in development and also following acute exacerbations in multiple sclerosis 12 , 13 , 14 . Main Phosphoinositides are low-abundance anionic membrane phospholipids that play critical roles in many physiological processes 15 , 16 . At the plasma membrane, a major phosphoinositide is PtdIns(4)P, which has direct signalling roles and in addition serves as the precursor of two other plasma membrane-enriched phosphoinositides with major signalling functions in this membrane, PtdIns(4,5)P 2 and PtdIns(3,4,5)P 3 (refs 4 , 17 ). In oligodendrocytes and Schwann cells, these lipids regulate several steps in the biogenesis and maintenance of myelin, including the recruitment of myelin basic protein (MBP) to the plasma membrane by PtdIns(4,5)P 2 and the promotion of myelin growth by PtdIns(3,4,5)P 3 (refs 8 , 9 , 10 , 11 ). Phosphorylation of phosphatidylinositol (PtdIns) to generate PtdIns(4)P at the plasma membrane is mediated by PtdIns 4-kinase Type IIIα (PI4KIIIα; Stt4 in yeast) 5 , 18 , 19 . The properties, targeting mechanisms and regulation of this enzyme have only recently come into focus. Two factors required for its localization at the plasma membrane have been described 5 , 7 and structurally characterized 20 : EFR3 and TTC7 (Efr3 and Ypp1 in yeast). To identify additional regulators of PI4KIIIα, we used quantitative interaction proteomics. We immunoprecipitated either stably expressed TTC7B–GFP or GFP, followed by protease digestion and mass spectrometry analysis, to identify candidate TTC7B-binding proteins ( Fig. 1a and Supplementary Table 1 ). These experiments identified the known TTC7B interaction partners PI4KIIIα, EFR3A and EFR3B. Among the other hits, two of the most significantly enriched candidates were the paralogous FAM126A, also known as hyccin, and FAM126B, both of which were also identified in similar experiments using EFR3A–GFP or EFR3B–GFP as the bait ( Supplementary Fig. 1a ). Figure 1: FAM126 is a component of the PI4KIIIα complex. ( a ) Volcano plots of proteins associating with TTC7B–GFP versus GFP alone (top) and FAM126A–GFP versus GFP alone (bottom), from a label-free proteomics analysis of anti-GFP immunoprecipitates of HEK 293T cells stably expressing TTC7B–GFP, FAM126A–GFP or GFP alone. The logarithmic ratios of protein intensities are plotted against negative logarithmic P values of two-tailed Student’s t -test, equal variance, performed from n = 3 independent experiments. The red dashed line (significance, 0.05) separates specifically interacting proteins (top right portion of plot) from background. Selected top hits are indicated with black dots (bait is indicated in green), and all specific interactors are reported in Supplementary Table 1 . ( b ) Immunoblot analysis of PI4KIIIα complex components in anti-GFP immunoprecipitates of HEK 293T cells stably expressing TTC7B–GFP (top), FAM126A–GFP (bottom) or GFP alone. ( c ) Domain structure, evolutionary conservation and predicted extent of disordered secondary structural elements of FAM126. ( d ) Immunoblot analysis of anti-GFP immunoprecipitates from HeLa cells transfected with 3xFLAG–PI4KIIIα, TTC7B–MYC, EFR3B–HA, and the indicated GFP-tagged construct. Arrows denote specific bands and arrowheads denote nonspecific antibody bands. Unprocessed original scans of blots are shown in Supplementary Fig. 6. Full size image Mutations in FAM126A that lead to loss of the FAM126A protein cause a recessive leukoencephalopathy termed hypomyelination and congenital cataract 3 (HCC). Manifestations of this condition, which include progressive neurological impairment, mild to moderate cognitive defects, and peripheral neuropathy, stem from hypomyelination in the central and peripheral nervous systems 3 , 21 , 22 . No molecular functions or activities have been ascribed to FAM126A, and the cellular and molecular mechanisms of HCC pathogenesis are unknown. Thus, the identification of FAM126A as a potential interaction partner of TTC7 and EFR3 led us to investigate its role in PI4KIIIα complex formation and function as a first step towards understanding whether defects in phosphoinositide metabolism may cause HCC pathology. We first explored and confirmed the interaction of FAM126A with TTC7, and more generally its association with the PI4KIIIα complex, in co-immunoprecipitation/immunoblot experiments ( Fig. 1b ). Two FAM126A immunoreactive species (relative molecular masses of 58,000 ( M r 58K) and 47,000 ( M r 47K)) were enriched in TTC7B–GFP immunoprecipitates, corresponding to two predicted splice forms of FAM126A ( Supplementary Fig. 1b ). Accordingly, these two bands are absent in tissues from FAM126A knockout (KO) mice ( Supplementary Fig. 1c ). Given that the longer, M r 58K band represents the main form in mammalian brain ( Supplementary Fig. 1d ), we chose to focus on it in our subsequent studies. We then performed a reciprocal proteomics experiment using FAM126A–GFP-expressing cells to assess whether PI4KIIIα, TTC7 and EFR3 are the main interaction partners of FAM126A. Indeed, all of these proteins were the strongest hits in analogous quantitative proteomics experiments ( Fig. 1a and Supplementary Table 1 ) and were highly enriched in immunoblot analysis of FAM126A–GFP immunoprecipitates ( Fig. 1b ). Examination of the primary amino acid sequence of FAM126A revealed a highly conserved, structured amino-terminal portion (FAM126A-N, residues 1–289) common to both splice forms and a poorly conserved carboxy-terminal tail predicted to be disordered (FAM126A-C, residues 290–521; Fig. 1c ). Co-immunoprecipitation experiments of GFP-tagged full-length FAM126A, FAM126A-N or FAM126A-C with differentially tagged PI4KIIIα, TTC7B and EFR3B revealed an interaction of all of these components with full-length FAM126A and FAM126A-N ( Fig. 1d ). Note that the apparent more robust interaction of these proteins with FAM126A-N than with full-length FAM126A ( Fig. 1d , lanes 6 and 7) reflects higher levels of FAM126A-N in the total lysate. Interestingly, overexpression of FAM126A-N led to a marked increase in levels of transfected TTC7B in total lysate ( Fig. 1d , lane 3), suggesting a stabilizing interaction between these two proteins, as further confirmed by experiments described below. We next examined whether, as would be expected for a direct TTC7 interactor 5 , FAM126A-N localizes to the plasma membrane when co-expressed with other PI4KIIIα complex subunits. Using confocal microscopy, we found that GFP-tagged FAM126A-N and full-length FAM126A were localized to the cytosol in HeLa and COS-7 cells ( Fig. 2a and Supplementary Fig. 2a ). Whereas co-expression of FAM126A-N with EFR3B, the membrane anchor for the PI4KIIIα complex ( Fig. 2 , bottom), did not change the FAM126A-N localization ( Fig. 2b ), co-expression of FAM126A-N or full-length FAM126A with both EFR3B and TTC7B resulted in a relocalization of FAM126A to the plasma membrane ( Fig. 2c and Supplementary Fig. 2b ). Further, co-expression of FAM126A-N, TTC7B, EFR3B and PI4KIIIα resulted in co-localization of all four proteins at the plasma membrane in a manner that was dependent on the presence of TTC7B ( Fig. 2d, e ). Omission of EFR3B revealed a cytosolic co-localization (but nuclear exclusion) of TTC7B, FAM126A-N and PI4KIIIα ( Fig. 2f ). Collectively, these data argue for a central role of TTC7B not only in bridging PI4KIIIα to EFR3B, the plasma membrane anchor, as reported previously 5 , 7 , 20 , but also in mediating the association of FAM126A to the plasma membrane, consistent with a direct interaction. Figure 2: FAM126A is recruited to the plasma membrane in a complex with TTC7B, PI4KIIIα and EFR3B. ( a – f ), Top: Confocal imaging of live HeLa cells transfected with GFP–FAM126A-N and the following additional plasmids: none ( a ); EFR3B–BFP ( b ); EFR3B–BFP and TTC7B–mCherry ( c ); EFR3B–BFP, PI4KIIIα–mCherry and TTC7B–3xFLAG ( d ); EFR3B–BFP and PI4KIIIα–mCherry ( e ); TTC7B–BFP and PI4KIIIα–mCherry ( f ). Bottom: Schematic representation of the transfected constructs and interpretation of the results. Scale bars, 20 μm. Full size image The localization experiments did not, however, allow us to conclusively determine whether, at the plasma membrane, FAM126A and PI4KIIIα can bind to TTC7B simultaneously, or, alternatively, whether FAM126A competes with PI4KIIIα for binding to TTC7B. To address this issue, we first confirmed the direct interaction between FAM126A-N and TTC7B by co-expression of the proteins in Escherichia coli and analysis of the resulting complex by size-exclusion chromatography ( Fig. 3a ). The two proteins co-eluted as a heterodimer. Importantly, protein–protein interaction experiments with purified proteins showed that the TTC7B/FAM126A-N subcomplex interacts directly with PI4KIIIα ( Fig. 3b ), ruling out a competition between FAM126A and PI4KIIIα for TTC7 binding. Figure 3: FAM126A binds directly to TTC7B, and both FAM126A and TTC7B bind to PI4KIIIα and stimulate its catalytic activity. ( a ) TTC7B and FAM126A-N comigrate as a heterodimer on a sizing column (Superdex 200), as shown by SDS–PAGE analysis of elution fractions. Proteins were visualized using Coomassie blue staining. ( b ) TTC7B/GST–FAM126A-N pulls down PI4KIIIα in a protein–protein interaction experiment with purified proteins. The GST-tagged proteins (GST, GST–FAM126A-N (126A-N) or the GST–FAM126A-N/TTC7B dimer (Complex)) were immobilized on glutathione-conjugated Sepharose and incubated with purified PI4KIIIα. Proteins were visualized using Coomassie blue staining. The experiment was performed in triplicate. Arrowhead denotes the chaperone Hsp70, which co-purified with PI4KIIIα. ( c ) Kinase activity assay using purified PI4KIIIα subcomplexes. Wild-type (WT) or kinase-dead (KD) PI4KIIIα was produced either alone or in complex with TTC7B or TTC7B/FAM126A-N as indicated ( Supplementary Fig. 3a ), and in vitro lipid kinase activity assays were performed using PtdIns-containing liposomes and γ 32 P-labelled ATP. The extent of PtdIns(4)P formation was assessed by thin-layer chromatography and autoradiography. Two-tailed Student’s t -test with unequal variance: ∗ P = 0.0001 ( n = 3 independent experiments). Error bars represent standard deviation. Unprocessed original scans of gels are shown in Supplementary Fig. 6. Full size image Given that the TTC7B/FAM126A-N heterodimer bound tightly to PI4KIIIα, we next examined the effects of each of these proteins on the catalytic activity of PI4KIIIα. We purified wild-type (WT) or kinase-dead PI4KIIIα alone or in complex with either TTC7B alone or both TTC7B and FAM126A-N ( Supplementary Fig. 3a ) and assessed the relative kinase activity in vitro by monitoring the formation of PtdIns(4)P from PtdIns in the presence of γ 32 P-labelled ATP. The PI4KIIIα/TTC7B complex was about twice as enzymatically active as PI4KIIIα alone, and the PI4KIIIα/TTC7B/FAM126A-N ternary complex was roughly fivefold more active ( Fig. 3c ). Thus, both TTC7B and FAM126A seem to play a role in stabilizing the PI4KIIIα fold and/or stimulating the intrinsic enzymatic activity of the kinase. To obtain further insights into the role of FAM126A in the kinase complex, we determined the crystal structure of the TTC7B/FAM126A-N dimer at 2.9 Å resolution ( Fig. 4a and Supplementary Fig. 3b ). Both proteins are almost entirely alpha-helical. In TTC7B, as in its yeast homologue Ypp1 (ref. 20 ; PDBID 4N5C), the alpha helices are arranged into a superhelix, with the very C-terminal peptide inserted into the centre of the superhelix ( Fig. 4b ). FAM126A-N is globular, except for a long hairpin (residues 121–165) that extends out and wraps around TTC7B ( Fig. 4b ). The proteins interact through an unusually large interface ( ∼ 6,000 Å 2 total occluded surface area, where >2,000 Å 2 is considered large), indicative of a high-affinity, stable interaction as would be expected if FAM126A is an intrinsic part of the kinase complex ( Fig. 4c ). Figure 4: Crystal structure of a TTC7B/FAM126A co-complex reveals an unusually large protein–protein interface and a conserved binding surface for PI4KIIIα. ( a ) Ribbon diagram for the TTC7B/FAM126A-N complex. Note the FAM126A-N hairpin structure that wraps around TTC7B like an arm. Point mutations in FAM126A that underlie HCC are indicated in Supplementary Fig. 3c . ( b ) Ribbon diagrams for TTC7B and FAM126A-N coloured from blue (N terminus) to red (C terminus). The ‘arm’ in FAM126A is green. Disordered residues absent from the model are indicated by dotted lines. ( c ) Space-filling models of the TTC7B/FAM126A-N complex (left) and TTC7B (middle and right) coloured by sequence conservation. Conserved surfaces on TTC7B/FAM126A-N (yellow outline) or TTC7B alone (orange outline) that may interact with PI4KIIIα are circled (left). The TTC7B surfaces at the interface with FAM126A-N are indicated by dotted yellow lines (middle and right). ( d ) Model for PI4KIIIα assembly at the plasma membrane. The EFR3 model is based on the structure of yeast Efr3 (ref. 20 ) (PDBID 4N5A). Full size image Interestingly, although the interacting surfaces in both proteins are well conserved in metazoans ( Fig. 4c ), the FAM126A-interacting surface of TTC7B is not conserved in fungi (including in Saccharomyces cerevisiae Ypp1; ref. 20 ). Indeed, we have not been able to identify a fungal FAM126A homologue, and it is possible that FAM126 is not conserved across eukaryotic evolution but is a modification to the PI4KIIIα system present only in higher eukaryotes. Nevertheless, the overall organization of EFR3, TTC7 and PI4KIIIα within the complex is most likely conserved ( Fig. 4d ). Thus, as the yeast homologue of TTC7 interacts with the kinase through conserved surfaces in the C-terminal lobe 20 , we reason that mammalian TTC7 interacts with PI4KIIIα in a similar way. We note that in the TTC7B/FAM126A-N subcomplex there is a conserved surface comprising the C-terminal portion of TTC7B and adjoining FAM126A residues, and we propose that these surfaces together form the binding site for PI4KIIIα in mammals ( Fig. 4c ). The structural data suggest a stabilizing function for FAM126A in the PI4KIIIα complex. To test this hypothesis, we first examined the levels of PI4KIIIα complex components in primary human skin fibroblasts from five HCC patients that are homozygous for either nonsense or missense mutations in FAM126A (refs 3 , 21 , 23 ). Cells from patients with nonsense mutations were completely devoid of the FAM126A protein ( Fig. 5a , lanes 3, 5 and 7), and cells from patients with missense mutations (L53P and C57R, see Supplementary Fig. 3c ) exhibited greatly reduced levels of FAM126A ( Fig. 5a , lanes 4 and 6), consistent with the misfolding and near-complete degradation of FAM126A in these cells. Indeed, immunoblot analysis revealed decreases in the levels of PI4KIIIα and its adaptors TTC7A, TTC7B and EFR3A (EFR3B was not detectable in the fibroblasts; Fig. 5a and Supplementary Fig. 4a ), suggesting a general destabilization and degradation of the PI4KIIIα complex components in the absence of FAM126A. Accordingly, reintroduction of GFP-tagged FAM126A into the patient fibroblasts using a lentiviral vector partially rescued this phenotype ( Supplementary Fig. 4b ). Although a compensatory increase in FAM126B, a paralogue of FAM126A, was also observed in patient fibroblasts ( Fig. 5a and Supplementary Fig. 4a ), the approximately tenfold lower expression of FAM126B messenger RNA in fibroblasts ( Supplementary Fig. 4c ) may explain why this increase is not sufficient to compensate for lack of FAM126A. Figure 5: Loss of FAM126A leads to defects in the PI4KIIIα complex in HCC patient fibroblasts and FAM126A KO mouse brain. ( a ) Immunoblot analysis of PI4KIIIα complex components in control and HCC patient fibroblasts. ( b ) HPLC analysis of total 3 H-inositol-labelled phosphoinositide content in control and HCC patient fibroblasts. Two-tailed Student’s t -test, unequal variance, ∗ P = 0.0019, n = 4 independent experiments. Error bars represent standard deviation. ( c ) Quantification of the plasma membrane PtdIns(4)P pool in control and HCC patient fibroblasts by immunofluorescence using an anti-PtdIns(4)P antibody. Two-tailed Student’s t -test, unequal variance, ∗ ∗ P = 0.0001. Box plot demarcates 25th and 75th percentile (middle line is median), and bars represent the minimum and maximum values ( n = 19 cells for control, n = 23 cells for patient, from a total of 4 independent experiments). Similar results were seen in a second HCC patient cell line. ( d ) Immunoblot analysis of PI4KIIIα complex components in WT primary cortical neuronal and oligodendrocyte (OLs) cultures. ( e , f ) Immunoblot analysis of PI4KIIIα complex components in total brain ( e ) or optic nerve ( f ) from male WT or FAM126A KO mice at age P30. See Supplementary Fig. 4a for quantification of the immunoblot data from a and d – f . Unprocessed original scans of blots are shown in Supplementary Fig. 6. Full size image High-performance liquid chromatography (HPLC) analysis of total phosphoinositide content revealed a decrease in total cellular PtdIns(4)P in HCC fibroblasts relative to control fibroblasts ( Fig. 5b ), and a specific decrease of the plasma membrane fraction of PtdIns(4)P in HCC fibroblasts relative to control fibroblasts was confirmed by immunofluorescence ( Fig. 5c and Supplementary Fig. 4d ). Thus, FAM126A loss specifically affects PI4KIIIα complex assembly and PI4KIIIα-mediated PtdIns(4)P synthesis at the plasma membrane. Global KO of PI4KIIIα in several model organisms leads to lethality 5 , 24 , 25 , 26 , whereas loss of FAM126A is permissive of life, although it causes defects in myelination in humans, leading to HCC (ref. 3 ). To investigate the potential role of FAM126A in promoting PI4KIIIα function in myelination, we compared the expression levels of PI4KIIIα complex subunits in mouse primary cortical neurons and oligodendrocytes. We found that FAM126A was expressed in both oligodendrocytes and neurons, whereas its paralogue FAM126B was expressed at much lower levels in oligodendrocytes relative to neurons ( Fig. 5d and Supplementary Fig. 4a ). Thus, in principle, a global loss of FAM126A could, in the brain, be more effectively compensated by FAM126B in neurons than in oligodendrocytes and thus more severely affect PI4KIIIα function in the latter cell type. To support this possibility, we evaluated the levels of PI4KIIIα complex components in brain tissue from FAM126A KO mice 27 . Although these mice do not exhibit an obvious abnormal phenotype or specific defects in myelination as evaluated by light and electron microscopy analysis ( Supplementary Fig. 5 ), immunoblot analysis of lysates from total brain and optic nerve (a pure white matter tract) revealed a selective decrease of TTC7A and EFR3A relative to WT ( Fig. 5e, f and Supplementary Fig. 4a ). These are the isoforms of TTC7 and EFR3 whose levels are more highly expressed in oligodendrocytes relative to neurons, compared with their respective paralogues TTC7B and EFR3B ( Fig. 5d and Supplementary Fig. 4a ). These results are further supported by immunoblot analysis of cellular fractions of different lineages that were generated from dissociated brain tissue by affinity purification from WT and FAM126A KO mice. Notably, there was a more pronounced decrease in PI4KIIIα, EFR3A/B and TTC7A/B in cells of the oligodendroglial lineage compared with those of the neuronal lineage ( Supplementary Fig. 4e ). Taken together, these data connect FAM126A to PI4KIIIα function in the nervous system and implicate the white matter as a region potentially susceptible to PI4KIIIα dysfunction following global loss of FAM126A, pointing to a mechanistic hypothesis for HCC disease pathogenesis. Although it seems surprising that the lack of FAM126A in mice does not produce the neurological phenotype observed in humans, we note that the greater diameter of axons and thickness of the myelin sheath around such axons in humans implies a larger surface area of oligodendrocytes 28 , 29 . Thus, a greater demand on the machinery responsible for myelin biogenesis and maintenance in humans may explain their enhanced sensitivity to loss of FAM126A and impaired PtdIns(4)P synthesis. Interestingly, FAM126A expression was reported to be repressed by activated β-catenin 30 , and canonical Wnt/β-catenin/TCF signalling was recently shown to be a powerful negative regulator of oligodendrocyte differentiation 31 , a step that precedes the massive plasma membrane biogenesis that occurs in myelination. Our data suggest that activation of PI4KIIIα may be an important downstream consequence of the relief of Wnt/β-catenin/TCF signalling that occurs in oligodendrocyte development. In sum, our results demonstrate that FAM126A (hyccin) is an intrinsic component of the PI4KIIIα complex and an important regulator of PtdIns(4)P production at the plasma membrane, the first step in the synthesis of the bulk of the downstream phosphoinositides PtdIns(4,5)P 2 and PtdIns(3,4,5)P 3 . We further provide evidence for an important role of FAM126A in oligodendrocytes, a cell type whose specialized function is to markedly expand its plasma membrane. PtdIns(4,5)P 2 is implicated in myelin compaction 8 , 9 , and PtdIns(3,4,5)P 3 helps drive myelin growth 10 , 11 . Collectively, our studies point to impaired production of these phosphoinositides as a mechanism through which absence of FAM126A results in a hypomyelinating leukoencephalopathy in humans, and they demonstrate the critical importance of plasma membrane phosphoinositide homeostasis in myelin development and potentially 14 also in remyelination after myelin loss. □ Methods Plasmids and cloning. TTC7B–mCherry, mCherry–PI4KIIIα (NCBI NP_477352.3) and EFR3B–HA were previously described 5 , and PM–mCherry was obtained from the De Camilli laboratory. 3xFLAG–PI4KIIIα was generated by subcloning PI4KIIIα into the p3xFLAG-CMV-10 vector (Sigma) using NotI and EcoRI. TTC7B–3xFLAG was generated by subcloning TTC7B into the p3xFLAG-CMV-14 vector (Sigma) using HindIII and XbaI. EFR3B–BFP and TTC7B–BFP were generated by subcloning EFR3B into the pTagBFP-N vector (Evrogen) using XhoI and EcoRI (for EFR3B) and XhoI (for TTC7B). TTC7B–MYC was generated by subcloning TTC7B into the pCMV-MYC-N vector (Clontech) using XhoI and NotI. FAM126A (isoform 1, NM_032581.2) was obtained from Origene and subcloned into the pEGFP-N1 vector (Clontech) using XhoI and BamHI to make FAM126A–GFP. GFP–FAM126A-N and GFP–FAM126A-C were generated by subcloning FAM126A 1−289 and FAM126A 289−521 , respectively, into the pEGFP-C1 vector (Clontech) using KpnI and BamHI. For generation of stable HEK 293T cell lines, TTC7B–GFP, FAM126A–GFP and GFP were subcloned into the pCDNA5/FRT or pCDNA5/FRT/TO vectors (Thermo Fisher) using NheI and NotI (FRT, TTC7B–GFP), BamHI and XhoI (FRT/TO, FAM126A), or EcoRV and NotI (FRT/TO, GFP). The coding sequence for TTC7B 9−843 was cloned into a modified pCOLADuet-1 vector, which introduces an N-terminal dodeca-histidine-SUMO tag. The sequence for FAM126A 2−308 was cloned into pGEX-6P-1 (GE Healthcare). The sequence for the 419-amino-acid splice form of FAM126A (isoform 2, corresponding to NCBI XP_005249951.1) was synthesized as a gBlock (Integrated DNA Technologies) and cloned into the pEGFP-N1 vector using SalI and BamHI. This GFP-tagged alternate splice form, FAM126A 1−419 -GFP, was subcloned into the pLL3.7 lentiviral vector (Addgene plasmid no. 11795, gift from M. Higley, Yale University, USA) downstream of the CAG promoter at the EcoRI restriction site using the InFusion system (Clontech). The lentiviral packaging vectors pMD2.G and psPAX2 (Addgene no. 12259 and 12260) were kindly provided by D. Trono (Ecole Polytechnique Fdrale de Lausanne, Switzerland). Antibodies. Working dilutions are listed in parentheses for all antibodies. An antibody against EFR3B (1:1,000) was previously reported 5 . Antibodies against mouse TTC7A (1:200) and mouse TTC7B (1:200) were generated by immunization of rabbits with KLH conjugates of the following peptides: Ac-CEASSPVLPFSIIAREL-NH 2 (TTC7A); Ac-RLETEIERCRSECQWERIPELC-NH 2 (TTC7B) (Cocalico Biologicals). Sera were affinity purified using the antigenic peptide immobilized on SulfoLink resin, following the manufacturer’s instructions (Thermo Fisher). An antibody against FAM126A (1:1,000) was generated by immunization of rabbits with a purified FAM126A fragment (residues 2–308), and sera were affinity purified by incubation with a nitrocellulose membrane onto which the FAM126A(2–308) fragment had been adsorbed, followed by elution with mild acid. The sources for the other antibodies are as follows: CNPase (1:1000, clone 11-5B, MAB326R, EMD Millipore), EFR3A (1:1,000, Ab2, Sigma), FAM126B (1:1,000, Novus Biologicals), FLAG (1:2,000, M2, Sigma), GAPDH (1:5,000, 1D4, Proteus Biosciences), GFP (1:5,000, Living Colors Monoclonal, Clontech), HA (1:5,000, 3F10, Roche), MBP (1:1,000, SMI 99, Covance), MYC (1:2,000, rabbit polyclonal, EMD Millipore), NeuN (1:1,000, clone A60, MAB377, EMD Millipore), human TTC7B (1:500, mouse IgG fraction, Sigma), PI4KIIIα (1:1,000, Cell Signaling Technology), PtdIns(4)P (1:100, Echelon), SYP (1:20,000, G95, Synaptic Systems), tubulin (1:10,000, B-5-1-2, Sigma). Cell culture. All cells were grown at 37 °C in a 5% CO 2 atmosphere (except for Expi293 cells (obtained from Thermo Fisher), which were grown at 37 °C in an 8% CO 2 atmosphere, rotating at 125 r.p.m. on an orbital shaker) in cell culture medium (Dulbecco’s modified Eagle medium supplemented with 10% fetal bovine serum and penicillin/streptomycin) and were tested to be negative for mycoplasma. HEK 293T cell lines stably expressing TTC7B–GFP, FAM126A–GFP or GFP were generated by transfection of Flp-In-293 cells (obtained from Thermo Fisher) with flippase (pOG44, Thermo Fisher) and either TTC7B–GFP, FAM126A or GFP in the pCDNA5/FRT or pCDNA5/FRT/TO vector, followed by selection according to the manufacturer’s instructions (Thermo Fisher) using hygromycin B (Sigma). Stable HEK 293T cell lines were maintained in cell culture medium supplemented with 10 μg ml −1 hygromycin B. Stable HeLa cell lines expressing EFR3A–GFP, EFR3B–GFP or GFP were generated by transient transfection of HeLa cells (obtained from ATCC) followed by selection using G418 (Thermo Fisher), clonal isolation, and expansion. HEK 293 cell lines (parental cell lines for Expi293 and Flp-In-293 cells) and HeLa cells were validated by short tandem repeat profiling analysis by ATCC. HCC patient and control primary human skin fibroblasts (obtained from the Cell Line and DNA Biobank from Patients Affected by Genetic Diseases, Gaslini Institute) were described and validated previously 21 , 27 . Following ethical guidelines, the samples were obtained for analysis and storage with written informed consent. Consent was sought using a form approved by the Gaslini Institute Ethics Committee. Transfection. HeLa and HEK 293T cells were transfected with appropriate plasmids using Fugene HD (Promega) according to the manufacturer’s instructions approximately 18–24 h before analysis. Expi293 cells were transfected with ExpiFectamine (Thermo Fisher) according to the manufacturer’s instructions. Immunoprecipitation and immunoblots. HEK 293T cells stably expressing TTC7B–GFP, FAM126A-N–GFP or GFP were collected, resuspended in lysis buffer (150 mM NaCl, 20 mM Tris, 1 mM EDTA, 1% Triton X-100, pH 7.4, supplemented with protease inhibitors (cOmplete, EDTA-free (Roche))), sonicated briefly, and centrifuged for 10 min at 16,000 g . The supernatant was immunoprecipitated in lysis buffer by addition of GFP-trap agarose (Chromotek) and rocking for 1 h at 4 °C. The resin was then isolated by centrifugation at 1,000 g , rinsed three times with lysis buffer, and analysed either by SDS–PAGE, immunoblot (with detection by chemiluminescence), or mass spectrometry (see below). Original (uncropped) immunoblots are shown in Supplementary Fig. 6 . Pulldown analysis by MS. Resin containing purified immunoprecipitates was rinsed three times with wash buffer (150 mM NaCl, 50 mM Tris, pH 7.4) and then denatured for 30 min with urea (8 M) in 0.1 M Tris, pH 7.4, 1 mM dithiothreitol before alkylation and pre-digestion with endoproteinase LysC (Wako Chemicals USA). After incubation for 3 h samples were diluted fourfold with ammonium bicarbonate (25 mM) and further digested with trypsin (Promega) overnight. Digestions were stopped by addition of trifluoroacetic acid (TFA, 1 μl) and the resulting peptides were loaded and desalted on C18 Stage Tips. LC–MS/MS analysis. Peptides were eluted from C18 Stage Tips with 60 μl of elution buffer (80% acetonitrile and 0.1% formic acid) and samples were dried down to 5 μl in a vacuum centrifuge. Peptides were then subjected to reversed phase chromatography on an Easy nLC 1000 system (Thermo Fisher Scientific) using a 50-cm column (New Objective) with an inner diameter of 75 μm, packed in-house with 1.9 μm C18 resin (Dr. Maisch). Peptides were eluted with an acetonitrile gradient (5–30% for 95 min at a constant flow rate of 250 nl min −1 ) and directly electrosprayed into a mass spectrometer (Q Exactive; Thermo Fisher Scientific). Mass spectra were acquired on the spectrometer in a data-dependent mode to automatically switch between full-scan MS and up to 10 data-dependent MS/MS scans. The maximum injection time for full scans was 20 ms, with a target value of 3,000,000 at a resolution of 70,000 at m / z = 200. The ten most intense multiple charged ions ( z ≥ 2) from the survey scan were selected with an isolation width of 3 Th and fragmented with higher-energy collision dissociation (HCD) with normalized collision energies of 25. Target values for MS/MS were set to 1,000,000 with a maximum injection time of 120 ms at a resolution of 17,500 at m / z = 200. To avoid repetitive sequencing, the dynamic exclusion of sequenced peptides was set to 20 s. MS data analysis. MS and MS/MS spectra were analysed using MaxQuant (version 1.4.0.5), using its integrated ANDROMEDA search algorithms 32 . Scoring of peptides for identification was carried out with an initial allowed mass deviation of the precursor ion of up to 6 ppm for the search for peptides with a minimum length of six amino acids. The allowed fragment mass deviation was 20 ppm. The false discovery rate (FDR) was set to 0.01 for proteins and peptides. Peak lists were searched against a local database for human proteome. Maximum missed cleavages were set to 2. The search included carbamidomethylation of cysteines as a fixed modification and methionine oxidation and N-terminal acetylation as variable modifications. All calculations and plots were performed as described previously 33 using the R software package. Imaging. For live-cell imaging, cells were grown in glass-bottom dishes (no. 1.5 thickness, MatTek Corporation). For immunofluorescence, cells were grown on cover slips (no. 1 thickness, neuVitro). Immunofluorescence labelling for PtdIns(4)P was performed as previously described 34 . Imaging experiments were performed on a spinning-disc confocal microscope, using the PerkinElmer UltraVIEW VoX system including a Nikon Ti-E Eclipse inverted microscope equipped with Perfect Focus, a temperature-controlled stage, a 14-bit electron-multiplying CCD (charge-coupled device) camera (Hamamatsu C9100-50), and a spinning-disc confocal scan head (CSU-X1, Yokogawa) controlled by Volocity software (PerkinElmer). All images were acquired through a 60× or 100× oil objective (1.4 NA, CFI Plan Apo VC). Blue fluorescence was excited with a 405 nm/50 mW diode laser (Melles Griot) and collected by a BP 445/60 filter. Green fluorescence was excited with a 488 nm/50 mW diode laser (Coherent) and collected by a BP 527/55 filter. Red fluorescence was excited with a 561 nm/50 mW diode laser (Cobolt) and collected by a BP 615/70 filter. Multicolour images were acquired sequentially. Image acquisition was performed using the Volocity software (PerkinElmer) and image analysis was performed using Fiji. For the quantification of plasma membrane PtdIns(4)P immunofluorescence, samples were blinded, and average intensity maximum projections were generated. The cell outline was drawn manually, and the mean pixel intensity was recorded for each sample. Protein expression (bacterial). Plasmids containing the coding sequences for His 12 –SUMO–TTC7B 9−843 and GST–FAM126A 2−308 were co-transformed into E. coli BL21(DE3) cells. The cells were grown to an attenuance of ∼ 0.7 at 600 nm at 37 °C and then shifted to 18 °C. Protein expression was induced at 18 °C by addition of isopropylthio-β-galactosidase (IPTG) to a final concentration of 0.5 mM. The cells were collected 20 h after induction and frozen at −80 °C. Protein purification. For the TTC7B 9−843 /FAM126A 2−308 complex, cells were resuspended in lysis buffer (20 mM Tris (pH 8.0 at 22 °C), 200 mM NaCl, 1 mM Tris(2-carboxyethyl)phosphine (TCEP), and 20 mM imidazole) supplemented with protease inhibitors (cOmplete, EDTA-free (Roche)), and lysed using a cell disruptor (Avestin). The complex was first isolated using Ni-NTA resin (QIAGEN). After elution, the His 12 -SUMO tag was cleaved off with SUMO protease. Protein was then bound to Glutathione-S-Sepharose 4B resin (GE Healthcare). Bound complex was eluted from the resin by cleavage with PreScission protease to remove the GST tags. Protein was further purified by size-exclusion chromatography on Superdex 200 (GE Healthcare) in buffer (20 mM Tris (pH 8.0 at 22 °C), 200 mM NaCl, 1 mM TCEP) and concentrated to 6 mg ml −1 . For protein–protein interaction experiments, GST-tagged FAM126A alone or complexed with TTC7B was isolated by Glutathione-S-Sepharose 4B resin and then eluted from the resin using glutathione. Protein was further purified by size-exclusion chromatography on Superdex 200 (GE Healthcare) in buffer (20 mM HEPES (pH 7.4), 150 mM NaCl, 1 mM TCEP). Selenomethionine-substituted proteins for structure determination were prepared similarly to native proteins as previously described 35 . Crystallization. Crystals of native and selenomethionine-substituted TTC7B 9−843 /FAM126A 2−308 complexes were grown by the hanging-drop method at 20 °C, mixing 1.35 μl each of protein solution (6 mg ml −1 ) and mother liquor (0.1 M HEPES (pH 7.0–7.4), 4–5% polyethylene glycol (PEG) 8,000) and 0.3 μl of 0.2 M 3-(1-pyridino)-1-propane sulphonate (NDSB-201). The crystals belong to space group P2 1 2 1 2 1 with two copies of the complex in the asymmetric unit. Data collection and structure determination. Crystals were serially transferred into mother liquor supplemented with ethylene glycol from 10 to 30%, loop mounted, and flash-frozen in liquid nitrogen. Diffraction data were collected at the selenium anomalous edge. Data for both native and selenomethionine-substituted crystals were collected at the NE-CAT beamline 24ID-E at the Advanced Photon Source (APS). The data were processed using HKL2000 (ref. 36 ; Supplementary Table 2 ). For phasing, we used the selenomethionine-substituted crystals in the SAD method 37 . Phasing was carried out using PHENIX (ref. 38 ). A representative electron density map is shown in Supplementary Fig. 3b (top). We built the model of the complex using COOT (ref. 39 ) and refined it against native data in PHENIX (ref. 38 ) using positional, translation/libration/screw motion, and individual B-factor refinement options and secondary-structure restraints ( Supplementary Table 2 and Supplementary Fig. 3b (bottom). The final structure has good geometry (98.3 and 1.7% of residues are in allowed and generously allowed regions of the Ramachandran plot, respectively). The N terminus of one of the two TTC7B molecules in the asymmetric unit is not ordered and was not modelled, and several loops were omitted in both copies of TTC7B (A: residues 33–36, 157–166, 197–212, 292–299, 343–358, 462–464, 619–686; B: 9–139, 154–170, 194–213, 286–318, 342–362, 460–464, 618–688) and FAM126A (A: residues 2–7, 18–23, 30–33, 93–94, 149–153, 288–308; B: 2–6, 16–24, 30–35, 49–52, 93–98, 114–120, 140–157, 288–308). Protein expression (mammalian) for pulldown and kinase assay. Expi293 cells (Thermo Fisher) were maintained and transfected with ExpiFectamine (Thermo Fisher), according to the manufacturer’s instructions, with the following combinations of plasmids: 3xFLAG–PI4KIIIα (wild-type); 3xFLAG–PI4KIIIα (D1957A kinase-dead mutant); 3xFLAG–PI4KIIIα (WT) and TTC7B–HA; or 3xFLAG–PI4KIIIα (WT), TTC7B–HA and GFP–FAM126A(1–289). Cells were collected, resuspended in lysis buffer (250 mM NaCl, 20 mM Tris pH 8.0, 10% glycerol, 1 mM TCEP, supplemented with protease inhibitors (cOmplete, EDTA-free (Roche))), sonicated briefly, and centrifuged for 10 min at 16,000 g . The supernatant was immunoprecipitated in lysis buffer by addition of M2 anti-FLAG agarose (Sigma) and rocking for 2 h at 4 °C. The resin was then isolated by centrifugation at 1,000 g , rinsed three times with lysis buffer, and protein complexes were eluted by incubation with 3xFLAG peptide (125 μg ml −1 in lysis buffer). PI4KIIIα lipid kinase assay. The relative activity of the different PI4KIIIα preparations was assayed as follows. Liposomes containing C16-PI (Echelon Biosciences) were generated by sonication of a 1 mg ml −1 solution in kinase buffer (100 mM Tris-HCl pH 7.5, 50 mM EGTA, 100 mM MgCl 2 )) and incubated (at 80 μM) in a 50 μl reaction with 400 ng of PI4KIIIα or PI4KIIIα complex (equal amounts of PI4KIIIα in each sample were ensured by the use of both the BCA assay and SDS–PAGE/Coomassie staining), γ 32 P-labelled ATP (10 μCi), and cold ATP (50 μM) in kinase buffer for 5 min at 37 °C. The reaction was quenched by the addition of 700 μl of 2:1 chloroform/methanol containing 10 μg ml −1 Folch fraction (brain phosphoinositides) and 400 μl of 0.1 M HCl. The organic extracts were dried, resuspended in a small amount of 1:1 chloroform/methanol, and equal amounts were analysed by thin-layer chromatography (mobile phase is 14:32:24:30:64 water/acetic acid/methanol/acetone/chloroform). PtdIns(4)P was identified by comparison with PtdIns(3)P generated in a parallel reaction using PI(3)K p110γ (Sigma) and quantified by autoradiography using a STORM 860 system (Molecular Dynamics). Phosphoinositide analysis. PtdIns(4)P and PtdIns(4,5)P 2 levels were quantified by 3 H-inositol metabolic labelling for 72 h. Lipid extraction, deacylation, and HPLC separation by anion-exchange with detection using a radiometric detector were performed as previously described 5 . Mouse husbandry and neuronal and oligodendrocyte cultures. To generate the FAM126A knockout mice, VelociGene technology was used to replace exons 2 and 3 of FAM126A (Ensembl Accession Number: ENSMUSG00000028995; GenBank Accession number: NM_053090) and the intervening intron, corresponding to an 8.4-kb genomic region (chr5: 23,991,590-23,999,988; UCSC-genome Browser GRCm38/mm10, Dec11 Assembly) with a β-galactosidase(LacZ)/neomycin (NeoR) cassette as previously described 27 . The F2 offspring created in the mixed 129/C57BL/6 genetic background were backcrossed for ten generations into the C57BL/6 pure strain. Animal care and use was carried out in accordance with institutional guidelines (Animal Care and Use Committee of Istituto de Fisiologia Clinica, CNR, Pisa, Italy). Primary cortical neuronal and oligodendroglial cultures were generated as previously described 40 . Samples used for immunoblot analysis ( Fig. 5 ) were collected at DIV16 (neurons) and DIV5 (oligodendrocytes). No statistical method was used to predetermine samples size, and the experiments were not randomized. Lentivirus production and rescue experiment. Lentiviral particles were purified as described herein. HEK 293T cells were transfected with either pLL3.7-FAM126A 1−419 -GFP or pLL3.7 empty vector in combination with the packaging plasmids pMD2.G and psPAX2 using Lipofectamine 2000 (Thermo Fisher) according to the manufacturer’s instructions. Seventy-two hours after transfection, the viral supernatant was collected, filtered through a 0.45 μm filter, and centrifuged in ultraviolet-sterilized tubes at 37693 g for 90 min at 4 °C using an SW28.1 rotor (Beckman). The liquid was decanted, and 250 μl of OptiMEM (Thermo Fisher) was added to the tube and incubated at 4 °C overnight. The following day, the virus was gently resuspended, virus titre was measured by ELISA (p24 HIV ELISA kit, Cell BioLabs), and aliquots were generated and stored at −80 ° C. HCC patient fibroblasts were infected with 0.5 × 10 12 viral particles ml −1 in antibiotic-free media in the presence of Polybrene (8 μg ml −1 , Sigma-Aldrich). After 24 h, the supernatant was replaced with fresh media. After an additional 6 d, the cells were rinsed, trypsinized, and lysates were generated and analysed by immunoblot. Quantitative RT–PCR analysis. Total RNA was isolated from control primary human skin fibroblasts using the RNeasy kit (Thermo Fisher) according to the manufacturer’s instructions and stored in nuclease-free water at −20 °C. Total RNA (2 μg) was reverse transcribed with an oligo(dT) primer using SuperScript polymerase (Thermo Fisher). cDNA was analysed in triplicate by qRT–PCR amplification using SYBR Green Supermix on a Bio-Rad CFX96 Real-Time PCR Detection System. PCR amplification conditions were as follows: 95 °C (2 min) and 4 cycles of 95 °C (5 s) and 60 °C (30 s). Primer pairs (FAM126A: 5′-CACGAGTCGAGGTCCTGC-3′ and 5′-TCCTCCACAACCCCTTTCTC-3′; FAM126B: 5′-TCCCCTCCTTATCCAAGCCT-3′ and 5′-ATGCTGACACAATGCCCCTT-3′) were designed to amplify mRNA-specific fragments, and unique products were tested by melt-curve analysis. PCR efficiencies were 101.6% (slope of −3.285 in a tenfold dilution series) for FAM126A and 96.7% (slope of −3.404 in a tenfold dilution series) for FAM126B using the indicated primers. Data were analysed using ΔΔCt, and values were normalized to the housekeeping gene ribosomal protein S26 (using 5′-CCGTGCCTCCAAGATGACAA-3′ and 5′-GCAATGACGAATTTCTTAATGGCCT-3′ as primers). Immunoisolation of oligodendrocyte and neuronal lineage cells. For acute cell isolation, cells were isolated using the MACS cell selection kit (Miltenyi Biotec). Briefly, brains from male FAM126A KO and littermate WT control mice (on the pure C57BL/6 background) at postnatal day 8 were isolated, minced, and mechanically dissociated with a gentleMACS dissociator according to the manufacturer’s instructions. The cells were then passed through a 45-μm strainer and incubated with the cell-specific beads for 15 min. The cells were then loaded on LS columns and separated on a quadroMACS magnet, first with anti-O4 beads (elution was labelled oligodendrocyte lineage), and then with anti-ACSA beads to deplete astrocytes, and the remainder (flow-through) was a sample enriched in cells of the neuronal lineage. The target cells were washed once in PBS, then pelleted and snap-frozen. Pellets were subsequently thawed, and lysates were generated for immunoblot analysis. Light and electron microscopic imaging of myelin in FAM126A KO mice. Male mice at age P15 (three FAM126A KO and two littermate WT controls, on a mixed 129/C57BL/6 background) were perfused transcardially with 2% paraformaldehyde, 2.5% glutaraldehyde in 100 mM cacodylate buffer pH 7.4. Samples were postfixed with 1% osmium tetroxide, 1.5% potassium ferrocyanide in 100 mM cacodylate buffer pH 7.4, en bloc stained with uranyl acetate, dehydrated in increasing concentrations of ethanol and propylene oxide, and finally embedded in Epon. Samples were cured at 60 °C in an oven for 48 h. Epon blocks were sectioned using a Leica EM UC7 ultramicrotome (Leica Microsystems). For light microscopy, semithin sections (1 μm) of optic nerve, corpus callosum (sagittal section of the central area), and spinal cord (ventral funiculus cervical region) were generated and stained with toluidine blue. Light microscopy was performed on a Zeiss Axio Imager equipped with a Plan-Apochromat 100×/1.4 oil objective with an AxioCam MRc 5 colour CCD camera. For electron microscopy, ultrathin sections (60 nm) were contrasted with 2% uranyl acetate and Sato’s lead solutions and observed with a LEO 912AB Zeiss Transmission Electron Microscope (Carl Zeiss). Digital micrographs were taken with a 2k × 2k bottom-mounted slow-scan Proscan camera (ProScan) controlled by the EsivisionPro 3.2 software (Soft Imaging System). Statistics and reproducibility. Statistical analysis was performed using either Microsoft Excel or Prism 6 software. Detailed statistical information (which statistical test was used, number of independent experiments, P values, definition of error bars) is listed in individual figure legends. All immunoblots were repeated at least three times except for some immunoblots shown in Fig. 5 , which were repeated twice (see legend for Supplementary Fig. 4a for details). All imaging experiments were repeated three times. No statistical method was used to predetermine samples size for animal experiments, and the experiments were not randomized. Accession numbers. Structure factors and coordinates for the TTC7/FAM126A-N complex have been deposited in the Protein Data Bank (accession number 5DSE ).

An inherited disease of myelin marked by slow, progressive neurological impairment is caused by mutations of a gene that controls lipid metabolism, a finding that may shed insight into mechanisms to control the course of multiple sclerosis (MS), a Yale team has found. Mutations in a single gene, called FAM126A, causes a panoply of pathologies, such as developmental delay, intellectual disability, peripheral neuropathy, and muscle wasting, in addition to congenital cataracts. Until now the precise function of the gene was unknown. The labs of Yale cell biologists Pietro De Camilli and Karin Reinisch found that the protein encoded by the gene, called hyccin, helps produce a lipid crucial to formation of the myelin sheaths that surround and protect the axons of neurons throughout the nervous system. Their labs, working with other groups in the United States, Italy, and Germany, analyzed cells from patients suffering from the disease known as Hypomyelination and Congenital Cataract and found that FAM126A mutations results in the destabilization of an enzyme complex crucial to production of myelin. In MS, the course of the disease is critically dependent upon the reformation of myelin sheaths after immune system attacks then destroys them, eventually leading to the death of the neurons. The researchers hypothesize that the lipid that hyccin helps generate may play a key role in creation of myelin sheaths in normal development as well as in recovering MS patients. Postdoc Jeremy Baskin (now at Cornell) and graduate student Xudong Wu (now at Harvard) led the study in the De Camilli and Reinisch labs, respectively. The research was published Nov. 16 in the journal Nature Cell Biology.

10.1038/ncb3271

Chemistry

Metabolically engineered bacterium produces lutein

Sang Lee, Metabolic engineering of Escherichia coli with electron channelling for the production of natural products, Nature Catalysis (2022). DOI: 10.1038/s41929-022-00820-4. www.nature.com/articles/s41929-022-00820-4 Journal information: Nature Catalysis

https://dx.doi.org/10.1038/s41929-022-00820-4

https://phys.org/news/2022-08-metabolically-bacterium-lutein.html

Abstract The biosynthesis of natural products often requires eukaryotic cytochrome P450s (P450s) in combination with P450 reductase, in physical proximity, to perform electron-transfer reactions. Unfortunately, functional expression of eukaryotic P450s in bacteria remains generally difficult. Here we report an electron channelling strategy based on the application of Photorhabdus luminescens CipB scaffold protein, which allows efficient electron transfer between P450s and reductases by bringing these enzymes in close proximity. The general applicability of this electron channelling strategy is proved by developing recombinant Escherichia coli strains producing lutein, (+)-nootkatone, apigenin and l -3,4-dihydroxyphenylalanine ( l -DOPA), each of which requires P450s in its biosynthetic pathway. The production titres are then further enhanced by increasing the haem pathway flux or by optimization of the culture conditions. Remarkably, the final lutein strain produced 218.0 mg l −1 of lutein with a productivity of 5.01 mg l −1 h −1 in fed-batch fermentation under optimized culture conditions. Main Natural products have been widely used in the pharmaceutical, food, cosmetic and health industries because of their antioxidant, anti-tuberculosis, anti-inflammatory, anticancer and immunomodulatory properties 1 . Many functionally useful natural products are obtained by extraction from plants or animals, a process that is quite inefficient, with difficulties in separating target natural products from other similar metabolites 2 . Also, given the complex structures of these natural products, which feature an array of chiral or reactive functional groups, total chemical synthesis is rather difficult and costly. Such problems have motivated researchers to produce natural products by means of the fermentation of metabolically engineered microorganisms. Escherichia coli is considered to be one of the best host strains because of its high growth rate and the availability of various techniques for genetic manipulation and high cell-density culture 3 . Among many enzymes involved in the complex biosynthetic pathways of natural products, eukaryotic cytochrome P450s (P450s) are often critical enzymes in determining production efficiency, because they catalyse a range of reactions, including hydroxylation, dehydrogenation, isomerization and epoxidation. However, the expression of eukaryotic P450s in bacteria in functionally active forms has been a major challenge. In eukaryotes, membrane-associated P450s characteristically bind to the membranes of intracellular organelles, but these structures are uncommon in bacteria 4 . Furthermore, the physical distance between membrane-bound P450s and P450 reductase appears to be another critical factor in determining the reaction efficiency, because electrons generated by the oxidation of nicotinamide adenine dinucleotide phosphate (NADPH) need to be transferred from the P450 reductase to P450 4 . In a multi-step biosynthetic pathway involving competing reactions or toxic intermediates, the strategy of substrate channelling by tethering two or more enzymes in proximity has been useful in enhancing the overall reaction rates and rapidly converting metabolites to a desired product 5 , 6 , 7 . It was thus thought that electron channelling by tethering such enzymes involved in electron transfer may similarly allow more efficient electron transfers. In previous reports, electrons extracted from NADPH oxidation by P450 reductase may be immediately transferred to P450s when the enzymes are physically assembled. A recent report has highlighted the production of caffeic acid in E. coli by employing a heterotrimer-forming proliferating cell nuclear antigen scaffold system to tether bacterial P450, ferredoxin and ferredoxin reductase 8 , 9 . Another study has reported the genetic fusion of a bacterial P450 enzyme with a reductase to facilitate interaction between two enzymes 10 , 11 , 12 , 13 , 14 . These results also support the importance of an efficient electron-transfer mechanism during the production of natural products. In this Article, we describe how electron channelling is applied using the Photorhabdus luminescens CipB scaffold protein 5 , which has a relatively small size (11.29 kDa) and clusters to form a stable protein crystalline inclusion (PCI). The natural proclivity of CipB proteins to form scaffolds can be applied to assemble multiple enzymes into a functional PCI. We demonstrate the validity and universal applicability of this electron channelling strategy through four case studies, respectively producing lutein, (+)-nootkatone, apigenin and l -3,4-dihydroxyphenylalanine ( l -DOPA), all involving P450s in their biosynthetic pathways. The construction of the E. coli strains producing the four natural products is first described, followed by application of the electron channelling strategy to each strain to streamline the electron transfer between the P450s and their partner reductases. Further metabolic engineering and scale-up fermentation were also performed to increase the product titres. Results Construction of the lutein-producing E. coli strain Lutein is one of the xanthophylls that is naturally abundant in egg yolk, fruits and green leafy vegetables. Lutein exists in the macula lutea membranes of the human eye, protecting the eye against oxidative damage and excess radiation 15 . Intake of lutein reduces the risk or progression of some eye diseases, including age-related macular degeneration and cataracts 16 . It is these health benefits that has led to the increased demand for lutein. Accordingly, there has been much interest in developing metabolically engineered bacteria capable of efficiently producing lutein to replace the inefficient plant-based production system. We constructed the lutein biosynthetic pathway in E. coli , which can produce the building blocks of carotenoids—isopentenyl diphosphate (IPP) and dimethylallyl diphosphate—via the inherent 1-deoxy- d -xylulose 5-phosphate (DXP) pathway, and then convert them to farnesyl diphosphate by means of farnesyl diphosphate synthase. To complete the lutein biosynthetic pathway in E. coli , eight genes constituting the pathway from farnesyl diphosphate to lutein were introduced from heterologous hosts (Fig. 1 ). First, a biosynthetic pathway from farnesyl diphosphate to lycopene was established by employing Pantoea ananatis -derived crtE , crtB and crtI encoding geranylgeranyl pyrophosphate synthase, phytoene synthase and phytoene dehydrogenase, respectively (see Supplementary Note 1 for details of the cloning procedure). Next, a biosynthetic pathway from lycopene to lutein was established by employing the Arabidopsis thaliana -derived genes encoding lycopene ε-cyclase ( LUT2 ) 17 , lycopene β-cyclase ( LCYB ) 17 , β-carotene 3-hydroxylase ( LUT5 ) 18 , carotene ε-monooxygenase ( LUT1 ) 19 and P450 reductase 2 ( ATR2 ) 20 (Supplementary Note 2 provides further details). For their functional expression, the LUT2 , LCYB , LUT5 and LUT1 genes were codon-optimized and their native N-terminal signal peptide sequences predicted by ChloroP software 21 were eliminated to result in truncated genes trLUT2 , trLCYB , trLUT5 and trLUT1 , respectively. Although ATR2 also harbours a predicted N-terminal signal peptide, it was not truncated because the activity of the ATR2 harbouring the native signal peptide is reportedly higher than that of N-terminal truncated ATR2 in E. coli 22 . The primary lutein biosynthetic pathway constructed here is shown in Fig. 1 , together with the branched competing pathway of β-carotene formation. Fig. 1: The primary lutein biosynthetic pathway. The heterologous lutein biosynthetic pathway constructed in this study and the enzymes involved in the pathway are shown. The bent arrow and T-shaped structure shown on the gene cassettes indicate the promoter and transcription terminator, respectively. 5-ALA, 5-aminolevulinic acid; DXP, 1-deoxy- d -xylulose 5-phosphate; FPP, farnesyl diphosphate; l -Glu-tRNA, l -glutamyl-tRNA; G3P, d -glyceraldehyde 3-phosphate; GGPP, geranylgeranyl diphosphate; TCA cycle, tricarboxylic acid cycle. Full size image Plasmids pLUT1, pLUT2 and pLUT3, which carries the genes comprising the lutein biosynthetic pathway, were transformed into the WLGB-RPP strain 23 , a previously developed engineered E. coli K-12 W3110 strain possessing enhanced metabolic flux towards the carotenoid biosynthetic pathway, to construct the LUT1 strain (Table 1 ; Supplementary Note 1 provides details of the procedure for strain construction). Glycerol was used as a carbon source (20 g l −1 for all flask cultures), as it was found to be superior to other carbon sources including glucose for the production of carotenoids in previous studies by the authors 24 , 25 . The expression of heterologous genes from the plasmids was fully induced by adding 1 mM of isopropyl β- d -1-thiogalactopyranoside (IPTG). However, the LUT1 strain only produced α-carotene (0.54 mg l −1 ) and slightly more β-carotene (0.61 mg l −1 ), while lutein production was not detected (Fig. 2 ). Table 1 List of key production strains used in this study Full size table Fig. 2: Results of flask culture for lutein production. Results of flask cultures of the LUT1–4 and LUT4M strains are shown. Data are presented as mean ± s.d. of three independent experiments ( n = 3). For the comparison of data groups, one-way analysis of variance (ANOVA) followed by Tukey’s multiple comparison test was used. P = 1.44 × 10 −9 . Different letters (A-D) indicate significant differences among data groups at a significance level of 0.05. Source data Full size image It was reasoned that the promiscuous enzyme trLCYB diverts the metabolic flux to a competing reaction converting lycopene to β-carotene, thereby reducing the metabolic flux towards the α-carotene biosynthetic pathway and resulting in low-level production of lutein. To reduce the metabolic flux towards the β-carotene pathway and consequently increase the flux towards α-carotene formation, a substrate channelling strategy was employed. Substrate channelling is a well-known strategy that has been widely used to streamline enzymatic reactions 6 , 26 . In this study, P. luminescens -derived CipA, which forms enzymatically active PCI in E. coli , was used. PCIs are also formed when CipA is fused to other enzymes 5 . Three different substrate channels using the CipA-scaffold protein were designed and tested to examine whether the flux towards α-carotene could be increased. The CipA protein was directly fused to the N-terminal region of the enzymes without using any linkers, based on ref. 5 . Three channels were designed to tether CrtI, trLUT2 and trLCYB; CrtI and trLUT2; and trLUT2 and trLCYB. The LUT2, LUT3 and LUT4 strains were correspondingly constructed (Table 1 and Supplementary Note 1 ). To test the three substrate channelling combinations, the LUT2, LUT3 and LUT4 strains were cultured under conditions identical to those used to culture LUT1, as described above. Among them, the LUT4 strain, which can assemble trLUT2 and trLCYB through CipA scaffolding, produced lutein to the highest titre of 0.84 mg l −1 (Fig. 2 ; further discussion is provided in Supplementary Note 3 ). Lutein produced by LUT4 was confirmed by a high-performance liquid chromatography–mass spectrometry (HPLC-MS) analysis (Supplementary Fig. 1 ). At the same time, 6.74 and 4.50 mg l −1 of α-carotene and β-carotene, respectively, were also produced (Fig. 2 ). When CipA was fused to only one of three enzymes among CrtI, trLUT2 and trLCYB, lutein was not produced. This result indicates that CipA fusion itself has not conferred a positive effect on enzyme activity other than the formation of the substrate channel (Supplementary Note 3 ). As the next step, given the production of a non-negligible amount of β-carotene from LUT4 (4.50 mg l −1 ), it was inevitable to improve the product selectivity of trLCYB towards α-carotene. It was reported that A. thaliana strain harbouring a mutant LCYB (G451E) produced more lutein and α-carotene compared with the wild type 27 . It was thus hypothesized that the introduction of this mutation might alter the product selectivity towards α-carotene. A flask culture of LUT4M in which the mutation was introduced into LCYB in LUT4 (Table 1 and Supplementary Note 1 ) produced 9.11 and 1.43 mg l −1 of α-carotene and β-carotene, respectively (Fig. 2 ). Introduction of the G404E mutation into trLCYB effectively increased the production ratio of α-carotene to β-carotene from 1.50 to 6.37. Accordingly, lutein production was also increased to 1.70 mg l −1 . However, most of the α-carotene was not converted to lutein, indicating that the reaction steps comprising two P450s are the major bottlenecks in the lutein biosynthetic pathway. This problem was solved by electron channelling, as described below after sections on the construction of strains producing the other three natural products. Construction of the E. coli strain producing (+)-nootkatone (+)-Nootkatone is a high-value sesquiterpene ketone that is responsible for aroma in cedar and grapefruit 28 . Due to its pleasant aroma to humans, (+)-nootkatone is commonly used in the flavour and fragrance industries. It has also been used as an ingredient in insect repellents 29 . Several studies have concentrated on (+)-nootkatone production using E. coli 30 , 31 , but none achieved one-step de novo production. (+)-Nootkatone is a terpenoid and shares the biosynthetic pathway with lutein up to farnesyl diphosphate. From farnesyl diphosphate, (+)-valencene is produced by valencene synthase (ValC), and (+)-nootkatone is produced from (+)-valencene via nootkatol by CYP71D55 32 , which is a P450 enzyme (Fig. 3 ). First, to construct the (+)-valencene producing E. coli strain VAL1, valC from Callitropsis nootkatensis 33 was introduced into the E. coli WLGB-RPP strain (Supplementary Note 1 ) 23 . Because (+)-valencene is a highly volatile compound having low solubility in water, a two-phase culture method was used for the in situ extraction of (+)-valencene produced during the cultivation of VAL1. As a result, 2.32 mg l −1 of (+)-valencene was produced (Supplementary Fig. 4 ). The (+)-valencene titre was further increased to 7.31 mg l −1 when the metabolic flux in the DXP pathway was strengthened by amplifying the dxs (encoding DXP synthase) and fni (encoding IPP isomerase) genes from Bacillus subtilis 34 and the codon-optimized and signal peptide-truncated GPPS2 (geranyl diphosphate synthase) encoding gene ( trGPPS2 ) from Abies grandis 35 (Supplementary Fig. 4 and Supplementary Note 1 ). Next, to convert (+)-valencene to (+)-nootkatone, the codon-optimized and N-terminal signal peptide-truncated V482I/A484I mutant CYP71D55 (trCYP71D55mut) from Hyoscyamus muticus 32 and the codon-optimized ATR2 from A. thaliana accompanied as a partner reductase were used. The N-terminal signal peptide region of CYP71D55 was identified by referring to UniProtKB annotation 36 . Additionally, because the conversion efficiency of (+)-valencene to (+)-nootkatone by CYP71D55 was reported to be low 32 , the zsd1 gene (encoding zerumbone synthase) from Zingiber zerumbet was introduced to promote the conversion of nootkatol to (+)-nootkatone 37 . The resultant strain NKT1 (Table 1 and Supplementary Note 1 ) successfully produced 0.61 mg l −1 of (+)-nootkatone in a flask culture (Fig. 4c ). However, the titre was far below than that of its precursor (+)-valencene produced from VAL2 (7.31 mg l −1 ), indicating the low activity of P450, which is the main actor in converting (+)-valencene to (+)-nootkatone. Fig. 3: Biosynthetic pathways of (+)-nootkatone, apigenin and L-DOPA. Heterologous (+)-nootkatone, apigenin and l -DOPA biosynthetic pathways constructed in E. coli and heterologous genes introduced in this study are shown. The bent arrow and T-shaped structure shown on the gene cassettes indicate the promoter and transcription terminator, respectively. DAHP, 3-deoxy- d -arabino-heptulosonic acid 7-phosphate; DXP, 1-deoxy- d -xylulose 5-phosphate; E4P, d -erythrose-4-phosphate; FPP, farnesyl diphosphate; G3P, d -glyceraldehyde 3-phosphate; l -DOPA, l -3,4-dihydroxyphenylalanine; PEP, phosphoenolpyruvate; L, linker. Full size image Fig. 4: Applying the electron channelling strategy to produce natural products in E. coli . a , Schematic of electron transfer between ATR2 and two P450s. The CipB scaffold protein fused ATR2, trLUT5 and trLUT1 are expected to be clustered together in the cell, and their close proximity would lead to more efficient electron transfer between them. b , Results of the flask cultures of the LUT4M and LUT5M strains. Data are presented as mean ± s.d. of three independent experiments ( n = 3). Statistically significant differences between two data groups were determined using a two-tailed Student’s t -test. *** P < 0.001; P = 5.07 × 10 −5 . c – e , Results of flask or test tube cultures for increasing (+)-nootkatone ( c ), apigenin ( d ) and l -DOPA ( e ) production in E. coli by employing electron channelling. (+)-Nootkatone and l -DOPA were produced from glycerol in flask cultures, while apigenin was converted from naringenin in test tube cultures. Data are presented as mean ± s.d. of three independent experiments ( n = 3). For the comparison of data groups, ANOVA followed by Tukey’s multiple comparison test was used. P = 1.20 × 10 −3 ( c ), P = 1 . 92 × 10 −7 ( d ), P = 1 . 75 × 10 −5 ( e ). A, B and C indicate significant differences among data groups at a significance level of 0.05. Source data Full size image Construction of the E. coli strain producing apigenin Apigenin is a yellow bioactive compound found in plants that is categorized as a flavonoid. It possesses various biological activities, including antioxidant, chemopreventive, anticancer and anti-invasive activities, suggesting its therapeutic potential for the treatment of many chronic diseases 38 . From tyrosine, a series of reactions catalysed by tyrosine ammonia-lyase (TAL), 4-coumarate-CoA ligase 1 (4CL1), chalcone synthase (CHS) and chalcone isomerase (CHI) produce coumaric acid, coumaroyl-CoA, naringenin chalcone and naringenin, respectively. Then, naringenin can be converted to apigenin by flavone synthase (FNS; Fig. 3 ). There are two types of FNS; type I has been used for the microbial production of apigenin as a cytoplasmic enzyme 39 , but type II has not been used owing to its intrinsic complexity as a eukaryotic-derived P450 40 . Here, the type II FNS (FNS II) was selected to examine the effectiveness of the electron channelling strategy. The NAR strain—the E. coli BTY5 strain 41 harbouring pTY13-HisTAL and pNAR, developed for naringenin production in our previous study 42 —was cultured in flasks for the confirmation of naringenin production; 65.0 mg l −1 of naringenin was produced. For the next step, the codon-optimized truncated FNS II free of N-terminal signal peptide (trCYP93B16) from Glycine max 43 was used for the conversion of naringenin to apigenin, and ATR2 from A. thaliana was used as its partner reductase. The N-terminal signal peptide region of CYP93B16 was identified by referring to UniProtKB annotation 36 . Before examining the feasibility of the de novo production of apigenin, the successful conversion of naringenin to apigenin was first examined using the AGN1 strain— E. coli BL21(DE3) harbouring pAGN1 (harbouring ATR2 and trCYP93B16 ) (Table 1 and Supplementary Note 1 ). In a test-tube culture, AGN1 produced 2.22 mg l −1 of apigenin from 100 mg l −1 of naringenin supplemented, with 30.1 mg l −1 of naringenin remaining unconverted (Fig. 4d ). Construction of the E. coli strain producing l -DOPA The final natural product tested was l -DOPA, an important precursor for many alkaloids and catecholamines, as well as melanin in plants 44 . It can be used as a therapeutic agent for treating Parkinson’s disease and other neurodegenerative disorders 45 . l -DOPA shares its metabolic pathway down to tyrosine with that of apigenin, and it is mainly converted from tyrosine via hydroxylation (Fig. 3 ). Recently, the F309L mutant CYP76AD1 from Beta vulgaris showing enhanced tyrosine hydroxylase activity was reported 46 . We thus exploited this mutant enzyme to examine the universal applicability of the electron channelling strategy. We used the codon-optimized truncated mutant CYP76AD1 free of the N-terminal signal peptide from B. vulgaris (trCYP76AD1mut) and ATR2 from A. thaliana as a P450 partner reductase. The N-terminal signal peptide region of CYP76AD1mut was identified by referring to the UniProtKB annotation 36 . The trCYP76AD1mut and ATR2 genes were introduced into the tyrosine-overproducing E. coli strain BTY5.13 41 to construct DP1 (Table 1 and Supplementary Note 1 ). The DP1 strain produced 1.32 mg l −1 of l -DOPA in flask culture (Fig. 4e ). Application of electron channelling for lutein production As described earlier, it was evident that the two P450s, trLUT5 and trLUT1, were the major bottlenecks in the lutein biosynthetic pathway. According to the sodium dodecyl sulfate polyacrylamide gel electrophoresis (SDS–PAGE) results, both P450s mostly existed in the cytoplasm as expected, whereas ATR2 existed in the membrane (Supplementary Fig. 2 ). It was thus hypothesized that the spatial distance of the cytosolic P450s and membrane-associated ATR2 might result in a poor electron transfer between the enzymes. To solve this problem, we designed a strategy of electron channelling, similar to substrate channelling, by tethering trLUT5, trLUT1 and ATR2 together using a scaffold protein (Fig. 4a ). The P. luminescens CipB scaffold protein, which is known to form its own PCI that is physically independent of the CipA-scaffold protein 5 , was employed. As in the case of CipA, CipB was directly fused to the enzymes without using any linkers (Supplementary Note 1 ). The resultant strain LUT5M produced 5.80 mg l −1 of lutein, which is 3.41-fold of that obtained with the LUT4M strain (1.70 mg l −1 ) (Fig. 4b ; Supplementary Note 4 provides a discussion on the demonstration of an electron channelling effect). In vivo formation of PCI comprising trLUT5, trLUT1 and ATR2 was confirmed by microscopic observation and SDS–PAGE analysis (Supplementary Fig. 2 ). The constituent of isolated PCI comprising the three enzymes was confirmed by western blotting (Supplementary Fig. 2 ). Next, the effects of CipB fusion on individual enzymes and the subsequent changes in lutein production profiles were compared. To see the effect of electron channelling only, five different LUT1-derived strains, in which CipB was fused with ATR2, trLUT5 and trLUT1 (LUT1B), ATR2 (LUT1B1), trLUT5 (LUT1B2), trLUT1 (LUT1B3), and trLUT5 and trLUT1 (LUT1B4), were constructed (Supplementary Fig. 3 and Supplementary Note 1 ). The LUT1 and LUT1B strains served as controls. Flask cultures of the LUT1, LUT1B, LUT1B1, LUT1B2, LUT1B3 and LUT1B4 strains produced 0, 2.46, 0.99, 1.97, 1.98 and 0.24 mg l −1 of lutein, respectively (Supplementary Fig. 2 ). The performances of individual enzymes were improved when CipB was fused, which led to increased lutein production in LUT1B1, LUT1B2 and LUT1B3 strains. When both trLUT5 and trLUT1 were fused to CipB (LUT1B4), however, lutein production decreased, which seems to be due to the reduced accessibility of the enzymes toward the substrates caused by the formation of PCIs consisting of multiple enzymes. Due to the characteristic that the substrate can only access enzymes presented on the surface of PCIs, the higher the number of enzymes anchored to PCIs, the less accessibility each enzyme has to the freely diffusing substrate 5 . Application of electron channelling for other products As the CipB scaffold-assisted electron channelling strategy allowed increased production of lutein, we examined whether this strategy is generally applicable to the production of other natural products, such as (+)-nootkatone, apigenin and l -DOPA examples in this study, which require electron transfer in their biosynthesis. The first committed step of (+)-nootkatone production requires oxidation of (+)-valencene, which is catalysed by CYP71D55, a P450 enzyme. For electron channelling, the two enzymes ATR2 and trCYP71D55mut were fused with the CipB scaffold protein without any linker in the NKT1 strain. The resultant NKT2 strain, however, produced even less (+)-nootkatone (0.16 mg l −1 ) than that of the control strain NKT1 (0.61 mg l −1 ) in the flask culture (Fig. 4c ). Because CipB may cause steric hindrance on passenger enzymes mainly due to its characteristic PCI formation, a flexible linker sequence (ASASNGASA) was inserted between CipB and ATR2/trCYP71D55mut to minimize this hindrance effect. The resulting NKT3 strain produced 4.43 mg l −1 of (+)-nootkatone (Fig. 4c ), which is 7.26 times that obtained with the control strain NKT1 (0.61 mg l −1 ). The (+)-nootkatone produced from NKT3 was confirmed by gas chromatography–mass spectrometry (GC-MS), and the in vivo formation of PCI comprising CipB-fused trCYP71D55mut and CipB-fused ATR2 was confirmed by microscopic observations and SDS–PAGE analysis (Supplementary Fig. 4 ). In the apigenin biosynthetic pathway, the conversion from naringenin to apigenin is catalysed by a P450 enzyme, CYP93B16. For electron channelling, CipB was fused with ATR2 and trCYP93B16 without any linker (Supplementary Note 1 ). Similar to the above (+)-nootkatone case, apigenin produced (1.66 mg l −1 ) by the resulting AGN2 strain was lower than that produced (2.22 mg l −1 ) by the AGN1 strain in the test-tube culture supplemented with naringenin (Fig. 4d ). Thus, the same flexible linker sequence (ASASNGASA) as used in the (+)-nootkatone case was inserted between CipB and the two enzymes trCYP93B16 and ATR2. The resulting AGN3 strain produced 5.66 mg l −1 of apigenin from 100 mg l −1 of naringenin, which is 2.55-fold of that (2.22 mg l −1 ) produced by the AGN1 strain in the test-tube culture (Fig. 4d ). Plasmid pAGN3 was then transformed into the NAR strain to construct AGN4. Flask culture of AGN4 resulted in de novo production of 1.05 mg l −1 of apigenin from glycerol. Apigenin production was confirmed by HPLC-MS, and the in vivo formation of PCI comprising trCYP93B16 and ATR2 was confirmed by microscopic observation and SDS–PAGE analysis (Supplementary Fig. 5 ). Biosynthesis of the last example product, l -DOPA, also requires a P450-catalysed reaction at the final step. The corresponding enzymes in the DP1 strain, trCYP76AD1mut from B. vulgaris and its redox partner ATR2, were then genetically fused with CipB without (DP2 strain) or with (DP3 strain) a flexible linker. In the flask culture, production of l -DOPA by DP3 increased to 24.50 mg l −1 , which is 18.56-fold of that obtained from the control strain DP1 (Fig. 4e ). This result indicated that the steric hindrance caused by the direct CipB fusion with the enzymes has inhibited l -DOPA production as in the (+)-nootkatone and apigenin production cases. l -DOPA production was confirmed by HPLC-MS, and in vivo formation of PCI comprising CipB-fused trCYP76AD1mut and CipB-fused ATR2 was confirmed by microscopic observation and SDS–PAGE analysis (Supplementary Fig. 6 ). It should be noted that the CipB-based electron channelling worked well without an additional attachment of linker to the enzymes (P450s and P450 reductase) in the lutein case, successfully increasing the production level, whereas the presence of a linker between fusion partners was needed to increase production of (+)-nootkatone, apigenin and l -DOPA by alleviating the steric hindrance effects (Supplementary Fig. 7 ). This difference might be due to several factors, such as structural changes affecting the flexibility of the N-terminal peptide of P450 fused to the C-terminus of CipB and variations in the expression yield caused by the changes in translation efficiency upon CipB fusion. It has been reported that the sequence or length of the linker between the two fused proteins affects the flexibility of the fusion complex, substantially influencing its activities 10 , 47 . Increasing haem production for enhancing P450 activity After verification of the CipB scaffold-assisted electron channelling strategy, further engineering was pursued to enhance the activity of P450. Given that P450s require haem as a coenzyme, acting as a mediator transferring electrons from NADPH to oxygen, limitation of haem availability would impair P450 function. To increase the metabolic flux in the C5 haem biosynthetic pathway (Fig. 5a ), the E. coli hemA fbr , hemL , hemB and hemH genes 48 encoding feedback-resistant glutamyl-tRNA reductase, glutamate-1-semialdehyde 2,1-aminomutase, porphobilinogen synthase and ferrochelatase, respectively, were selected for overexpression (see Supplementary Note 5 for further details). First, the haem pathway flux was amplified in the lutein-producing strain (Supplementary Note 1 ). Flask cultures of LUT5M, LUT5MH1 (LUT5M overexpressing hemA fbr and hemL ) and LUT5MH2 (LUT5M overexpressing hemA fbr , hemL , hemB and hemH ) produced 6.03, 10.34 and 6.04 mg l −1 of lutein, respectively (Fig. 5b ; see Supplementary Note 5 for further discussion), suggesting that the increased production of haem by overexpression of the hemA fb and hemL genes was beneficial for enhancing the conversion of α-carotene to lutein catalysed by haem-requiring trLUT5 and trLUT1. Fig. 5: Enhancement of lutein production by increasing haem supply. a , The biosynthetic pathways of lutein and haem and a schematic diagram of electron transfer from NADPH to oxygen via ATR2 and trLUT5/trLUT1. 5-ALA, 5-aminolevulinic acid; DXP, 1-deoxy- d -xylulose 5-phosphate; l -Glu-tRNA, l -glutamyl-tRNA; G3P, d -glyceraldehyde 3-phosphate; TCA cycle, tricarboxylic acid cycle. b , Flask culture results of the LUT5M, LUT5MH1 and LUT5MH2 strains and the corresponding images of flasks at the end of the culture. Because haem and its precursors are brown, the amount of them produced can be predicted by examining the colour of the culture broth. Data are presented as mean ± s.d. of three independent experiments ( n = 3). For the comparison of data groups, ANOVA followed by Tukey’s multiple comparison test was used. P = 8.52 × 10 −5 . Different letters (A, B) indicate significant differences among data groups at a significance level of 0.05. Source data Full size image Because the overexpression of hemA fbr and hemL effectively increased lutein production, these two genes were also overexpressed in the strains producing (+)-nootkatone (NKT3) and apigenin (AGN4). In the flask-culture results, apigenin production increased to 2.09 mg l −1 , two times higher than that obtained with AGN4 (1.05 mg l −1 ). On the other hand, (+)-nootkatone production decreased with the overexpression of genes involved in the haem biosynthetic pathway (Supplementary Fig. 8 ; see Supplementary Note 6 for further details). Optimization of culture conditions Next we examined different culture conditions for the enhanced production of natural products using the strains developed thus far. The best lutein producer strain LUT5MH1 was cultured under different conditions: different media (MR and R/2), cultivation temperatures (22, 25, 28, 30 and 37 °C) and IPTG concentration (0, 0.01, 0.02, 0.05, 0.1, 0.2, 0.5 and 1 mM) were examined (Supplementary Note 7 ). Because the strain grown on R/2 medium led to a higher lutein production level than the MR medium (Fig. 6a ), the R/2 medium was used subsequently. Lowering the cultivation temperature to 28 °C increased lutein production to 16.29 mg l −1 in R/2 medium (Fig. 6b ). The lutein titre was further increased to 23.50 mg l −1 by using a lower concentration (0.05 mM) of IPTG for induction in R/2 medium at 28 °C (Fig. 6b ). Under this condition, 3.04 and 0.66 mg l −1 of α-carotene and β-carotene, respectively, were produced, showing notably reduced by-product formation. The different cultivation temperatures and IPTG concentrations were also examined for (+)-nootkatone and apigenin production. In both cases, the lower temperature led to increased production, but a lower concentration of IPTG did not significantly change the production level (Supplementary Fig. 8 ; see Supplementary Note 6 for further details). Fig. 6: Enhancement of lutein production by optimization of culture conditions. a , R/2 and MR media were compared. Strain LUT5MH1 was cultured in R/2 or MR medium at 30 °C, and gene expression was induced with 1 mM IPTG. Data are presented as mean ± s.d. of three independent experiments ( n = 3). Statistically significant differences between two data groups were determined using a two-tailed Student’s t -test. *** P < 0.001; P = 3.87 × 10 −5 . b , Various culture temperatures were tested for lutein production. LUT5MH1 was cultured in R/2 medium at 37, 30, 28, 25 or 22 °C and gene expression was induced with 1 mM IPTG. The effects of IPTG concentrations used for induction on lutein production were also compared. LUT5MH1 was cultured in R/2 medium at 28 °C, and IPTG was added to a final concentration of 0, 0.01, 0.02, 0.05, 0.1, 0.2, 0.5 or 1 mM when the OD 600 reached ~0.4–0.6. Data are presented as mean ± s.d. of three independent experiments ( n = 3). For a comparison of data groups, ANOVA followed by Tukey’s multiple comparison test was used. P left = 3.06 × 10 11 , P right = 2.27 × 10 −15 . Different letters (A–E) indicate significant differences among data groups at a significance level of 0.05. Source data Full size image Fed-batch fermentation Next, fed-batch fermentations were performed under the conditions optimized in the previous section. Fed-batch culture of the LUT5MH1 strain was performed in R/2 medium as follows. At the end of the batch phase, a nutrient was automatically added by a pH-stat feeding strategy 3 using a feeding solution containing 800 g l −1 of glycerol (see Methods for the composition). Based on the results of the flask cultures above, the temperature was maintained at 28 °C during the entire culture period, including the seed culture step. IPTG was added to a final concentration of 0.05 mM for induction when the optical density at 600 nm (OD 600 ) reached 20–30 (corresponding to 8–11 gDCW l −1 ; DCW, dry cell weight). The initial fed-batch fermentation of LUT5MH1 produced 25.47 mg l −1 (0.44 mg gDCW −1 ) of lutein with a productivity of 0.64 mg l −1 h −1 (Supplementary Fig. 9 ). Surprisingly, the lutein titre obtained from the fed-batch fermentation was not much higher than that (23.50 mg l −1 ; 5.8 mg gDCW −1 ) obtained by the flask culture, suggesting that the optimal condition for fed-batch culture is different from that of flask culture. Thus, the fed-batch fermentation condition was subsequently optimized. The cultivation temperature before the addition of IPTG was increased from 28 to 37 °C to shorten the lag phase. Also, higher IPTG concentrations (0.2, 0.5 and 1 mM) were examined for proper induction considering the higher cell density during the fed-batch culture (see Supplementary Note 8 for further details). Among the various fermentation conditions examined, the most efficient lutein production was observed when cells were grown until the OD 600 reached 23.4 at 37 °C, followed by shifting the temperature down to 28 °C immediately before induction with 0.5 mM IPTG; the lutein titre and productivity increased notably to 133.44 mg l −1 (2.17 mg gDCW −1 ) and 2.72 mg l −1 h −1 , respectively (Supplementary Fig. 9 ). Pictures of the fermentation broth taken during the fed-batch fermentation of LUT5MH1 are shown in Fig. 7a . Fig. 7: Fed-batch fermentation of LUT5MH1 for lutein production. a , Pictures of the fermenter taken during the fed-batch fermentation of LUT5MH1 are shown. b , Time profiles of the fed-batch fermentation of the LUT5MH1 strain with carbon starvation are shown. The temperature was maintained at 37 °C, then reduced to 28 °C after the addition of 0.5 mM IPTG. Cells were starved for 2 h (~12–14 h) after the batch phase. The black arrowhead denotes the time point of IPTG addition. Source data Full size image In one of the fed-batch fermentations, we accidentally forgot to provide a feeding solution for 2 h at the end of the batch phase, causing cells to undergo carbon starvation, after which feeding solution was provided via the pH-stat feeding strategy as in the previous fed-batch fermentations. Surprisingly, lutein production increased notably to 194.20 mg l −1 (3.38 mg gDCW −1 ) with a productivity of 3.35 mg l −1 h −1 (Supplementary Fig. 9 ). To check the reproducibility of this unexpected phenomenon, a fed-batch fermentation was conducted under identical conditions with carbon starvation for 2 h at the end of the batch phase, but deliberately this time. As a result, 218.0 mg l −1 (4.01 mg gDCW −1 ) of lutein was produced with productivity of 5.01 mg l −1 h −1 (Fig. 7b ). To understand this phenomenon, the transcriptomes of carbon-starved cells and normally grown cells were analysed and compared to examine the changes in the metabolic state of the cells (Supplementary Fig. 10 and Supplementary Tables 2 and 3 ). According to the results of the transcriptomic analysis and subsequent experiments, it is likely that the changes in the expression levels of stress-response genes were related to the final titre of lutein. More details are described in Supplementary Note 9 . Fed-batch cultures of the strains producing (+)-nootkatone and apigenin were also performed under the previously set cultivation conditions, yielding 15.73 mg l −1 of (+)-nootkatone and 26.17 mg l −1 of apigenin, respectively (Supplementary Fig. 11 ; see Supplementary Note 6 for further details). Conclusions In this Article, we describe the development of the CipB scaffold-assisted electron channelling system to facilitate electron transfers between P450s and their partner reductases, and demonstrate its general applicability by the enhanced production of lutein, (+)-nootkatone, apigenin and l -DOPA in metabolically engineered E. coli . No linker was needed in fusing CipB and the enzymes (P450s and P450 reductase) in the lutein case, but a linker was required between the fusion partners in the cases of (+)-nootkatone, apigenin and l -DOPA. Such a difference seems to be due to the geometrical changes caused by the fusion of CipB and enzymes and the changes of expression levels upon CipB fusion 10 , 47 . It is also worthwhile noting that substrates are only accessible to the enzymes present on the surface of a PCI when using the CipB-mediated electron channelling system 5 . This means that the smaller ratio of the surface area to the volume of PCI led to the lower enzymatic conversion efficiency. Thus, when multiple P450s and P450 reductases need to be tethered, it will be helpful to employ other scaffold proteins, such as CipA, as used for the substrate channelling in this study, together with CipB. After validating the efficacy of the electron channelling strategy, the strain performances were further improved by metabolic engineering. Last, but not least, the fermentation conditions, including the cultivation temperature, inducer concentration and the serendipitously found carbon starvation after the batch phase, among others, were optimized, greatly affecting the final titre, content and productivity of the natural products. The beneficial effect of carbon starvation on increased lutein production suggests that properly inducing cellular stress responses can be another important strategy for enhancing the production of carotenoids and, potentially, some other natural products, requiring the prevention of oxidative stress. By taking these steps of developing strains and fermentation processes, we have developed an engineered E. coli strain producing lutein at high titre (218.0 mg l −1 ; 4.01 mg gDCW −1 ) and productivity (5.01 mg l −1 h −1 ), and achieved de novo production of (+)-nootkatone, apigenin and l -DOPA using plant-derived P450s in E. coli . Although the lutein content (20 mg gDCW −1 ) in marigold flowers is higher than that of the engineered E. coli strain developed in this study, the lutein produced in plants is difficult to extract and purify because it exists mostly in ester forms, with the longer chemical structures making it firmly embedded within the cell wall 49 . Also, it takes ~205 days to harvest the flower, which is not economically feasible considering the productivity 49 . In the best microalgal system, the lutein content reached was 13.18 mg gDCW −1 after 12 days of cultivation 50 . Thus, metabolically engineered E. coli producing 218.0 mg l −1 of free lutein in less than two days is superior to other lutein production systems with respect to both productivity and extraction efficiency. Eukaryotic P450s often play critical roles in the biosynthesis of natural products, but the difficulty of expressing them in functionally active forms in bacteria and the inefficient electron transfer between P450s and P450 reductases have been substantial problems. The electron channelling strategy reported here will be generally useful to solve such problems and enhance the production of natural products requiring P450s for their biosynthesis, as successfully demonstrated by the increased production of four example natural products, lutein, (+)-nootkatone, apigenin and l -DOPA. Methods Strains and media All strains constructed and used in this study are listed in Supplementary Table 1 , and the key production strains are listed in Table 1 with brief descriptions. DH5α, the E. coli strain used for routine gene-cloning works, was aerobically cultivated in lysogeny broth (LB; 10 g tryptone, 5 g yeast extract and 10 g NaCl per litre) at 37 °C while being shaken at 220 r.p.m. The WLGB-RPP and BTY5 strains—engineered E. coli strains that were developed in our previous work for lycopene 23 and tyrosine 41 production, respectively—were used to develop metabolically engineered strains for lutein, (+)-nootkatone, apigenin and l -DOPA production. Cells were first cultivated in a 14-ml test tube containing 3 ml of Terrific broth (TB) medium (20 g tryptone, 24 g yeast extract, 4 ml of glycerol, 0.017 M KH 2 PO 4 and 0.072 M K 2 HPO 4 per litre) 51 at 30 °C while being shaken at 220 r.p.m., until reaching an OD 600 of ~1–2, then were transferred to a 250-ml baffled flask containing 20 ml of TB medium. After cultivation at 30 °C while shaking at 220 r.p.m. until an OD 600 of ~1–2, cells were transferred to a 250-ml baffled flask containing 50 ml of R/2 or MR medium (R/2 unless otherwise mentioned) supplemented with 20 g l −1 of glycerol and 3 g l −1 of yeast extract, and cultured at 30 °C (unless otherwise mentioned) while being shaken at 200 r.p.m. for 48 h after IPTG induction. For the cultivation of BTY5-derived strains, 3 g l −1 of (NH 4 ) 2 SO 4 was additionally supplemented to the R/2 medium, and for VAL strains, 10 ml of dodecane was added after IPTG induction for two-phase cultivation. In the experiment for the in vivo conversion of naringenin to apigenin, E. coli BL21(DE3)-derived strains were transferred to a 14-ml test tube containing 3 ml of R/2 medium supplemented with 20 g l −1 of glycerol and 3 g l −1 of yeast extract after the first cultivation in 3 ml of TB medium, followed by cultivation at 30 °C while being shaken at 200 r.p.m. for 24 h after IPTG induction. At the induction point, 100 mg l −1 of naringenin was supplied as a substrate. Unless otherwise noted, 1 mM IPTG was used for the induction of heterologous genes when the OD 600 reached ~0.4–0.6. The R/2 medium (pH 6.8) contains 2 g (NH 4 ) 2 HPO 4 , 6.75 g KH 2 PO 4 , 0.85 g citric acid, 0.7 g MgSO 4 ·7H 2 O and 5 ml of a trace metal solution (10 g FeSO 4 ·7H 2 O, 2.25 g ZnSO 4 ·7H 2 O, 1 g CuSO 4 ·5H 2 O, 0.5 g MnSO 4 ·5H 2 O, 0.23 g Na 2 B 4 O 7 ·10H 2 O, 2 g CaCl 2 ·2H 2 O and 0.1 g (NH 4 ) 6 Mo 7 O 24 per litre of 5 M HCl) per litre 52 . The MR medium (pH 6.8) contains 4 g (NH 4 ) 2 HPO 4 , 6.67 g KH 2 PO 4 , 0.8 g citric acid, 0.8 g MgSO 4 ·7H 2 O and 5 ml of a trace metal solution 30 . For selective pressure, 25 mg l −1 of kanamycin (Km), 34 mg l −1 of chloramphenicol (Cm) 100 mg l −1 of ampicillin (Ap) and/or 100 mg l −1 of spectinomycin (Spc) were used when required. To monitor the cell density, OD 600 was measured with a spectrophotometer (Ultrospec 3000; Amersham Biosciences). Construction of plasmids The designed plasmids were constructed by traditional cloning methods or the Gibson assembly method 53 . The backbone plasmids and the primers used for plasmid construction are listed in Supplementary Tables 1 and 4 , respectively. Many plasmids were constructed in this study, and the restriction sites of the plasmids into which genes were cloned are presented in Supplementary Table 1 . The crtE , crtB and crtI genes from P. ananatis were amplified from pCar184 23 . The trLUT2 , trLCYB , trLCYBmut , ATR2 , trLUT5 and trLUT1 genes from A. thaliana , trCYP76AD1mut from B. vulgaris , valC from Callitropsis nootkatensis , zsd1 from Zingiber zerumbet , trGPPS2 from Abies grandis , trCYP71D55mut from H. muticus , trCYP93B16 from G. max and the cipB gene from P. luminescens were codon-optimized and synthesized (Integrated DNA Technologies). The cipA gene from P. luminescens was synthesized without codon-optimization (Integrated DNA Technologies). The sequences of codon-optimized synthetic genes are presented in Supplementary Table 5 . The dxs and fni genes were amplified from the genomic DNA of B. subtilis 168. The hemA fbr , hemL , hemB , hemH , zraP , yodD , lsrR , astD , bssS , yohC , yhfG , ychH , ycgB , glgS , mntS and yegP genes were amplified from the genomic DNA of E. coli K-12 W3110. Two lysine residues to make the gene feedback inhibition-resistant were inserted into the hemA gene during the polymerase chain reaction (PCR) 48 , 54 . The dsrA and sdsR genes were synthesized with an additional transcription terminator and trc promoter at the 5′ end. The abbreviations of enzymes and genes encoding the corresponding enzymes are listed in Supplementary Table 6 . To construct pLUT1, the crtE_F/R and crtI_F1/crtB_R primer sets were used for amplifying the crtE and crtI-crtB genes, respectively, from pCar184. The amplified products were overlapped by PCR and cloned into pTac15k. To construct pLUT4, the cipA and crtI genes were amplified by the primer sets cipA_F1/R1 and crtI_F2/R, respectively. Overlapping PCR was performed to generate cipA-crtI , which was then cloned into the same site of crtI in pLUT1 by Gibson assembly. The plasmid backbone was prepared by PCR amplification using the primers crtB_F1/crtE_R. To construct pLUT2, the trLUT2 and trLCYB genes were amplified using the primer sets trLUT2_F/R and trLCYB_F/R, respectively, and cloned into pBBR1Tac. To construct pLUT5, the cipA gene fragments amplified using the primer sets cipA_F2/R2 and cipA_F3/R3 were each fused with trLUT2 and trLCYB by PCR to generate cipA-trLUT2 and cipA-trLCYB , respectively, which were then cloned into pBBR1Tac. To construct pLUT6, the cipA-trLUT2 and trLCYB genes were cloned into the same plasmid. To construct pBBR1Tac-trLUT2-cipA-trLCYB, the trLUT2 and cipA-trLCYB genes were cloned into pBBR1Tac. To construct pLUT5M, trLCYBmut was cloned instead of trLCYB , using the same primers and cloning procedure. To construct pLUT4-His, the cipA-crtI-His gene amplified from pLUT4 using the primer set cipA_F1/crtI-His_R and the plasmid backbone amplified from pLUT1 using the primer set crtB_F2/crtE_R, were assembled by Gibson assembly. To construct pLUT5-His, the His-cipA-trLCYB-His gene amplified from pLUT5 using the primer set His-cipA_F/trLCYB-His_R and the plasmid backbone amplified from pLUT5 using the primer set pBBR1Tac_F/trLUT2_R2 were assembled by Gibson assembly. For the construction of pLUT6-His, the cipA-trLUT2-His gene amplified from pLUT6 using the primer set cipA_F2/trLUT2-His_R and the trLCYB gene amplified using the primer set trLCYB_F/R were cloned into pBBR1Tac. Before constructing pLUT3, three pTrc99a-based plasmids harbouring each of the ATR2 , trLUT5 and trLUT1 genes were constructed. Each gene was amplified using the primer sets ATR2_F1/R1, trLUT5_F/R1 and trLUT1_F/R1, and cloned into pTrc99a, generating pTrc-ATR2, pTrc-trLUT5 and pTrc-trLUT1, respectively. The trLUT5 gene with the trc promoter was amplified from pTrc-trLUT5 using the primers pTrc_F1/trLUT5_R2 and the trLUT1 gene with the trc promoter was amplified from pTrc-trLUT1 using the primers pTrc_F2/trLUT1_R2. The PCR products were cloned into pTrc-ATR2 to construct pLUT3. To construct pLUT7, the cipB gene fragments amplified using the primer sets cipB_F1/R1, cipB_F1/R2 and cipB_F1/R3 were each fused with the ATR2 , trLUT5 and trLUT1 genes, which were amplified using the primer sets ATR2_F1/R1, trLUT5_F/R1 and trLUT1_F/R1, respectively. The PCR products were cloned into pTrc99a, generating pTrc-cipB-ATR2, pTrc-cipB-trLUT5 and pTrc-cipB-trLUT1, respectively. The cipB-trLUT5 and cipB-trLUT1 genes with trc promoter were each amplified using the primer sets pTrc_F1/trLUT5_R2 and pTrc_F2/trLUT1_R2, respectively, and cloned into pTrc-cipB-ATR2 to construct pLUT7. To construct pLUT3B1, the trLUT5 and trLUT1 genes with the trc promoter were cloned into pTrc-cipB-ATR2. For the construction of pLUT3B2, the cipB-trLUT5 and trLUT1 genes with the trc promoter were cloned into pTrc-ATR2. To construct pLUT3B3, the trLUT5 and cipB-trLUT1 genes with the trc promoter were cloned into pTrc-ATR2. For the construction of pLUT3B4, the cipB-trLUT5 and cipB-trLUT1 genes with the trc promoter were cloned into pTrc-ATR2. The trLUT5 and cipB-trLUT5 genes were amplified using primer sets pTrc_F1/trLUT5_R2 from pTrc-trLUT5 and pTrc-cipB-trLUT5, respectively, while the trLUT1 and cipB-trLUT1 genes were amplified using primer sets pTrc_F2/trLUT1_R2 from pTrc-trLUT1 and pTrc-cipB-trLUT1, respectively. To construct pLUT7-His, the cipB-ATR2-His , cipB-trLUT5-Hi s and cipB-trLUT1-His genes amplified using the primer sets cipB_F1/ATR2-His_R, cipB_F1/trLUT5-His_R and cipB_F1/trLUT1-His_R from pTrc-cipB-ATR2, pTrc-cipB-trLUT5 and pTrc-cipB-trLUT1, respectively, were each cloned into pTrc99a, generating pTrc-cipB-ATR2-His, pTrc-cipB-trLUT5-His and pTrc-cipB-trLUT1-His, respectively. The cipB-trLUT5-His and cipB-trLUT1-His genes with the trc promoter were then each amplified using the primer sets pTrc_F1/His_R1 and pTrc_F2/His_R2 from pTrc-cipB-trLUT5-His and pTrc-cipB-trLUT1-His, respectively, and cloned into pTrc-cipB-ATR2-His. To construct pHEM1, the hemA fbr and hemL genes were amplified using the primer sets hemA_F1/R and hemL_F/R1, respectively, and cloned into pTrcCDF. Before the construction of pHEM2, the plasmid backbone amplified using the primers pTrcCDF_F/R, hemB amplified using the primers hemB_F/R, and hemH amplified using the primers hemH_F/R1 were assembled by Gibson assembly, generating pTrc-hemBH. Using pTrc-hemBH as a template, the hemBH operon with the trc promoter was amplified using the primers pTrc_F3/hemH_R2 and cloned into pHEM1. Plasmid pHEM3 was constructed by the insertion of the hemA fbr and hemL genes amplified using hemA_F2/hemL_R2 from pHEM1 into pBBR1Tac by Gibson assembly. For the construction of pHEM1-derived target gene overexpressing plasmids, the zraP , yodD , lsrR , astD , bssS , yohC , yhfG , ychH , ycgB , glgS , mntS , yegP , dsrA and sdsR genes were amplified using the primer sets zraP_F/R, yodD_F/R, lsrR_F/R, astD_F/R, bssS_F/R, yohC_F/R, yhfG_F/R, ychH_F/R, ycgB_F/R, glgS_F/R, mntS_F/R, yegP_F/R, rrnb_F/dsrA_R and rrnb_F/sdsR_R, respectively, and cloned into pHEM1. Plasmid pTacRSF was generated by Gibson assembly of a part of linearized pRSFDuet (amplified using the primers pRSFDuet_F/R) and a part of linearized pTac15k (amplified using the primers pTac15k_F/R). Plasmid pVAL1 was then constructed by cloning the valC gene amplified using the primers valC_F/R into pTacRSF by Gibson assembly. To construct pVAL2, zsd1 amplified using the primers zsd1_F/R was cloned into the downstream of valC . To construct pTac-dxs-fni-trGPPS2, dxs amplified using the primers dxs_F/R and fni amplified using the primers fni_F/R were first cloned into pTacCDF by Gibson assembly. Then, trGPPS2 amplified using the primers trGPPS2_F/R was inserted into the downstream of fni . To construct pNKT1, the trCYP71D55mut gene was amplified using the primers pTrc_F4/trCYP71D55mut_F1/R and cloned into pTrc-ATR2 by Gibson assembly. To construct pNKT2, the cipB and trCYP71D55mut genes amplified using the primer sets cipB_F2/cipB_R4 and trCYP71D55mut_F2/R, respectively, were fused by PCR using the primers pTrc_F4/trCYP71D55mut_R, and then cloned into pTrc-cipB-ATR2 by Gibson assembly. Before constructing pNKT3, pTrc-cipB-L-ATR2 was first constructed. The cipB and ATR2 genes amplified using the primer sets cipB_F1/R4 and ATR2_F2/R1, respectively, were fused by PCR, and then inserted into pTrc99a. Then, the cipB and trCYP71D55mut genes amplified using the primer sets cipB_F2/cipB_R4 and trCYP71D55mut_F3/R, respectively, were fused by PCR using the primers pTrc_F4/trCYP71D55mut_R, and then inserted into pTrc-cipB-L-ATR2 by Gibson assembly. Plasmid pAGN1 was constructed by cloning of the trCYP93B16 gene, amplified using the primers pTrc_F5/trCYP93B16_F1/R, into pTrc-ATR2 by Gibson assembly. To construct pAGN2, the cipB and trCYP93B16 genes amplified using the primer sets pTrc_F5/cipB_F2/R4 and trCYP93B16_F2/R, respectively, were cloned into pTrc-cipB-ATR2 by Gibson assembly. To construct pAGN3, the cipB and trCYP93B16 genes amplified using the primer sets pTrc_F5/cipB_F2/R4 and trCYP93B16_F3/R, respectively, were cloned into pTrc-cipB-L-ATR2 by Gibson assembly. To construct pDP1, the trCYP76AD1mut gene was amplified using the primers pTrc_F4/trCYP76AD1mut_F1/R and cloned into pTrc-ATR2 by Gibson assembly. To construct pDP2, the cipB and trCYP76AD1mut genes amplified using the primer sets cipB_F2/cipB_R4 and trCYP76AD1mut_F2/R, respectively, were fused by PCR using the primers pTrc_F4/trCYP76AD1mut_R, and then cloned into pTrc-cipB-ATR2 by Gibson assembly. To construct pDP3, the cipB and trCYP76AD1mut genes amplified using primers cipB_F2/cipB_R4 and trCYP76AD1mut_F3/R, respectively, were fused by PCR using the primers pTrc_F4/trCYP76AD1mut_R, and cloned into pTrc-cipB-L-ATR2 by Gibson assembly. Cell preparation for protein analysis To prepare protein samples, cells were first inoculated into a 14-ml test tube containing 3 ml of TB medium and cultivated at 30 °C at 220 r.p.m. overnight. Then, 1 ml of the culture was transferred to a 250-ml baffled flask containing 50 ml of R/2 medium supplemented with 20 g l −1 of glycerol and 3 g l −1 of yeast extract, and cultured at 30 °C at 220 r.p.m. For the cultivation of E. coli BL21(DE3) and BTY5-derived strains, 3 g l −1 of (NH 4 ) 2 SO 4 was additionally supplemented to the R/2 medium. When the OD 600 reached ~0.4–0.6, 1 mM IPTG was added for induction. Antibiotics were added to ensure plasmid stability. Cells were collected 24 h after IPTG induction by centrifugation at 16,000 g for 1 min. The cell pellet was washed with ice-cold phosphate-buffered saline (PBS) and resuspended in the same buffer. The cell suspension was then disrupted by sonication (VCX800, Sonics) and used for further analysis. PBS (pH 7.2) purchased from Enzynomics (Republic of Korea) contains 137.0 mM NaCl, 2.7 mM KCl, 4.3 mM Na 2 HPO 4 and 1.4 mM KH 2 PO 4 . The sonicator was equipped with a 3-mm tapered microtip set on 30 W. The 3-min sonication cycle comprised 3-s pulse/6-s off. Membrane protein isolation and analysis After disruption of the collected cells, the membrane fraction was isolated as previously described 55 . Briefly, the insoluble fraction was removed by centrifugation at 21,000 g at 4 °C for 10 min, and repeated to collect the supernatant. The membrane fraction in the supernatant was separated by ultracentrifugation at 150,000 g at 4 °C for 45 min, and recovered by incubating the collected pellet in PBS at 37 °C for 1 h. The supernatant was collected for analysis of the cytoplasmic fraction. Protein samples were analysed by SDS–PAGE. PCI isolation and analysis To isolate PCI, cells disrupted by sonication were centrifuged at 4,000 g at 4 °C for 15 min. The cell pellet was washed with ice-cold PBS three times to remove impurities. The pellet was then washed with ice-cold washing buffer (200 mM NaCl, 1 mM ethylenediaminetetraacetic acid and 1% TritonX-100 in PBS) three times to remove membrane-bound proteins. The remaining washing buffer was washed with ice-cold PBS three times. The finally collected PCIs were resuspended in PBS and analysed by SDS–PAGE. For western blotting, the proteins loaded on the SDS–PAGE gel were transferred to a polyvinyl difluoride membrane (Roche) at a constant electric current of 200 mA for 2 h. The membrane was then blocked with 5% (wt/vol) skimmed milk in Tris-buffered saline (TBS) solution at room temperature for 2 h. The TBS solution contained 25 mM Tris, 192 mM NaCl and 10% methanol. The blocked membrane was incubated for 1 h at room temperature with 1:8,000 diluted horseradish peroxidase-conjugated anti-His antibody (Abcam) in TBS solution containing 5% (wt/vol) skimmed milk for 1 h at room temperature. After intermittent washing six times with TBS solution, the membrane was soaked with an enhanced chemiluminescence reagent (ECL Prime, GE Healthcare) for visualization of the target protein bands, which were then imaged with the ChemiDoc MP imaging system (Bio-Rad). For microscopic observation of the cells, an Eclipse Ti-U (Nikon) or Optiphot-2 microscope (Nikon) was used. The microscopic images were taken with the NIS-Elements software. Cell broth (3 μl) adjusted to an OD 600 of 7 was loaded onto the glass slide for microscopic observation. Fed-batch fermentation Cells were first grown in 250-ml baffled flasks containing 20 ml of TB medium at 28 or 37 °C and 200 r.p.m. until reaching an OD 600 of ~1–2. Then, 1 ml of the culture was transferred to a 250-ml baffled flask containing 50 ml of R/2 medium supplemented with 3 g l −1 of yeast extract and 20 g l −1 of glycerol. The second seed culture was prepared by growing cells at 28 or 37 °C and 200 r.p.m. until reaching an OD 600 of 4. Then, 40 ml of the culture was transferred to a 6.6-l (Bioflo 320, Eppendorf) or a 5-l (MARADO-05D-PS, BioCNS) jar fermenter containing 1.6 l of air-saturated R/2 medium. Unlike that used for flask culture, the R/2 medium used for fed-batch fermentation contained 7.5 ml of trace metal solution per litre, and was supplemented with 5 g l −1 of yeast extract and 30 g l −1 of glycerol. For cultivation of the AGN4H strain, 3 g l −1 of (NH 4 ) 2 SO 4 was additionally supplemented to R/2 medium in both flask culture and fermentation. For the cultivation of the NKT3 strain, 320 ml of dodecane was added after IPTG induction. The pH was controlled at 6.95, except for a temporary period of pH perturbation upon carbon source depletion, using 28% (vol/vol) ammonia solution. The dissolved oxygen level was kept at 40% of the air saturation value by automatically shifting the agitation speed from 200 r.p.m. to 1,000 r.p.m., and by automatically providing a mixture of air and pure oxygen at a flow rate of 2 l min −1 ; when the temperature was changed during fermentation, the air saturation value set at the initial temperature was kept. The feeding solution containing 800 g glycerol, 4 ml of 3 M HCl, 6 ml of trace metal solution and 12 g of MgSO 4 ·7H 2 O per litre was supplied to the fermentation broth by the pH-stat method 56 . For the cultivation of the AGN4H strain, 40 g l −1 of (NH 4 ) 2 SO 4 was supplemented to the feeding solution. For the starvation culture for lutein production, the feeding solution began to be supplied 2 h after glycerol was depleted. For the induction of gene expression, 0.05, 0.2, 0.5 or 1 mM IPTG was supplied to the fermenter when the OD 600 reached ~20–30. Antibiotics were supplemented at the same concentrations as those in flask cultures. For the measurement of dry cell weight (DCW), cells were washed twice with distilled water and dried at 75 °C for two days. Analytical procedures for carotenoids Cells were collected from the culture sample by centrifugation at 16,000 g for 1 min. Carotenoids were extracted from the cell pellet using acetone, as described in our previous study 24 . Briefly, 1 ml of analytical-grade acetone was added to the pellet, and the cells were resuspended by vigorous vortexing. The cells were disrupted by sonication (SD-200H, Sungdong Ultrasonic) for 30 min in an ice-cold water bath. The disrupted cell suspension was centrifuged at 16,000 g for 1 min, and the supernatant containing extracted carotenoids was isolated. The authentic standards of lutein and zeaxanthin were purchased from Cayman, and those of lycopene, β-carotene and α-carotene were purchased from Merck. HPLC analysis for the quantification of extracted carotenoids was performed using a 1260 Infinity II HPLC system (Agilent) equipped with a diode array detector. A YMC carotenoid C30 column (250 mm × 4.6 mm, 5 μm, YMC) was used, and the separation of carotenoids was monitored at an absorbance of 450 nm. The solvent was flowed at a rate of 0.6 ml min −1 with gradients of buffer A (90% methanol, vol/vol) and B ( tert -butylmethylether). For the analysis of carotenoids except for lycopene, the mobile phase comprising 90 vol% A/10 vol% B was in an isocratic mode for 3 min, then gradually changed to 30 vol% A/70 vol% B for another 12 min. After keeping constant conditions for the following 8 min, the composition was gradually changed to 90 vol% A/10 vol% B for 1 min. To confirm the production of lutein by MS, an HPLC-MS system (1100 series and LC/MSD VL, Agilent) equipped with an Eclipse XDB-C18 column (4.6 mm × 150 mm, 1.8 μm, Agilent) was used, and absorbance was monitored at 338 nm. The gradient elution was performed using a mobile phase comprising buffer C (0.1 vol% formic acid) and buffer D (30 vol% methanol). After keeping the isocratic condition at 70 vol% C/30 vol% D for 1 min at a flow rate of 1 ml min −1 , buffer C in the mobile phase was gradually substituted by buffer D for the following 9 min. Buffer D was then constantly flowed for another 9 min. For the last 1 min, the mobile-phase composition was returned to 70 vol% C/30 vol% D. After separation, the m / z values of the samples were scanned using electrospray ionization positive ion mode for the mass analysis. The capillary voltage was 2.5 kV and the nebulizer pressure was 30 psig. The nitrogen gas at 350 °C was flowed at a rate of 12 l min −1 for drying solvents. Analytical procedures for (+)-valencene and (+)-nootkatone To analyse the (+)-valencene and (+)-nootkatone produced from two-phase cultivation, the upper dodecane layer was taken after centrifugation of the culture broth at 16,000 g . To analyse (+)-nootkatone produced from single aqueous phase flask culture, the supernatant was first collected by centrifugation of the culture broth at 4,000 g for 15 min. The (+)-nootkatone in the supernatant was extracted using the same volume of ethyl acetate by shaking the mixture at 250 r.p.m. and 40 °C for 1 h. The ethyl acetate layer containing (+)-nootkatone was isolated after centrifugation at 4,000 g for 10 min. The authentic standards of (+)-valencene and (+)-nootkatone were purchased from Merck. For the determination of (+)-valencene concentration, samples were analysed by a GC system (6890N, Agilent) equipped with an Agilent 7683 automatic injector, a flame ionization detector and a fused-silica capillary column (ATTM-Wax, 30 m, internal diameter 0.53 mm, film thickness 1.20 μm, Alltech). The inlet temperature was maintained at 230 °C. The oven was held at 50 °C for 3 min, heated to 130 °C at 5 °C min −1 and then heated to 225 °C at 25 °C min −1 for 12 min. Peak detection was performed with a flame ionization detector (280 °C). (+)-Nootkatone was analysed by GC-MS (6890N and 5973Network, Agilent) in scan mode. The oven was held at 80 °C for 3 min and then heated to 310 °C at 12 °C min −1 for 10 min. The amount of (+)-nootkatone was determined by data acquired in the selected ion monitoring mode. The ions monitored were as follows: m / z 41, 79, 91, 121, 133, 147, 161 and 218. The concentrations of (+)-valencene and (+)-nootkatone produced in two-phase fed-batch fermentation were calculated considering the change in vol/vol ratio of dodecane to the culture broth—calculated by the sum of the initial volume and the added volume of feeding solution and ammonia solution—at each time point. Analytical procedures for naringenin, apigenin and l -DOPA The supernatant was collected by centrifugation of the culture sample at 16,000 g for 1 min. To extract naringenin and apigenin in the supernatant, the same volume of ethyl acetate was added, followed by vortexing at 1,500 r.p.m. and 40 °C for 10 min using a Thermo Shaker TS100 device (Ruicheng). The ethyl acetate layer was then isolated after centrifugation at 16,000 g for 1 min. The authentic standards of naringenin, apigenin and l -DOPA were purchased from Merck. The HPLC analyses for the quantification of naringenin, apigenin and l -DOPA were performed using an Agilent 1100 series system equipped with a diode array detector, and a Poroshell 120 EC-C18 column (150 mm × 4.6 mm, 4 μm, Agilent). Absorbance at 260 nm was monitored for the separation and detection of naringenin and apigenin, and at 220 nm for l -DOPA. The sample was flowed at a rate of 0.6 ml min −1 with gradients in the buffer E (0.1 vol% trifluoroacetic acid) and F (acetonitrile). The mobile phase comprising 95 vol% E/5 vol% F was in isocratic mode for 3 min, then gradually changed to 30 vol% E/70 vol% F for the next 37 min. The apigenin and l -DOPA in samples were then analysed for confirmation using an HPLC-MS system (1100 series and LC/MSD VL, Agilent) equipped with an XBridge C18 column (4.6 mm × 150 mm, 5 μm, Waters). The mobile phase comprising buffer G (20 mM ammonium acetate, pH 9.6) and buffer H (acetonitrile) was flowed at 0.4 ml min −1 . The isocratic condition at 95 vol% G/5 vol% H was kept for 3 min and then gradually changed to 30 vol% G/70 vol% H for the next 12 min. After that, the ratio was gradually changed to 10 vol% G/90 vol% H for another 5 min and was maintained for the last 10 min. After separation, the m / z values of samples were scanned using an electrospray ionization negative ion mode for the mass analysis. The capillary voltage was 2.5 kV and the nebulizer pressure was 30 psig. The nitrogen gas at 350 °C was flowed at a rate of 12 l min −1 for drying the solvents. Transcriptome analysis Cells grown under normal and starved conditions were collected at the same time point (after 2 h of carbon starvation). RNA isolation and transcriptome data analysis were performed by EBIOGEN. Briefly, total RNA was extracted using Trizol reagent (Invitrogen) according to the manufacturer’s instructions. The quality of RNA was checked by an Agilent 2100 bioanalyser (Agilent) and rRNA was removed using a Ribo-Zero Magnetic kit (Epicentre). With purified RNA, a library was constructed using the SMARTer Stranded RNA-Seq Kit (Takara Bio) according to the manufacturer’s protocol. The mean fragment size of libraries was determined using the Agilent 2100 bioanalyser (Agilent), and quantification was performed by using a StepOne Real-Time PCR System (Life Technologies). The high-throughput sequencing was performed using HiSeq 2500 (Illumina) and the reads were aligned using the Bowtie 2 software. The gene expression levels were determined based on an alignment file using EdgeR within R 57 using Bioconductor 58 . To compare the expression levels between samples, the quantile normalization method was used. Genes were classified according to DAVID ( ) searches. Statistics and reproducibility Data are shown as the averages of triplicates ± standard deviations (s.d.). Pairwise comparison between two data groups was carried out using the two-tailed Student’s t -test. Differences between the two data groups were considered as statistically valid if * P < 0.05, ** P < 0.01 or *** P < 0.001. For the comparison of more than two data groups, ANOVA followed by Tukey’s multiple comparison test was used. Differences among data groups were considered statistically valid if P < 0.05. The experiment was repeated three times independently with similar results for SDS–PAGE, western blotting and microscopic analysis. Details of parameters used for multiple comparison of the data set are shown in Supplementary Table 7 . Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this Article. Data availability The data to reproduce the findings in this study are presented in the Paper and the Supplementary Information or are available from the authors upon reasonable request. Source data are provided with this Paper.

Lutein is classified as a xanthophyll chemical that is abundant in egg yolk, fruits, and vegetables. It protects the eye from oxidative damage from radiation and reduces the risk of eye diseases including macular degeneration and cataracts. Commercialized products featuring lutein are derived from the extracts of the marigold flower, which is known to harbor abundant amounts of lutein. However, the drawback of lutein production from nature is that it takes a long time to grow and harvest marigold flowers. Furthermore, it requires additional physical and chemical-based extractions with a low yield, which makes it economically unfeasible in terms of productivity. The high cost and low yield of these bioprocesses has made it difficult to readily meet the demand for lutein. These challenges inspired the metabolic engineers at KAIST, including researchers Dr. Seon Young Park, Ph.D. Candidate Hyunmin Eun, and Distinguished Professor Sang Yup Lee from the Department of Chemical and Biomolecular Engineering. The team's study was published in Nature Catalysis on August 5, 2022. This research details the ability to produce lutein from E. coli with a high yield using a cheap carbon source, glycerol, via systems metabolic engineering. The research group focused on solving the bottlenecks of the biosynthetic pathway for lutein production constructed within an individual cell. First, using systems metabolic engineering, which is an integrated technology to engineer the metabolism of a microorganism, lutein was produced when the lutein biosynthesis pathway was introduced, albeit in very small amounts. To improve the productivity of lutein production, the bottleneck enzymes within the metabolic pathway were first identified. It turned out that metabolic reactions that involve a promiscuous enzyme, an enzyme that is involved in two or more metabolic reactions, and electron-requiring cytochrome P450 enzymes are the main bottleneck steps of the pathway inhibiting lutein biosynthesis. To overcome these challenges, substrate channeling, a strategy to artificially recruit enzymes in physical proximity within the cell in order to increase the local concentrations of substrates that can be converted into products, was employed to channel more metabolic flux towards the target chemical while reducing the formation of unwanted byproducts. Furthermore, electron channeling, a strategy similar to substrate channeling but differing in terms of increasing the local concentrations of electrons required for oxidoreduction reactions mediated by P450 and its reductase partners, was applied to further streamline the metabolic flux towards lutein biosynthesis, which led to the highest titer of lutein production achieved in a bacterial host ever reported. The same electron channeling strategy was successfully applied for the production of other natural products including nootkatone and apigenin in E. coli, showcasing the general applicability of the strategy in the research field. "It is expected that this microbial cell factory-based production of lutein will be able to replace the current plant extraction-based process," said Dr. Seon Young Park, the first author of the paper. She explained that another important point of the research is that integrated metabolic engineering strategies developed from this study can be generally applicable for the efficient production of other natural products useful as pharmaceuticals or nutraceuticals. "As maintaining good health in an aging society is becoming increasingly important, we expect that the technology and strategies developed here will play pivotal roles in producing other valuable natural products of medical or nutritional importance," explained Distinguished Professor Sang Yup Lee.

10.1038/s41929-022-00820-4

Medicine

Study identifies novel epigenetic changes in pediatric brain cancer

Sibo Zhao et al, Epigenetic Alterations of Repeated Relapses in Patient-matched Childhood Ependymomas, Nature Communications (2022). DOI: 10.1038/s41467-022-34514-z Journal information: Nature Communications

https://dx.doi.org/10.1038/s41467-022-34514-z

https://medicalxpress.com/news/2022-12-epigenetic-pediatric-brain-cancer.html

Abstract Recurrence is frequent in pediatric ependymoma (EPN). Our longitudinal integrated analysis of 30 patient-matched repeated relapses (3.67 ± 1.76 times) over 13 years (5.8 ± 3.8) reveals stable molecular subtypes (RELA and PFA) and convergent DNA methylation reprogramming during serial relapses accompanied by increased orthotopic patient derived xenograft (PDX) (13/27) formation in the late recurrences. A set of differentially methylated CpGs (DMCs) and DNA methylation regions (DMRs) are found to persist in primary and relapse tumors (potential driver DMCs) and are acquired exclusively in the relapses (potential booster DMCs). Integrating with RNAseq reveals differentially expressed genes regulated by potential driver DMRs ( CACNA1H, SLC12A7, RARA in RELA and HSPB8, GMPR, ITGB4 in PFA) and potential booster DMRs ( PLEKHG1 in RELA and NOTCH, EPHA2, SUFU, FOXJ1 in PFA tumors). DMCs predicators of relapse are also identified in the primary tumors. This study provides a high-resolution epigenetic roadmap of serial EPN relapses and 13 orthotopic PDX models to facilitate biological and preclinical studies. Introduction Ependymoma (EPN) is the third most common malignant brain tumor of childhood, accounting for up to 12% of intracranial tumors in children. Current therapy includes maximal surgical resection and focal radiation, resulting in a 5-year overall survival (OS) and progression-free survival (PFS) of 70% and 57%, respectively 1 , 2 , 3 , 4 , 5 . However, nearly half of patients will experience late relapses 1 , 3 , 4 . Despite repeated treatment, most children with relapsed tumors eventually succumb to the disease. Ten-year OS and PFS decrease further to 50 ± 5% and 29 ± 5%, respectively 1 . The use of chemotherapeutic agents for EPN has been extensively studied for decades, however, survival benefit remains controversial 6 , 7 . In-depth understanding of the biology of EPN recurrence is needed. While it is well established that tumor location is important in EPN biology 8 , recent studies have shown that epigenetic changes mediated by DNA methylation play an important role in EPN tumorigenesis 9 , 10 , 11 as gene mutations in EPN are much less frequent than in adult cancers 9 , 10 , 11 . Indeed, non-mutational epigenetic reprogramming has recently been incorporated as an emerging hallmark of cancer 12 . Similar to several other types of pediatric cancers 13 , 14 , 15 , DNA methylation analysis has successfully subclassified pediatric EPNs into nine molecularly subgroups with distinct clinical outcomes 16 , 17 . Although oncogenic drivers have been identified for primary EPN 8 , 11 , 15 , 16 , 18 , 19 , 20 , 21 , 22 , 23 , understanding of recurrent EPN biology is still at its infancy. One of the challenges is the difficulties of obtaining relapsed tumor tissues. Although longitudinal analysis of consecutive, serially relapsing patient tumor samples will enable the separation of serially conserved potential epigenetic driver(s) from transient alterations, it remains very difficult and requires a committed collaborative team of physicians, neuropathologist and tumor biologists. Using single recurrent EPN tumor with or without patient-matched primary tumors is a frequent approach. The second challenge lies in the reliable detection of subtle genetic/epigenetic changes. As a recent study revealed that morphological changes in recurrent EPNs was not sufficiently explained by epigenetic changes detected with 450k DNA methylation array 24 , improved DNA methylation site coverage combined with integrated analysis of gene expression patterns 9 , 10 , 25 , 26 , 27 , 28 should facilitate the discovery of genetic interactions and new therapeutic opportunities. A third challenge in studying EPN relapse is the limited availability of animal models. After the establishment of the first supratentorial EPN patient-derived orthotopic xenograft (PDOX, or orthotopic PDX) model by our group 29 , we and others have shown that direct implantation of patient tumors into the matching locations in mouse brains promoted the establishment of clinically relevant animal models that replicate histopathological features, invasive/metastasis phenotype and genetic profiles of the original patients 29 , 30 , 31 , 32 , 33 , 34 , 35 . While several genetically engineered animal models have been developed 22 , 23 , 36 , correlating tumorigenicity of patient EPN tumors during tumor progression, i.e., from diagnosis till consecutive, serial relapses of each patient, should serve as an in vivo functional assay to determine key tumorigenic drivers and identify druggable targets for EPN relapses. Here, we report a deep longitudinal analyses of DNA methylation landscape using single-base-resolution DNA methylome in a cohort of 30 patient-matched primary and serially relapsing EPN tumors together with integrated analysis of RNA-seq profiles and functional examination of PDOX tumorigenicity for this set of matched primary and relapsed EPNs. Our goal is to apply this integrated analyses and clinically relevant patient-matched PDOX mouse models from the same cohort of patients to identify the potential drivers of relapse (abnormal DNA methylation present in primary tumor and sustained in all relapsed EPNs), the potential boosters of relapse (DNA methylation newly acquired in relapsed tumors and persisted with tumor progression), and potential predictors of relapse in the primary tumors (DNA methylation in the primary tumors which predicts future/subsequent relapse). Results Maintenance of molecular subtypes during serial relapses of EPN The distinct molecular subtypes of pediatric EPN at diagnosis can impact clinical treatment 13 , 14 , 16 , 37 , 38 . While analysis of single recurrences suggested the maintenance of molecular subtypes, it remains to be determined if the molecular subtypes of EPNs will change over repeated relapses of a long period of time. Despite 50% EPN patients develop tumor recurrences 1 , 3 , 4 , relapsed tumor samples, particularly the repeated recurrences, are difficult to obtain. In our study, we collected a total of 30 serially relapsing EPN tumor samples from 10 of 110 (9.1%) pediatric patients who were followed up over 13 years. Consent to publish clinical information potentially identifying individuals was obtained. Many recurrent tumor samples were obtained after the patients were treated with chemo- and/or radiation therapies (Fig. 1A ). The time from diagnosis to last recurrence ranged from 2.75 to 13 years, and the number of recurrences per patient ranged from 1 to 7 times (3.67 ± 1.76 times) (Fig. 1A , Table 1 ). Fig. 1: DNA methylation and transcriptomic landscape of recurrent ependymoma tumors. A Time-course summary of surgery, radio- and chemotherapy treatment and tumorigenicity information of the 10 sets of recurrent ependymomas. Images of intra-cerebellar ( ICb ) and intra-cerebral ( IC ) orthotopic xenograft tumor formation from primary ( p ) or recurrent ( r ) PFA and RELA tumors (model ID in round brackets) were inserted to show tumor ( T ), hydrocephalus ( * ) and CSF spread. B Orthotopic xenograft mouse models of recurrent PFA EPNs. Tumor formation can be seen on gross (top) mouse brains (outlined). H&E staining (middle) showing histological comparison between the originating patient tumor and the PDOX tumors. Changes of animal survival times (lower panels) during serial in vivo transplanations of intra-cerebellar ( ICb ) PDOX models from passage I ( I ) to V ( V ) and the impact of the implanted different cell numbers (from 1000 cells per mouse to 100,000 cells) on animal survival times of two PDOX models were shown ( n = 10 mice/group). Scale bar = 50 µM. C Heatmap showing the DNA methylation ratios of the top variable 20,000 CpGs across all EPN tumor samples. The patient ID is labeled vertically on the top of the heatmap. The black and gray bars under the patient IDs are used to separate tumors from the same patient. D Phylogeny tree construction using the top 6000 CpGs with variable DNA methylation ratios in RELA (red circle) and PFA (blue circle) tumors. E Principle component analysis of RNA-seq data from EPN tumor samples using age matched childhood normal brain tissues as references. F Heatmap representing the PFA (left ) and RELA (right) primary tumor signature genes’ expression levels in all samples, including multiple relapses. The primary signature genes were selected from a previously published database GSE64415 16 , 39 , 88 , and those overlapped with our consistently decreased (green) and increased (red) genes were highlighted. G CA methylation ratios in PFA (left) and RELA (lower) primary and recurrent tumors as compared with normal childhood cerebellar and cerebral tissues. Boxplots indicate median, first and third quartiles (Q1 and Q3), whiskers extend to the furthest values; the uppermost and lowest line indicates the maximum and minimum values, respectively. Marked on the top of each boxplot is the number of samples analyzed including Cerebellum ( n = 2); Cerebrum ( n = 3); PFA-P ( n = 2); PFA-R1 ( n = 5); PFA-R2 ( n = 3); PFA-R3 ( n = 1); PFA-R4 ( n = 1); RELA-P ( n = 3); RELA-R1 ( n = 4); RELA-R2 ( n = 2); RELA-R3 ( n = 2); RELA-R4 ( n = 2); RELA-lateR ( n = 2). Full size image Table 1 Summary of clinical information of the ependymoma patients and the autopsied normal tissues Full size table To achieve high-resolution analysis of DNA methylation, we performed genome-wide DNA methylation sequencing using four normal pediatric brain tissues (2 cerebellar and 2 cerebral tissues) (Table 1 ) procured from warm (<6 hr) autopsy of two children as controls (Fig. 1A , Table 1 ). In all the samples, we identified >2 × 10 6 CpGs each covered by at least five reads with a mean DNA methylation ratio at 51% (Supplementary Data 1 ). The EPN tumors exhibited DNA methylation profiles distinct from the normal tissues (Supplementary Fig. 1A, B ). Similar to their primary tumors, the recurrent EPNs were subclassified into 5 RELA and 5 PFA tumors that displayed clear segregation in phylogenetic construction using the top 6000 variant CpGs (Fig. 1D, E and Supplementary Fig. 1C, D ) as well as the maintenance of key DNA methylation signatures (Fig. 1C ) and gene expression patterns that were previously identified in primary EPN tumors (GSE64415) 39 (Fig. 1F , Supplementary Data 2 ). These set of data demonstrated the maintenance of EPN molecular subtypes during repeated relapses (≥2) during years of chemo- and/or radiation therapies. Additional analysis of CpG islands detected an increase of DNA methylation levels in the relapsed tumors compared to the primary EPNs, of which the levels of CpG islands were higher than the normal brain tissues (Supplementary Fig. 1E ). In contrast, no consistent DNA methylation changes on CpG shore regions (flanks CpG islands up to 2 kb away from the CpG islands) were identified in the relapsed tumors (Supplementary Fig. 1E ). Non-CpG methylation is recently recognized as a layer of epigenetic information assembled at the root of vertebrates and plays new regulatory roles independent of the ancestral form of the CpG methylation 40 . Its patterns are often tissue-spcific 40 , 41 , 42 , 43 . In the present study, the primary and recurrent tumors were found to have a dramatically decreased mCpA levels in both RELA and PFA tumors (Fig. 1G ) when compared with the normal tissues. Although the differences between RELA and PFA tumors were not significant, this finding suggested that the reduction of mCpA can be a potentially epigenetic signature of EPN tumor of which the functional roles warrant further examination. Convergence of DNA methylation landscape during repeated relapses It remains unknown how EPN progression and repeated relapses affect cellular subpopulations. To understand the impact of selective, adaptive and progressive pressures on epigenetic reprogramming during long-term serial relapses of EPN, we calculated the Pearson correlation coefficient of DNA methylation ratios between the consecutively relapsing tumors during tumor relapses for each patient. In the event an adjacent previous relapsing tumor was missing, a previously available sample was used to calculate the correlation. Unlike the low-level correlations between primary tumor and normal brain tissues (average r of 0.24 for RELA and 0.23 for PFA), the Pearson correlation between two longitudinal consecutively relapsing tumors increased significantly, reaching 0.75 in 8/11 pair-wise comparisons in RELA recurrences, and 0.83 in 6/7 PFA recurrences (Fig. 2A , Supplementary Fig. 2A – C ), respectively. For example, in RELA1 the correlation coefficient ( r ) increased to 0.68 between the 3rd and 2nd recurrence, 0.70 between the 4th/3rd recurrence, and 0.77 between the 7th/5th recurrence, demonstrating a trend of convergence during repeated recurrences (recurrent times ≥ 2). A similar pattern was also exemplified for PFA1 (Fig. 2B ). Parallel analysis of transcriptional data detected a similar trend of increased correlation (Supplementary Fig. 2D ) as well. As no previous samples/studies were available, our data provided more insight on EPN relapse. This finding is also important, as the epigenetically homogeneous tumors can theoretically be more effectively targeted than the widely heterogeneous tumors. Unlike other cancers exhibiting mutational divergent at relapses 44 , this result indicated that pediatric EPNs harbor epigenetic lesions that are selectively enriched by clinical treatments (chemo- and/or radio-therapies) and biological evolution in driving the progression of tumor recurrences. Fig. 2: Progressive convergence of DNA methylation profiles during repeated ependymoma relapses. A Line charts showing the changes of DNA methylation correlations between adjacent recurrent tumors during repeated recurrences of RELA (upper panel) and PFA (lower panel) tumors. Each dot represents one tumor sample and graphed to the time of recurrence. B Representative smoothed density scatterplots displaying the increased correlation coefficient ( r ) of DNA methylation profiles between the adjacent recurrent tumors of RELA1 (upper panel) and PFA1 (lower panel) patients. C FISH analysis of chromosome 1q gain showing the locations of FISH probes in 1p (red) and 1q (green) (top), representative images of 1q ( G : green) gain relative to 1p ( R : red) (middle) in matching pairs of patients ( Pt ) and PDOX tumors. The number of cells counted (n) was marked on top of the column of every sample (red bar indicates 1p count, and green bar indicates 1q count). Statistical analysis was performed through two-sided Student t -test. ** P < 0.01. P -values = 7.7186E-30 (Pt-2002), 5.32303E-34 (ICb-2002EPN), 1.17953E-13 (Pt-0614), 1.5575E-08 (ICb-0614EPN), 4.53091E-21 (Pt-4423), 5.4252E-14 (ICb-4423EPN). Data are presented as mean values ± SD. (Magnification:×100). D Preservation of DNA methylation profiles of patient tumors in their matching PDOX models either from the primary ( P ) or recurrent tumors from the first ( R1 ) up to the 7th ( R7 ) recurrences. Global Pearson correlation ( r ) of the DNA methylation profiles from the matched patient tumor and PDOX tumors were labeled above the connected line. The numbers of CpGs used in the correlation analysis were 2,211,714 (RELA1-R7), 2,156,125 (RELA4-R1), 2,075,950 (RELA5-R1), 1,994,148 (PFA1-R3), 2,473,879 (PFA2-R2), 1,728,999 (PFA4-P), and 1,200,936 (PFA4-R1). Majority of the patient DMRs were maintained in their matching PDOX tumors (lower panel). The total numbers of hyperDMRs that were used in the analysis in patient and PDOX tumors are 6919 and 6265 (RELA1-R7), 10,868 and 12,825 (RELA4-R1), 11,406 and 14,333 (RELA-R1)m 5815 and 6653 (PFA-R3), 13,541 and 15,036 (PFA2-R2), 4107 and 8309 (RFA4-P), 3155 and 8373 (RFA4-R1); whereas the total numbers of hypoDMRs were 15,147 and 27,594 (RELA1-R7), 5797 and 8506 (RELA4-R1), 4361 and 7009 (RELA-R1), 12,051 and 15,519 (PFA-R3), 5824 and 5847 (PFA2-R2), 5373 and 10,917 (RFA4-P), 8969 and 10,050 (RFA4-R1), respectively. E Unsupervised clustering of primary/recurrent tumors from the first ( -R1 ) up to the 7th (-R7 ) relapse as well as matching PDOX tumors at specific passages (passage). Full size image Late EPN relapses exhibited increase of tumorigenicity in the brains of SCID mice We next investigated if the progressive and convergent epigenetic changes of the recurrent tumors are functionally important in promoting increased tumorigenic capacities. Despite the success in developing PDOX models for pediatric malignant brain tumors 29 , 31 , 32 , 34 , 35 , 45 , 46 , 47 , 48 , 49 by our group and others, childhood EPN is notoriously known for low tumor take rate 29 , 50 , 51 . With our discovery of the progressive and convergent nature of epigenetic changes during serial EPN relapses, we hypothesized that such epigenetic reprogramming enabled the recurrent tumors to acquire tumorigenic capabilities. Therefore, we systematically implanted 27 EPN tumors (6 primary and 21 recurrent) into the matching locations in the brains of SCID mice following our standardized protocols (same locations, same depth, both sexes of animals of similar age) 29 , 30 , 31 , 34 , 52 (Fig. 1A ). All the mice received the same number (1 × 10 5 ) of viable tumor cells. The animals were closely monitored for up to 15 months and euthanized if they develop neurological deficits or became moribund. Formation of intra-cerebral (IC) or intra-cerebellar (ICb) xenografts was confirmed via gross and histological examinations in 13 patient tumors (48.1%). Among the 5 patients whose primary tumors ( n = 1 patient: PFA3) or early recurrent tumors ( n = 4 patients: RELA1, PFA5, PFA2, PFA1) did not form xenografts, their late recurrent tumor(s) formed PDOX tumors. In RELA1, tumor formation was confirmed in the 7th recurrence after repeated failure of tumor growth at 2nd, 3rd, 4th, and 5th relapse (no patient tumor tissues were available at 1st and 6th recurrence). In PFA1 and PFA2 (Fig. 1A ), PDOX formation was confirmed in the 2nd recurrence. Although tumorigenicity can be affected by many factors, our unique approach of testing serially relapsed tumor tissues from the same patients and using the standardized tumor implantation protocol provided functional data to support the malignant progression of tumor recurrence and suggested a role of the convergent epigenetic reprogramming in promoting tumorigenicity in EPN relapses. Many of the models have been sub-transplanted in vivo in mouse brains for up to five passages while exhibiting a reverse correlation between tumor cells implanted (1000–100,000 cells/mouse) and animal survival times (Fig. 1B ). This set of models will provide a much-needed platform to translate biological findings to functional and preclinical testing for relapsed EPNs. To determine their molecular fidelity, we compared the global mean DNA methylation ratios and DMRs in 7 pairs of patient-xenograft tumors and detected a high-level of similarities ( r > 0.9) in 6/7 models (Fig. 2D, E , Supplementary Fig. 2A – D ) even during subtransplantations (RELA5 and PFA4), demonstrating the faithful recapitulation of patient epigenetic profiles in their matched PDOX models (Fig. 2E ). Gain of chromosome 1q has been associated with recurrence and poor prognosis in PFA tumor 1 , 50 , 53 . To examine if 1q gains might also be associated with tumorigenicity, we applied CNVkit to infer CNV status using our RRBS data. Due to the technique limitations, i.e., the identical reads and uneven coverage in tumors resulted from the enzyme digestion, there was not definitive identification of 1q gain in PFA patient and PDOX tumors (Supplementary Fig. 3 ). To supplement this approach, we applied FISH and detected 1q gain in paraffin sections of three sets of patient and xenograft tumors of PFA ependymoma (Fig. 2C ), suggesting that 1q gain in patient tumors were preserved in the PDOX tumors as well. These data support the analysis of additional PFA tumors, both tumorigenic and non-tumorigenic, to establish the role of 1q gain in PFA tumorigenicity. Potential DNA methylation drivers of EPN recurrence identified To determine the fate of abnormal DNA methylation in the primary tumors and identify potential epigenetic drivers of recurrences, we performed a deep longitudinal analysis of the repeated recurrences to segregate the differential methylated sites (DMCs) consistently present during serial relapses from random DMCs in-transit (Fig. 3A , Supplementary Data 3 ). We used alluvia plot to show the dynamic changes of DNA methylation in the repeated relapses (≥2). Most of the DMCs underwent single direction changes while less than 2% of DMCs switched between HyperDMCs and HypoDMCs (Fig. 3B , Supplementary Fig. 4A ). When the tumors were analyzed individually, they exhibited wide-ranges of variabilities, i.e., 28,791 to 99,702 (54,380 ± 24,495) HyperDMC and 13,662 to 84,714 (44,672 ± 21,560) HypoDMC in RELA and PFA recurrences. To identify the DMCs shared by >2 patients (Fig. 3C , vertical histogram) or maintained in each patient during multiple relapses (Fig. 3C , horizontal histogram), we applied UpsetR to plot the numbers and identified 8155 HyperDMCs and 2324 HypoDMCs shared by all the five RELA patients (HyperDMC Shared and HypoDMC Shared ) (Fig. 3C ); and 6845 HyperDMC shared and 6925 HypoDMC shared shared by the five PFA patients. Many of these shared DMCs, ranging from 61.4 to 83.8%, were located in proximal regulatory regions of genes. To further increase the stringency of potential DNA methylation driver discovery, we focused on the DMCs that exhibited DNA methylation ratio differences greater than >0.3 between tumor and normal cerebellum/cerebrum tissues. From the DMC shared we identified 57 consistent-HyperDMC shared and 51 consistent-HypoDMC shared in RELA tumors; and 148 and 118 in PFA tumors, respectively (Fig. 3D , Supplementary Data 4 ). These DMCs persisted from the primary tumors to late relapses, thereby constituting potential DNA methylation driver signatures of EPN relapse. Fig. 3: Identification of potential DNA methylation drivers for ependymoma relapse. A Schematic illustration of dynamic analysis of time-course DNA methylation to identify consistent DMCs that were present from primary to all recurrent tumors in each individual patient, and the shared DMCs that were consistently present in all RELA or PFA patients. B Representative alluvia plots ( left panel ) showing the dynamic changes of DNA methylation for RELA4 (left upper panel) and PFA1 (left lower panel) tumors during repeated recurrences. Normal brain tissues were used as reference to determine the CpG status, i.e., Hyper-, Hypo-methylation and No Change. Seven different patterns (categories) and the numbers of CpG changes were listed with different colors. Graphs (right panel) showed percentages of the 7 different categories of CpG changes of each RELA (right upper panel) and PFA patient (right lower panel). C UpSet R plots showing the number of consistent Hyper- and HypoDMCs ( y- axis) that were shared among RELA (left panels) and PFA patients (right panels) (connected dots with lines) as well as numbers of consistent DMCs for each patient (the horizonal histograms). Consistent DMCs that were shared by all RELA or PFA accounted for a small fraction and were highlighted in red, respectively. D Heatmaps showing the DNA hyper- (red) and hypo-methylation (blue) ratios of potential DNA methylation drivers (CpGs) for RELA (left panels) and PFA (right panels) recurrent tumors. E Schematic illustration of the data analysis steps to identify potential driver genes regulated by potential DNA methylation drivers. Differentially expressed genes (DEGs) (log2 fold change between tumor and normal tissues) were extracted from RNA-seq of the same set of tumors. DEGs discovered in patient tumors but absent in the matching PDOX tumors were filtered out to identity genes that contributed to PDOX tumorigenicity. F Correlation of expression and DNA methylation of potential genes were shown in the scatterplots for RELA (upper panel) and PFA (the lower panel). Those negatively correlated in RELA were highlighted in red (upper panel) and in PFA in blue (lower panel) and listed to the right of the plot. Full size image Potential DNA methylation drivers regulated a set of differentially expressed genes To identify the target genes of the potential driver DMCs, we examined the differentially methylated regions (DMRs), which are groups of neighboring DMCs that play a more important role in regulating gene expressions than the single DMCs. In RELA primary and recurrent tumors, we detected an average of 9730 HyperDMRs and 8802 HypoDMRs, slightly higher than the 7910 HyperDMRs and 7977 HypoDMRs in PFA tumors (Supplementary Data 5 ) that were distributed on different functional elements (Supplementary Fig. 4B ). Recognizing the impact of locations on the biology, we compared the primary and recurrent tumors between RELA and PFA EPNs (Supplementary Fig. 5 ) and identified 2131 and 976 consistent Hyper/HypoDMAs presented only in all recurrent samples. Cross examination with a public dataset (GSE65362) 16 revealed a high-level similarity of DNA methylation ratios of the CpGs within our DMRs (Supplementary Fig. 5B ). DNA methylation in promoters can repress gene expression 54 . To identify functionally important DMRs, i.e., those associated with genes, we focused on the most significant DMRs that were located within gene transcription start sites (TSS) up/downstream 5 kb of known target genes (Supplementary Fig. 4C, D ). In the RELA patients, 588 potential driver genes (529 associated with the Hyper-DMRs and 59 with the Hypo-DMRs) were shared by all the patients (Supplementary Data 6 ); whereas in PFA patients, there were 736 potential driver genes (522 associated with the Hyper-DMRs and 214 with the Hypo-DMRs) (Supplementary Data 7 ). To further identify potential DMR drivers that were required for the PDOX tumor formation, we filtered potential patient DMR drivers with PDOX DMRs. In RELA patient tumors, 90.5% (479/529) of hyperDMR- and 89.8% (53/59) of hypoDMR-associated potential driver genes were preserved in their PDOX models, whereas in PFA tumors, it was 95.4% (522/547) and 93.8% (214/228), respectively (Fig. 3E , Supplementary Data 8 ). This integrated analysis of patient tumors with their matching PDOX tumors represent a strategy for the discovery of functionally important target genes of potential DMR drivers. To determine if the levels of these target gene expressions were actually regulated by the potential driver DMRs in EPN relapses, we analyzed the RNA-seq data from the same batch of samples to detect the differentially expressed genes (DEGs) (Supplementary Fig. 6A ) that were significantly different ( FDR < 0.05) from the normal cerebella or cerebra, followed by matching the DEGs with the potential DMR drivers that displayed average differences of DMR ratios between cerebella/cerebra and tumors >0.3. Comparison between PFA and RELA tumors revealed that the PFA tumors shared more DEGs (upregulated = 1303 and down = 1290) and by three groups of tumors (from PFA-P to PFA-R1 and PFA-R20, which were remarkably higher than that in RELA tumors (up = relapses by three groups of tumors from PFA-P to PFA-R1 and PFA-R2 were significantly higher than those in RELA tumors (upregulated = 208 and downregulated = 281) that were selectively shared by RELA-P and RELA1 only (Supplementary Fig. 6A ). The differences of cell-of-origin between RELA and PFA tumors 20 , 21 , 22 , 23 , 38 , 55 may have contributed to the differences of DEG panels and the relatively conserved DEGs during PFA relapses. In RELA tumors, we identified 34/38 (89.5%) downregulated target genes (including SYN2, PHACTR3, KCNJ9, RIMS4, FBOX41 ) of the consistent-hyperDMRs Shared , and 5/9 (55.6%) upregulated target genes ( CACNA1H, SLC12A7, CSPG4, RARA, and ZNF423 ) of the consistent-HypoDMRs shared (Fig. 3F ). These genes accounted for 2.4% of the 208 upregulated genes and 12.1% of the 281 downregulated genes in the RELA tumors. Similarly, in PFA tumors, 81/151 (53.6%) downregulated genes (including FAT1, MYT1L, GRM4, KCNK9, PCDHA5, and KCNA1 ) were regulated by the consistent-hyperDMRs shared and 54/76 (71.1%) upregulated genes (including FAM92B, HSPB8, GMPR, ITGB4, FHAD1, and FXYD1 ) by the consistent-hypoDMRs shared (Fig. 3F , Supplementary Data 5 ). They accounted for 4.1% of 1303 upregulated genes and 6.3% of the 1290 downregulated genes in the PFA tumors (Supplementary Fig. 6A ). Although the differences of cell-of-origin between RELA and PFA tumors 20 , 21 , 22 , 23 , 38 , 55 may have contributed to the differences of DEG panels, the numbers of the shared upregulated ( n = 1303) and downregulated ( n = 1290) DEGs during PFA relapses from PFA-P to PFA-R1 and PFA-R2) were significantly higher than those in RELA tumors Functionally, these genes were associated with neuron development, neuron differentiation and neurogenesis (Supplementary Fig. 6B ). The strong negative correlations between the DMRs and their target genes (Fig. 3F ) indicated a potential regulatory role of the potential DMR drivers in regulating the gene expressions. Most of the aforementioned DMR regulated genes were not previously discovered for EPN relapses. Their potential as relapse driver genes was further enhanced by the fact that many of them, including CACNA1H 56 , SLC12A7 57 , CSPG4 58 , 59 , RARA 60 in RELA , and HSPB8 61 , ITGB4 62 , FAT1 63 , 64 in PFA tumors, have previously been associated with human cancers or ependymoma tumor dependency gene ( CACNA1H ) 11 , 65 Despite the biological differences between RELA and PFA tumors, there were 7 downregulated genes shared by the two types of tumors ( ARRB1, KCNJ9 , KIAA0513, KIF5C, SNAP25, SPTBN2, and TNR ) and they were all regulated by DNA hypermethylation (Supplementary Data 6 and 7 ). Worthy of note is that we also found over-expression of WEE1 , of which multiple inhibitors have entered clinical trials 66 , 67 , in 4/4 PFA and 3/4 RELA tumor sets (Supplementary Figs. 5C , 7B ). Compared with RELA tumors, PFA tumors had nearly 3 folds more potential driver genes, most probably due to the differences of cell-of-origin. Altogether, this longitudinal DNA methylation analysis not only identified potential drivers (DMRs and genes) of relapses that were specific to or shared by RELA and PFA tumors, but also revealed the underlying mechanisms of the altered expressions in a set of relapse-related potential driver genes. Potential DNA methylation boosters of relapse were discovered in the serial relapses of ependymoma Our collection of the multiple serially relapsing tumors from the same patients also presented an opportunity to examine DMCs that were newly acquired in the relapsed tumors. The DMCs that persisted in all the recurrences may have sustained (and boosted) tumor relapse and contributed to the increased tumorigenicity. To improve the level of confidence in discovering recurrent-specific DMCs, we included an additional four PFA and four RELA primary tumors from a previous study (GSE87779) 16 as a validation set (Fig. 4A, B , Supplementary Fig. 7D ) and identified a set of RELA and PFA recurrent-specific DMCs in the relapse tumors in this public database (Supplementary Fig. 7D ). Following our analysis of individual tumors, we extracted the DMCs that were shared by the 5 sets of RELA or the five sets of PFA recurrences (hereafter referred as potential DMC Booster ), and identified 296 HypoDMC Booster and 38 HyperDMC Booster in RELA; 165 HypoDMC Booster and 323 HyperDMC Booster in PFA recurrences (Fig. 4A ). We applied the same strategy as detailed in Fig. 3 to identify the genes regulated by DMC Boosters . In RELA tumors, we found 115 HyperDMRs Booster associated genes and 35 HypoDMRs Booster associated genes); whereas in PFA tumors 124 HyperDMRs Booster and 219 HypoDMRs Booster associated genes (Supplementary Data 9 ). Many of these genes could have been missed if only examine one recurrent tumor. Fig. 4: Identification of potential DNA methylation booster of recurrent ependymoma. A Heatmap showing the DMCs that were newly acquired in the recurrent ( R ) RELA (upper panel) and PFA (lower panel) tumors but absent in the primary ( P ) ependymomas. B Representative UCSC genome browser showing regions that was selectively hypermethylated in all recurrent tumors of a RELA recurrent tumors ( n = 7) (upper panel) and PFA ( n = 4) ( lower panel ) but not in their matching primary tumors. C Scheme showing the identification of recurrent-specific DMR (DNA methylation booster) associated genes. Hyper- and hypoDMR associated genes found in the patient tumors but not preserved in the matching PDOX models were filtered out. Differential expressions genes (DEGs) that were negatively correlated with DNA methylation were shown in the scatterplots with relative levels of change for RELA (red in upper panel) and PFA (blue in the lower panel) together with a list of top candidate genes. Full size image Since many of the relapsed tumors were not tumorigenic until late recurrences, we reasoned that some of the potential booster related genes contributed to the elevated tumorigenicity. Direct comparison between patient tumors and their PDOX tumors showed that only 21.7% (25/115) hyperDMR Booster and 28.6% (10/35) hypoDMR Booster related genes in RELA tumors were maintained in the PDOX tumors, whereas in PFA tumors, it was 89.5% (111/124) and 89.8% (188/219 genes) (Supplementary Data 9 ), which was significantly higher than that in RELA tumors ( P < 0.001). These data suggested that the number of genes regulated by DMR Boosters was affected by the difference of cell-of-origin and played a more important role in PFA recurrences, and PDOX tumor formation can be particularly helpful in identify tumorigenic DMR Booster related genes in RELA tumors. To identify potential DMR booster driven genes with high stringency, we focused on the DEGs consistently present in all the relapsed tumors. Unlike the potential DMC-driver related genes, there were no shared potential booster genes between RELA and PFA tumors. In the RELA recurrent tumors, the expression of 10/10 (100%) hyperDMR-regulated genes ( NCDN, DES, MF12, CYB5A, BANK1, DLGAP4, SMTN, UCK1, WAC, and WWWC3 ) 68 and 1/3 (33.3%) hypoDMR regulated genes ( PLEKHG1) were negatively correlated with the corresponding DNA methylation changes (Fig. 4C , upper panel), accounting for 3.4% and 0.48% of the total down- and upregulated genes in RELA tumors. More importantly, all these genes were maintained in the PDOX tumors (Supplementary Data 9 and 10 ) which further supported their role in tumorigenicity. Functionally, some of the downregulated genes by hyperDMR have been associated with human cancers, including advanced stage of cancers (SMTN 69 and DES 70 , BANK1 71 ), autophage related death and prognosis of pancreatic cancer (CYB5A) 72 , 73 and aggressive phenotype of glioma and colorectal cancer recurrence (MFI2) 74 , 75 . As the only upregulated potential booster gene, PLEKHG1 appeared to be an attractive therapeutic target. Although its biological function is not fully understood, a reverse correlation of PLEKHG1 expression with poor survival in low grade gliomas has been noted (Supplementary Fig. 8A, B ). In PFA tumors, we detected 41/66 (62.1%) hyperDMR down-regulated genes (including HTR1A, GRM5, FGF5, GPR25 and SHH ) and 58/89 (65.2%) hypoDMR upregulated genes (including CAPS, ALDH3A1, FAM74A3, FOXJ1, EHF, ITGB5, NOTCH1, EPHA2, and SUFU ) (Fig. 4C , lower panel) (Supplementary Data 10 ), accounting for 3.2% and 4.5% of the total down- and upregulated genes in the PFA tumors. In addition to NOTCH 76 , EPHA2 77 and SUFU 78 that are known to be involved in EPN biology, many of these genes ( FOXJ1 79 , ALDH3A1 80 , EHF 81 ) were associated with human cancers or brain tumors (Supplementary Fig. 7 ). Although the small number of dysregulated genes limited our capacity of detailed biological enrichment analysis, our discovery of their potentially roles in promoting EPN relapses is exciting and warrants future functional validation and drug development. Relapse predictors of DNA methylation can be identified from primary tumors at diagnosis Since not all EPNs relapse, it is highly desirable to develop DNA methylation markers that can predict tumor recurrences when the tumor is diagnosed. We subtracted the primary tumor DMCs of EPNs that eventually recurred (i.e., primary-tumor Eventually Recurred ) from those in the primary tumors that did not relapse (primary tumor Not-replapsed ) over 10 years follow-up. Using UpsetR, we plotted the numbers of DMCs shared by primary-tumor Eventually Recurred or unique in primary tumor Not-replapsed (Fig. 5A , vertical histogram) as well as the DMCs from each primary tumor (Fig. 5A , horizontal histogram). We further hypothesized that the active predictors identified in the primary tumors should also be present in the subsequent recurrent tumors. As shown in Fig. 5B , the frequency of the DMCs specific to the primary tumor Eventually Recurred were significantly higher than that in the primary tumors Not-relapsed . To increase the stringency of our data analysis, we focused on the DMCs that were selectively present in all (12 of 12) of the recurrent RELA tumors and all (10 of 10) recurrent PFA tumors (Fig. 5B, C ), respectively; and narrowed the candidate predictor DMCs to 2207 hyper- and 791 hypoDMCs in RELA and 1305 hyper- and 2029 hypoDMCs in PFA tumors (Fig. 5D ). From this list, we further ranked the DMCs (Fig. 5E ) based on the differences of DNA methylation ratios between tumor and normal tissues and identified 7 DMCs in RELA and 22 DMCs in PFA tumors (with differences of DNA methylation ratio ≥0.8) as the top candidates of relapse predictor DMCs (Fig. 5E ). Worthy of note is that some markers with the highest confidence, such as chr17-80943940-80943942 and chr5-92910051-92910053, exhibited wide-range of DNA methylation ratios, suggesting the possibility of sample variations. Despite the relatively small sample size, our discovery of this set of “recurrence-bound” Hyper- and HypoDMCs in the primary tumors is very encouraging. They provided proof-of-principle to support future extended studies on this important topic. New diagnostic predictors of recurrence can potentially cause a paradigm shift in the clinical care of childhood EPNs. Fig. 5: Identification of DNA methylation predictors of recurrence in primary ependymoma tumors. A UpSet R plot showing the hyper- and hypo-methylated CpGs sites (DMCs) in RELA (upper panel) and PFA (lower panel) primary tumors ( -P ) as compared with normal cerebrum or cerebellum tissues. The horizonal histogram represents the number of DMCs in each comparison between primary tumor and normal brain tissues; the vertical histogram represents the number of DMCs shared by tumors marked by connected dots. Red and green bar highlights the hyper- and hypoDMCs that were shared by all the tumors that eventually recurred (primary-tumor Eventually recurred ) but not in the normal brain tissues and the non-recurrent reference tumors. B Bar graph showing the frequencies of hyper- and hypoDMCs specific to ependymoma primary tumors that eventually recurred ( -P-Specific ) and that did not relapse (- NonRecur Specific ) in the RELA recurrent ( n = 12) and PFA ( n = 10) recurrent tumors. Boxplots indicate median, first and third quartiles (Q1 and Q3), whiskers extend to the furthest values; the uppermost and lowest line indicates the maximum and minimum values, respectively. In RELA tumors, the numbers of HyperDMCs analyzed are RELA6- NonRecurSpecific ( n = 10,309 CpGs), RELA-P-specific ( n = 357,141 CpGs); wherease of the HypoDMCs are RELA6-NonRecurSpecific ( n = 17,188 CpGs) and RELA-P-specific ( n = 507,987 CpGs). In PFA tumors, the HyperDMCs analyzed are PFA6-NonRecurSpecific ( n = 24,296 CpGs) and PFA-P-specific ( n = 349,746 CpGs); and HypoDMCs PFA6NonRecurSpecific ( n = 40,525 CpGs); PFA-P-specific ( n = 451,426 CpGs). C Heatmap showing the DNA methylation ratios of DMCs specific to RELA (upper panel) or PFA (lower panel) primary tumors Eventually recurred exist in all recurrent tumors. Black dash box shows that these DMCs have similar DNA methylation ratios among RELA-/PFA-NonRecur primary tumor, normal cerebellum and cerebrum, but different from the primary tumors Eventually recurred. D Venn diagram representing the number of overlapped and specific hyper/hypoDMCs (from C ) between RELA and PFA. E Heatmap showing the top CpGs’ DNA methylation ratios with highest confidence that can potentially predict recurrence from primary tumors of RELA (upper panel) and PFA (lower panel) ependymoma. Full size image Discussion In this study, we improved the understanding of pediatric EPN relapse by using a collection of serially relapsing EPN tumors matched with primary tumor at diagnosis. Using a high-resolution analysis of DNA methylation together with RNA-seq, we showed that the molecular subtypes were maintained during long-term consecutive serial relapses and discovered that their epigenetic profiles progressively converged during serial relapses. Systematic analyses on animal tumorigenicity derived from the same panel of tumor samples revealed a significantly increased PDOX tumorigenicity in some late relapse tumors, which provided functional data supporting the progressive nature of EPN relapses. Parallel analysis of these PDOX models also bridged a gap between the epigenetic reprogramming and the increased tumorigenicity by fine-tuning the potential DMC drivers and boosters critical for EPN relapses. Patient tumors represent the most reliable source for biological studies. One of the major challenges in understanding tumor relapse is the very limited availability of recurrent tumor tissues 78 . It requires a strong collaborative team and a long-term commitment. In the current study, we proactively collected and carefully followed 110 EPN patients over 13 years and successfully collected 10 (9%) sets of relapsed tumors. However, when compared with a relapse rate of 50% in childhood EPNs, we only captured <20% the recurrences. One major reason is that the treatment of EPN relapses does not include surgery or biopsy as the need of making diagnosis has already been met in the primary tumors and there is currently no sufficient justification for a routine “second look” of tumor pathology. Our discovery of epigenetic reprogramming in the recurrent tumors provided biological evidence to support biopsy or surgical resection of recurrent EPNs for updated molecular diagnosis and informed clinical treatment of recurrent EPNs. The epigenetic convergence during repeated recurrences of pediatric EPNs indicates a decreased cellular heterogeneity of DNA methylation in the recurrent tumors, a result different from previous reports on the increased or expanded gene mutation loads in the recurrent tumors of other type of cancers 44 . One possible cause of this phenomenon is that clinical therapies, particularly radiation therapy, which remains the mainstay of clinical treatment in pediatric EPNs 1 , 2 , selected or conditioned a subpopulation of surviving tumor cells for relapse. This result may be clinically important as it suggested a possibility of targeting a small set of epigenetic drivers of recurrence for significantly improved efficacy. Tumor recurrence is propelled by a cascade of genetic and epigenetic events. Our identification of potential relapse drivers (the DMCs and genes that persisted from primary to all the relapsed tumors) and boosters (the DMCs and genes that only emerged in the relapsed tumors) deciphered a detailed long-range roadmap of pediatric EPN recurrences. However, childhood brains are often in different differentiation status, e.g., from 2–10 year old as in our cohorts. When combined with different time frame of recurrence, ranging from 1 to 13 years, some of the potential DMC drivers may be attributed to the patient specific-differentiation status of cerebrum and cerebellum. In addition to discovering a set of potential driver and booster genes for EPN relapses, one encouraging aspect of our finding is that many of the genes have already been involved in human cancer biology. For example, CACNA1H , a voltage-gated calcium channel, has been detected in breast cancer 56 , 82 ; RARA , a retinoic acid receptor alpha, has exhibited important roles in leukemia and recently in medulloblastoma and glioma 83 ; and HSPB8 ( heat shock protein beta-8) promotes glioma growth and metastasis 61 , 84 . As the only potential booster gene in RELA relapses, PLEKHG1 (pleckstrin homology and RhoGEF domain containing G1) was significantly upregulated in gastric cancer plasma and associated with poor overall survival 85 . Identification of NOTCH, EPHA2 and SUFU as potential booster genes of PFA relapse not only suggested potential roles of these genes in recurrences, but also suggested a set of potential druggable targets for the difficult-to-treat PFA tumors 76 , 86 . Identification of markers that can predict tumor recurrence at the time of diagnosis is highly desirable. Similar to difficulties in obtaining tumor tissues from recurrent EPNs, it is also difficult to locate surgical samples from patients who remain tumor free for at least >5 years from diagnosis. Despite extensive efforts, we were only able to obtain 1 sample each of RELA and PFA tumors that did not recur 5–10 years. Our identification of DMC predictors of recurrence may help patient stratification for biologically based rational selection of treatment strategies. This set of proof-of-principle data should ignite broad interest in this field by analyzing larger collections of recurrent and non-recurrent samples to improve and validate the list of predictors. Despite sample variations of candidate predicators, it is highly desired that the number of predictors for EPN recurrence for each current or future molecular subtypes will decrease or be clinically applicable. To translate biological studies from bench to clinic, we systematically implanted the same panel of tumor tissues into the matching locations in the mouse brains. Although it is time-consuming requiring nearly 13 years of continued efforts, we demonstrated the power of this strategy by discovering the elevated tumorigenicity of late recurrences, filtering out 6–10% the DMCs that were not directly involved in PDOX formation, and functionally validating potential DMC drivers that sustained the progression of recurrent EPNs. These set of data extended previous and our findings of chromosome 1q gain in promoting poor prognosis and potentially driving PFA tumorigenicity by providing a broader and higher resolution molecular signatures. This panel of clinically relevant animal models mayfacilitate the biological and preclinical studies of EPN relapses. There are some limitations of our study. The number of patients is relatively small due to the rarity of pediatric EPNs, and we hope our finding will facilitate the biopsy or surgery on recurrent tumors to better understand tumor biology and increase the availability of recurrent tumors. Tumorigenicity can be affected by multiple factors. Although we have tried to standardize the protocol, there might be other determinants of PDOX tumor formation in addition to the genetic/epigenetic changes that were addressed in our study. Analysis of single cells may shed light on the cellular heterogeneity and driver cells of EPN recurrences. Emerging data have suggested potential roles of non-CpG methylation in brain development and cancer biology 40 , 41 , 43 , 87 . In addition, more cases and higher resolutions (e.g., whole-genome methylation sequencing) are needed to support our finding regarding the decreased mCpA in PFA and RELA EPN (regardless of relapses or not). In summary, we defined a roadmap of the epigenetic progression in patient-matched serial EPN recurrences, discovered convergent DNA methylation profile as a feature of serial relapses, and suggested a set of potential DNA methylation drivers and boosters that sustained and promoted recurrence. We also uncovered an increased tumorigenicity in late recurrences that paralleled the epigenetic progression, established a panel of PDOX models and demonstrated the power of these models in fine-tuning the list of potential drivers and boosters of EPN relapses. Our findings may improve the insight about the underlying epigenetic mechanisms of EPN recurrences. Methods Tumor tissues from childhood EPN patients Signed informed consent was obtained from the patient or their legal guardian prior to sample acquisition following an Institutional Review Board (IRB) of Baylor College of Medicine approved protocol (H-4844). Thirty-three freshly resected EPN tumor specimens from 10 patients undergoing serial surgical resections of primary and recurrent tumors (3.67 ± 1.76/per patient) over a period of 13 years (5.8 ± 3.8) at Texas Children’s Hospital were obtained for this study. Four samples of normal brain tissues (two cerebral and two cerebellar tissues) obtained from warm (<6 h) autopsy of children were included as the normal references. The patients’ demographic and clinical information is described in Table 1 . All the samples were subjected to pathological diagnosis and graded following the WHO system. Tumor tissues were divided into two portions for processing. One part was snap frozen in liquid nitrogen and preserved in the −80 °C freezer. The 2nd part of fresh tumor tissues was washed and minced with fine scissors into small fragments. Single cells and small clumps (3–5 cells per clump) of tumor cells were collected with a 35 μ cell strainer, resuspended in DMEM growth medium to achieve a final concentration of 1 × 10 8 live cells per ml, as assessed by trypan blue staining, and transferred to animal facility on ice. Global reduced representation bisulfite sequencing (RRBS) Genomic DNA was extracted with Allprep DNA/RNA mini kit (Qiagen). and concentrations measured using the Qubit® dsDNA BR Assay Kit (Thermo Fisher Scientific), followed by DNA quality assessment with the Fragment Analyzer TM and either the DNF-487 Standard Sensitivity or the DNF-488 High sensitivity genomic DNA Analysis Kit (Advanced Analytical). RRBS libraries were prepared using the Premium Reduced Representation Bisulfite Sequencing (RRBS) Kit (Diagenode Cat# C02030033), according to the manufacturer’s protocol. 100 ng of genomic DNA were used to start library preparation for each sample. Following library preparation, samples were pooled together either by 8 (mouse contamination up to 10%), 7 (mouse contamination up to 25%), 5 (mouse contamination up to 40%), or 4 (mouse contamination up to 55%). In total, 16 pools were prepared. PCR clean-up after the final library amplification was performed using a 1.5× beads:sample ratio of Agencourt® AMPure® XP (Beckman Coulter). RRBS library pools quality control was performed by measuring DNA concentration of the pools using the Qubit® dsDNA HS Assay Kit (Thermo Fisher Scientific), and the profile of the pools was checked using the High Sensitivity DNA chip for 2100 Bioanalyzer (Agilent). Each RRBS library pool was deep sequenced on one lane HiSeq3000 (Illumina) using 50 bp single-read sequencing (SR50). RRBS data generation and processing Raw RRBS FASTQ files were mapped to NCBI Human Reference Genome Build GRCh37 (hg19) using BSMAP (v2.9) RRBS mode 20 . DNA methylation ratio and differential methylated cytosine (DMCs/DMRs) were analyzed by using MOABS (v1.2.9) 21 . CpG sites with five or more reads covered were used for downstream analysis. Bisulfite conversion rates were estimated on the basis of lambda phage genome spike-ins. The bedGraph files including single base pair DNA methylation ratios were transformed to bigwig file format which can be visualized using the UCSC genome browser. DNA methylation heatmaps were plotted using R package heatmap3 by taking the shared CpGs among all the samples as input. DNA methylation phylogenetics analysis was performed by using R package ape 22 . To compare multiple groups’ DMCs, we merge all the DMCs in all two-group comparisons (union DMC sets). The UpSetR 23 package was used to visualize the union of DMC sets among multiple two-group comparisons. To analyze dynamic changes of DMCs among tumor recurrent, we first separate DMCs to three categories (Hyper; Hypo; NoChange) based on adjacent two recurrent stages. Then, by considering four adjacent recurrent stages (P vs. cerebellum; R1 vs. cerebellum; R2 vs. cerebellum and R3 vs. cerebellum), we filtered out those DMCs with hyper/hypo and hypo/hyper switch between any two adjacent stages due to the small numbers. We finally separated DMCs into seven categories: consistent Hyper; consistent Hypo; Gain Hyper; Gain Hypo; Loss Hyper; Loss Hypo and switch (between hyper/hypo and NoChange). R alluvial package ( ) was used to visualize the dynamic changes of DMCs along tumor recurrence. Colored density scatterplot of DNA methylation ratios was performed by using R package smoothScatter ( ). GREAT 24 was used to predict DMRs’ functions. The analysis codes are available at . To infer DNA copy number status, particularly chromosome 1q, we applied CNVkit ( ) to infer CNV using RRBS data for PFA human and PDX samples. The CNVkits calculate normalized coverage in bin-level then it removes the systemic bias (such as CG content) use circular binary segmentation (CBS) to infer discreate copy number regions as segments. RNA-seq analysis RNA-seq libraries for transcriptome analysis were prepared using the TruSeq RNA Sample Preparation Kit (Illumina) and Agilent Automation NGS system per manufacturers’ instructions. Sample prep began with 1 µg of total RNA from each sample. Poly-A RNA was purified from the sample with oligo dT magnetic beads, and the poly(A) RNA was fragmented with divalent cations. Fragmented poly-A RNA was converted into cDNA through reverse transcription and were repaired using T4 DNA polymerase, Klenow polymerase, and T4 polynucleotide kinase. 3’ A-tailing with exo-minus Klenow polymerase was followed by ligation of Illumina paired-end oligo adapters to the cDNA fragment. Ligated DNA was PCR amplified for 15 cycles and purified using AMPure XP beads. After purification of the PCR products with AMPure XP beads, the quality and quantity of the resulting. FastQC ( ) was used to do quality checks for raw fastq files. Raw FASTQ files were aligned to NCBI Human Reference Genome Build GRCh37 (hg19) using HISAT2 16 with default settings. The uniquely mapped reads were used for downstream analysis. HTSeq 17 was used to count the reads count mapped in exon regions for each gene. Read counts matrix (row as genes; column as samples) were inputted to DESeq2 18 to identify differentially expressed genes (DEGs). We consider genes with ≤FDR ≤ 0.05 and fold change ≥ 2 folds as DEGs. Principal component analysis of DEGs was performed using R package DESeq2. DEGs’ function enrichment was using GSEA 19 . Development of patient-derived orthotopic xenograft (PDOX) mouse models The SCID mice, NOD.129S7(B6)- Ragl tm1Mom /J (Jax Laboratory), mice were bred and housed in a specific pathogen-free (SPF) animal facility at Texas Children’s Hospital in Houston or Lurie Children’s Hospital in Chicago. All the experiments were conducted using Institutional Animal Care and Use Committee (IACUC) approved protocols from Baylor College of Medicine or Northwestern University. The Office of Laboratory Animal Welfare in both institutions are fully accredited from the Association for Assessment and Accreditation of Laboratory Animal Care International (AAALACI). The vivarium water, temperature, and light cycles are controlled by centralized computers. The vivarium is staffed by full time veterinarians and support personnel that administer a complete program of veterinary care. Tumor tissues from primary and surgical transplantation of tumor cells into the mouse brain was performed using a free hand implantation strategy 29 , 30 , 31 . Both male and female mice, aged 5–8 weeks (to simulate the developing brain in children), were anesthetized with inhaled isoflurane and/or pentobarbital (50 mg/kg, i.p . injections). Each animal will be given a pain medication at time of surgery before the intracranial injection of tumor cells. Tumor cells (1 × 10 5 ) were suspended in 2 µl of culture medium and injected into the right frontal/temporal region (1 mm to the right of the midline, 2.5 mm anterior to the lambdoidal suture) or right cerebellar lobe ((1 mm to the right of the midline, 1 mm posterior to the lambdoidal suture) and 3 mm deep via a 10 µl 26-gauge Hamilton Gastight 1701 syringe needle 29 , 30 , 31 . After tumor injections, animals will be monitored daily for 3–4 days. After the initial period of daily observation, all tumor-bearing animals will be monitored for symptoms of discomfort or pain. A veterinary intervention will be elicited if an animal is unable to eat or move or showing behavior such as huddled posture or self-mutilation, or any signs of infection (such as eye or ear). If an animal develops torticollis, uncontrolled circling or other signs of neurologic deficits such as limb paralysis, or loss of body weight (>15%), or become moribund, the animal will be euthanized through intraperitoneal injection of Euthasol (pentobarbital sodium and phenytoin sodium) at 150 mg/Kg to introduce deep anesthesia before the whole mouse brain is removed for histopathologic examination. Those mice without any neurological deficit after 12 months were euthanized and examined for tumor development. To perform serial subtransplantations, whole brains of donor mice were aseptically removed, coronally cut into halves, and transferred back to the tissue culture laboratory. Tumors were then dissected under the microscope, mechanically dissociated into cell suspensions, counted and injected into the brains of recipient SCID mice as described above 30 . Fluorescence in situ hybridization (FISH) analysis FISH analysis was performed on 5 μm paraffin embedded sections slides 48 using Vysis/Abbott Molecular (Des Plaines, IL) dual color probes targeting chromosome 1p36.3/TP73 and 1q25.2/ANGPTL loci, with the 1p36.3/TP73 locus labeled with spectrum red and the 1q25.2/ANGPTL labeled with spectrum green, for the detection of copy number alterations of both loci, following standard laboratory procedures in the clinical cytogenetics laboratory. A total of 100 non-overlapping cells were evaluated by two technologists independently. The average signals for both 1p36.3/TP73 and 1q25.2/ANGPTL were calculated. The signal ratio of 1q/1p ≥ 2.0, is interpreted as gain or amplification of 1q25.2/ANGPTL. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Data availability The raw RNA-seq and methylation data generated by this study are available from the NCBI under accession number GSE156619 . The RELA and PFA signature gene list is from public dataset GSE64415 ) 16 , 39 , 88 . Additional RELA and PFA primary tumor WGBS data are from public dataset GSE87779 10 and public DNA methylation array data for RELA and PFA relapse samples is from GSE65362 16 . Source data are provided as a Source Data file. The remaining data are available within the Article, Supplementary Information or Source Data file. Source data are provided with this paper. Change history 22 December 2022 A Correction to this paper has been published:

Investigators have identified previously unknown sets of epigenetic changes in pediatric brain tumors, which could serve as novel therapeutic targets and provide alternative treatment options for patients, according to a Northwestern Medicine study published in Nature Communications. Xiao-Nan Li, MD, Ph.D., the Rachelle and Mark Gordon Professor of Cancer Research and a member of the Robert H. Lurie Comprehensive Cancer Center of Northwestern University, was senior author of the study. Pediatric ependymoma, the third most common type of brain cancer in children, currently has a recurrence rate of 50% and most patients will eventually succumb to the disease due to poor treatment response. While the cancer is aggressive and recurrence is common, the tumors grow slower in comparison to other types of brain tumors, offering a small window for patient care teams to try different treatment strategies, including the resection of recurrent tumors. Although surgery and radiation can help prolong patient survival, more effective and long-term treatment options are urgently needed, according to Li. In the current study, Li's team performed DNA methylation sequencing together with RNA sequencing to identify epigenetic alterations using patient-matched primary and recurring ependymoma tumor tissue samples from 10 pediatric patients who had been treated with radiation and experienced cancer recurrence multiple times over a 13-year period. From this unique set of patient tumors, the investigators developed a novel panel of patient-derived orthotopic xenograft mouse models, which are critically needed for improving the understanding of tumor biology and for the testing of new therapies, according to Li. "Ependymoma is a very interesting tumor, and unlike many other human cancers, mutations are not a major cause. In this tumor, we call the mutations 'silent' because there's not a whole lot of mutations. The changes are mostly on the epigenetics side, and that's why we focus on DNA methylation," said Li, who is also a professor of Pediatrics in the Division of Hematology, Oncology, and Stem Cell Transplantation. Using the xenograft models, the investigators identified three unique sets of epigenetic mutations. The first set included differentially expressed genes that were regulated by potential "driver" DNA methylation regions (DMRs), or genomic regions containing different DNA methylation patterns, that carried over from the primary tumors to recurring tumors. The second set contained differentially expressed genes that were regulated by potential "booster" DMRs, which were found only in recurring tumors. In particular, the investigators discovered that the gene PLEKHG1 was expressed in all the recurrent tumor models which, according to Li, suggests that this gene could be a potential therapeutic target. The third set of mutations identified in primary tumors were "predictors of relapse," or a set of epigenetic markers that could help predict whether a tumor has a high risk of recurrence. These predictors could help providers avoid giving unnecessary treatment to patients which could lead to further complications down the road, according to Li. "The pediatric brain is still developing, and if you treat them unnecessarily too harsh, it will delay development and decease quality of life. If there's any way we can predict and improve their quality of life, that's so important," Li said. Furthermore, screening these drivers and boosters could help determine potential therapeutic targets and precision medicine-based treatment strategies. "The important things to know are that recurrent tumors and not exactly the same as primary tumors—they change a lot—and for this type of study, we need more collaboration. We need to work with as many hospitals and institutions as possible," Li said.

10.1038/s41467-022-34514-z

Medicine

Using stem cells to regenerate the heart

Julia Junghof et al, CDH18 is a fetal epicardial biomarker regulating differentiation towards vascular smooth muscle cells, npj Regenerative Medicine (2022). DOI: 10.1038/s41536-022-00207-w

http://dx.doi.org/10.1038/s41536-022-00207-w

https://medicalxpress.com/news/2022-02-stem-cells-regenerate-heart.html

Abstract The epicardium is a mesothelial layer covering the myocardium serving as a progenitor source during cardiac development. The epicardium reactivates upon cardiac injury supporting cardiac repair and regeneration. Fine-tuned balanced signaling regulates cell plasticity and cell-fate decisions of epicardial-derived cells (EPCDs) via epicardial-to-mesenchymal transition (EMT). However, powerful tools to investigate epicardial function, including markers with pivotal roles in developmental signaling, are still lacking. Here, we recapitulated epicardiogenesis using human induced pluripotent stem cells (hiPSCs) and identified type II classical cadherin CDH18 as a biomarker defining lineage specification in human active epicardium. The loss of CDH18 led to the onset of EMT and specific differentiation towards cardiac smooth muscle cells. Furthermore, GATA4 regulated epicardial CDH18 expression. These results highlight the importance of tracing CDH18 expression in hiPSC-derived epicardial cells, providing a model for investigating epicardial function in human development and disease and enabling new possibilities for regenerative medicine. Introduction Cardiovascular diseases are the leading cause of death worldwide. The human heart is incapable of restoring itself and myocardial repair and regeneration processes are still poorly understood. In recent years the epicardium has emerged as a therapeutic target, given its ability to re-activate upon cardiac injury and promote cardiac repair 1 . The epicardium is a mesothelial layer covering the myocardium and serves as a progenitor source supporting cardiac development, repair, and regeneration 2 , 3 , 4 . Epicardiogenesis is evolutionary conserved and absolutely essential for cardiac development, with failure of epicardial formation being embryonic lethal 5 , 6 . After the formation of the proepicardium (PE), a transient cauliflower-like structure at the venous pole of the looping heart, cells of the PE migrate over and cover the looping heart tube completely, subsequently forming the epicardium. The embryonic epicardium is active, epicardial cells are proliferative and have the ability to undergo epicardial epithelial-to-mesenchymal transition (EMT). Those epicardial-derived cells (EPDCs) can proliferate and migrate, invading the underlying myocardium, where they subsequently differentiate into various cardiac cell types like vascular smooth muscle cells (SMC), cardiac fibroblasts (CF), and, to a lesser degree, endothelial cells (EC) 7 , 8 , 9 . The epicardium and EPDCs exhibit extensive developmental plasticity, which is crucial for cardiogenesis, by contributing to coronary vessel formation as well as myocardial growth and maturation 1 , 2 , 4 , 10 , 11 . During adulthood, the epicardium enters quiescence and serves as a protective layer. Only upon cardiac injury, the epicardium re-activates and contributes as a quick transient progenitor hub of derivative cells to the cardiovascular system regeneration 4 , 12 . This fact provides the epicardium with a pivotal role in mediating regenerative responses. Unlike other species, mammals lack the ability to restore the heart. Nonetheless, re-activation of the epicardium is required for cardiac scar formation as well as coronary vessel growth 13 . Not only promotes the epicardium myocardial regeneration via paracrine signaling, but it also mediates inflammatory responses 14 , 15 . Moreover, engineered patches carrying epicardial follistatin-like 1 (FSTL-1) were able to enhance cardiac regenerative responses 16 . However, despite ongoing investigation on the epicardium as a therapeutic target tissue, processes governing epicardial development, re-activation and engraftment ability are not well understood, mainly due to lacking comprehension of the fundamental biology behind the epicardium itself, including the expression dynamics of epicardial epitopes mediating tissue responses to homeostasis disruptions and cellular signaling upon cardiac damage. The epicardium coordinates a complex network of surface remodelers in order to respond to different stimuli driven by organ development and tissue repair 11 , 17 . The cell surfaceome describes the whole proteome of surface and transmembrane proteins. During tissue morphogenesis, homeostasis, and regeneration, the surfaceome, particularly surface remodelers, plays an essential role in expansion, migration, and invasion, forming complex cellular structures 18 . However, cell-surface markers defining functional epicardial cells or regulating epicardial cell-fate decisions have been difficult to identify. The cadherin family of cell–cell adhesion proteins are important for tissue morphogenesis 19 . Even though cadherins may have originated to facilitate mechanical cell-cell adhesion, their functions are pleiotropic and have evolved to be imperative for many other aspects, including cell recognition, coordinated cell movements, cell-fate decisions and the maintenance of structural tissue polarity, by controlling diverse signaling pathways 20 , 21 , 22 , 23 , 24 , 25 . In this study, we identified type II classical cadherin CDH18, formerly known as cadherin 14 (CDH14), as a specific biomarker expressed in the fetal-stage epicardium, defining cellular specification. Human pluripotent stem cells (hPSCs) 26 , 27 and their capacity to differentiate into cardiac cells allow recapitulation of epicardiogenesis in vitro 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 . Here we generated human-induced (hi)PSC-derived epicardial-like (EPI) cells. Epicardial identity was associated with sustained expression levels of key epicardial genes and retained CDH18 expression in correlation to GATA4 expression, an essential transcription factor proven to be required to form PE in vivo 37 . The loss of CDH18 expression led to the activation of cell-fate specific EMT towards SMC differentiation, confirming an important biological function of CDH18 , not only to define progressive epicardial lineage specification but also to regulate downstream signaling that drives EPDC derivation. Our study sets a basis for a reproducible model to investigate epicardial function in human cardiac development and disease, ultimately enabling new possibilities in regenerative medicine. Results Defined transcriptome of developing epicardial surfaceome revealed the enrichment of CD22 and CDH18 in late-stage populations To investigate the expression profiles of the epicardium, we compared RNA-Seq dataset of induced epicardial-like cells at day (d) 12 (Epi12) and 48 (Epi48) representing a developmental early- and late-stage respectively and investigated their relationship with the adult quiescent epicardium 28 . Whole mRNA transcriptome-based principal-component analysis (PCA) demonstrated 84.75% variance between Epi12 and Epi48 populations (Fig. 1a ), with the biggest difference observed between Epi12 and adult—quiescent—cells (Supplementary Fig. 1a, b ). The active embryonic epicardium is defined by the expression of the transcription factors Wilms’ tumor 1 ( WT1 ) and T-Box 18 ( TBX18 ), followed by aldehyde dehydrogenase 1 member A2 ( ALDH1A2 ), distinguishing it from the myocardium and endocardium, as well as the adult epicardium which lies quiescent (Supplementary Fig. 1c ). All three genes were upregulated in correlation to each other over time (Supplementary Fig. 1d, e ), confirming the embryonic epicardial identity of the transcriptome profiles. Comparative analysis of genes in correlation to WT1, TBX18 and ALDH1A2 , unveiled 2221 common genes in their intersection (Fig. 1b ), with 49 genes identified as surface protein-encoding. We also found 825 differentially expressed genes (DEGs) between Epi12 and Epi48 populations (Supplementary Fig. 1f ) allowing us to subsequently define a putative transcriptome of epitopes of the active epicardium (Fig. 1c ). Fig. 1: Defined surfaceome transcript of the developing epicardium revealed the enrichment of CD22 and CDH18 in late-stage populations. a PCA using the GSE84085 RNA-Seq expression dataset of d12 early epicardial-like cells (Epi12, red) (19-9-7-Epi, n = 2), d48 late epicardial-like cells (Epi48, brown) (H9-Epi, ES03-Epi and 19-9-11-Epi, n = 6) and human adult epicardium (dark violet) (donors 9605, 9633, 9634 and 9635, n = 8). b Venn-Diagram showing genes in significant absolute correlation to the epicardial markers WT1 (light green), TBX18 (light pink) and ALDH1A2 (light purple) (Pearson correlations; R ≥ 0.6) as well as cell surface marker enrichment (target) . c Overlap of the target gene set from b (gray) against DEGs (Supplementary Fig. 1e ) (light brown) for the surfaceome transcript identification of the epicardium. d Heatmap showing correlations (Pearson, R , 1 to −0.7, color scale) of surface marker expression relative to the expression of endothelial cell markers CDH5 and PECAM1 . e Fold change of surface marker expression in Epi48 (brown) relative to Epi12 (red). f Heatmap showing correlations (Pearson, R , 0.6–1, color scale) of the top three genes from ( e ) to epicardial markers. g Retrospective analysis showing a heatmap depicting active epicardial markers and CD22 and CDH18 expression in embryonic and adult cells. h Analysis for CD22 and CDH18 expression in CF (left) and SMC (right) with representative CF genes (yellow) and SMC genes (light blue) [error bars indicate standard derivation (SD)]. i Heatmap showing correlations (Pearson, R , 0 to −1, color scale) of CD22 and CDH18 for CF markers PDGFRA , FAP and POSTN and SMC markers CNN1 , ACTA2 , and TAGLN . j GSE106168 retrospective analysis for CDH18 expression in different cardiac cellular clusters [(C1—5W whole heart ( n = 257); C2—Cardiomyocytes ( n = 1492); C3—fibroblast-like cells ( n = 786); C4—endothelial cells ( n = 445); C5—heart valves ( n = 427); C6—epicardial -EPI- cells ( n = 46); C7—immune cells ( n = 27); C8—macrophages ( n = 308); C9—T and B lymphocytes ( n = 58)] k GSE106168 retrospective analysis for CDH18 expression in epicardial-EPI-cluster [ UPK3B high WT1 high TBX18 high ALDH1A2 high ] ranging different time points of gestation [5 weeks, n = 30; 22 weeks, n = 46]. Full size image Next, we excluded genes that had no significant negative correlation to the endothelial cell markers CDH5 and PECAM1 (Fig. 1d ) and selected late-stage upregulated genes (Fig. 1e ) that showed a highly positive correlation to epicardial genes (Fig. 1f ), obtaining CD22 and CDH18 as candidate cell surface biomarkers in the active epicardium. We then confirmed that both candidates were no longer expressed in adult tissue similar to active epicardial key genes (Fig. 1g and Supplementary Fig. 1g ). Next, we verified their absence in two important epicardial derivatives. While CF and SMC markers were upregulated in their respective cellular identities, no mRNA expression of CD22 or CDH18 was detected (Fig. 1h, i ). Moreover, we confirmed tissue specificity using independent datasets: Heart CD22 expression ranked #2 after central nervous system (CNS), and heart CDH18 expression ranked #3 after CNS and testis (Supplementary Fig. 1h ), two known CDH18-expressing tissues 38 , 39 , 40 . Finally, as an increase of CDH18 is difficult to observe during heart development of mouse embryos due to cellular heterogeneity in the bulk population (Supplementary Fig. 1i ), we organotypically cultured explanted epicardial tissue from E14 hearts and documented CDH18 expression in epicardial explants (Supplementary Fig. 1i ). We confirmed the epicardial specificity of CDH18 expression in a physiological context of cardiac cells derived from human embryos ranging from different time points of gestation (Fig. 1j, k ). CDH18 but not CD22 is an epicardial biomarker defining progressive lineage specification To define the expression pattern of CD22 and CDH18 , we first recapitulated epicardiogenesis in vitro. We modified previously reported induction methods for hiPSC-derived epicardial-like (EPI) cells 28 , 32 , 35 to robustly yield WT1 + cells at a higher rate, and long-term culture of cells was achieved by permanent inhibition of transforming growth factor (TGF)-β (Fig. 2a ). As epicardial cells and atrial cardiomyocytes share a common progenitor pool during development 3 , we excluded the presence of any residual myocyte traces in our system, by generating EPI cells from MYH6 reporter line (MYH6-EIP4) 41 , obtaining no cardiomyocytes (Supplementary Fig. 2a–c ). Moreover, the absence of endothelial markers CD31 and CD144 ensured that no cells of endothelial nature were present (Supplementary Fig. 2d–f ). EPI cells showed typical epicardial cobblestone-like morphology (Fig. 2b ), and key active epicardial genes ( WT1 and TBX18) started to be upregulated as early as d12 continuing so until d24 (Fig. 2c ). In agreement with previous models, cells did not upregulate ALDH1A2 —a gene highly expressed in the epicardium but not PE—until d24, indicating that the cells at d12 represented an early stage during development comparable to PE cells and specified into later stage epicardium at d24 (fetal-like) (Fig. 2c, d ). We found all genes expressed in correlation to each other (Supplementary Fig. 3a ) and did not observe significant differences in their expressions after d24 (Supplementary Fig. 3b, c ), endorsing that cells had completed differentiation into fetal-stage epicardium by d24. We also confirmed WT1 protein expression on this day by immunofluorescence (Supplementary Fig. 3d–f ). Fig. 2: CDH18 but not CD22 is an epicardial biomarker defining progressive lineage specification. a Schematic illustration depicting the epicardial differentiation from hiPSCs and representative developmental stages. b Phase contrast microscopy showing the morphology of EPI cells from d6-48. c qRT-PCR analysis of WT1 (light green), TBX18 (light red) and ALDH1A2 (light purple) during induction normalized to d6. [d6 n = 7, d12 n = 10, d24 n = 10; * p < 0.05, ** p < 0.005, *** p < 0.0005]. d Heatmap showing PE and primitive epicardium gene signature enriched in d12 and d24 EPI cells respectively. e Time course of CDH18 expression levels (orange) normalized to d6 [d6 n = 4, d12 n = 4, d24 n = 8, d48 n = 8; ** p < 0.01, *** p = 0.0008]. f Flow cytometry analysis for CDH18 (orange) [gray, unstained control]. g , Nonlinear-regression fit curve depicting WT1 (green, squares) and CDH18 (orange, triangles) co-expression in EPI cells [ n = 3]. h Western blot analysis showing the co-expression of WT1 and CDH18 in EPI cells during induction. i Immunocytochemistry of the EPDC markers POSTN (CF) and αSMA (SMC) and the EPI markers ZO1 and WT1 for EMT validation [fluorescence microscopy images were pseudo-colored using BZ-X analyzer software]. j Heatmap of epicardium, EMT, SMC and CF gene signatures [GSE165450]. k , l CDH18 expression analyzed by k qRT-PCR in induced derivatives (light orange) from d24 cells (left) and d12 cells (right), normalized to the respective EPI control cells (orange) [ n = 3; one-way ANOVA with Dunnett’s multiple comparison test, *** p < 0.0005] with l corresponding western blot analysis. [statistical analysis performed by Mann–Whitney-test, unless otherwise indicated; all error bars represent standard error of the mean (SEM); scale bars 100 µm]. Full size image We then checked CD22 expression levels during epicardial differentiation. We found CD22 expression increased but failed to detect CD22 protein in d24 EPI cells, suggesting protein translation insufficiency, and therefore invalidating CD22 as a suitable biomarker (Supplementary Fig. 4 ). Next, we investigated CDH18 levels and could verify its expression in EPI cells derived from different hiPSC lines (Fig. 2f and Supplementary Fig. 5a ). Both mRNA and protein (Fig. 2e–g ) levels increased exponentially over time of epicardial differentiation and CDH18 was co-expressed with WT1 in a highly correlated fashion (Fig. 2g, h and Supplementary Fig. 5 ), implying a potential role during epicardial specification. Epicardial cells have the potential to transit into EPDCs via EMT and subsequently differentiate into CF and SMC during development and also after re-activation upon cardiac injury. This functionality has also been well reported in vitro by treatment with basic fibroblast growth factor (bFGF) and TGF-β 29 , 31 , 34 . We initiated an EMT response (Supplementary Fig. 6a ) and observed changed morphology as well as loss of WT1 and ZO1 indicating loss of epicardial identity and demonstrating the functionality to transition into EPDCs (Fig. 2i ). The identity of induced CFs (iCFs) and induced SMCs (iSMCs) was validated by expression of POSTN and αSMA respectively (Fig. 2i ). We also sequenced the whole transcriptome of iCFs and iSMCs, verifying distinct cellular identities by PCA (Supplementary Fig. 6b ), as well as enrichment of respective gene signatures (Fig. 2j ). We then analyzed CDH18 expression in epicardial derivatives from both PE-like d12 and fetal-like d24 EPI cells. CDH18 was lost in iCF and iSMC (Fig. 2k, l and Supplementary Fig. 6c–f ), demonstrating that reduced levels of CDH18 correlated to a loss of EPI identity. Loss of CDH18 upon EMT was also confirmed in murine epicardial MEC1 cell line (Supplementary Fig. 6d ). Taken together, our results demonstrate that CDH18 is a specific active epicardial biomarker and suggest a causal role in specification and maintenance of epicardial identity. Loss of CDH18 triggers cell-fate specific EMT towards SMCs To explore the biological relevance of CDH18 , we next silenced CDH18 gene expression in d24 EPI cells (d24•si18) (Fig. 3a ). The downregulation of CDH18 (Fig. 3b–d ) led to visible morphological changes with a fraction of cells becoming elongated and losing the cobblestone-like morphology (Fig. 3d , yellow arrowheads) indicating a switch to mesenchymal cells 42 . Cells that retained epicardial morphology were in close contact, suggesting that cell–cell interactions might play a role in maintaining EPI identity. Similar results were obtained using a second siRNA (Supplementary Fig. 7a–c ). CDH18- silenced cells showed reduced growth (Fig. 3e ) suggesting that CDH18 downregulation in fetal-stage epicardium may confer either a proliferative disadvantage impairing epicardium homeostasis and expansion or define the triggering signal for EMT initiation. We observed a gain of SNAI1 expression (Fig. 3f ) upon CDH18 downregulation, thus indicating the initiation of EMT. Confirming this functional onset, we observed a strong downregulation of CDH1 (E-cadherin). However, upregulation of CDH2 (N-cadherin) (Fig. 3g ) was only modest. Fig. 3: Loss of CDH18 triggers cell-fate specific EMT towards smooth muscle cells. a Illustration of the experimental workflow. b and c Downregulation of b CDH18 [ n = 3; unpaired Students t -test, *** p = 0.0003] and c CDH18 protein expression after 8 days of CDH18 silencing in d24 EPI cells (d24•si18). d Phase contrast microscopy displaying d24•si18 cells with lost cobblestone morphology and elongated shape (yellow arrowheads). e Growth curve of d24 cells silenced for CDH18 (green) and control (scr) (gray) for 2-8 days [ n = 3; ** p = 0.0038, **** p < 0.0001]. f SNAI1 western blot of d24•si18 cells. g CDH1 and CDH2 expression in d24•si18 cells [control (scr); n = 3; * p = 0.0141, **** p < 0.0001]. h Western blot of CDH18 and SNAI1 after 8 days of CDH18 silencing in d24 EPI cells (d12•si18). i Microscopy analysis of d12•si18 showing changed morphology (yellow arrowheads) and loss of Ki67 (pink) expression [scale bar 50 µm]. j Expression of CDH1 and CDH2 in d12•si18 EPI cells compared to their control cells [ n = 3; * p = 0.0231, *** p = 0.0001]. k Boyden chamber assay evaluating cell migration capacity in d12•si18 cells [ n = 3; ** p = 0.002, *** p = 0.0004]. Right panel shows invaded cells stained by crystal violet. l Western blot analysis showing reduction of CDH18 and TCF21, increased LEF1 and β-catenin with decreased phosphorylated (p)-β-catenin. m Expression analysis of SMC markers ACTA2 and CNN1 in d12•si18 cells and iSMC [ n = 3; unpaired Students t -test: ns = not significant]. n Immunocytochemistry analysis of αSMA and ZO1 expression in d12•si18 cells [scale bar 50 µm]. [statistical analysis performed by two-way ANOVA with Sidak’s multiple comparison test, unless otherwise indicated; all error bars represent SEM; scale bars 100 µm, unless otherwise indicated]. Full size image A previous report showed that epicardial cells are more prone to undergo EMT during the early stages of development 43 , 44 . We, therefore, questioned, if fetal-like d24 EPI cells were more resistant to initiate EMT than d12 EPI cells, a time-point that portrays a more PE-like state. Accordingly, we silenced CDH18 in d12 EPI cells (d12•si18) finding similar morphofunctional changes as in d24 EPI cells (Fig. 3h–j and Supplementary Fig. 7d–f ). Loss of CDH18 was accompanied by gain of SNAI1 (Fig. 3h ) and d12•si18 cells showed changed morphology, reduced growth as well as decreased levels of Ki67 (Supplementary Fig. 7f, g and Fig. 3I ), thus indicating the onset of EMT. Notably, the switch from CDH1 to CDH2 was much stronger in d12 silenced cells, implying the acquisition of mesenchymal identity (Fig. 3j ). Furthermore, CDH18 -silenced cells showed enhanced migration (Fig. 3k ), endorsing that the downregulation of CDH18 leads to a loss of epicardial identity towards the EPDC-like state. To exclude that our observations were biased due to an early loss of cells we also confirmed that cells displayed reduced proliferation marker expression Ki67, upregulation of SNAI1 and loss of ZO1 3–4 days after silencing (Supplementary Fig. 7h–o ), further strengthening, that cells were undergoing EMT upon CDH18 downregulation. Several pathways are reported to be involved in the regulation of EMT 45 , 46 , 47 , 48 , 49 , 50 , 51 , 52 , 53 , 54 , with Wnt pathway regulation by Wt1 being linked to epicardial development and cell differentiation 6 , as well as cardiac regeneration 55 . Cadherins are known to exert their functions through phosphorylation-mediated protein stabilization of β-catenin in the Wnt signaling pathway 56 , 57 . Moreover, a previous report has shown the direct interaction of Cdh18 with β-catenin 58 . We, therefore, hypothesized that CDH18 might manage its impact on epicardial EMT via the regulation of β-catenin. Indeed, we observed that loss of CDH18 leads to increased levels of β-catenin with decreased phosphorylated (p)-β-catenin as well as an increase in LEF1 (Fig. 3l and Supplementary Fig. 7m, n ), indicating the activation of the Wnt pathway. Interestingly, we found reduced levels of TCF21 upon CDH18 -silencing (Fig. 3l ). TCF21 acts as an epicardial marker playing pivotal roles during EMT and EPDC-fate regulation between CF and SMC 59 , 60 , 61 . EPI cells showed increased TCF21 expression at the mRNA and protein levels (Supplementary Fig. 8a, b ). Reduced levels of TCF21 imply differentiation towards an SMC rather than a CF fate 7 , 60 . To confirm this hypothesis, we compared the expression levels of SMC genes in CDH18 -silenced EPI cells and iSMCs. We found that CDH18 -silenced cells recapitulated the iSMC expression levels of ACTA2 and CNN1 especially in d12 silenced cells and to a lesser degree also in d24-silenced cells (Fig. 3m and Supplementary Fig. 8c–e ). Canonical TGFβ-pathway activation seemed to play a noteworthier role for EMT onset in fetal-like d24 compared to PE-like d12 cells, as TGFβ inhibitor withdrawal does enhance the expression levels of ACTA2 in fetal-like d24 cells (Supplementary Fig. 8e ). To assess the role of non-canonical activation, we tested the impact of ROCK, inhibitor impairing RhoA-mediated pathway, on CDH18 downregulation in d12 cells. While no significant change could be observed in SNAI1 expression, ACTA2 expression was reduced by ROCK inhibition, albeit only when TGFβ signaling was not blocked (Supplementary Fig. 8f ). The acquisition of an SMC-like identity in CDH18 -silenced cells was further verified by the loss of ZO1 and the expression of α-SMA (Fig. 3n ). Next, we aimed to investigate the transcriptional contribution of CDH18 -downregulation to the SMC phenotype and the enhanced potential to derive SMC-like cells from d12 PE-like cells compared to d24 fetal-like cells. We, therefore, performed RNA-Seq and showed a transcriptome distribution of up to 99.4% combined variance by PCA (Fig. 4a ). PE-like CDH18 -silenced cells were allocated within the same arm as iSMCs, mapped closer to iSMC and mapped more DEGs (Fig. 4a, b and Supplementary Fig. 9a, b ). The downregulation of CDH18 suppressed epicardial identity both at the fetal-like stage and PE-like stage and activated the SMC gene program, while CF-related genes remained unaffected (Fig. 4c, d and Supplementary Fig. 9c ). Specifically, in d12 PE-like stage cells silencing of CDH18 influenced gene categories supporting SMC acquisition (Fig. 4c ) and showed a greater expression of important markers of EMT activation (Fig. 4f ). The altered expression of important cell cycle checkpoint regulators confirmed reduced proliferation upon CDH18 downregulation (Fig. 4g ). Finally, we analyzed gene signatures to identify activated signaling pathways, finding that TGF-β and Wnt pathways, which are known to contribute to the SMC phenotype, were affected and activated upon CDH18 downregulation (Fig. 4h, i ). Fig. 4: Transcriptional profiling of CDH18 -silenced EPI cells. a PCA of CDH18 -silenced d12 and d24 EPI cells as well as induced EPDCs compared to unspecific siRNA-treated control (scr) [d24•si18, n = 2; d12•si18, n = 2; iSMC, n = 1; CF; n = 1]. b Agglomerative hierarchical sample cluster: The Agnes dendrogram was plotted by applying Manhattan distances between samples. c Chord diagram representing the flow and the detailed relationship between d12•si18 DEGs (left semicircle perimeter) and their enriched GO biological processes (right semicircle perimeter). d Heatmap showing the expression analysis of the gene set enriched for epicardial, SMC and CF genes in iSMCs, CDH18 -downregulated and control (scr) cells. e Heatmap showing the expression analysis of the gene set enriched for epicardial, SMC and CF genes in CDH18 -downregulated and control (scr) cells under different culture conditions. f , g Expression levels for f EMT specific transcription factors (EMT-TFs) ZEB1 , ZEB2 , SNAI2 and TWIST1 and g CDKN1A , CDKN1B , CDKN2A , BUB1 , MKI67 , MCM6 , MCM4 and MCM2 in CDH18 -downregulated cells. h IPA analysis of CDH18 -silenced d12 EPI cells (si18) versus control (scr) cells showing the activation pattern of signature genes involved in major important developmental pathways. i Pathway enrichment analysis showing the proportional correlation of genes affected in important developmental pathways contributing to SMC differentiation. [analysis of Fig. 4 was based on RNA-Seq dataset GSE165450]. Full size image Altogether, we confirmed that CDH18 is an important regulator of the SMC phenotype. CDH18 downregulation led to a loss of epicardial cell identity under the retained inhibition of TGF-β, a situation that normally sustains epicardial hallmarks. The role of CDH18 is essential in epicardial fate and plasticity determination during human development, especially if the loss of CDH18 occurs at an early stage of epicardial specification. CDH18 is unable to block TGF-β-driven EMT towards SMC differentiation To investigate the potential of CDH18 expression to regulate epicardial maintenance and SMC differentiation, we evaluated the capacity of CDH18 to inhibit the initiation of EMT towards SMC differentiation. We ectopically overexpressed human CDH18 cDNA (Supplementary Fig. 10a ) in d12 EPI cells, which we assessed via mCherry expression, and subsequently induced EMT towards iSMC formation (Fig. 5a ). Unfortunately, EPI cells were difficult to transfect, consistent with a previous report on mouse epicardial-like cells 34 , 48 . Optimization of the transfection efficiency (Supplementary Table 1 ) led to ~25% of cells expressing mCherry at d3 post-transfection (Fig. 5b, c and Supplementary Fig. 10b ). To obtain a pure CDH18 -overexpressing population, we then sorted mCherry-expressing cells (mCherry+) (Supplementary Fig. 10c ). The gain of CDH18 expression was confirmed in unsorted and sorted settings (Fig. 5d ), and CDH18 protein gain was confirmed in mCherry+ cells (Fig. 5e ). Fig. 5: CDH18 is unable to block TGF-β-driven EMT towards SMC differentiation. a Schematic illustration depicting the experimental workflow. b Fluorescence microscopy pictures showing mCherry-expressing cells 3 days’ post transfection (dpt) of d12 EPI cells. c Flow cytometry analysis showing mCherry expression in EPI cells overexpressing CDH18 3 dpt [gating set using empty-transfected cells (Supplementary Fig. 5 b)]. d , e Validation of ectopic CDH18 cDNA overexpression demonstrating the increase of d CDH18 by qRT-PCR and e CDH18 protein by western blot analysis upon plasmid transfection and sorting. f Phase contrast pictures of un-/sorted CDH18 -overexpressing cells induced towards iSMC (+TGF-β + bFGF) or cultured in EPI maintaining conditions (+SB4321542). g , h Expression of g CDH18 and h ACTA2 in un-/sorted CDH18 -overexpressing cells and control cells under both EPI maintaining (+SB4321542) and iSMC inducing (+TGF-β + bFGF) culture conditions. i Expression analysis based on RNA-Seq dataset [GSE165450] of selected SMC marker genes. j Boyden chamber assay evaluating cell migration capacity in d12•si18 cells [ n = 3; paired Students t -test; * p = 0.0336, ** p < 0.0054]. [unsorted transfected cells (OE), empty-transfected cells (neg), sorted mCherry-expressing cells (mCherry+), sorted non-mCherry-expressing cells (mCherry-); statistical analysis performed by Repeated measure (RM) one-way ANOVA unless indicated otherwise: **** p < 0.0001, *** p < 0.0005, ** p < 0.005, * p < 0.05; all error bars represent SEM; scale bars 100 µm]. Full size image Next, we analyzed the effect of CDH18 overexpression during iSMC induction. We did not observe morphological differences between conditions (Fig. 5f ). Nevertheless, CDH18 -overexpressing cells showed fewer iSMCs (Fig. 5f ), hinting at a potential role of CDH18 in the maintenance of the active epicardium. All populations showed less CDH18 expression upon treatment with TGF-β and bFGF (Fig. 5g ), but transfected cells retained CDH18 expression even after undergoing EMT. In agreement with the morphological SMC phenotype (Fig. 5f ), all populations showed increased ACTA2 expression upon iSMC induction (Fig. 5h ). Notably, ACTA2 levels were significantly reduced in cells still expressing CDH18 . RNA-Seq analysis revealed the reduced expression of several SMC markers in CDH18-overexpressing SMCs (Fig. 5i ), highlighting an important role of CDH18 in preventing the fully SMC identity acquisition. During development, not only the differentiation of epicardial cells into SMC, but also their migration is important for successful coronary artery formation 62 , 63 , 64 . To investigate the functional properties of CDH18 -overexpressing cells, we next tested their invasion ability. We compared EPI cells to cells undergoing spontaneous EMT in absence of TGF-β inhibitor and CDH18-overexpressing SMCs (Fig. 5j ). In all conditions, CDH18 -overexpressing cells showed reduced invasion capacity. We, therefore, concluded, that although ectopic CDH18 overexpression cannot fully block the TGF-β-driven differentiation to SMC, it partially impairs the acquisition of SMC identity and their functional properties. CDH18 expression is under the control of GATA4 Finally, to elucidate the regulation of CDH18 in epicardiogenesis, we identified 23 transcription factors (TFs) out of 522 genes that are in correlation with CDH18 and enriched in the late-stage active epicardium (Fig. 6a, b ). Among these TF encoding genes, only GATA4 showed a high relative score to bind the CDH18 promoter (Fig. 6c and Supplementary Fig. 11a ). During EPI induction GATA4 is expressed in associative correlation to CDH18 (Supplementary Fig. 11b, c and Fig. 6d ). GATA4 is a well-characterized regulator at early cardiogenesis playing a pivotal role in PE formation. GATA4 -null mice embryos fail to form the epicardium due to impaired PE development, thus being an essential gene in epicardiogenesis 37 . We, therefore, silenced GATA4 using two siRNAs (siGata4#1 and siGata4#2) at different time points during epicardial differentiation: at d5 of epicardial induction marking the beginning of cell development towards PE-like stage and at d12 marking the specification towards fetal-like epicardium (Supplementary Fig. 11d ). GATA4 -silenced cells showed reduced cell survival (Fig. 6e , upper panel), but surviving clones did not alter GATA4 levels (Supplementary Fig. 11e ) and were able to reconstitute after 7 days (Fig. 6e , lower panel) showing no significant reduction of either GATA4 or CDH18 expression (Fig. 6e, f ). Thus, implying that upon efficient silencing of GATA4 , cells were not able to survive. We, therefore, concluded that GATA4 is necessary for cells at d5 to differentiate towards the PE-like state. Next, we downregulated GATA4 in d12 EPI cells and observed some morphological changes after siRNA administration over time (Fig. 6g , yellow arrowheads). The downregulation of GATA4 was accompanied by CDH18 downregulation in a similar fashion on both mRNA (Fig. 6h and Supplementary Fig. 11f ) and protein level (Fig. 6i ). Thus, these results promote GATA4 as a putative regulator of CDH18 expression, highlighting their relevance for the formation and specification of the active epicardium. Fig. 6: CDH18 expression is under the control of GATA4 . a Venn diagram depicting the interface of positively correlative genes to CDH18 (R ≥ 0.6) (light orange) and DEGs between early (Epi12) and late (Epi48) stage epicardium (target) (light brown) [GSE84085] revealed a total of 522 genes within their intersection. b Classification of 23 transcription factors (TFs) by overlapping the list of TFs (bright green) and genes identified in a (gray). c Binding site analysis showing the predicted binding sequence of GATA4 and relative binding score. d Co-expression network analysis of the qRT-PCR data of GATA4 (Supplementary Fig. 6c ) and CDH18 (Fig. 2f ) expression in EPI cells over time. e , f Silencing of GATA4 in d5 cells using two different siRNAs (siGata4#1 and siGata4#2) [ n = 3] show e a reduction in epicardial cell density after 3 and 7 days and f the expression levels of GATA4 and CDH18 after 7 days [ns = not significant]. g – i Silencing of GATA4 in d12 cells using two different siRNAs (siGata4#1 and siGata4#2) [ n = 3] shows g morphological changes (blue arrowheads) after 3 and 7 days and h reduction of GATA4 and CDH18 expression, as revealed by qRT-PCR [*** p < 0.0005, **** p < 0.0001], as well as i loss of GATA4 and CDH18 protein expression, as shown by western blot analysis. [statistical analysis performed by one-way ANOVA with Dunnett’s multiple comparison test, unless indicated otherwise; all error bars represent SEM; scale bars 100 µm]. Full size image Discussion The epicardium is essential for cardiogenesis and its re-activation upon injury during adulthood is indispensable for cardiac regeneration. Exploiting epicardial targeting to model cell therapy-based approaches for cardiac repair and regeneration requires detailed knowledge of the mechanisms that regulate the active epicardium. Here, we recapitulated the human epicardiogenesis using hiPSCs and defined a surfaceome of human active epicardium that is later inactivated in the quiescent (adult) epicardium. We define CDH18 as an epicardial biomarker that is exclusively expressed in epicardial cells compared to their derivatives and other cardiac cell types. The epicardium specificity of CDH18 expression can be used as a specific, rather than associated, active epicardial biomarker to both define biological identity and modulate EMT-specific processes towards SMC. Previous identification of epicardial cells was based on the expression of several transcription factors such as WT1 , TBX18 and TCF21 . However, their expressions are not restrictive to the epicardium. Our findings provide a molecular alternative for the identification of epicardial cells by tracing CDH18 expression, a biomarker of the active fetal epicardium. Cadherin 18 (CDH18), formerly named cadherin 14, is a type II classical cadherin predominantly expressed in the central nervous system, contributing significantly to axonal development and maintenance 38 , 39 , 40 . This study succeeded in documenting murine CDH18 expression in fetal heart explant culture, known to be derived from the epicardium. Given the thin-shaped nature of the epicardium, its density represents a notable minority in the context of the human heart. We did not detect protein expression in whole trypsinised whole-heart lysates, as mostly cardiomyocytes were isolated. In addition, our study suggests that CDH18 expression shapes the fetal active epicardium transcriptome, but is downregulated in quiescent adult tissue 4 , 65 , thus explaining why CDH18 remained undetected in the heart 40 . Investigating the expression of CDH18 in human embryonic epicardial cells in vivo is technically difficult due to the lack of technical approaches to trace epicardial development as well as the limited availability of human fetal cells. The use of hiPSC overcomes this issue by enabling recapitulation of epicardiogenesis in vitro. Whereas our study limits our main conclusions to in vitro differentiated epicardial cells and their derivatives, we believe that our key discoveries are physiologically relevant, since RNA-Seq based meta-analysis shows that CDH18 is expressed in the human heart during the third month of gestation, a time point of active development of the cardiovascular system in the human embryo, being in line with the hiPSC-based model shown in this study. Moreover, we documented murine embryonic CDH18 expression in epicardial explant cultures from E14. Recent reports demonstrate an essential role for WT1 in epicardial cell specification and maturation from the PE, suggesting that junction remodeling is crucial for this transition 66 , 67 . We showed that CDH18 is expressed in correlation to WT1 during epicardial development, leading us to hypothesize that CDH18 might play a critical role in the specification and maintenance of epicardial cell identity, similar to WT1 . During human heart formation as well as during cardiac repair epicardial cells also contribute to coronary vessel formation by providing necessary vascular SMC and pericytes, which invade the myocardium. Notably, we provide compelling evidence to indicate that CDH18 is critical for the modulation and establishment of epicardial-derived SMCs. We proved that CDH18 downregulation modulates the expression of EMT markers and impacts epicardial identity. Recent studies characterized a CDH18 tumor-suppressor role in glioma carcinogenesis and progression, demonstrating an active role in invasion and cell migration 38 , important processes that also relate to cardiac regeneration. The ectopic overexpression of CDH18 was reported in relation to decreased migration capacity in accordance with our study. Most importantly, our study gained mechanistic insights into the biological regulation during the initiation of EMT in epicardial cells. The loss of CDH18 decreased proliferation, CDH1 expression and β-catenin. Interaction with and regulation of stabilization of β-catenin via cadherins is critical for the cellular organization, polarity and development 56 , 57 . Indeed, a previous study showed that Cdh14 interacts with and stabilizes β-catenin, similar to Cdh1 58 . Another study reported, that type II Cdh14 (Cdh18) along with type I E-cadherin (Cdh1) and N-cadherin (Cdh2) can bind to G12 family proteins 68 , which are known to regulate β-catenin translocation from the cell surface, a function associated with Rho-dependent cytoskeletal rearrangement. In our study, we demonstrated that the loss of CDH18 increased β-catenin levels, indicating Wnt activation and the promotion of EMT. The activation of Wnt signaling via β-catenin has long been deemed important for epicardial EMT and SMC differentiation which known to be regulated by Wt1 6 , 7 , 29 , 53 , 69 , 70 . Although controversial 52 , our study supports the necessity of β-catenin for SMC differentiation in vitro 70 . TCF21 is another key regulator of epicardial plasticity and EMT, promoting CF over SMC fate 60 . It is possible that rather than the activation of specific signaling pathways, the expression of TCF21 is of far greater importance for cell-fate decisions. Thus, investigating Wnt and TGF-β signaling pathways and their influence on TCF21 could shed more light on this area, which will also allow understanding cell-fate decisions during cardiac regeneration, in particular those physiological stimuli that presumably force cells to assume a CF rather than pericyte phenotype. In addition, TCF21 is known to be induced by retinoic acid (RA), thereby delaying SMC differentiation 60 , 71 , 72 . We showed that the expression of ALDH1A2 is not upregulated until d24 and is low in d12 EPI cells. Thus, our observation of d24-silenced cells expressing lower levels of the SMC marker ACTA2 could be linked to the higher presence of RA and TCF21 in fetal-like cells compared to PE-like cells. During cardiac repair, the expression of ALDH1A2 is re-activated, which might explain the preferred CF fate of epicardial cells that undergo EMT upon re-activation. Here, we showed loss of TCF21 upon loss of CDH18, suggesting a role for β-catenin-mediated Wnt activation in SMC differentiation. Furthermore, this observation occurred under the inhibition of the TGF-β co-receptor ALK5, which is known to impair EMT 73 . Our findings provide new insights into the regulation of the EMT process as a crucial step mediating epicardial-driven regenerative responses. The exact mechanisms governing epicardial EMT and subsequent cell-fate decision of EPDCs in active epicardium are not well understood. Most experiments have used chick, avian or mouse models, which can differ considerably from humans. We observed the activation of downstream TGF-β signal targets, confirmed by RNA-Seq, upon CDH18 silencing albeit treatment with SB431542. Our study endorses the link between TGF-β signaling and SMC phenotype 29 , 31 , 34 , 70 , as we observed SMC differentiation and TGF-β activation upon lower CDH18 expression. Although CDH18 overexpression could not completely block TGF-β-driven SMC differentiation, we found a partial prevention in the acquisition of SMC identity, indicated by lower levels of ACTA2 . Altogether, our findings indicate a causal role of CDH18 in TGF-β pathway during the acquisition of SMC identity. CDH18 is a cell surface protein linked to G12 family proteins 68 , and RhoA pathway-mediated non-canonical TGF-β pathway, which has been linked to EMT and cell invasion 45 , 48 , 49 , 50 , 74 , 75 , might be activated upon the silencing. Interestingly, one study showed that inhibition of the RhoA pathway does not influence canonical TGF-β activation but still impairs SMC formation 76 . In our study, however, for those experiments that kept ALK5 inhibition conditions, ROCK inhibitor did not alter SMC marker expression upon CDH18 silencing. Only when ALK5 inhibitor was absent, did inhibition of ROCK lead to a decrease in ACTA2 . Recently, a new study linked a TGFβ-independent causal role of the extracellular matrix (ECM) protein agrin to epicardial EMT mediated by dystroglycan 17 . This report also showed that their findings occur via Wnt signaling activation, a phenotype that could have common denominators with our study. Future studies are needed to shed more light into the different behavior of cells during human epicardial development. Such studies will also help deepen our understanding of epicardium re-activation upon cardiac injury and advance the field of cardiac regenerative medicine. Finally, we demonstrated the correlative relationship between CDH18 and GATA4 expression, proposing GATA4 as a putative regulator of CDH18 . This effect likely occurs by static binding to the CDH18 promoter near the transcriptional starting site, reinforcing GATA4 as an important activator of the embryonic epicardium program. CDH18 is a highly conserved protein in higher mammals and species that evolutionary developed a double closed circulatory system, maintaining an amino acid alignment of over 95% identity and 100% query sequence coverage relative to humans. The conserved sequence suggests that CDH18 has remained relatively unchanged far back up the phylogenetic tree in species developing a higher compartmentalized system to create a physical separation of oxygenated and deoxygenated blood (Supplementary Fig. 12 ). Higher demand for SMC might explain CDH18 conservation in development (Supplementary Fig. 13 ). Altogether, these findings indicate a specific role for CDH18 in epicardial regeneration and development, both processes that are under the control of evolutionarily conserved pathways In conclusion, our work defines a biological function of CDH18 in the epicardial context, enabling roads for the manipulation and therapeutic gene modulation of the active epicardium for cardiac repair and regeneration. Methods Cell lines and culture conditions The hiPSC lines 201B7 and 409B2 (retroviral reprogramming by Yamanaka factors) 27 as well as a MYH6-eGFP reporter cell line (MYH6-EIP4) 41 were cultured in ReproCell ES media containing 4 ng/ml bFGF on irradiated MEF feeder cells. For removal of the feeder layer, the cells were treated with CTK. Feeder-free 201B7 were cultured in complete StemFit® AK02N media on iMatrix-511 (Matrixome) coated dishes. All cell lines used were female. MEC1 Mouse Embryonic Epicardial Cell Line SCC187 (Merck Millipore) was cultured and maintained in DMEM supplemented with 10% FBS. Mycoplasma testing was performed on a regular basis to exclude contamination. Epicardial induction For epicardial differentiation, single cell suspension of hiPSC was generated using Accutase and subsequently cells were plated onto low-attachment HEMA-coated plates (6000-8000/96-well) to form embryoid bodies (EBs). Differentiation media was composed of complete StemPro®-34 media supplemented with 50 μg/ml ascorbic acid, 2 mM L-glutamine, 0.4 μM monothioglycerol and 150 mg/ml transferrin. To initiate differentiation, 0.5% Matrigel, 10 μM Y-27632 and 2 ng/ml human recombinant (hr)BMP4 were added to the differentiation media. After 24 h (h) EBs were cultured in differentiation media with a final concentration of 10 ng/ml hrBMP4, 2 ng/ml Activin A and 5 ng/ml hrbFGF. After 84 h, EBs were collected, dissociated using Accutase, and plated onto 0.1% gelatin-coated dishes (0.3 × 105 cells/cm 2 ) in differentiation media containing 3 mM CHIR99021, 30 ng/ml hrBMP4, 5 ng/ml hrVEGF and 10 µM SB431542. From d7 onwards, induced hiPSC-derived epicardial-like (EPI) cells were maintained in maintenance media (DMEM containing 10% FBS and 10 µM SB431542). For cell passaging, EPI cells were detached by Accutase treatment for 5 min and replated as described above. Cell transfection For the silencing experiments, Silencer ® Select siRNAs (ThermoFisher) were diluted to a 10 µmol stock solution: siCdh18#1 (s2816), siCdh18#2 (s2817), siGata4#1 (s535120) and siGata4#2 (s535121). To silence cells, 10 µl siRNA stock was diluted in 500 µl Opti-MEM, and 10 µl RNAiMax was added (silencing solution) followed by 10 min incubation at room temperature. The solution was added dropwise into one well of a six-well plate containing 2 ml maintenance media and 0.2–0.3 × 10 5 cells seeded a day prior. The media was exchanged 24 h later, and the cells were cultured as usual thereafter. For the immunocytochemistry experiments, 260 µl silencing solution was added dropwise into one well of a 12-well plate containing 1 ml maintenance media and 0.05 × 10 5 cells seeded a day prior to silencing. For CDH18 cDNA overexpression experiments, the cells were seeded to be 60–70% confluent at least one day prior to the transfection. For optimization of transfection protocol, transfection reagents and ratios as well as the plasmid amount were used as indicated in Supplementary Table 1 . Unless otherwise stated, 6 µg plasmid per one well of a six-well plate or 34 µg plasmid per 10 cm dish were transfected using FuGENE ® 6 at an agent-to-DNA ratio of 2:1. Amounts of reagents were calculated using FuGENE ® HD Protocol Database web tool. Cell sorting was performed by FACS via assessment of mCherry expression (Texas-Red positive cells). Induction of EMT For the induction of EMT, EPI cells were passaged one day prior, at a density of 0.1–0.2 × 10 5 per well of a six-well plate. For differentiation towards CF, the cells were treated with 10 ng/ml bFGF in 10% FBS DMEM for at least 8 days. For differentiation towards SMC cells were treated with 5 ng/ml TGF-β for 4 days followed by 10 ng/ml bFGF for another 4 days in 10% FBS DMEM. Isolation of mouse fetal hearts and whole-heart lysate Embryos at E14 were collected into pre-warmed PBS and any extraembryonic tissue was removed using tweezers. Embryos were decapitated, the chest cavity was opened by ripping the anterior side apart and fetal hearts were isolated mechanically into a separate dish with pre-warmed PBS. Whole heart protein lysate was prepared by firstly cutting hearts into small pieces and treatment with trypsin for 20 min at 37 °C. The cell suspension was re-suspended, treated for an additional 20 min at 37 °C and supernatant was collected after centrifugation at 300 × g for 5 min. Ex vivo explant assay Collected hearts at E14 were cut into 4–8 pieces and placed onto 0.1% gelatin-coated dished in DMEM (low glucose) supplemented with 15% FBS. Outgrowth cultures formed as early as 24 h post placing. After 72 h heart pieces were manually removed with tweezers. Explant cells were passage by trypsinization the following day and cultured in 10% FBS-DMEM (low glucose) supplemented with 10 µM SB431542. Cells were collected 1–3 days after passage. EMT was induced as described above in 10% FBS-DMEM (low glucose). Quantitative RT-PCR For RNA extraction, live cells were collected as du- or triplicate in QIAzol lysis reagent, and total RNA isolation was performed using the RNAeasy micro kit (Qiagen) according to the manufacturer’s manual. The generation of cDNA was performed by using the ReverTra Ace system (Toyobo BIOTECH) according to the manufacturer’s manual. QRT-PCR was performed in du- or triplicate either using TaqMan gene expression assays ( WT1 Hs01103750_g1; TBX18 Hs01385457_m1; ALDH1A2 Hs00108254_m1; TCF21 Hs00162646_m1) in TaqMan master mix solution or SYBR green (ThermoFisher) with primers against identified candidates (Supplementary Table 2 ) in SYBR green master mix solution according to the manufacturers’ respective manuals. Data acquisition was carried out by StepOne Plus (AppliedBiosystems). Analysis was performed using Microsoft Excel and GraphPad Prism. Values without readout were not included. The distribution of values was verified by a box plot depiction to determine the appropriate statistical analysis. For direct comparison of dataset normally distributed data were analyzed by students t-test, as indicated in figure legends, otherwise, the Mann–Whitney test was used. Comparisons of more than two datasets were carried out via ANOVA, according to experimental conditions, as indicated in the figure legends. All error bars represent the standard error of the mean (SEM). For co-expression analysis, qRT-PCR data was used to generate a regression line by GraphPad Prism v7 and higher. Immunocytochemistry Cells were fixed by 4% PFA treatment for 15 min and stored in PBS at 4 °C. Blocking was performed for 30–45 min in blocking buffer: 1% BSA and 0.5% Triton X for nuclear staining or 0.1% Tween 20 for surface or intracellular staining and 0.1 M glycine in PBS. Following three PBS washes, primary antibody was added in blocking buffer without glycine and incubated overnight at 4 °C. The primary antibodies used are as follows: Anti-WT1 1:200 (Abcam; ab89901), Anti-ZO1 1:100–300 (Invitrogen; ZO1-1A12), Anti-POSTN 1:200 (ThermoFisher; PA5-98301), Anti-α-SMA 1:200 (Abcam; ab7817) and Anti-Ki67 1:200 (BioLegend; 350502). The next day, the cells were washed thrice with PBS, secondary antibody (Invitrogen, 1:1000–2000) was added in 1% BSA-PBS and incubated for 2 h at room temperature. The secondary antibodies were used are as follows: goat-anti-mouse-Alexa488 (A11001), goat-anti-rabbit-Alexa546 (A11010), goat-anti-mouse-Alexa546 (A11030), goat-anti-mouse-Alexa594 (A11032), goat-anti-rabbit-Alexa647 (A21245), goat-anti-mouse-Alexa647 (A21236). Following three washes with PBS, Hoechst or DAPI (5 ng/ml, 1:10,000) was added for nuclear counterstaining. Flow cytometry analysis/FACS The cell was dissociated into single-cell suspension by Accutase treatment, washed twice and fixed using 4% PFA for 15 min. For nuclear staining, the cells were permeabilized by 0.1% Triton X solution. Conjugated antibodies were diluted 1:50 in FACS buffer (5% FBS-PBS) and incubated at room temperature for 30–45 min. The conjugated antibodies used are as follows: WT1-Alexa488 (Abcam; ab202635), rb-IgG-Alexa-488 (Abcam; ab199091), CDH18-FITC (Biorbyt Ltd; orb7854), CD22-APC (BD; 562860) CD31-APC (BioLegend; 303116) and CD144-FITC (BD; 560874). For sorting, a single cell suspension was generated as described above, and counterstaining with 1:1,000 DAPI (5 ng/ml) was performed for dead cell exclusion. Sorting gates were set as described. Analysis was performed using BD FACSDiva v6 or v8 and FlowJo v10. The cell population was identified by FSC/SSC gating and doublets discrimination was performed. Negative gates were established using unstained samples or isotype controls and negative control samples. Positive gates were set to contain no negative signal. Immunoblotting Cells were detached using Accutase and lysed in Mammalian Protein Extraction Reagent (M-PER) (Thermo Scientific; 78501) buffer. The amount of protein was determined by the Bradford assay using BSA as the standard. The primary antibodies (dilution 1:1000) used are as follows: Anti-CDH18 antibody (Proteintech 13091-1-AP), Anti-GATA4 (CST; 369665S (D3A3M)), Anti-SNAI1 (Abcam; ab63371), Anti-α-SMA (Abcam; ab11952), Anti-β-actin (Sigma; A5441), Anti-WT1 (Abcam; ab89901), Anti-CD22 (Abcam; ab207727), Anti-TCF21 (Abcam; ab32981), Anti-β-catenin (CST; 8814), Anti-phosporylated-β-catenin (S33/S37/T41) (CST; 9561), Anti-TNNI1 (Abcam; ab203515). The secondary antibodies (dilution 1:5000) used were as follows: goat anti-rabbit-horseradish peroxidase (HRP) (Abcam; ab97051), rabbit anti-mouse-HRP (Abcam; ab97046). Blots derive from the same experiment and were processed in parallel. Unprocessed blots are shown in Supplementary Figs. 14 – 18 . Proliferation assay A time-course curve of parental (scr) and siRNA-expressing cells was generated by seeding 4.5 × 10 3 cells into 6-cm bottom-well diameter dishes. After 24 h, fresh medium was replaced and cells were fixed to be referred as a standard for relative growth (day 0). Relative growth was assessed every 48 h (2 days), fixing in 2% PFA and stained with 1% crystal violet (Sigma; C6158-50G). After extensive cell washing, crystal violet was solubilized in 20% acetic acid (Sigma) and quantified absorbance at 595 nm as a relative assessment of cell number (PerkinElmer; EnVision 2104 Multilabel Reader). Relative values in Fig. 4 represent cell growth of daughter the indicated medium. Percentage zero (0%) refers to initial growth at day 0. Invasion assay Cell were silenced as described above and collected 4 days later. 2.5 × 10 4 cells were seeded onto a Corning ® Biocoat ® Matrigel® Invasion chamber and incubated for 24 or 48 h. The removal of non-invasive cells was carried out as stated in the Corning ® Biocoat ® Matrigel ® Invasion manual, and the cells were fixed for 2 min in methanol followed by staining using crystal violet overnight and finally washed with distilled water. Pictures were taken using BZ-X710 (Keyence) and cell counting was performed with ImageJ. Cell counting was performed blinded using ImageJ v1.52. Images were converted to 8-bit greyscale format, and the same threshold was set for all images to highlight cell structures. Potential cell clusters were separated by the watershed function, and only structures 0.01-0.1 pixel 2 in size were counted. Graphs were generated using GraphPad Prism v8. Bioinformatics Retrospective analysis of CDH18 gene expression was done using GSE84085 (GEO) for epicardium development and GSE7307 for cardiovascular-derived tissues (subset of aorta, coronary artery, heart, heart atrium and heart ventricle) [GEO and R2; genomic analysis and visualization platform ( )]. Promoter sequences were identified by Ensembl and then screened for transcription factor binding sites with PROMO-ALGGEN (based on TRANSFAC v8.3) and JASPAR 2016. Image acquisition, processing and analysis Microscopy images were taken by BZ-X710 and processed by BZ-X Analyzer (Keyence); pseudo-coloring was used as indicated in figure legends. For phase-contrast and fluorescence pictures, a representative section was chosen, cropped and magnified. Western blot data were recorded by LAS4000 (Cytiva). General image analysis was performed using Microsoft powerpoint as well as ImageJ v1.52. RNA-sequencing Data normalization was carried out using NOISeq. The final processed data and raw fastq files generated de novo in this study were submitted to Gene Expression Omnibus (GEO) under the accession code GSE165450. The raw data were analyzed by using RStudio for further analysis on gene expression. Gene expression data for reading, exploring and pre-processing were conducted using the Bioconductor package NOISeq pipeline to perform data and differential expression analysis for RNA-Seq. A hierarchical derived cluster dendrogram was generated by using hclust, stats package, and agnes in the cluster package. Distances were assessed by using Manhattan city-block distance algorithm. K -means were calculated with the kmeans function. Distance and correlation matrices were computed and plotted by using get_dist , fviz_dist included in the factoextra package. The fviz_cluster function was used to compute cluster scatter plots. Heatmaps were depicted to cluster the expression data for DEGs using R script. Statistical analysis and visualization of the functional profiles of genes as well as gene clusters for GO terms were conducted using the DOSE and clusterProfiler . Ingenuity pathway analysis (IPA) was used to develop an upstream pathway analysis as well as pathway activity patterns. Chord diagrams were generated by using GOplot. Material availability The CDH18 overexpression plasmid used in this study is available from VectorBuilder [VB200615-1011cyh]. Other materials are available from the corresponding author upon reasonable request, but we may require payment and/or a Materials Transfer Agreement. Animal experiments All experiments involving animals were approved by the Kyoto University Animal Experimentation Committee and carried out in accordance to the Guidelines for Animal Experiments of Kyoto University and Guide to the Care and Use of Laboratory Animals by the Institute of Animal Resources. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability All data supporting this study are available within this publication and its supplementary information or can be provided upon request. The RNA-Seq dataset generated in this study is deposited at Gene Expression Omnibus (GEO) with the accession code GSE165450.

Heart disease remains the leading cause of death in the world. One reason is that unlike other tissues, such as bone and skin, the heart has remarkably poor regenerative capability after an injury such as a heart attack. Scientists have therefore searched for heart cells that have regenerative properties. A new study by the Yoshinori Yoshida laboratory reports the use of iPS cells to produce one such cell type, epicardial cells. The epicardium describes the most outer layer of the heart. As the heart forms, epicardial cells migrate and differentiate into different heart cells. However, after birth, the epicardium takes a quiescent state and only reactivates following injury to produce fibrosis and scarring. While this response is not ideal and fails to fully recover heart function, in some species, such as zebrafish, epicardium reactivation is a critical step for heart regeneration. In general, the epicardium of a growing fetal heart behaves differently from that of a fully formed adult heart. "The embryonic epicardium is active. Its cells are proliferative and undergo an epithelial-to-mesenchymal transition that results in various cardiac cell types. This property is not seen in adult heart," said Julia Junghof, who is the first author of the study and is expected to earn her Ph.D. To study these differences, she and colleagues differentiated human iPS cells into epicardial cells. They found the protein cadherin 18 (CDH18) was expressed in only the fetal type. "We normally rely on the expression of genes to identify epicardial cells but do not have reliable surface markers. Cadherins are best known for cell-to-cell adhesion, but they also are involved in cell-fate decisions," said CiRA Associate Professor Yoshinori Yoshida, who led the study. Artificially expressing genes in iPS cells is technically difficult and expensive. Using surface markers reduces the labor and cost of the experiments and also results in safer cells that can be used later for patient care. The study shows cadherin 18 (CDH18) is such a surface marker. CDH18 is better known for its role in axon development and maintenance in the nervous system, but the study showed its usefulness for identifying iPS cells that successfully differentiated into epicardial cells. Notably, while CDH18 was expressed in fetal epicardial cells, it was not expressed in adult epicardial cells, which could reflect the different regenerative properties of the heart. The loss of CDH18 expression resulted in epicardial cells losing their ability to multiply and caused an epicardial-to-mesenchymal transition that resulted in smooth muscle cells for the heart, providing new insights on how the epicardium can be activated following injury to promote heart regeneration. Thus, measuring the expression of CDH18 on differentiating iPS cells may provide an effective means for producing cells with heart regeneration ability in the lab. CiRA scientist Dr. Antonio Lucena-Cacace, a senior author of the study, added that the findings also give clues on how heart regeneration can be turned on and off in the body. "Our findings show how regulation of the epicardial-to-mesenchymal transition affects the cell fate and regenerative responses of epicardial cells," he said.

10.1038/s41536-022-00207-w

Physics

Spin-to-charge conversion achieves 95% overall qubit readout fidelity

Qi Zhang et al, High-fidelity single-shot readout of single electron spin in diamond with spin-to-charge conversion, Nature Communications (2021). DOI: 10.1038/s41467-021-21781-5 Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-021-21781-5

https://phys.org/news/2021-04-spin-to-charge-conversion-qubit-readout-fidelity.html

Abstract High fidelity single-shot readout of qubits is a crucial component for fault-tolerant quantum computing and scalable quantum networks. In recent years, the nitrogen-vacancy (NV) center in diamond has risen as a leading platform for the above applications. The current single-shot readout of the NV electron spin relies on resonance fluorescence method at cryogenic temperature. However, the spin-flip process interrupts the optical cycling transition, therefore, limits the readout fidelity. Here, we introduce a spin-to-charge conversion method assisted by near-infrared (NIR) light to suppress the spin-flip error. This method leverages high spin-selectivity of cryogenic resonance excitation and flexibility of photoionization. We achieve an overall fidelity > 95% for the single-shot readout of an NV center electron spin in the presence of high strain and fast spin-flip process. With further improvements, this technique has the potential to achieve spin readout fidelity exceeding the fault-tolerant threshold, and may also find applications on integrated optoelectronic devices. Introduction Resonance fluorescence method has become a commonly used method to achieve the single-shot readout of various solid-state spins such as quantum dot 1 , 2 , rare-earth ions in crystals 3 , 4 , silicon-vacancy center 5 , 6 , and nitrogen-vacancy (NV) center 7 in diamond. Under spin-selective excitation of optical cycling transition, the spin state is inferred according to collected spin-dependent fluorescence photon counts. However, the accompanying spin non-conservation processes usually limit the optical readout window for photon collection and induce the spin state flip error. This effect has become a significant obstacle for achieving high-fidelity single-shot readout, in particular, to exceed the fault-tolerant threshold 8 , 9 , 10 , 11 , 12 . A powerful method to suppress this effect is to explore optical structures for the emitters. The microstructure, such as a solid-state immersion lens, is widely used to enhance the fluorescence collection efficiency 7 , 13 , 14 , 15 , 16 . High-quality nano-cavities strongly coupled to these quantum emitters could even enhance the photon emission rate by orders of magnitude 3 , 4 , 5 , 6 . Despite these significant achievements, the practical application of such a high-quality cavity remains technically challenging. Extensive engineering works are required to obtain the high-quality cavity, place the emitter into the optimal cavity position, and tune the frequency on-demand. Besides, the fabrication process introduces unwanted strain and surface defects 17 , which may degrade the spin and optical properties 7 . Here, we demonstrate a new method to achieve a single-shot readout of NV center electron spin by combing a spin-selective photoionization process. The spin state is on-demand converted into charge state before the spin-flip relaxation becomes significant (Fig. 1 a, b). Then the charge state is measured with near unity fidelity thanks to their stability under optical illumination. The essence of this approach is to enhance the ratio of ionization rate (Γ i on ) to the spin-flip rate (Γ flip ). Fig. 1: Single-shot readout scheme based on SCC. a Energy levels used to achieve SCC. Qubit is encoded in the ground state \(\left|0\right\rangle\) and \(\left|1\right\rangle\) , and the \(\left|-1\right\rangle =\left|{\rm{AUX}}\right\rangle\) state acts as the auxiliary level. The magnetic field of 585 G aligned to NV axis lifts the degeneracy between \(\left|-1\right\rangle\) and \(\left|+1\right\rangle\) . Note that the magnitude and direction of the magnetic field used here is not special. The coherent manipulation between \(\left|0\right\rangle\) and \(\left|\!\pm \!1\right\rangle\) can be realized by resonant microwave, labeled by blue arrows. E y ( E 1,2 ) corresponds to the optical transition of the m S = 0 ( m S = ±1) state. The counts rate is proportional to the excited state emission rate and the fluorescence photon collection efficiency. The key part of SCC is to ionize (dark red arrow) the excited states of m S = 0 before it substantially relaxes to the ground \(\left|\!\pm \!1\right\rangle\) states through the spin-flip relaxation process (gray dashed arrow). CB is the conduction band of diamond and VB the valance band. Γ ion denotes the ionize rate, and Γ flip denotes the spin-flip rate from the E y excited state to the ground \(\left|\!\pm \!1\right\rangle\) states. A more detailed model is in Supplementary information (SI) . b A schematic diagram of SCC readout. Under the illumination of 637 nm laser, NV − keeps fluorescing stably for a long time, while NV 0 is not excited. c The excitation spectrum of the NV center used here at cryogenic temperature of 8 K. Frequency is given relative to 470.4675 THz (637.2225 nm). The non-axial strain ( δ ) induces a splitting of 2 δ = 11.8 GHz between E y and E x transitions 45 . d Spin-flip process induces the photoluminescence (PL) decay under E y excitation (5.7 nW, saturation power ~13 nW) with NV initially prepared in \(\left|0\right\rangle\) . At the final equilibrium of PL decay curve, the NV spin is pumped into \(\left|\!\pm \!1\right\rangle\) . The solid line is the simulation according to the model described in SI , with the best-fitted spin-flip rate Γ flip = 0.75 ± 0.02 MHz. Inset: PL decay for NV initialized to \(\left|\!\pm \!1\right\rangle\) under E 1,2 excitation (4.2 nW, saturation power ~ 34 nW). From the PL decay curves, the spin initialization fidelity is estimated to be 99.7 ± 0.1% for \(\left|\!\pm \!1\right\rangle\) subspace and 99.8 ± 0.1% for \(\left|0\right\rangle\) (SI). Full size image Results The experiments are performed on a bulk NV center inside a solid immersion lens at a cryogenic temperature of 8 K. The measurement scheme utilizes the cycling transition E y that connects excited and ground states with spin projection m S = 0 (Fig. 1 a), and the E 1,2 transition connecting states with spin projection m S = ±1. The corresponding optical transitions is shown in Fig. 1 c. The fabrication of the solid immersion lens introduced non-axial strain δ = 5.9 GHz to the NV center used. Therefore, a spin-flip rate Γ flip of 0.75 ± 0.02 MHz is observed (Fig. 1 d), much faster than previously reported 0.2 MHz with low strains 7 . Under selective excitation of E y , spin state \(\left|0\right\rangle\) could be pumped to the excited state, and be further ionized to charge state NV 0 under another NIR laser excitation (1064 nm, Fig. 1 a). In contrast, \(\left|\!\pm \!1\right\rangle\) will not be excited and stay at charge state NV − . Such a deterministic SCC differs from previous work using non-resonant excitation to enhance the readout efficiency of NV center 18 , 19 , 20 , 21 , 22 , 23 . To verify the photoionization process, we first characterize the charge state readout. Under simultaneous excitation of E y and E 1,2 transitions, NV − emits photons regardless of the spin state, while leaving NV 0 in the unexcited dark state. The charge state can thus be determined from the detected photon number during the integration window. We evaluate the charge readout fidelity by measuring the correlation between two consecutive readouts (Fig. 2 a). The correlation results with an integration window of 500 μs is shown in Fig. 2 b and the statistical distribution of the photon number is shown in Fig. 2 c. As expected, the NV − state is distinguishable from the NV 0 state according to the photon counts (Fig. 2 c). More importantly, a strong positive correlation is observed, except for six anti-correlation cases. And all these anti-correlation cases (circles in Fig. 2 b) come from initial NV − transforming to NV 0 . This indicates a unity readout fidelity for NV 0 state and 99.92 ± 0.03% readout fidelity for NV − state. To understand the tiny readout imperfection for NV − state, we measure its lifetime under the continuous optical readout sequence. As shown in Fig. 2 d, one observes a lifetime of 400.7 ± 9.7 ms for NV − state, which causes a charge conversion error of 0.12% during the charge state readout, comparable to the observed imperfection. The average non-demolition charge readout fidelity is 99.96 ± 0.02%. Fig. 2: Non-demolition readout of charge state and ionization rate of NIR light. a Pulse sequence for the charge readout fidelity evaluation. A 3 μs pulse of 532 nm laser reset the population of NV − to be 78%, according the results in ( b ) and ( c ). Both of the two charge readings use an integration window of 500 μs. b The correlation between the two consecutive charge readouts. A total of 10,000 tests are performed. Among them there are 7771 cases of (NV − , NV − ), 6 cases of (NV − , NV 0 ), 2223 cases of (NV 0 , NV 0 ), and 0 cases of (NV 0 , NV − ). The orange circles mark the cases with anti-correlation. The charge state is judged to be NV 0 when the collected photon number ≤11. The dashed gray lines mark the threshold (the same for c ). c The photon number distribution of NV 0 and NV − . The charge readout fidelity F charge = 99.96%. d The lifetime of the charge state of NV − under E y + E 1,2 (6 + 5 nW) illumination, 400.7 ± 9.7 ms. e Pulse sequence for measuring the ionization rate under simultaneous illumination of E y and NIR light. f The ionization curves of NV − at different powers of 1064 nm. The solid lines are simulations based on different ionization rates. g The dependence of the NIR ionization rate on its power. The solid line is a linear fit to the data points, with a coefficient of 67.0 ± 6.7 kHz/mW. The three arrows correspond to the three ionization curves shown in ( f ). Full size image With the non-demolition charge readout, we investigate the ionization by various NIR illumination. We first initialize the charge state to NV − by a 532 nm laser pulse and measurement-based charge state post-selection. Then a 20 μs pulse of E 1,2 initializes the spin to state \(\left|0\right\rangle\) . After the charge and spin initialization, the SCC process is applied, followed by a charge state readout (Fig. 2 e). In contrast to the long charge lifetime of 400.7 ms observed in the absence of NIR laser (Fig. 2 d), the NV − population decays fast on the timescale of microseconds after simultaneous illumination of E y and NIR light (Fig. 2 f). However, the NV − population saturation level does not reach at 0, indicating that in some cases \(\left|0\right\rangle\) goes through the spin-flip process and gets trapped in \(\left|\!\pm \!1\right\rangle\) , which does not ionize. As the NIR power increases, the NV − population decay faster and saturates at lower levels. To estimate the ionization rate Γ ion , we develop an extensive model including a more complicated energy structure as described in Supplementary information . The model uses independently measured quantities and one free parameter Γ ion to fit the data shown in Fig. 2 f. The extracted ionization rate is proportional to the NIR laser power (Fig. 2 g). This indicates that the NV center is most likely to be ionized from the excited state by absorbing a single 1064 nm photon. The obtained coefficient of 67.0 ± 6.7 kHz/mW is much lower than the 1.2 ± 0.33 MHz/mW previously estimated at room temperature 24 , which requires further study in the future. The highest Γ ion obtained is 2.79 ± 0.08 MHz, only 3.7 times of Γ flip = 0.75 ± 0.02 MHz. One limitation is the output power of current CW NIR laser. The other is the high loss of laser power density on NV center due to transmission reduction and chromatic aberration of the objective. The resulting single-shot fidelity is 89.1 ± 0.2% (blue line in Fig. 3 c). To improve the conversion efficiency ( \(\left|0\right\rangle \to\) NV 0 ) under current conditions, we consider a correction scheme by utilizing the auxiliary level m S = −1. As shown in Fig. 3 a, the leakage population from \(\left|0\right\rangle\) to the AUX state, is transferred back to \(\left|0\right\rangle\) state through an MW AUX π pulse. With this correction, the \(\left|0\right\rangle\) is converted into NV 0 with higher efficiency, while conversion of state \(\left|1\right\rangle\) is not affected (Fig. 3 b). The resulting single-shot fidelity is shown in Fig. 3 c. With about 10 μs SCC duration, the average fidelity reaches its maximum of F avg = 1/2 ( \({F}_{\left|0\right\rangle }\) + \({F}_{\left|1\right\rangle }\) ) = 95.4 ± 0.2 %. The corresponding histogram is given in Fig. 3 d. We also compare the SCC method with the resonance fluorescence method for the single-shot readout. Due to the sizeable spin-flip rate, the optimal average fidelity with resonance fluorescence method is 79.6 ± 0.8% (Fig. 3 c, d), much lower than previous reports with low-strain NV centers 7 , 14 , 15 , 25 . Fig. 3: Single-shot readout of NV electron spin state via SCC. a Pulse sequence and diagram illustrating NV spin and charge dynamics for \(\left|0\right\rangle\) readout fidelity evaluation. As the spin-flip process traps some populations in \(\left|1\right\rangle\) and AUX state, an MW AUX π pulse rescues the part in AUX state back to \(\left|0\right\rangle\) so that they can be ionized in the next round. The SCC pulse and AUX correction (AUX Corr.) pulse are repeated for n rounds to get the optimal ionization. The sequence for evaluating \(\left|1\right\rangle\) readout fidelity only differs in the spin initialization part, which is an E y + MW AUX pulse of 200 μs. b NV − population dependence on SCC duration (2 μs × n ). The solid lines for \(\left|0\right\rangle\) are simulations. The solid lines for \(\left|1\right\rangle\) are linear fits to the data. c Average fidelity dependence on E y illumination time for different readout methods. In SCC methods E y illumination time equals the SCC duration, and in resonance fluorescence method it equals the read window. Blue and orange solid lines are the average of the corresponding lines in ( b ). The yellow line is an exponential fit to the results of the resonance fluorescence (Res. Fluor.) method. d Photon number distribution of the charge readout with the SCC method. This is obtained from 20,000 measurement repetitions with NV spin initially prepared in the \(\left|0\right\rangle\) (blue) and \(\left|1\right\rangle\) (orange). Inset: photon distribution for the resonance fluorescence method. Full size image Discussion The main limiting factor for our single-shot readout fidelity is the SCC efficiency. It depends on both the ionization rate and the spin-flip rate. Figure 4 a shows the simulation results using our model ( Supplementary information ). The larger ratio Γ ion /Γ flip is, the higher efficiency could be achieved. In practice, Γ flip has a lower bound solely determined by the intrinsic property of NV center. In contrast, Γ ion is convenient to increase by using high power NIR laser and good transmission objective. For a lower Γ flip ~ 0.2 MHz 7 , a modest NIR power > 1 W on the diamond could achieve an average single-shot readout fidelity exceeding 99.9% (Fig. 4 b), meeting the requirement for fault-tolerant quantum computing and networks 9 , 26 , 27 , 28 , 29 . Fig. 4: Exceeding the 99.9% fault-tolerant threshold. a Effects of ionization rate and spin-flip rate on SCC efficiency, which is indicated by the number next to each curve. b Dependence of overall spin readout fidelity on ionization rate at two different spin-flip rates. The orange line corresponds to spin-flip rate observed in this work. The blue line is a prediction for an NV center with a low spin-flip rate reported in ref. 7 . Full size image SCC readout is a demolition method for electron spins. Projective readout is still feasible for nuclear spins weakly coupled to the NV center, as their polarization is more robust to the perturbation from optical pumping and ionization 30 , 31 , 32 . The SCC scheme also has the potential for applications on integrated quantum devices 33 , 34 , 35 , 36 , 37 . At present, the photoelectric detection of single NV centers relies on measuring photocurrent from multiple ionizations 37 . The deterministic SCC opens the possibility for achieving optoelectronic single-shot readout of solid spins, potentially utilizing the single-electron transistor as charge reading head 38 , 39 . Another promising application of single-shot SCC is high-efficiency quantum sensing as discussed in a recent work 40 . Because most of the bio-molecules are rarely affected by the NIR light, the NIR-assisted SCC demonstrated here is helpful to avoid photo-damage on the bio-samples 41 , 42 , 43 , 44 . In summary, we demonstrate a NIR-assisted SCC method for the singe-shot readout of electron spin with fidelity of 95.4%. Different from previous methods which requires careful engineering to improve the emission rate and photon collection efficiency, our method only need an additional NIR beam. By directly controlling the NIR power, the above calculations suggest that the NIR-assisted SCC is an experimentally feasible approach toward spin readout exceeding the fault-tolerant threshold. We would like to note 40 , which makes use of a similar scheme to achieve single-shot readout with poor optics, using visible, rather than infrared light. Data availability All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary information. Additional data related to this paper may be requested from the authors.

A team led by Professor Du Jiangfeng and Professor Wang Ya from the Chinese Academy of Sciences (CAS) Key Laboratory of Microscale Magnetic Resonance of the University of Science and Technology of China put forward an innovative spin-to-charge conversion method to achieve high-fidelity readout of qubits, stepping closer towards fault-tolerant quantum computing. Quantum supremacy over classical computers has been fully exhibited in some specific problems, yet the next milestone, fault-tolerant quantum computing, still requires the accumulated logic gate error and the spin readout fidelity to exceed the fault-tolerant threshold. Du's team has resolved the first requirement in the nitrogen-vacancy (NV) center system [Nat. Commun. 6, 8748 (2015)] previously and this work targeted at high-fidelity readout of qubits. Qubit state, such as spin state, is fragile: a common readout approach may cause the flip between the 0 and 1 states for even a few photons resulting in a reading error. The readout fidelity of traditional resonance fluorescence method is strictly limited by such property. Since the spin state is difficult to measure, researchers blazed a trail to replace it with an easy-to-readout and measurable property: the charge state. They first compared the optical readout lifetime of the charge state and spin state, finding that the charge state is more stable than the spin state by five orders of magnitude. Experiment results showed that the average non-demolition charge readout fidelity reached 99.96%. Then the team adopted near-infrared (NIR) light (1064 nm) to induce the ionization of the excited spin state, transforming the spin state 0 and 1 to the "electrically neutral" and "negatively charged" charge states respectively. This process converted the spin readout to the charge readout. The results indicated that the error of traditional resonance fluorescence method reached 20.1%, while the error of this new method can be suppressed to 4.6%. The article was published in Nature Communications. This new method is compatible with tradition methods, provisioning a spin readout fidelity exceeding the fault-tolerant threshold in real applications. Thanks to the less damage of NIR light to biological tissues and other samples, this method will also effectively improve the detection efficiency of quantum sensors.

10.1038/s41467-021-21781-5

Space

Scientists verify trap-release-amplify model by reproducing electromagnetic waves on Mars

Shangchun Teng et al, Whistler-mode chorus waves at Mars, Nature Communications (2023). DOI: 10.1038/s41467-023-38776-z Journal information: Nature Communications

https://dx.doi.org/10.1038/s41467-023-38776-z

https://phys.org/news/2023-06-scientists-trap-release-amplify-electromagnetic-mars.html

Abstract Chorus waves are naturally occurring electromagnetic emissions in space and are known to produce highly energetic electrons in the hazardous radiation belt. The characteristic feature of chorus is its fast frequency chirping, whose mechanism remains a long-standing problem. While many theories agree on its nonlinear nature, they differ on whether or how the background magnetic field inhomogeneity plays a key role. Here, using observations of chorus at Mars and Earth, we report direct evidence showing that the chorus chirping rate is consistently related to the background magnetic field inhomogeneity, despite orders of magnitude difference in a key parameter quantifying the inhomogeneity at the two planets. Our results show an extreme test of a recently proposed chorus generation model and confirm the connection between the chirping rate and magnetic field inhomogeneity, opening the door to controlled plasma wave excitation in the laboratory and space. Introduction Whistler-mode chorus waves are electromagnetic emissions frequently found in space and at other planets 1 , 2 . When the electromagnetic signal of chorus is converted to an acoustic signal, these waves sound like birds chirping at dawn, hence the name. Chorus waves are known to play a dominant role in accelerating relativistic electrons in Earth’s radiation belts 3 , 4 . They have also been demonstrated to scatter electrons with energies of a few hundred eV to a few keV into the atmosphere to form diffuse and pulsating auroras 5 , 6 , 7 , playing a crucial role in the energy and mass coupling between the ionosphere and magnetosphere 8 and to be an embryonic source for plasmaspheric hiss 9 , 10 . Satellite observations show that chorus emissions are narrowband and quasi-coherent, consisting of discrete chirping elements with their central frequency changing rapidly as a function of time (i.e., chirping) 1 , 11 . The frequency chirping is not unique to whistler-mode chorus but has also been observed in electromagnetic ion cyclotron waves in space plasmas or Alfvén waves in fusion plasmas 12 , 13 . Hence there is a broad interest in understanding the fundamental physical mechanism of chorus chirping, in addition to the consensus on the importance of chorus. However, despite the early establishment of nonlinear wave–particle interactions in the generation of chorus, the theoretical mechanism of how nonlinear interactions lead to chorus chirping has been under intensive debate for over 70 years 14 , 15 , 16 , 17 , 18 . A recently proposed “Trap-Release-Amplify” (TaRA) model 19 shows that nonlinear wave–particle interactions and background magnetic field inhomogeneity work together to produce the chirping of chorus in space, unifying two different ways of estimating chorus chirping rate from previous theoretical models 14 , 16 . As a self-consistent model, electrons interact with the quasi-coherent wave packet, resulting in quasi-coherent rather than the stochastic motion of a small group of resonant electrons. The quasi-coherent motion leads to significant changes in the electron’s energy and momentum. Because the phase space density along an electron’s trajectory is constant, there is a difference between the electron’s phase space density and its neighboring at the new location, leading to the formation of phase-space structures such as phase space holes. This nonlinear interaction process results in a frequency chirping rate proportional to the wave amplitude at the equator, a conclusion verified for chorus events at Earth 20 , 21 and is widely accepted 22 , 23 . The main unique feature of the TaRA model is that it further predicts that the chirping rate is proportional to the local magnetic field inhomogeneity when the quasi-coherent phase-space structure is released from the wave packet and leads to selective amplification of new narrowband emissions. This prediction of the TaRA model is consistent with one of the previous theoretical models of chorus 14 and the statistical dependence of the chirping rate of chorus on the radial distance at Earth 24 . However, different from the nonlinear chirping rate, which is a defining feature of chorus, the dependence of the chirping rate on the magnetic field inhomogeneity is still under debate 22 . Note that the change of inhomogeneity of the background magnetic field at Earth is relatively mild. A direct experimental test with orders of magnitude variation in magnetic field inhomogeneity is not possible using wave observations only at Earth. In this work, we conduct a detailed analysis of a previously reported chorus-like emission observed by the Mars Atmosphere and Volatile EvolutioN (MAVEN) mission 25 . Despite the vastly different magnetic field and plasma parameters, we demonstrate that the emission observed by MAVEN at Mars shares the same nonlinear nature with chorus waves at Earth and exhibits features consistent with the predictions made by the TaRA model. Results Observation MAVEN detected the wave event on 12 July 2015, and linear instability analysis confirms that the wave is of whistler mode 26 . However, linear analysis can neither confirm this event as chorus nor explain its frequency chirping, which is fundamentally nonlinear. Figure 1 a shows the trajectory of the MAVEN satellite and the magnetic field line it crosses while observing the event, whose frequency–time spectrogram is shown in Fig. 1 c. Note that Mars does not have a global magnetic field like Earth. The closed magnetic field line is part of the crustal magnetic field of Mars. As a comparison, we also show in Fig. 1 b the global magnetic field and a chorus event of Earth observed by the Van Allen Probes mission 27 (Fig. 1 d). The crustal magnetic field line at Mars encountered by MAVEN has a much smaller scale than that of typical magnetic field lines at Earth. The inhomogeneity parameter ( ξ ) of a given magnetic field line, relevant to studies of chorus, can be obtained by approximating the magnetic field strength near the equator (minimum B ) by a parabolic function; i.e., B = B 0 (1 + ξ s 2 ), where s is the distance from the equator along a field line. Using corresponding magnetic field models, we obtain that, for the chorus event at Earth, ξ E ≈ 5 × 10 −9 km −2 , and for the event at Mars, ξ M ≈ 1.4 × 10 −4 km −2 . The five orders of magnitude difference between the inhomogeneity factor ξ provide a unique opportunity for an extreme test of the dependence of chorus properties on background magnetic field inhomogeneity. Fig. 1: Magnetic field and chorus emissions at Mars and Earth. Traced magnetic field lines, represented by blue lines, at Mars ( a ) and Earth ( b ) are shown. The scale bars indicate the size of each planet. The yellow arrow in panel ( a ) denotes the trajectory of MAVEN, and the black dot in panel ( b ) indicates where chorus waves shown in panel ( d ) were observed. Frequency–time spectrograms of chorus waves observed at Mars ( c ) by MAVEN and Earth ( d ) by Van Allen Probe B are shown. Color-coded is the power spectrum density calculated using the wave electric field ( c ) and magnetic field ( d ). The time coordinate indicates the number of seconds since 2015-07-12/06:00:53 UT for panel ( c ) and 2012-10-08/06:03:57 UT for panel ( d ). The white dashed lines in ( c ) and ( d ) indicate half electron cyclotron frequency. Black dots in panel ( c ) denote the maximum power spectral density at a given time, and the white line represents the linear least-squares fitting, whose slope gives the frequency chirping rate shown in normalized units. Full size image Nonlinear chirping rate To prove that the waves presented in Fig. 1 c at Mars are indeed chorus emissions, we need to demonstrate that the frequency chirping is due to nonlinear wave–particle interactions. Because of the instrument resolution limitations, directly identifying nonlinear phase-space structures in the observed electron distribution is not possible. Therefore, we perform a self-consistent computer simulation using the observed particle distribution and the local magnetic field line model (see “Methods”, subsection “Computer simulation setup”). Figure 2 a shows the simulated wave event, which exhibits the same chirping characteristics as the observed event in Fig. 1 c. The chirping rate of the element from the computer simulation is \(4.6\times 1{0}^{-4}{\Omega }_{e0}^{-2}\) , whereas the chirping rates of the two observed elements are 2.5 × 10 −4 \({\Omega }_{e0}^{-2}\) (or 306 Hz/s) and 1.7 × 10 −4 \({\Omega }_{e0}^{-2}\) (or 205 Hz/s). Here Ω e 0 ≡ e B 0 / m is the electron angular cyclotron frequency at the equator, with e the elementary charge and m electron mass. Therefore, the chirping rate from simulation and observation differs by about a factor of two to three. Figure 2 b shows the wave electric field from simulation and observation in physical units. The electric field amplitude from the simulation is consistent with that from observation within a factor of three. Note that the computer simulation does not reproduce the second element from observation, which is expected and commonly seen in other simulations of chorus, because no free energy is re-supplied to the simulation system after the generation of the first element 28 , 29 . The slight difference in chirping rate between the first and second elements may indicate a change in electron distribution, as noted in previous studies 30 , 31 . However, due to the limited time resolution (2 s) of the SWEA instrument onboard MAVEN, this difference cannot be resolved. Overall, the simulation results exhibit good consistency with the observed data for both the chirping rate and wave amplitude, despite uncertainties in the simulation parameters. Therefore, the successful reproduction of the observed event by computer simulation enables us to obtain parameters that are not possibly available from direct observation. Fig. 2: Computer simulation of the chorus event at Mars. a The simulated rising-tone chorus waves obtained using the electric fields at s / d e = 0.38, corresponding to the position of the observed chorus event. Color-coded is the power spectral density of the wave electric field in normalized units. The black dots and white line are the same as those in Fig. 1 c. b , c Comparison of waveform from simulation and observation. Full size image Now we confirm that the frequency chirping observed by MAVEN and from the computer simulation is caused by a nonlinear process, a characteristic feature of chorus waves. Nonlinear wave–particle interaction theories 16 , 17 , 18 , 19 predict a chirping rate (Γ NL ), which is proportional to the wave magnetic amplitude ( δ B ) at the equator. This chirping rate is also proportional to a parameter R , which characterizes the nonlinear wave–particle interactions and is typically within the range of 0.2 ~ 0.8 (see “Methods”, subsection “Theoretical estimate of the chirping rate”). Note that the event from observation is off the equator; therefore, the theoretical chirping rate cannot be directly calculated using the measured wave amplitude. Nevertheless, we may test if the nonlinear chirping rate is of the same order as that from observation. Using the average electric field from observation and the whistler wave dispersion relation, the wave magnetic field amplitude is estimated to be approximately δ B / B 0 ≈ 10 −2 . Using plasma parameters determined from MAVEN observation, we estimate that for ω = 0.3Ω e 0 , Γ NL ≈ 4.5 × 10 −4 \({\Omega }_{e0}^{-2}\) and 1.8 × 10 −3 \({\Omega }_{e0}^{-2}\) with R = 0.2 and 0.8, respectively. Correspondingly, considering the crude nature of the estimate and that the event is off the equator, the theoretical nonlinear chirping rate is in the same order as the observed event. A more accurate determination of the nonlinear nature of the chirping is via the phase-space dynamics of electrons from simulation. Incidentally, this can also verify the nonlinear chirping rate, because the parameter R is defined by the chirping rate and can be estimated directly from the electron phase-space structure from the computer simulation of the event 16 , 17 , 32 . Figure 3 shows the electron phase-space distribution at the equator near the resonance velocity, which clearly confirms the presence of the characteristic phase-space hole associated with the generation of a rising-tone chorus. By performing a rough fitting to the boundary of the phase-space hole, we obtain directly that R ≈ 0.45. Correspondingly, the nonlinear chirping rate is naturally valid with this value of R at the equator. The analysis of the electron phase-space dynamics confirms the nonlinear chirping rate predicted by several models 15 , 16 , 17 , 18 . More importantly, it demonstrates that the wave event observed by MAVEN is generated by a nonlinear process, similar to chorus emissions observed at Earth. Fig. 3: Electron phase-space structure. The v ∥ - ζ phase-space distribution at s = 0 and v ⊥ = 0.0085 c , with ζ the angle between particle perpendicular velocity and wave magnetic field, from the computer simulation of the chorus event at Mars. The phase-space hole is clearly seen with the center near v ∥ = −0.005 c . A fitting to the boundary of this hole is performed using the motion of resonant electrons with R = 0.45. The black line indicates the separatrix and the red dot marks the corresponding center. Full size image Chirping rate from magnetic field inhomogeneity A unique feature of the TaRA model is that, besides the nonlinear chirping rate at the equator, it also predicts that the chirping rate is related to the magnetic field inhomogeneity at the source location, denoted by s 0 . Physically this is because of the phase-locking condition, which requires the balance between the wave chirping, characterized by a term R 1 , and the background magnetic field inhomogeneity, characterized by a term R 2 . The result of this balance is a theoretical chirping rate, Γ IN , that is proportional to the magnetic field inhomogeneity at the source location s 0 where the wave amplitude is negligibly small (see “Methods”, subsection “Theoretical estimate of the chirping rate”). The location s 0 could be roughly given by the interaction region size based on the linear motion of electrons or more accurately estimated directly from simulation. Again, we first use a Mars crustal magnetic field model to test if using the magnetic field inhomogeneity gives the correct order of magnitude estimate of the chirping rate. For this rough estimate, we use \({s}_{0}={(2\pi {v}_{r}/\xi {\Omega }_{e0})}^{1/3}\) , which corresponds to a shift of π radian in wave–particle interaction phase angle from the equator by assuming that electrons move adiabatically along the background magnetic field 14 . Here v r is the electron resonance velocity. For the chorus event at Mars, this estimate gives s 0 ≈ −0.28 d e with d e ≡ c /Ω e 0 , and the theoretical chirping rate Γ IN is about \(2\times 1{0}^{-3}{\Omega }_{e0}^{2}\) . This value is comparable to the estimate of the nonlinear chirping rate with R = 0.8 and larger than the chirping rate from the simulation by about a factor of four or the observed one by a factor of ten. As shown below, the discrepancy between the theoretical and the observed or simulated chirping rate for the case of Mars is mainly due to the rough nature of s 0 estimated from the linear motion of electrons. For a more direct comparison of the Mars event, we focus on the simulated event and use the information from computer simulation to find a better estimate of the source location s 0 . Figure 4 shows the wave propagation and the effective growth rate for the event of interest from simulation. The source location s 0 could be estimated from the figure of the effective growth rate γ eff as the location between smoothly varying regions and noisy-like regions due to background thermal noise. Clearly, for this particular case, the above crude estimate of s 0 based on the linear motion of electrons is too large; it should be roughly between −0.05 d e and −0.1 d e for the simulated case. Fig. 4: Wave propagation and effective growth rate. a The wave magnetic field strength and b the effective growth rate as a function of s and t for the wave propagating in the s direction, from the computer simulation of the chorus event at Mars. The orange dashed lines mark s / d e = −0.28, while the blue dashed and dotted lines mark s / d e = −0.1 and s / d e = −0.05, respectively. Full size image Figure 5 displays wave spectrograms for the interval s / d e = −0.1 to s / d e = 0.05 to support the estimated value of s 0 discussed above. At s / d e = 0 (Fig. 5 c) and 0.05 (Fig. 5 d), clear chirping elements are observed around t Ω e 0 = 300, characterized by chirping rates of \(7.8\times 1{0}^{-4}{\Omega }_{e0}^{-2}\) and \(8.2\times 1{0}^{-4}{\Omega }_{e0}^{-2}\) , respectively. The chirping rate at s / d e = 0.38 differs from that at s / d e = 0 due to wave packet distortion during propagation. It should be noted that the stronger signal near t Ω e 0 = 500 in these spectrograms should not be interpreted as a falling tone in Fig. 5 c–d or bi-directional chirping elements in other panels. No corresponding nonlinear phase-space structure can be found in the electron distribution, and these spectral structures do not maintain a consistent shape. At s / d e = −0.05, a weak chirping structure around t Ω e 0 = 300 is visible. To establish a connection between this structure and the chorus element at the equator, we present wave spectrograms at intermediate locations between s / d e = −0.05 and 0 in Fig. 5 e–h. These spectrograms clearly illustrate the amplification of the chirping element as it propagates from s / d e = −0.05 to s / d e = 0, supporting that the source region lies upstream of −0.05. Furthermore, the chirping elements exhibit consistent chirping rates; hence, we will use the chirping rate at the equator for the following analysis. On the other hand, at s / d e = −0.1 (Fig. 5 a), the wave signal is comparable to background noise between t Ω e 0 = 200 and 300. Taken together, these spectrograms show the chorus element’s source location is between s / d e = −0.05 and −0.1. Fig. 5: Wave magnetic field spectrogram upstream from the equator. Black dots mark the maximum wave power spectral density for a given time for the simulated chorus event at Mars. White lines denote a linear least-squares fitting to the corresponding black dots starting from the fifth black dot. The fitting for the spectrograms at s / d e = −0.04 and −0.05 uses seven data points in total due to their weak intensity. In general, the chirping rate estimate is more reliable for chorus elements with a stronger wave signal. Full size image To further constrain the value of s 0 and verify the balance of R 1 and R 2 or the theoretical chirping rate Γ IN , we analyze the electron phase-space dynamics and compare the chirping rate in Fig. 6 . In Fig. 6 a, we plot the ratio of a nonlinearity term \({\omega }_{tr}^{2}\) (see “Methods”, subsection “Theoretical estimate of the chirping rate”) and the inhomogeneity term R 2 to the wave chirping term R 1 as a function of s / d e between −0.1 and 0, together with the parameter R . The wave amplitude is estimated from the square root of the integrated wave power spectral density. As can be seen from the plot, the ratio \({\omega }_{tr}^{2}/{R}_{1}\) decreases from approximately 2 at s = 0 to around 0.2 at s / d e ~ −0.1. Note that due to simulation noise, this term can never reach zero in a particle simulation. On the other hand, the ratio − R 2 / R 1 increases rapidly from 0 at s / d e = 0 to about 1.04 at −0.1 due to the extremely large inhomogeneity factor of the background magnetic field. The parameter R varies between 0.5 and 0.6 between s / d e = −0.07 and 0 and decreases to 0 at approximately s / d e = −0.096, where R 1 = − R 2 . To support the estimated value of R in Fig. 6 a, we display electron phase-space distributions between s / d e = −0.05 and −0.1 in Fig. 6 b–g, which exhibit good consistency with the variation of R . These analyses suggest that the source location s 0 lies within the range of −0.1 ≲ s 0 / d e ≲ −0.09. Using this information about s 0 , we show the ratio of the chirping rates from nonlinear theory (Γ NL ) and inhomogeneity (Γ IN ) to those from simulation in Fig. 6 h. The comparison suggests that the chirping rate from inhomogeneity agrees with the rates obtained from both simulation and the well-established nonlinear theory. Fig. 6: Electron phase-space dynamics and comparison of chirping rates upstream from the equator. a Variation of \({\omega }_{tr}^{2}/{R}_{1},-{R}_{2}/{R}_{1}\) , and R as a function of s . Panels b – g show the electron phase-space ( v ∥ - ζ ) distribution between s / d e = −0.05 and −0.1. Panel h displays the ratio of theoretical chirping rates to the chirping rate from simulation between s / d e = −0.05 and 0 for the wave event at Mars. The nonlinear chirping rate is estimated using the general definition of R for off-equatorial locations, with R = 0.5 and 0.6. The chirping rate from inhomogeneity is estimated using s 0 / d e = −0.09 and −0.1. Full size image Discussion The typical inhomogeneity factor of the global magnetic field of Earth is about five orders of magnitude smaller than that of the crustal magnetic field of Mars used by this study. However, a previous analysis involving 1106 chorus elements has shown that the theoretical chirping rate from inhomogeneity exhibits good statistical agreement with the observed chorus elements at Earth 24 . For the Earth chorus event shown in Fig. 1 d, a theoretical estimate, with observed plasma parameters, gives that the observed chirping rate (Γ obs ) is approximately twice the theoretical rate obtained from inhomogeneity (Γ IN ) (see “Methods”, subsection “The frequency chirping rate calculation of Earth chorus waves”). These previous statistical results at Earth and the current result at Mars demonstrate that the chorus chirping rate is consistently related to the background magnetic field inhomogeneity, despite a few orders of magnitude difference in the inhomogeneity factor. Therefore, the combined results of these studies suggest that the physics involved in obtaining the chirping rate from inhomogeneity is as important as those involved in nonlinear wave–particle interactions for chorus emissions at these planets. The above comparison between observation, simulation, and theory proves the presence of chorus emissions at Mars, with fundamentally the same nonlinear nature as those at Earth. Furthermore, it presents an extreme test of one of the TaRA model predictions that the chirping rate is related to the background magnetic field inhomogeneity besides being proportional to wave amplitude at the equator. It demonstrates that the chirping of chorus results from a complex interplay between nonlinear wave–particle interactions and the magnetic field inhomogeneity. The consistency in the two kinds of chirping rate between observation and theory further establishes the validity of the recently proposed TaRA model for the event. Future studies involving a large number of chorus events are needed to statistically test the TaRA model. The relation between the chirping rate and magnetic field inhomogeneity may be used to infer the background magnetic field inhomogeneity of Mars or any other planets from chorus observation, while the nonlinear chirping rate can be used to estimate the chorus wave amplitude. These results allow a more controlled setup in future active experiments of plasma waves or radiation belt remediation through chorus emissions. Methods Computer simulation setup The study employs the self-consistent particle simulation code DAWN 33 to simulate a system along the magnetic field line, considering only parallel-propagating waves. The simulation treats cold electrons as a fluid and models hot electrons using the nonlinear δ f method 34 to minimize simulation noise. Particle boundaries are set as reflecting, while wave boundaries are absorbing. A grid size of 6 × 10 −4 d e is used to ensure the proper resolution of wavelength, which is approximately 0.04 d e at the equator for the emission observed by MAVEN. A time step of \(3\times 1{0}^{-4},{\Omega }_{e0}^{-1}\) is used to satisfy the Courant condition. The simulation uses 4000 cells and 2000 simulation particles per cell for each electron population. To improve computational efficiency while representing the observed electron distribution by MAVEN, our simulation employs two populations of hot electrons. Figure 7 a shows the electron phase-space densities (PSDs) as a function of energy obtained from the SWEA measurements 35 at 2015-07-12/06:00:53 UT. Linear instability analysis shows that the growth rate dominates in the parallel direction. We obtained the equatorial distributions of hot electrons by fitting the electron velocity distribution function with the sum of two bi-Maxwellian functions. The function form is given by $$f\left({u}_{\parallel },{u}_{\perp }\right)=\frac{1}{{(2\pi )}^{3/2}{w}_{\parallel }{w}_{\perp }^{2}}\exp \left(-\frac{{u}_{\parallel }^{2}}{2{w}_{\parallel }^{2}}-\frac{{u}_{\perp }^{2}}{2{w}_{\perp }^{2}}\right),$$ (1) in which u ∥ and u ⊥ are velocities parallel and perpendicular to the background magnetic field, and w ∥ and w ⊥ are corresponding thermal velocities in the non-relativistic limit. Fitting results reveal that the first component has a plasma frequency of 48.6Ω e 0 and thermal temperatures of T ∥ = 2.9 eV and T ⊥ = 4 eV, while the second component has a plasma frequency of ω p e = 168.5Ω e 0 and thermal temperatures of T ∥ = 20 eV and T ⊥ = 41 eV. We set the plasma frequency of the cold electrons to 128Ω e 0 to ensure the total electron number density is consistent with the observed value ( n e = 112 cm −3 ). Fig. 7: Electron distribution and linear growth rate. a Comparison of measured and fitted electron PSD as a function of energy at different pitch angles. Dots represent measured values, while lines show fitted results. b Two-dimensional electron PSD as a function of pitch angle ( α ) and energy from MAVEN measurements. c Two-dimensional electron PSD from fitting. d Two-dimensional linear growth rate calculation, showing that the linear growth rate peaks for parallel-propagating waves with frequencies near 0.21 ~ 0.23Ω e 0 . According to chorus wave excitation theories 17 , 19 , the frequency of the maximum linear growth rate roughly corresponds to the starting frequency of the chorus element. Correspondingly, the linear growth rate calculation is consistent with observation. Full size image Theoretical estimate of the chirping rate The TaRA model 19 estimates the chirping rate based on the following equation for the wave–particle interaction phase angle ζ ≡ 〈 v ⊥ , δ B 〉, $$\frac{{{{{{{{{\rm{d}}}}}}}}}^{2}\zeta }{{{{{{{{\rm{d}}}}}}}}{t}^{2}}={\omega }_{tr}^{2}\sin \zeta -({R}_{1}+{R}_{2}).$$ (2) Here \({\omega }_{tr}^{2}\equiv k{v}_{\perp }e\delta B/m\) is the phase trapping frequency squared, with k the wave number and v ⊥ the component of the resonant electron velocity perpendicular to the local magnetic field. The two parameters R 1 and R 2 are defined by $${R}_{1}={\left(1-\frac{{v}_{r}}{{v}_{g}}\right)}^{2}\frac{\partial \omega }{\partial t},$$ (3) $${R}_{2}=\left(\frac{k{v}_{\perp }^{2}}{2{\Omega }_{e}}-\frac{3{v}_{r}}{2}\right)\frac{\partial {\Omega }_{e}}{\partial s},$$ (4) where v g is the wave group velocity, and Ω e is the electron cyclotron frequency at s . The parameter R , which characterizes the nature of electron dynamics, is defined by \(R\equiv ({R}_{1}+{R}_{2})/{\omega }_{tr}^{2}\) . At the equator, where the background magnetic field inhomogeneity is negligible ( R 2 → 0), the nonlinear chirping rate can be obtained from the definition of R , i.e., $${\Gamma }_{{{{{{{{\rm{NL}}}}}}}}}={\left.R{\left(1-\frac{{v}_{r}}{{v}_{g}}\right)}^{-2}{\omega }_{tr}^{2}\right|}_{s=0}.$$ (5) Based on previous theories on nonlinear wave–particle interactions 16 , 17 , 18 , the value of R typically falls between 0.2 and 0.8 to maximize wave–particle power transfer. On the other hand, at the wave source location ( s 0 ) upstream, the amplitude of the wave is comparable to that of the background noise and \({\omega }_{tr}^{2}\ll {R}_{2}\) . By the principle of selective amplification, the TaRA model requires that the phase-locking condition (d 2 ζ /d t 2 = 0) is satisfied at s 0 . Correspondingly, we obtain a chirping rate proportional to the magnetic field inhomogeneity at s 0 , i.e., $${R}_{1}\approx -{R}_{2}\Rightarrow {\Gamma }_{{{{{{{{\rm{IN}}}}}}}}}\approx -{\left(1-\frac{{v}_{r}}{{v}_{g}}\right)}^{-2}\left(\frac{k{v}_{\perp }^{2}}{2{\Omega }_{e}}-\frac{3{v}_{r}}{2}\right){\left.\frac{\partial {\Omega }_{e}}{\partial s}\right|}_{s={s}_{0}}.$$ (6) Therefore, the TaRA model recovers the chirping rate originally proposed by Helliwell 14 , besides the nonlinear chirping rate. When estimating the two theoretical chirping rates Γ NL and Γ IN , the perpendicular velocity v ⊥ of the resonant particle is obtained with a pitch angle of 70° from nonlinear wave–particle interaction theories and the wave number k is determined by the whistler wave dispersion relation. Determining the background magnetic field inhomogeneity at Mars Figure 8 a shows that MAVEN was at an altitude of approximately 300 km during the chorus event observation. The magnetic field measured by MAVEN’s MAG instrument matches well with the widely used Morschhauser crustal magnetic field model 36 , as shown in Fig. 8 b. This indicates that the crustal magnetic field of Mars dominates during the event of interest. To determine the inhomogeneity factor ξ at Mars, we use the Morschhauser magnetic field model and trace the magnetic field line in both directions from the observing location (altitude: 325 km, longitude: 211°, and latitude: −31. 8°). At the time of observation (about 06:00:54 UT), the tracing result reveals that the observation point is not too far away from the magnetic field minimum (B \({}_{\min }\) ). We fit the magnetic field magnitude near the magnetic field minimum as a function of distance, shown in Fig. 8 c. For this fitting, we conclude that the normalized background magnetic field inhomogeneity parameter is \(\xi=1.4\,{d}_{e}^{-2}\) . Fig. 8: MAVEN altitude and magnetic field from observation and model. a The altitude of MAVEN is shown, while b compares the magnetic field strength from MAVEN observations to the Morschhauser crustal magnetic field model. The vertical dashed line indicates the time of the analyzed chorus event. c The traced (dots) and fitted (solid line) magnetic field strengths near the equator (minimum B) are compared, with the inhomogeneity factor ξ determined from the coefficient of the parabolic fitting. Full size image The frequency chirping rate calculation of Earth chorus waves For the chorus event observed at Earth shown in Fig. 1 d, the observed frequency chirping rate is about 6000 Hz/s. The measured background magnetic field is 152 nT, and the electron density is 2 cm −3 . The inhomogeneity of the background magnetic field is estimated to be 5.45 × 10 −9 km −2 using T89 magnetic field model 37 . With wave number k determined from the whistler wave dispersion relation for f = 0.35 f c e ≈ 1500 Hz and v ⊥ determined from the resonance velocity with a pitch angle of 70°, the chirping rate Γ IN is calculated to be approximately 3529 Hz/s, agrees with the observed one within a factor of two. Data availability The datasets generated during the current study have been deposited in a Zenodo repository and are openly available at 38 . MAVEN wave data are publicly available through . MAVEN particle data are publicly available through . MAVEN magnetic field data are publicly available through . Van Allen Probes wave data are obtained from . Source data are provided with this paper. Code availability The DAWN simulation code is available from the corresponding author upon request. The GEOPACK package for the Earth’s magnetic field is publicly accessible from . The MAVEN toolkit used to calculate the magnetic field of Mars is publicly accessible from .

Chinese scientists have reproduced the observed whistler mode chorus waves on Mars using data from the Mars Atmosphere and Volatile Evolution (MAVEN) mission and compared them with phenomena on the Earth. They found that both Mars and Earth exhibit whistler mode chorus waves triggered by nonlinear processes with the key role played by background magnetic field inhomogeneity in frequency sweeping. The study, published in Nature Communications, provides crucial support for understanding chorus waves in the Martian environment and verifies the previously proposed "Trap-Release-Amplify" (TaRA) model under more extreme conditions. Whistler-mode chorus waves are electromagnetic wave emissions widely presented in planetary magnetospheres. When their electromagnetic signals are converted into sound, they resemble the harmonious chorus of birds in the early morning, hence the name "chorus waves." Chorus waves can accelerate high-energy electrons in space through resonance, leading to a rapid increase in electron flux in Earth's radiation belts during geomagnetic storms. Additionally, they scatter high-energy electrons into the atmosphere, creating diffuse and pulsating auroras. One characteristic of chorus waves is their narrowband frequency sweeping structure. The excitation mechanism of this sweeping structure has been of great interest for decades, and scientists have proposed various theoretical models. However, there has been ongoing debate regarding why frequency sweeping occurs in chorus waves and how to calculate the sweeping frequency. One major point of contention is whether the background magnetic field inhomogeneity plays a crucial role in frequency sweeping and how it affects the sweeping phenomenon. The TaRA model, previously proposed by a team from the University of Science and Technology of China of the Chinese Academy of Sciences, is based on modern plasma physics theories and suggests that the frequency sweeping of chorus waves in the magnetosphere is the result of the combined effects of nonlinear processes and background magnetic field inhomogeneity. The model provides a corresponding formula for calculating the sweeping frequency. However, the variation in magnetic field inhomogeneity in Earth's magnetosphere is limited, making it difficult to test the TaRA model in a larger parameter space. Computer simulation of the chorus event at Mars. (a) The simulated rising-tone chorus waves obtained using the electric fields at s/de = 0.38, corresponding to the position of the observed chorus event. Color-coded is the power spectral density of the wave electric field in normalized units. The black dots and white line are the same as those in Fig. 1c. (b, c) Comparison of waveform from simulation and observation. Credit: Nature Communications (2023). DOI: 10.1038/s41467-023-38776-z There exist distinct magnetic field environments between Mars and Earth. The Earth possesses a global dipole-like magnetic field, while Mars only has localized remnant magnetization. In the remnant magnetization environment of Mars, similar chorus wave events have also been observed by the MAVEN satellite. The calculations reveal a difference of five orders of magnitude in background magnetic field inhomogeneity between Mars and Earth. By comparing wave events observed on Earth and Mars, the previously proposed TaRA model can be tested under more extreme conditions. To validate this model, in this study, scientists from the University of Science and Technology of China and the collaborators observed the particle distribution on Mars using the MAVEN satellite and combined it with the corresponding Martian crustal remnant magnetic field model. Employing a first-principles particle simulation method, scientists reproduced the observed chorus wave phenomena on Mars. Through the analysis of particle phase space distribution, they confirmed that the sweeping process of these waves is consistent with that of chorus waves on Earth, both triggered by nonlinear processes. Furthermore, scientists used two different methods provided by the TaRA model to calculate the sweeping frequency of chorus waves and compared them with the observation and simulation results. The results demonstrated a high degree of consistency between the sweeping frequencies calculated based on nonlinear processes and background magnetic field inhomogeneity, and the simulation results. These findings indicated that although Mars and Earth possess distinct magnetic and plasma environments, the observed chorus wave phenomena on Mars follow the same fundamental physical processes as those in Earth's magnetosphere. This study validated the wide applicability of the TaRA model in describing the sweeping physical processes of chorus waves under extreme conditions with a five-order difference in magnetic field inhomogeneity, which confirms the existence of chorus waves on Mars, and provides support for testing and applying the TaRA model under extreme conditions.

10.1038/s41467-023-38776-z

Computer

AI training: A backward cat pic is still a cat pic

Peter Koo et al, EvoAug: improving generalization and interpretability of genomic deep neural networks with evolution-inspired data augmentations, Genome Biology (2023). DOI: 10.1186/s13059-023-02941-w Journal information: Genome Biology

https://dx.doi.org/10.1186/s13059-023-02941-w

https://techxplore.com/news/2023-05-ai-cat-pic.html

Abstract Deep neural networks (DNNs) hold promise for functional genomics prediction, but their generalization capability may be limited by the amount of available data. To address this, we propose EvoAug, a suite of evolution-inspired augmentations that enhance the training of genomic DNNs by increasing genetic variation. Random transformation of DNA sequences can potentially alter their function in unknown ways, so we employ a fine-tuning procedure using the original non-transformed data to preserve functional integrity. Our results demonstrate that EvoAug substantially improves the generalization and interpretability of established DNNs across prominent regulatory genomics prediction tasks, offering a robust solution for genomic DNNs. Background Uncovering cis -regulatory elements and their coordinated interactions is a major goal of regulatory genomics. Deep neural networks (DNNs) offer a promising avenue to learn these genomic features de novo through being trained to take DNA sequences as input and predict their regulatory functions as output [ 1 , 2 , 3 ]. Following training, these DNNs have been employed to score the functional effect of disease-associated variants [ 4 , 5 ]. Moreover, post hoc model interpretability methods have revealed that DNNs base their decisions on learning sequence motifs of transcription factor (TF) binding sites and dependencies with other TFs and sequence context [ 6 , 7 , 8 , 9 , 10 ]. For DNNs, generalization typically improves with more training data. However, the amount of data generated in a high-throughput functional genomics experiment is fundamentally limited by the underlying biology. For example, the extent to which certain TFs bind to DNA is constrained by the availability of high-affinity binding sites in accessible chromatin. To expand a finite dataset, data augmentations can provide additional variations on existing training data [ 11 , 12 ]. Data augmentations act as a form of regularization, guiding the learned function to be invariant to symmetries created by the data transformations [ 13 , 14 ]. This approach can help prevent a DNN from overfitting to spurious features and improve generalization [ 15 ]. The main challenge with data augmentations in genomics is quantifying how the regulatory function changes for a given transformation. With image data, basic affine transformations can translate, magnify, or rotate an image without changing its label. However, in genomics, the available neutral augmentations are reverse-complement transformation [ 16 ] and small random translations of the input sequence [ 17 , 18 ]. With the finite size of experimental data and a paucity of augmentation methods, strategies to promote generalization for genomic DNNs are limited. Here, we introduce EvoAug, an open-source PyTorch package that provides a suite of evolution-inspired data augmentations. We show that training DNNs with EvoAug leads to better generalization performance and improves efficacy with standard post hoc explanation methods, including filter interpretability and attribution analysis, across prominent regulatory genomics prediction tasks for well-established DNNs. Results and discussion Evolution-inspired data augmentations for sequence-based genomic DNNs To enhance the effectiveness of sequence-based models, data augmentations should aim to increase genetic diversity while maintaining the same biological functionality. Evolution provides a natural process to generate genetic variability, including random mutations, deletions, insertions, inversions, and translocations, among others [ 19 ]. However, these genetic changes often have functional consequences that expand phenotypic diversity and aid in natural selection. While the addition of homologous sequences to a dataset could achieve the goal of increasing sequence diversity while preserving biological function, identifying regulatory regions with similar functions throughout the genomes across species is difficult. Alternatively, synthetic perturbations that do not alter the function can be applied, but it is crucial to have prior knowledge to ensure that features such as motifs and their dependencies are not affected. Therefore, formulating new data augmentation strategies for genomics remains a significant challenge. In this study, we present a suite of evolution-based data augmentations and a two-stage training curriculum to preserve functional integrity (Fig. 1 a). In the first stage, a DNN is trained on sequences with EvoAug augmentations applied stochastically online during training, using the same training labels as the wild-type sequence. The goal is to enhance the model’s ability to learn robust representations of features, such as motifs, by exposing it to expanded (albeit synthetically generated) genetic variation. While each augmentation has the potential to disrupt core motifs in any given perturbation, we expect the overall effect to preserve motifs on average. However, the specific data augmentations employed may introduce a bias in how these motif grammars are structured. Thus, in the second stage, the DNN is fine-tuned on the original, unperturbed data to refine these features and guide the function towards the observed biology, thereby removing any bias introduced by the data augmentations (see Methods). Fig. 1 EvoAug improves generalization and interpretability of Basset models. a Schematic of evolution-inspired data augmentations (left) and the two-stage training curriculum (right). b Generalization performance (area under the precision-recall curve) for Basset models pretrained with individual and combinations of augmentations, i.e., Noise+Ins+RC (Gaussian noise, insertion, reverse-complement) and all augmentations (Gaussian noise, reverse-complement, mutation, translocation, deletion, insertion), and fine-tuned on Basset dataset. Standard represents no augmentations during training. c Comparison of the average hit rate of first-layer filters to known motifs in the JASPAR database (top) and the average q -value of the filters with matches (bottom). d Comparison of the average Pearson correlation between model predictions and experimental data from CAGI5 Challenge. b – d Each box-plot represents 5 trials with random initializations Full size image EvoAug data augmentations introduce a modeling bias to learn invariances of the (un)natural symmetries generated by the augmentations. For instance, random insertions and deletions assume that the distance between motifs is not critical, whereas random inversions and translocations promote invariances to motif strand orientation and the order of motifs, respectively. Nevertheless, the bias created by the augmentations can lead to poor generalization when the introduced bias does not accurately reflect the underlying biology. Therefore, the fine-tuning stage is critical as it provides an avenue to unlearn any biases not supported by the observed data. EvoAug improves generalization and interpretability of genomic DNNs To demonstrate the utility of EvoAug, we analyzed several established DNNs across three prominent types of regulatory genomic prediction tasks that span a range of complexity. First, we applied Evoaug to the Basset model and dataset [ 20 ], which consists of a multi-task binary classification of chromatin accessibility sites across 161 cell types/tissues. We trained the Basset model with each augmentation applied independently and in various combinations. We conducted a hyperparameter sweep to determine the optimal settings for each augmentation (Additional file 1 : Figs. S1-S5). From hyperparameter sweeps, we observed that the inversion augmentation improved performance up to the sequence length, which is essentially a reverse-complement transformation (Additional file 1 : Figs. S1, S3, and S4). Hence, inversions were excluded to reduce redundancy. Remarkably, EvoAug-trained DNNs outperformed standard training with no augmentations (Fig. 1 b). The best results were achieved when multiple augmentations were used together. Additionally, we found that fine-tuning on the original data further improved performance, even when augmentation hyperparameters were poorly specified (Additional file 1 : Fig. S1). Notably, specific EvoAug augmentations, such as random mutations and combinations of data augmentations, had a profound impact on improving the motif representations learned by the first-layer convolutional filters (Fig. 1 c). The convolutional filters capture a wider repertoire of motifs and their representations better reflect known motifs, both quantitatively and qualitatively, when compared with convolutional filters of models trained without augmentations. This suggests that EvoAug augmentations can help DNNs learn more accurate and informative representations of the sequence motifs. A major downstream application of genomic DNNs is to score the functional consequences of non-coding mutations. By evaluating the zero-shot prediction capabilities of each DNN on saturation mutagenesis data of 15 cis -regulatory elements from the CAGI5 Challenge [ 21 ], we found that models trained with EvoAug outperformed their standard training counterpart (Fig. 1 d). Notably, Basset’s performance was comparable to other DNNs based on binary predictions [ 17 ]; however, its overall performance was lower than more sophisticated DNNs and top competitors in the CAGI5 challenge [ 2 ]. Interestingly, we observed that DNNs pretrained with Gaussian noise or random mutagenesis augmentations did not perform well. These augmentations impose flatness locally in sequence-function space, effectively reducing the effect size of nucleotide variants. However, fine-tuning these models improved their variant effect predictions beyond what was achieved with standard training, thus demonstrating the effectiveness of the two-stage training curriculum. To further demonstrate the benefits of EvoAug, we trained DeepSTARR models as a multi-task quantitative regression to predict enhancer activity from self-transcribing active regulatory region sequencing (STARR-seq) data [ 9 ], where each task represents a different promoter from a developmental or housekeeping gene in Drosophila S2 cells. Most EvoAug augmentations resulted in improved performance, except for reverse-complement and random mutations (Fig. 2 a and Additional file 1 : Figs. S3-S5). As before, we observed additional performance gains when augmentations were used in combination. Furthermore, the attribution maps generated by EvoAug-trained models were more interpretable, with identifiable motifs and less spurious noise (Additional file 1 : Fig. S6). In addition, we found that the EvoAug-trained DNNs consistently outperformed DNNs with standard training on various single-task binary classification tasks for TF binding across multiple chromatin immunoprecipitation sequencing (ChIP-seq) datasets (Fig. 2 b). Interestingly, we did not observe any significant improvement in performance after fine-tuning, suggesting that the implicit prior imposed by EvoAug augmentations was appropriate for these tasks; the underlying regulatory grammars for these TFs are not complex. Fig. 2 Generalization of EvoAug on additional models and datasets. a Box-plot of regression performance for DeepSTARR models pretrained with individual or combination of augmentations (i.e., insertion + translocation + deletion; all augmentations) and fine-tuned on original STARR-seq data for two promoters: developmental (top) and housekeeping (bottom). Standard represents no augmentations during training. b Box-plot of classification performance (area under the receiver-operating-characteristic curve) for DNNs trained on ChIP-seq datas. c Average classification performance for ChIP-seq experiments downsampled to different dataset sizes. Shaded region represents the standard deviation of the mean. a , b Each box-plot represents 5 trials with random initializations Full size image To further investigate the impact of EvoAug on small datasets, we retrained each DNN on down-sampled versions of two abundant ChIP-seq datasets. We found that EvoAug-trained DNNs exhibit a greater improvement in performance for smaller datasets compared to standard training (Fig. 2 c). This result suggests that EvoAug can be particularly useful in scenarios where the available training data is limited. Training with EvoAug adds a computational cost, depending on the augmentations chosen and their settings (Additional file 1 : Tables S2 and S3). Nevertheless, EvoAug stabilized training (Additional file 1 : Fig. S7), leading to smoother convergence and improved generalization overall. Conclusion EvoAug greatly expands the set of available data augmentations for genomic DNNs. Our study demonstrated that EvoAug’s two-stage training curriculum is effective in improving generalization performance. Moreover, EvoAug-trained models learned better representations of consensus motifs, as evidenced by filter visualization and attribution analysis. Our findings support previous arguments for using evolution as a natural source of data augmentation [ 22 ]. Interestingly, the impact of synthetic evolutionary perturbations was not excessively disruptive, and performance even improved before fine-tuning in most cases. This functional robustness appears to be a characteristic of the non-coding genome [ 23 ]. Data augmentations are a commonly used technique to balance bias and variance in machine learning models. However, their effectiveness is expected to decrease as the dataset size increases. Nevertheless, EvoAug still improved performance on the already large Basset dataset. Other methods that can enhance generalization include multitask learning [ 24 ], contrastive learning [ 25 , 26 ], and language modeling [ 27 ]. Even though Basset and DeepSTARR are already trained in a multitask framework, EvoAug improved their performance. Multitasking can introduce class imbalance, but EvoAug provides additional examples with pseudo-positive labels, which can mitigate this issue. EvoAug also provides different views of the data, which can be useful for contrastive learning. Importantly, EvoAug is a lightweight and effective strategy that only requires the original data. The optimal combination of augmentations and their hyperparameter choices depends on the model and dataset. While we performed hyperparameter grid searches in this study, more advanced search strategies such as population-based training [ 28 ] using Ray Tune [ 29 ] could improve efficiency. In the future, we plan to investigate EvoAug’s potential in cross-dataset generalization and variant effect predictions, including expression quantitative trait loci. EvoAug is a PyTorch package that is open-source [ 30 ], easy to use, extensible, and accessible via pip ( ) and GitHub ( ), with full documentation provided on ReadtheDocs.org ( ). In time, we plan to extend EvoAug functionality to TensorFlow [ 30 ] and JAX [ 31 ]. We anticipate that EvoAug will have broad utility in improving the efficacy of sequence-based DNNs for regulatory genomics. Methods Models and datasets Basset The Basset dataset [ 20 ] consists of a multi-task binary classification of chromatin accessibility sites across 161 cell types/tissues. The inputs are genomic sequences of length 600 nt and the output are binary labels (representing accessible or not accessible) for 161 cell types measured experimentally using DNase I hypersensitive sites sequencing (DNase-seq). We filtered sequences that contained at least one N character and the data splits (training; validation; test) reduced from (1,879,982; 70,000; 71,886) to (437,478; 16,410; 16,703). This “cleaned” dataset [ 32 ] was analyzed using a Basset-inspired model, which is given according to: Input x \in \{0,1\}^{600 \times 4} x \in \{0,1\}^{600 \times 4} (one-hot encoding of 600 nt sequence) 1D convolution (300 filters, size 19, stride 1) BatchNorm + ReLU Max-pooling (size 3, stride 3) 1D convolution (200 filters, size 11, stride 1) BatchNorm + ReLU Max-pooling (size 4, stride 4) 1D convolution (200 filters, size 7, stride 1) BatchNorm + ReLU Max-pooling (size 2, stride 2) Fully-connected (1000 units) BatchNorm + ReLU Dropout (0.3) Fully-connected (1000 units) BatchNorm + ReLU Dropout (0.3) Fully-connected output (161 units, sigmoid) BatchNorm represents batch normalization [ 33 ], and dropout [ 34 ] rates set the probability that neurons in a given layer are temporarily removed during each mini-batch of training. DeepSTARR The DeepSTARR dataset [ 9 ] consists of a multi-task regression of enhancer activity for two promoters, well-known developmental and housekeeping transcriptional programs in D. melanogaster S2 cells. The inputs are genomic sequences of length 249 nt and the output is 2 scalar values representing the activity of developmental enhancers and housekeeping enhancers measured experimentally using STARR-seq. Sequences with N characters were also removed, but this minimally affected the size of the dataset (i.e., reduced it by approximately 0.005%). This dataset [ 32 ] was analyzed using the original DeepSTARR model, given according to: Input x \in \{0,1\}^{249 \times 4} x \in \{0,1\}^{249 \times 4} 1D convolution (256 filters, size 7, stride 1) BatchNorm + ReLU Max-pooling (size 2, stride 2) 1D convolution (60 filters, size 3, stride 1) BatchNorm + ReLU Max-pooling (size 2, stride 2) 1D convolution (60 filters, size 5, stride 1) BatchNorm + ReLU Max-pooling (size 2, stride 2) 1D convolution (120 filters, size 3, stride 1) BatchNorm + ReLU Max-pooling (size 2, stride 2) Fully-connected (256 units) BatchNorm + ReLU Dropout (0.4) Fully-connected (256 units) BatchNorm + ReLU Dropout (0.4) Fully-connected output (2 units, linear) ChIP-seq Transcription factor (TF) chromatin immunoprecipitation sequencing (ChIP-seq) data was processed and framed as a binary classification task. The inputs are genomic sequences of length 200 nt and the output is a single binary label representing TF binding activity, with positive-label sequences indicating the presence of a ChIP-seq peak and negative-label sequences indicating a peak for a DNase I hypersensitive site from the same cell type but one that does not overlap with any ChIP-seq peaks. Nine representative TF ChIP-seq experiments in a GM12878 cell line and a DNase-seq experiment for the same cell line were downloaded from ENCODE [ 35 ]; for details, see Additional file 1 : Table S1. Negative sequences (i.e., DNase-seq peaks that do not overlap with any positive peaks) were randomly down-sampled to match the number of positive sequences, keeping the classes balanced. The dataset was split randomly into training, validation, and test set according to the fractions 0.7, 0.1, and 0.2, respectively [ 32 ]. A custom convolutional neural network was employed to analyze these datasets, given according to: Input x \in \{0,1\}^{200 \times 4} x \in \{0,1\}^{200 \times 4} 1D convolution (64 filters, size 7, stride 1) BatchNorm + ReLU Dropout (0.2) Max-pooling (size 4, stride 4) 1D convolution (96 filters, size 5, stride 1) BatchNorm + ReLU Dropout (0.2) Max-pooling (size 4, stride 4) 1D convolution (128 filters, size 5, stride 1) BatchNorm + ReLU Dropout (0.2) Max-pooling (size 2, stride 2) Fully-connected layer (256 units) BatchNorm + ReLU Dropout (0.5) Fully-connected output layer (1 unit, sigmoid) Evolution-inspired data augmentations EvoAug is comprised of a set of data augmentations given by the following: Mutation: a transformation where single nucleotide mutations are randomly applied to a given wild-type sequence. This is implemented as follows: (1) given the hyperparameter of the fraction of nucleotides in each sequence to mutate ( mutate_frac ), the number of mutations for a given sequence length is calculated; (2) a position along the sequence is randomly sampled (with replacement) for each number of mutations; and (3) the selected positions are mutagenized to a random nucleotide. Since our implementation does not guarantee that a nucleotide selected will be mutated to a different nucleotide than it originally was, we take approximate account for silent mutations by dividing the user-defined mutate_frac by 0.75 so that on average the fraction of nucleotides in each sequence mutated to a different nucleotide is equal to mutate_frac . Translocation: a transformation that randomly selects a break point in the sequence (thereby creating two segments) and then swaps the order of the two sequence segments. An equivalent statement of this transformation is a “roll”—shifting the sequence forward along its length a randomly specified distance and then reintroducing the part of the sequence shifted beyond the last position back at the first position. This is implemented as follows: (1) given the hyperparameters of the minimum distance ( shift_min , default 0) and maximum distance ( shift_max ) of the shift, the integer-valued shift length is chosen randomly from the interval {\left[ -\texttt {shift\_max}, -\texttt {shift\_min}\right] \cup \left[ \texttt {shift\_min}, \texttt {shift\_max}\right] } {\left[ -\texttt {shift\_max}, -\texttt {shift\_min}\right] \cup \left[ \texttt {shift\_min}, \texttt {shift\_max}\right] } , where a negative value simply denotes a backward shift rather than a forward shift, and (2) the shift is applied to the sequence with a roll() function in PyTorch. Insertion: a transformation where a random DNA sequence (of random length) is inserted randomly into a wild-type sequence. This is implemented as follows: (1) given the hyperparameters of the minimum length ( insert_min , default 0) and maximum length ( insert_max ) of the insertion, the integer-valued insertion length is chosen randomly from the interval between insert_min and insert_max (inclusive), and (2) the insertion is inserted at a random position within the original sequence. Importantly, to maintain a constant input sequence length to the model (i.e., original length plus insert_max ), the remaining amount of length between the insertion length and insert_max is split evenly and placed on the 5’ and 3’ flanks of the sequence, with the remainder from odd lengths going to the 3′ end. Whenever an insertion augmentation is employed in combination with other augmentations, all sequences without an insertion are padded with a stretch of random DNA of length insert_max at the 3′ end to ensure that the model processes sequences with a constant length for both training and inference time. Deletion: a transformation where a random, contiguous segment of a wild-type sequence is removed, and the shortened sequence is then padded with random DNA sequence to maintain the same length as wild-type. This is implemented as follows: (1) given the hyperparameters of the minimum length ( delete_min , default 0) and maximum length ( delete_max ) of the deletion, the integer-valued deletion length is chosen randomly from the interval between delete_min and delete_max (inclusive); (2) the starting position of the deletion is chosen randomly from the valid positions in the sequence that can encapsulate the deletion; (3) the deletion is performed on the designated stretch of the sequence; (4) the remaining portions of the sequence are concatenated together; and (5) random DNA is used to pad the 5′ and 3′ flanks to maintain a constant input sequence length, similar to the procedure for insertions. Inversion: a transformation where a random subsequence is replaced by its reverse-complement. This is implemented as follows: (1) given the hyperparameters of the minimum length ( invert_min , default 0) and maximum length ( invert_max ) of the inversion, the integer-valued inversion length is chosen randomly from the interval between invert_min and invert_max (inclusive); (2) the starting position of the inversion is chosen randomly from the valid position indices in the sequence; and (3) the inversion (i.e., a reverse-complement transformation) is performed on the designated subsequence while the remaining portions of the sequence remain untouched. Reverse-complement: a transformation where a full sequence is replaced with some probability rc_prob by its reverse-complement. Gaussian noise: a transformation where Gaussian noise (with distribution parameters \texttt {noise\_mean} = 0 \texttt {noise\_mean} = 0 and noise_std ) is added to the input sequence; a random value drawn independently and identically from the specified distribution is added to each element of the one-hot input matrix. Pretraining with data augmentations Training with augmentations requires two main hyperparameters: first, a set of augmentations to sample from; and second, the maximum number of augmentations to be applied to a sequence. For each mini-batch during training, each sequence is randomly augmented independently. The number of augmentations to be applied to a given sequence has two possible settings in EvoAug: (1) hard , always equal to the maximum number of augmentations, or (2) soft , randomly select the number of augmentations for a sequence from 1 to the maximum number. Our experiments with Basset and DeepSTARR use the former setting, while our experiments with ChIP-seq datasets use the latter setting. Then, the subset of augmentations to be applied to the sequence is sampled randomly without replacement from the user-defined set of augmentations. After a subset of augmentations is chosen, the order in which multiple augmentations are applied to a single sequence is given by the following priority: inversion, deletion, translocation, insertion, reverse-complement, mutation, noise addition. Each augmentation is then applied stochastically for each sequence. For the Basset and DeepSTARR models, each augmentation has an optimal setting that was determined from a hyperparameter search independently using the validation set (Additional file 1 : Figs. S1, S3, and S4). For the Basset models, the hyperparameters were set to: mutation: mutate_frac = 0.15 translocation: shift_min = 0, shift_max = 30 insertion: insert_min = 0, insert_max = 30 deletion: delete_min = 0, delete_max = 30 reverse-complement: rc_prob = 0.5 noise: noise_mean = 0, noise_std (standard deviation) = 0.3 For the DeepSTARR models, the hyperparameters were set to: mutation: mutate_frac = 0.05 translocation: shift_min = 0, shift_max = 20 insertion: insert_min = 0, insert_max = 20 deletion: delete_min = 0, delete_max = 30 reverse-complement: rc_prob = 0 noise: noise_mean = 0, noise_std = 0.3 When augmentations were used in combinations, the maximum number of augmentations was set to 3 for Basset and 2 for DeepSTARR. The same hyperparameter settings used in DeepSTARR analyses with all augmentations were used for the ChIP-seq analysis. For models trained with combinations of augmentations, the hyperparameters intrinsic to augmentations were set at the values identified above and the maximum number of augmentations per sequence was also determined through a hyperparameter sweep for each dataset (Additional file 1 : Figs. S2 and S5). Unless otherwise specified, all models were trained (with or without data augmentations) for 100 epochs using the Adam optimizer [ 36 ] with an initial learning rate of 1 \times 10^{-3} 1 \times 10^{-3} and a weight decay ( L_2 L_2 penalty) term of 1 \times 10^{-6} 1 \times 10^{-6} ; additionally, we employed early stopping with a patience of 10 epochs and a learning rate decay that decreased the learning rate by a factor of 0.1 when the validation loss did not improve for 5 epochs. For each model trained, the version of the model with the highest-performing weights during its training, as measured by validation loss, is the version of the model whose performance is reported here. Fine-tuning Models that completed training with data augmentations were subsequently fine-tuned on the original dataset without augmentations. Fine-tuning employs the Adam optimizer with a learning rate of 1 \times 10^{-4} 1 \times 10^{-4} and a weight decay ( L_2 L_2 penalty) term of 1 \times 10^{-6} 1 \times 10^{-6} for 5 epochs. The model that yields the lowest validation loss was used for test time evaluation. Evaluation When evaluating models on validation or test sets, no data augmentations were used on input sequences. For models trained with an insertion augmentation (alone or in combination with other augmentations), each sequence is padded at the 3′ end with a stretch of random DNA of length insert_max . Interpretability analysis Filter interpretability We visualized the first-layer filters of various Basset models according to activation-based alignments [ 37 ] and compared how well they match motifs in the 2022 JASPAR nonredundant vertebrates database [ 38 ] using Tomtom [ 39 ], a motif search comparison tool. Matrix profiles MA1929.1 and MA0615.1 were excluded from filter matching to remove poor quality hits; low information content filters tend to have a high hit rate with these two matrix profiles. Hit rate is calculated by measuring how many filters matched to at least one JASPAR motif. Average q -value is calculated by taking the average of the smallest q -values for each filter among its matches. Attribution analysis SHAP-based [ 40 ] attribution maps (implemented with GradientShap from the Captum package [ 41 ]) were used to generate sequence logos (visualized by Logomaker [ 42 ]) for sequences that exhibited high experimental enhancer activity for the Developmental promoter (i.e., task 0 in the DeepSTARR dataset). One thousand random DNA sequences were synthesized to serve as references for each GradientShap-based attribution map. A gradient correction [ 43 ] was applied to each attribution map. For comparison, this analysis was repeated for a DeepSTARR model that was trained without any augmentations and a fine-tuned DeepSTARR model that was pretrained with all augmentations (excluding inversions) with two augmentations per sequence. CAGI5 challenge analysis The CAGI5 challenge dataset [ 21 ] was used to benchmark model performance on variant effect predictions. This dataset contains massively parallel reporter assays (MPRAs) that measure the effect size of single-nucleotide variants through saturation mutagenesis of 15 different regulatory elements ranging from 187 nt to 600 nt in length. We extracted 600 nt sequences from the reference genome centered on each regulatory region of interest and converted it into a one-hot representation. Alternative alleles were then substituted correspondingly to construct the CAGI test sequences. For a given Basset model, the output predictions of two input sequences, one with a centered reference allele and the other with an alternative allele, are made. The cell type-agnostic approach employed in this study uses the mean across these values to calculate a single scalar value, functional activity across cell types. The effect size is then calculated with the log-ratio of this single value for the alternative allele and reference allele, according to: \log \log (alternative value/reference value). To evaluate the variant effect prediction performance, Pearson correlation was calculated within each CAGI5 experiment between the experimentally measured and predicted effect sizes. The average of the Pearson correlation across all 15 experiments represents the overall performance of the model. Availability of data and materials EvoAug Python package is deposited on the Python Package Index (PyPI) repository with documentation hosted on . The open-source project repository is available under the MIT license at GitHub [ 30 ] ( ). The code to reproduce analyses in this paper is available under the MIT license at GitHub, . Processed data, including DeepSTARR [ 9 ], Basset [ 20 ] and ChIP-seq analysis, are available at Zenodo [ 32 ] ( doi.org/10.5281/zenodo.7265991 ). Model weights and code [ 44 ] are also available at Zenodo ( doi.org/10.5281/zenodo.7767325 ).

Genes make up only a small fraction of the human genome. Between them are wide sequences of DNA that direct cells when, where, and how much each gene should be used. These biological instruction manuals are known as regulatory motifs. If that sounds complex, well, it is. The instructions for gene regulation are written in a complicated code, and scientists have turned to artificial intelligence to crack it. To learn the rules of DNA regulation, they're using deep neural networks (DNNs), which excel at finding patterns in large datasets. DNNs are at the core of popular AI tools like ChatGPT. Thanks to a new tool developed by Cold Spring Harbor Laboratory Assistant Professor Peter Koo, genome-analyzing DNNs can now be trained with far more data than can be obtained through experiments alone. "With DNNs, the mantra is the more data, the better," Koo says. "We really need these models to see a diversity of genomes so they can learn robust motif signals. But in some situations, the biology itself is the limiting factor, because we can't generate more data than exists inside the cell." If an AI learns from too few examples, it may misinterpret how a regulatory motif impacts gene function. The problem is that some motifs are uncommon. Very few examples are found in nature. To overcome this limitation, Koo and his colleagues developed EvoAug—a new method of augmenting the data used to train DNNs. EvoAug was inspired by a dataset hiding in plain sight—evolution. The process begins by generating artificial DNA sequences that nearly match real sequences found in cells. The sequences are tweaked in the same way genetic mutations have naturally altered the genome during evolution. Next, the models are trained to recognize regulatory motifs using the new sequences, with one key assumption. It's assumed the vast majority of tweaks will not disrupt the sequences' function. Koo compares augmenting the data in this way to training image-recognition software with mirror images of the same cat. The computer learns that a backward cat pic is still a cat pic. The reality, Koo says, is that some DNA changes do disrupt function. So, EvoAug includes a second training step using only real biological data. This guides the model "back to the biological reality of the dataset," Koo explains. Koo's team found that models trained with EvoAug perform better than those trained on biological data alone. As a result, scientists could soon get a better read of the regulatory DNA that write the rules of life itself. Ultimately, this could someday provide a whole new understanding of human health. The research was published in Genome Biology.

10.1186/s13059-023-02941-w

Biology

Human eggs remain healthy for decades by putting 'batteries on standby mode'

Elvan Böke, Oocytes maintain ROS-free mitochondrial metabolism by suppressing complex I, Nature (2022). DOI: 10.1038/s41586-022-04979-5. www.nature.com/articles/s41586-022-04979-5 Journal information: Nature

https://dx.doi.org/10.1038/s41586-022-04979-5

https://phys.org/news/2022-07-human-eggs-healthy-decades-batteries.html

Abstract Oocytes form before birth and remain viable for several decades before fertilization 1 . Although poor oocyte quality accounts for most female fertility problems, little is known about how oocytes maintain cellular fitness, or why their quality eventually declines with age 2 . Reactive oxygen species (ROS) produced as by-products of mitochondrial activity are associated with lower rates of fertilization and embryo survival 3 , 4 , 5 . Yet, how healthy oocytes balance essential mitochondrial activity with the production of ROS is unknown. Here we show that oocytes evade ROS by remodelling the mitochondrial electron transport chain through elimination of complex I. Combining live-cell imaging and proteomics in human and Xenopus oocytes, we find that early oocytes exhibit greatly reduced levels of complex I. This is accompanied by a highly active mitochondrial unfolded protein response, which is indicative of an imbalanced electron transport chain. Biochemical and functional assays confirm that complex I is neither assembled nor active in early oocytes. Thus, we report a physiological cell type without complex I in animals. Our findings also clarify why patients with complex-I-related hereditary mitochondrial diseases do not experience subfertility. Complex I suppression represents an evolutionarily conserved strategy that allows longevity while maintaining biological activity in long-lived oocytes. Main Human primordial oocytes are formed during fetal development and remain dormant in the ovary for up to 50 years. Despite a long period of dormancy, oocytes retain the ability to give rise to a new organism after fertilization. Decline in oocyte fitness is a key contributor to infertility with age 2 . However, little is known about how oocytes maintain cellular fitness for decades to preserve their developmental potential, complicating efforts to understand the declining oocyte quality in ageing women. Oocytes remain metabolically active during dormancy 6 , 7 , and thus must maintain mitochondrial activity for biosynthesis of essential biomolecules 8 . Yet, mitochondria are a major source of ROS, generating them as by-products of mitochondrial oxidative metabolism. Although ROS can function as signalling molecules 9 , at high concentrations ROS promote DNA mutagenesis and are cytotoxic. Indeed, ROS levels are linked to apoptosis and reduced developmental competence in oocytes and embryos 3 , 4 , 5 . However, the mechanisms by which oocytes maintain this delicate balance between mitochondrial activity and ROS production have remained elusive. Mitochondrial ROS in early oocytes Early human oocytes can be accessed only through invasive surgery into ovaries. Therefore, biochemical investigations into oocyte biology have historically been hindered by severe sample limitations. As a consequence, mitochondrial activity in primordial oocytes remains largely unstudied. Here we overcome challenges imposed by human oocytes by utilizing an improved human oocyte isolation protocol recently developed in our laboratory 6 , which we combine with a comparative evolutionary approach using more readily available Xenopus stage I oocytes (both referred to as early oocytes hereafter; Extended Data Fig. 1a,b ). This approach allowed us to generate hypotheses using multi-species or Xenopus -alone analyses, and subsequently test those hypotheses in human oocytes. We began our studies by imaging live early human and Xenopus oocytes labelled with various mitochondrial probes that quantify ROS levels. Neither Xenopus nor human early oocytes showed any detectable ROS signal, whereas mitochondria in somatic granulosa cells surrounding the oocytes exhibited ROS and served as positive controls (Fig. 1a–c and Extended Data Fig. 1c–g ). ROS induction in oocytes also served as a positive control for live ROS probes (Extended Data Fig. 1h,i ). Fig. 1: Early oocytes have undetectable levels of ROS. a , Live-cell imaging of human and Xenopus early oocytes, both with attached granulosa cells. The ROS level was measured using MitoTracker Red CM-H 2 XRos (H2X), a reduced mitochondrial dye that does not fluoresce until it is oxidized by ROS. The boxed area is magnified in the top right image. Xenopus granulosa cells were imaged at the basal plane of the oocyte. DIC, differential interference contrast. Scale bars, 15 µm (human oocytes), 50 µm ( Xenopus oocytes), 3 µm (human granulosa cells) and 10 µm ( Xenopus granulosa cells). b , c , Quantification of the mean fluorescence intensity (MFI) of H2X in the oocyte and in the population of granulosa cells surrounding the equatorial plane of the oocyte for human ( b ) and Xenopus ( c ) oocytes. The data represent the mean ± s.e.m. of three biological replicates, shown in different colours. ** P = 0.0001 and *** P = 4.13 × 10 −11 using a two-sided Student’s t -test. d , Overnight survival of oocytes at the indicated stages of oogenesis after treatment with menadione, N -acetyl cysteine (NAC) or the combination of both (see Extended Data Fig. 1j for experimental design). At least ten oocytes were incubated per condition. The data represent the mean ± s.e.m. across four biological replicates. * P = 1.94 × 10 −9 , ** P = 3.77 × 10 −18 and *** P = 2.37 × 10 −19 compared with the untreated condition using a two-sided Student’s t -test with Šidák–Bonferroni-adjusted P values for multiple comparisons. Source data Full size image To distinguish between the possibilities that low ROS probe levels resulted from low ROS production or, alternatively, a high scavenging capacity to eliminate ROS, we treated Xenopus oocytes with menadione and assessed their survival (Extended Data Fig. 1j ). Mild treatment with menadione promotes the formation of ROS (ref. 10 ) but does not affect survival negatively in cell lines and fruit flies 11 , 12 . However, most early oocytes (78.3%) died when they were left to recover overnight after menadione treatment, in contrast to what was observed for late-stage oocytes (Fig. 1d and Extended Data Fig. 1j). Treatment with an antioxidant that quenches ROS was able to rescue oocyte survival (Fig. 1d ). These results indicate that evasion of ROS damage in oocytes involves tight control of ROS generation, rather than a higher scavenging capacity of oocytes against ROS. Mitochondrial respiration in oocytes Using dyes that sense membrane potential (tetramethylrhodamine ethyl ester perchlorate (TMRE) and the cyanine dye JC-1), we found that mitochondria in human and Xenopus early oocytes exhibit lower membrane potentials compared to those of neighbouring granulosa cells, which served as positive controls (Fig. 2a,b and Extended Data Fig. 2a–d ). Undetectable ROS levels and low membrane potential suggest that the mitochondrial electron transport chain (ETC) activity in early oocytes is either low or absent. To differentiate between these two possibilities, we measured respiration rate in Xenopus oocytes. Early oocytes stripped of granulosa cells exhibited a low basal respiration rate but a similar maximal respiration rate compared to those of growing oocytes (Fig. 2c and Extended Data Fig. 2e,f). This respiration was efficiently coupled to ATP synthesis, resulting in an undetectable proton leak (Extended Data Fig. 2e ). Therefore, we conclude that mitochondria in early oocytes have a functional ETC, with low activity. Fig. 2: OXPHOS is low, but essential, in early oocytes. a , b , Live-cell imaging of human and Xenopus early oocytes with attached granulosa cells labelled with tetramethylrhodamine ethyl ester perchlorate (TMRE) to detect mitochondrial membrane potential (Δ Ψ m ; a ) and JC-1, a membrane potential sensitive binary dye ( b ). Green JC-1 fluorescence is a sign of low membrane potential; red fluorescence indicates JC-1 aggregation inside mitochondria, and thus high membrane potential. The insets in the Xenopus images show granulosa cells imaged in the basal plane of the oocyte. DIC, differential interference contrast. Scale bars, 10 µm (human oocytes) and 50 µm ( Xenopus oocytes). Representative images are shown (see Extended Data Fig. 2 for quantification of independent experiments). c , The basal oxygen consumption rate in early (stage I) and growing (stage III) Xenopus oocytes, normalized for total protein per sample ( n = 17 for stage I and n = 43 for stage III). The data represent mean ± s.e.m. *** P = 2.98 × 10 −8 using a two-sided Student’s t -test. d , Overnight survival of early (stage I) and late (stage VI) oocytes after treatments with mitochondrial poisons: complex I (CI) to V (CV) inhibitors and an ionophore (5 µM rotenone, 50 mM malonic acid, 5 µM antimycin A, 50 mM KCN, 200 µM N , N ′-dicyclohexylcarbodiimide (DCCD) or 30 µM carbonyl cyanide m -chlorophenyl hydrazone (CCCP), respectively). At least 50 early and 10 late-stage oocytes were incubated per condition. Δ Ψ m , mitochondrial membrane potential. The data represent the mean ± s.e.m. across three biological replicates. Source data Full size image To assess the importance of individual complexes of the oxidative phosphorylation (OXPHOS) machinery for oocyte health, we exposed Xenopus oocytes to inhibitors specific for each OXPHOS complex. We found that both early and late-stage oocytes died after treatment with inhibitors of complexes II, III, IV and V (malonate, antimycin A, KCN and N , N ′-dicyclohexylcarbodiimide (DCCD), respectively). Although late-stage oocytes died after treatment with the complex I inhibitor rotenone, 78% of early oocytes survived exposure to rotenone (Fig. 2d and Extended Data Fig. 2g ). The insensitivity of early oocytes to complex I inhibition indicates that they do not utilize complex I as an essential entry port for electrons. Mitochondrial proteome in oocytes Mitochondria in early oocytes have an apparent lack of ROS, low membrane potential, low basal respiration rates and rotenone resistance in culture. We next investigated the mechanistic basis of this unusual mitochondrial physiology. To do this, we purified mitochondria from early and late-stage Xenopus oocytes isolated from wild-type outbred animals, and performed proteomics using isobaric-tag-based quantification including muscle mitochondria as a somatic cell control (Extended Data Fig. 3a ). Our efforts identified 80% of all known mitochondrial proteins (Extended Data Fig. 3b,c and Supplementary Table 1 ). Most ETC subunits showed a lower absolute abundance in early oocytes compared to that in late-stage oocytes (Fig. 3a ), and to muscle (Extended Data Fig. 3d ), which is expected owing to the presence of fewer cristae in mitochondria of early oocytes 13 , 14 , 15 and compatible with their NADH levels 16 . In support of our findings with the ETC inhibitors (Fig. 2d and Extended Data Fig. 2g ), the depletion of complex I in early oocytes was the most pronounced of all ETC complexes (Fig. 3a and Extended Data Fig. 3e ). We reinforced this result by repeating proteomics with heart, liver and white adipose tissues (Extended Data Fig. 3f–h and Supplementary Table 2 ). Fig. 3: The mitochondrial proteomes of Xenopus and human oocytes. a , A volcano plot showing P values versus fold changes of mitochondrial proteins between early (stage I) and late (stage VI) oocytes. The subunits of the mitochondrial OXPHOS machinery are indicated in colour, according to the key in the plot. Other mitochondrial proteins significantly changing ( q value < 0.05, >1.5-fold change) are depicted in black. n = 3 outbred animals, P values were calculated using two-sided Student’s t -test, and q values were obtained by multiple-comparison adjustment. b , The early Xenopus oocyte proteome ranked by protein abundance. The inset shows data for the top 5% most abundant proteins, corresponding to the grey area of the graph. UPR mt proteins are indicated in red. The data are the mean ± s.e.m. from n = 3 outbred animals. c , The human primordial follicle proteome ranked by protein abundance. The inset shows data for the top 5% most abundant proteins, corresponding to the grey area of the graph. Oocytes were collected from ovaries of two patients and pooled together. UPR mt proteins are indicated in red. d , e , Scatter plots comparing mitochondrial ( d ) and OXPHOS ( e ) protein abundance in human primordial follicles and ovarian somatic cells. The dashed line represents the identity line x = y and the solid line shows the linear regression estimate relating protein abundance between mitochondrial proteomes of primordial follicles and ovarian somatic cells. IM, import machinery. Full size image Furthermore, among the most abundant proteins in the mitochondria of early oocytes were mitochondrial proteases and chaperones (Fig. 3b and Extended Data Figs. 3i,j and 4a ). These proteins are upregulated after the activation of the mitochondrial unfolded protein response (UPR mt ) 17 , 18 , 19 , which is often triggered by an imbalance of ETC subunits in mitochondria. Consistent with an active UPR mt (ref. 20 ), nuclear transcripts encoding complex I subunits were downregulated in early oocytes whereas mitochondrially encoded transcripts of complex I did not show significant changes compared to those of late-stage oocytes (Extended Data Fig. 3k,l ). We next examined whether complex I subunits were also depleted in human oocytes. Early oocytes and ovarian somatic cells were isolated from ovarian cortices of patients, and analysed by label-free proteomics. We identified 40% of all known mitochondrial proteins (Supplementary Table 3 ). The upregulation of proteins related to UPR mt was conserved in human early oocytes, and further confirmed with immunofluorescence (Fig. 3c and Extended Data Fig. 4b ). An analysis of the OXPHOS machinery comparing oocytes and ovarian somatic cells revealed that, in line with the Xenopus data, many complex I subunits were either at very low levels or not identified in human oocytes (Fig. 3d,e and Extended Data Fig. 5a ). In conclusion, our proteomic characterization of mitochondria revealed an overall reduction of ETC subunits in early oocytes of human and Xenopus , with complex I levels exhibiting the strongest disproportionate depletion. Absence of complex I in early oocytes Taken together, the results of our proteomics and survival experiments suggest that both early human and Xenopus oocytes remodel their ETC to decrease complex I levels to an extent that complex I becomes unnecessary for survival. This result is unexpected, because no other animal cell type with functioning mitochondria has been shown to be able to dispense with complex I in physiological conditions, and only one other multicellular eukaryote, the parasitic plant mistletoe, is known to dispense with complex I entirely 21 . Therefore, we directly assayed complex I assembly status and function in early oocytes, using colorimetric, spectrophotometric and metabolic assays. We first investigated the assembly status of complex I in oocytes, which is tightly linked to its function 22 . Complex I is an approximately 1-MDa complex composed of 14 core and 31 accessory subunits in humans, some of which are essential for its assembly and function 23 . We first examined our proteomics data for any specific downregulation of a particular complex I module in early oocytes. However, levels of subunits belonging to the four major functional modules of complex I, namely N, Q, PP and PD modules, were not significantly different between Xenopus early and late-stage oocytes (Extended Data Fig. 6a ). The size of complex I in native protein gels has been used as a tool to reveal the assembly status of the complex 22 , 24 , 25 . Thus, we compared mitochondria isolated from early oocytes to those from late-stage oocytes, and from muscle tissue of Xenopus and mice as somatic cell controls, by blue native polyacrylamide gel electrophoresis (BN-PAGE) followed by complex I in-gel activity assays or by an immunoblot against a complex I core subunit, Ndufs1. Notably, complex I neither was fully assembled nor exhibited any in-gel activity in early oocytes (Fig. 4a and Extended Data Fig. 6b,c ). Denaturing SDS–PAGE gels also verified comparable mitochondrial loading and very low protein levels of complex I subunits in early oocytes (Extended Data Fig. 6d ). To rule out any possibility of immunoblotting detection problems, areas corresponding to assembled complex I and complex II from BN-PAGE gels were analysed by proteomics (Extended Data Fig. 6e ). Although complex II subunits were detected at comparable levels in all samples, most complex I subunits were not detected in early oocytes (Extended Data Fig. 6f and Supplementary Table 4 ). Thus, we conclude that complex I is not fully assembled in early oocytes. Fig. 4: Complex I is not assembled in early oocytes. a , Mitochondrial fractions solubilized in n -dodecyl-β- d -maltoside (DDM) were resolved by BN-PAGE and complex I activity was assayed by reduction of nitro blue tetrazolium chloride (NBT) in the presence of NADH. n ≥ 3 (see Extended Data Fig. 6b for quantifications). b , Spectrophotometric analysis of complex I (green, rotenone-specific activity) and complex IV (orange, KCN-specific activity) activities in mitochondrial extracts from early (stage I) and late (stage VI) oocytes and muscle. cyt c , cytochrome c ; abs, absorbance; a.u., arbitrary units. The data represent the mean ± s.e.m.; n = 3 biological replicates. c , Flavin mononucleotide (FMN) and flavin adenine dinucleotide (FAD) levels in early (stage I) and late (stage VI) Xenopus oocytes. The data represent the mean ± s.e.m.; n = 6. *** P = 6.92 × 10 −9 and ** P = 3.57 × 10 −5 using two-sided Student’s t -test with Šidák–Bonferroni-adjusted P values for multiple comparisons. Source data Full size image In-gel activity assays detect the presence of flavin mononucleotide (FMN)-containing (sub)assemblies of complex I, but do not detect the physiological activity of the assembled complex. Therefore, we measured NADH:CoQ oxidoreductase activity in isolated mitochondrial membranes from early and late-stage oocytes, as well as muscle tissue, to measure substrate consumption by complex I, which reflects physiological activity of complex I (Fig. 4b ). We also measured complex IV and citrate synthase activities to confirm the presence of mitochondrial activity in these samples. Complex IV and citrate synthase activities were detected in all three samples (Fig. 4b and Extended Data Fig. 6g ). However, complex I activity was absent in early oocyte samples, in contrast to the findings for late-stage oocyte samples and muscle samples (Fig. 4b ). Finally, to validate the absence of complex I in early oocytes, we checked the levels of FMN, an integral part of complex I in early and late-stage oocytes. Although levels of another flavin nucleotide, flavin adenine dinucleotide (FAD), were within a 2-fold range between these stages, FMN levels were about 200-fold higher in late-stage oocytes, compared to the low levels detected in early oocytes (Fig. 4c ). The remarkable depletion of FMN is complementary evidence supporting complex I deficiency in early oocytes. The absence of complex I could also explain the reduced activity of other ETC complexes in early oocytes by affecting the stability of supercomplexes 26 . Assessment of supercomplex distribution showed no supercomplex formation in early oocytes, in contrast to the findings in late-stage oocytes and muscle (Extended Data Fig. 6h,i ). Thus, we conclude that the absence of complex I impedes the formation of supercomplexes, which might contribute to the overall reduction of ETC activity in early oocytes. Complex I and ROS throughout oogenesis We then reasoned that an absence of complex I, one of the main ROS generators in the cell, might be sufficient to explain the undetectable ROS levels in early oocytes 27 . Therefore, we studied the relationship between complex I abundance and ROS levels throughout oogenesis. First, we investigated the assembly of complex I during oogenesis. Complex I activity was barely detectable in stage II oocytes, but peaked and plateaued in maturing (stage III) oocytes (Fig. 5a ). We then assessed the survival of oocytes in the presence of rotenone throughout oogenesis. The overnight survival of oocytes in rotenone was consistent with their levels of assembled complex I: stage I and II oocytes survived in the presence of rotenone whereas maturing and mature oocytes died (Fig. 5b ). Hence, we conclude that complex I is assembled and fully functional in maturing (stage III) and late-stage oocytes but absent in early oocytes. Fig. 5: Complex I and ROS levels correlate throughout oogenesis. a , Mitochondrial fractions from early (stage I), maturing (stage II and stage III) and late-stage (stage VI) Xenopus oocytes and muscle solubilized in n -dodecyl-β- d -maltoside (DDM) were resolved by BN-PAGE and complex I activity was assayed. One representative gel from three independent experiments is shown. CS, citrate synthase. b , Overnight survival of early (stage I), maturing (stage II and III) and late-stage (stage VI) Xenopus oocytes after treatment with the complex I inhibitor rotenone (5 µM). At least 10 oocytes were incubated per condition. The data represent the mean ± s.e.m.; n = 3 biological replicates. * P = 0.0028 and ** P = 0.0002 using two-sided Student’s t -test with Šidák–Bonferroni-adjusted P values for multiple comparisons. c , Prdx3 dimer/monomer ratio assessed in oocytes in the indicated stages of oogenesis. The data represent the mean ± s.e.m; n = 4 biological replicates. NS, not significant ( P = 0.1128), * P = 0.0376 and ** P = 0.0164 using two-sided Student’s t -test with Šidák–Bonferroni-adjusted P values for multiple comparisons. Source data Full size image Second, we investigated whether the assembly of complex I throughout oogenesis was accompanied by the production of ROS in oocytes. The opacity of maturing Xenopus oocytes impedes the use of most fluorescent ROS markers. Therefore, we turned to known metabolic and protein 'sentinels' of ROS levels in cells and evaluated the redox state of glutathione 28 , 29 , 30 and mitochondrial peroxiredoxin 3 (Prdx3) in oocytes. We found that ratio of reduced glutathione to oxidized glutathione was 20-fold higher in early oocytes compared to that in late-stage oocytes (Extended Data Fig. 7a ), indicating a reduced cellular redox state in early oocytes, consistent with oocytes having undetectable levels of ROS. Next, we checked the redox state of Prdx3 during oogenesis. Peroxiredoxins dimerize in the presence of peroxide, and thus, the ratio of peroxiredoxin dimers to monomers correlates with the level of cellular peroxide 31 , 32 , 33 . Prdx3 dimerization increased throughout oogenesis, from negligible levels in early (stage I) oocytes to the highest measured level in late-stage (stage VI) oocytes (Fig. 5c and Extended Data Fig. 7b ). Stage II oocytes, in which complex I activity is very low (Fig. 5a ), showed a nonsignificant increase in dimer/monomer ratio (Fig. 5c ). The timing of complex I assembly and the increase in ROS levels correlate: ROS start to build up as soon as complex I is assembled in oocytes. On the basis of these results, we speculate that the maturation of oocytes involves a slow, gradual transition to a metabolism that involves a functional complex I. Combining the in vivo evidence with proteomics and biochemical assays in vitro, our results demonstrate that early oocytes avoid ROS by eliminating one of the main ROS generators in the cell, mitochondrial complex I. Complex I subunits are reduced to such low levels that complex I cannot be fully assembled, nor can its activity be detected in early oocytes. This reveals a new strategy used by Xenopus and most likely human oocytes to maintain a low-ROS-producing mitochondrial metabolism. Although quiescence is associated with ETC remodelling in Drosophila oocytes 34 , to our knowledge, vertebrate early oocytes are the first and only physiological cell type in animals that exist without a functional mitochondrial complex I. Discussion Here we have shown that dormancy involves survival with an inactive mitochondrial complex I. By shutting down complex I and keeping the rest of the OXPHOS system active, early oocytes keep their mitochondria polarized to support the synthesis of haeme, essential amino acids and nucleotides, while keeping their activity low to avoid ROS. Other quiescent cells, such as neuronal and haematopoietic stem cells, exhibit similarly low ROS levels, and reduced ETC activity 9 , 35 , raising the possibility that this regulatory mechanism might be utilized by other cell types. Furthermore, UPR mt is activated in early oocytes (Fig. 3b,c and Extended Data Fig. 4 ), probably in response to an imbalance of ETC complexes caused by the absence of complex I. Given that UPR mt activation itself is sufficient to increase the lifespan of Caenorhabditis elegans and mouse 17 , 18 , 19 , we speculate that complex I inhibition further enhances the longevity of oocytes through its downstream activation of UPR mt . The causal relationships between these interacting factors and oocyte lifespan remain a fascinating future direction to investigate. Severe sample limitations prevent biochemical assays of human oocytes—30 thousand donor ovaries would be required for one experiment to directly measure complex I function using current technologies. Ideally, future methodological developments will allow direct evaluation of complex I activity in human oocytes. It would also be interesting to investigate whether similar mechanisms apply in the oocytes of other mammals such as mice. Until then, we rely on proteomics, imaging and the activation of downstream pathways (UPR mt ) that suggest that complex I is also absent in human primordial oocytes. Moreover, the absence of complex I in early oocytes can also explain why complex-I-related mitochondrial pathologies (such as Leber's hereditary optic neuropathy) do not lead to subfertility or selection against homoplasmic mitochondrial DNA mutations that occur in other types of ETC dysfunction 36 , 37 , 38 . As the oogenic mitochondrial bottleneck occurs in early oogenesis 39 , there would not be a selective pressure against mutations affecting an inactive complex I. Our findings reveal yet another unique aspect of physiology that oocytes have evolved to balance their essential function of beginning life with the requirement for longevity. This raises the question whether complex I deficiency in primordial oocytes can be exploited for other purposes. Some cancers seen in young women are highly treatable; however, their treatment leads to a severe reduction of the ovarian reserve and reduced prospects of motherhood. Drugs against complex I exist, and are already proposed for cancer treatments 40 . Future studies will show whether repurposing complex I antagonists can improve chemotherapy-related infertility, and thus life quality of young female cancer survivors. Methods Ethics Ethical committee permission to work with primordial oocytes from human ovary samples was obtained from the Comité Étic d’Investigació Clínica CEIC-Parc de Salut MAR (Barcelona) and Comité Ético de Investigación Clínica–Hospital Clínic de Barcelona with approval number HCB/2018/0497. Written informed consent was obtained from all participants before their inclusion in the study. Animals used in this study were housed in the Barcelona Biomedical Research Park, accredited by the International Association for Assessment and Accreditation of Laboratory Animal Care. Animal euthanasia was performed by personnel certified by the competent authority (Generalitat de Catalunya) and conformed to the guidelines from the European Community Directive 2010/63 EU, transposed into Spanish legislation on RD 53/2013 for the experimental use of animals. Animal models Xenopus laevis adult females of between 2 and 4 years old were purchased from Nasco and maintained in water tanks in the following controlled conditions: 18–21 °C, pH 6.8–7.5, O 2 4–20 ppm, conductivity 500–1,500 µs, ammonia <0.1 ppm. The C57BL/6J mice used in the experiments were purchased from Charles River Laboratories and maintained in the Animal Facility of the Barcelona Biomedical Research Park under specific-pathogen-free conditions at 22 °C with 40–60% humidity, in a 12 h light/dark cycle, and with access to food and water ad libitum. Female mice of 7 weeks of age were used for extracting muscle tissue. Oocyte isolation and culture Human primordial oocytes Ovaries were provided by the gynaecology service of Hospital Clinic de Barcelona, from women aged 19 to 34 undergoing ovarian surgery and were processed as previously described 6 . Briefly, ovarian cortex samples were digested in DMEM containing 25 mM HEPES and 2 mg ml −1 collagenase type III (Worthington Biochemical, LS004183) for 2 h at 37 °C with occasional swirling. Individual cells were separated from tissue fragments by sedimentation, and collagenase was neutralized by adding 10% FBS (Thermo, 10270106). Follicles were picked manually under a dissecting microscope. All human oocyte imaging experiments were conducted in DMEM/F12 medium (Thermo, 11330-032) containing 15 mM HEPES and 10% FBS (Thermo, 10270106). Xenopus oocytes Ovaries were dissected from young adult (aged 3 to 5 years) female X. laevis that had undergone euthanasia by submersion in 15% benzocaine for 15 min. Ovaries were digested using 3 mg ml −1 collagenase IA (Sigma, C9891-1G) in Marc's modified Ringer's (MMR) buffer by gentle rocking until dissociated oocytes were visible, for 30 to 45 min. The resulting mix was passed through two sets of filter meshes (Spectra/Mesh, 146424 and 146426). All washes were performed in MMR. For live-imaging experiments with intact granulosa cells, oocytes were transferred to oocyte culture medium (OCM) 41 at this stage. For the rest of the experiments, oocytes were stripped of accompanying granulosa cells by treatment with 10 mg ml −1 trypsin in PBS for 1 min, followed by washes in MMR. Removal of granulosa cells was confirmed by Hoechst staining of a small number of oocytes. HeLa cell culture HeLa cells were obtained from ATCC (CCL2), authenticated based on morphological inspection and confirmed to be mycoplasma negative. Cells were grown in DMEM (Thermo, 41965039) supplemented with 1 mM sodium pyruvate (Thermo, 11360070) and 10% FBS (Thermo, 102701060). Live-cell imaging Human or Xenopus early oocytes were labelled in their respective culture medium (see above). Human oocytes were imaged using a 63× water-immersion objective (NA 1.20, Leica, 506346) with an incubation chamber maintained at 37 °C and 5% CO 2 . Frog oocytes were imaged using a 40× water-immersion objective (NA 1.10, Leica, 506357) in OCM at room temperature and atmospheric air, unless stated otherwise. All images were acquired using a Leica TCS SP8 microscope with the LAS X software (Leica, v3.5.5.19976). Mean fluorescence intensities in granulosa cells and oocytes were quantified using Fiji software. ROS probes Oocytes and associated granulosa cells were incubated in 500 nM MitoTracker Red CM-H2Xros (Thermo, M7513) for 30 min, 5 µM MitoSOX Red (Thermo, M36008) for 10 min, or 5 µM CellROX for 30 min. Cells were then washed and imaged in 35-mm glass-bottom MatTek dishes in culture medium, except for CellROX labelling, for which MMR was used for imaging to satisfy the manufacturer’s instructions. Mitochondrial membrane potential probes Oocytes and associated granulosa cells were labelled for 30 min in 500 nM tetramethylrhodamine ethyl ester perchlorate (TMRE) (Thermo, T669), or 45 min in 4 µM JC-1 (Abcam, ab141387). Cells were then washed and imaged in 35-mm glass-bottom MatTek dishes. Oxygen consumption rate Oxygen consumption rate (OCR) of Xenopus oocytes was measured using a Seahorse XFe96 Analyser (Agilent) with Seahorse Wave software (Agilent, v2.6). Granulosa-cell-stripped oocytes were placed in XFe96 culture plates immediately after their isolation in Seahorse XF DMEM medium pH 7.4 supplemented with 10 mM glucose, 1 mM pyruvate and 2 mM glutamine (Agilent; 103015-100, 103577-100, 103578-100 and 103579-100). A cartridge was loaded with concentrated inhibitor solution to achieve 5 µM oligomycin, 2 µM carbonyl cyanide 4-(trifluoromethoxy) phenylhydrazone or a combination of 0.5 µM rotenone and 0.5 µM antimycin A. Mock medium injections were performed to account for inhibitor-independent decline in OCR. For each sequential injection, at least 4 measurement cycles were acquired consisting of 20 s mix, 90 s wait and 3 min measure, in at least 3 replicates. For basal and maximal respiration rates, assay-independent OCR decline was corrected, and non-mitochondrial respiration (resistant to rotenone–antimycin mix) was subtracted. OCR measurements for growing oocytes (stage III; with a diameter of 450–600 µm (ref. 42 )) had to be performed statically because the probe of the analyser compressed and destroyed these large oocytes in long-term measurements. For growing (stage III) oocytes, OCR was acquired during 5 cycles per well, each cycle being 20 s mix, 90 s wait and 3 min measure, in at least 4 replicates. The well size imposed a technical limitation on the maximum number of oocytes per well (100 early and 8 growing oocytes); thus, respiration data were normalized for the total protein amount per sample. Treatments with OXPHOS inhibitors At least 50 stage I and stage II, 20 stage III and 10 stage VI oocytes were assayed per condition. Oocytes were placed in 35-mm glass-bottom dishes (MatTek) and incubated for 16 h at 18 °C in OCM with or without the addition of the indicated mitochondrial inhibitors at the following concentrations: 5 µM rotenone (Sigma, R8875), 50 mM malonic acid (Sigma, M1296), 5 µM antimycin A (Abcam, ab141904), 50 mM potassium cyanide (KCN; Merck Millipore, 1049670100), 200 µM N , N ′-dicyclohexylcarbodiimide (DCCD) (Sigma, D80002) and 30 µM carbonyl cyanide m -chlorophenyl hydrazone (CCCP) (Abcam, ab141229). Survival was assessed by counting the number of oocytes with intact morphology before and after treatments. Cell death in stage III to VI oocytes was recognized by the development of a mottling pattern in the pigmentation 43 . Images were taken by a Leica IC90 E stereoscope. Early (stage I) oocytes were treated with 10 µM menadione (Sigma, M5625) or left untreated, for 2 h in OCM, and washed into fresh OCM. Untreated oocytes were labelled with wheat germ agglutinin 488 (Biotium, 29022-1) to mark their plasma membrane and mixed with menadione-treated oocytes in a glass-bottom MatTek dish 4 h after menadione was removed. The mixed population of oocytes were then labelled with MitoSOX and imaged. At least 50 stage I and II oocytes and at least 10 stage III and VI oocytes were treated with 10 µM menadione (Sigma, M5625) in the presence or in the absence of 10 mM N -acetyl cysteine (NAC) (Sigma, A9165). After 2 h, menadione was removed and N -acetyl cysteine was retained for an overnight incubation. Survival was determined by counting the number of oocytes immediately before menadione treatment ( t = 0) and after 16 h in recovery. Mitochondrial-enriched extracts Mitochondrial-enriched fractions were obtained as described previously for gastrocnemius muscle and with minor adaptations for oocyte samples 44 . Freshly isolated early oocytes from Xenopus were lysed in mitochondria buffer (250 mM sucrose, 3 mM EGTA, 10 mM Tris pH 7.4), and spun at low speed to remove debris. The resulting supernatant was centrifuged at 20,000 g for 20 min at 4 °C. Late-stage oocytes were spin-crashed, and yolk-free fraction was combined 1:1 with mitochondria buffer and centrifuged at 20,000 g for 20 min at 4 °C to pellet mitochondria. Mitochondrial pellets from early and late-stage oocytes were resuspended in mitochondria buffer and subjected to DNase treatment for 10 min and proteinase K treatment for 20 min. Phenylmethylsulfonyl fluoride was added to stop proteolytic activity and samples were centrifuged again at 20,000 g for 20 min at 4 °C. Protein concentration was estimated and aliquots of crude mitochondria were stored at −80 °C until use. Spectrometric assessment of enzymatic activities of mitochondrial complexes The specific activities of mitochondrial complex I, complex IV and citrate synthase were determined as described before with minor modifications 45 . Briefly, mitochondrial extracts were subjected to three freeze–thaw cycles in hypotonic buffer (10 mM Tris-HCl) before activity analysis using an Infinite M200 plate reader (Tecan) with Tecan i-control software (Tecan, v3.23) in black-bottom 96-well plates (Nunc) at 37 °C. For complex I NADH:CoQ activity assessment, reaction solutions (50 mM KP pH 7.5, 3 mg ml −1 BSA, 300 µM KCN and 200 µM NADH) with or without rotenone (10 µM) were distributed into each well first. Mitochondrial extracts were then added and NADH absorbance at 340 nm was measured for 2 min to establish baseline activity. The reaction was then started by the addition of ubiquinone (60 µM). NADH absorbance was recorded for 15 min every 15 s. For complex IV activity assessment, reaction solutions (50 mM KP pH 7, 60 µM reduced cytochrome c ) with or without KCN (600 µM) were distributed into each well first, and absorbance of reduced cytochrome c at 550 nm was recorded for 2 min to establish baseline oxidation. Mitochondrial extracts were then added and absorbance was measured for 15 min every 15 s. For citrate synthase activity, reaction solution (100 µM Tris pH 8, 0.1% Triton X-100, 100 µM DTNB and 300 µM acetyl CoA) was distributed into each well first. Mitochondrial extracts were then added and absorbance at 410 nm was measured for 2 min to set the baseline; then the reaction was started by addition of the substrate oxaloacetic acid (500 µM). Production of TNB (yellow) was recorded by measuring the absorbance at 410 nm for 15 min every 15 s. Enzymatic assays were plotted with the baseline represented as 1 for simplicity. Denaturing SDS gel electrophoresis Oocytes were collected after isolation, frozen in liquid nitrogen and kept at −80 °C until further use. Samples were processed as described previously 46 . Gastrocnemius total homogenates were obtained as described previously 47 . HeLa cells were lysed in RIPA buffer (50 mM Tris-HCl, 150 mM NaCl, 1% Nonidet P-40, 01% SDS and 1 mM EDTA, supplemented with protease inhibitor cocktail (Complete Roche Mini, 1 tablet per 50 ml)) and spun at 20,000 g to eliminate cell debris. Oocyte lysates for determination of the redox state of peroxiredoxin were protected against artefactual oxidation by alkylation as described previously 48 , but in OCM. Cell lysates or mitochondrial-enriched fractions were resolved by SDS–PAGE using 4–12% NuPAGE Bis-Tris gels. BN-PAGE electrophoresis, and in-gel activity assays Mitochondrial content in samples of different cell types (different stages of oocytes and muscle tissue) was first assessed by western blotting for their citrate synthase levels (Supplementary Figs. 1b and 2c,d ). Next, similar amounts of mitochondrial fractions were solubilized in 1% n -dodecyl-β- d -maltoside (DDM) or digitonin, and were resolved in the native state using NativePAGE 3–12% Bis-Tris (Thermo, BN1001BOX) gradient gels as described before 49 . The left part of the gel was cut and stained with Coomassie (InstantBlue, Sigma) after BN-PAGE to reveal the native protein molecular weight marker protein (Supplementary Figs. 1a,b and 2a,c,d ). Complex I and complex IV activity in-gel assays were performed as described previously 24 . Briefly, immediately after the run, BN-PAGE gels were incubated in assay solution: for complex I in 2 mM Tris pH 7.4, 0.1 mg ml −1 NADH and 2.5 mg ml −1 nitro blue tetrazolium chloride (NBT) to asses NADH:FMN electron transfer, denoted by the appearance of dark purple colour; and for complex IV in 10 mM phosphate buffer pH 7.4, 1 mg ml −1 cytochrome c and 0.5 mg ml −1 of 3,3′-diaminobenzidine (DAB) in the presence or absence of 0.6 mM KCN to assess the specific cytochrome c oxidation, denoted by the appearance of dark brown colour. The intensities of reduced nitro blue tetrazolium chloride (NBT) were normalized to citrate synthase levels of the same samples, detected by SDS–PAGE followed by immunoblotting. Gels were imaged using an Amersham Imager (GE Healthcare; Supplementary Figs. 1 and 2 ). Intensity measurements were performed using Fiji software. Immunoblot analysis Denaturing SDS–PAGE gels were transferred to nitrocellulose membranes through wet transfer using a Mini Trans-Blot Cell (Bio-Rad). Membranes were blocked in Intercept (TBS) Blocking Buffer (LI-COR), and incubated overnight at 4 °C with primary antibodies diluted in Intercept 0.05% Tween-20 as follows: anti-ATP5A1 (Abcam; ab14748; 1:1,000), anti-citrate synthase (Abcam, ab96600; 1:1,000), anti-GAPDH (Thermo, AM4300; 1:5,000), anti-HSPE1 (Thermo, PA5-30428; 1:1,000), anti-NDUFB8 (Abcam, ab110242; 1:1,000), anti-NDUFS1 (Abcam, ab169540; 1:1,000), anti-PRDX3 (Abcam, ab73349; 1:1,000) and anti-SDHB (Abcam, ab14714; 1:1,000). Primary antibodies were washed with TBS-T (0.05% Tween-20) and membranes were incubated in the secondary antibodies anti-mouse IgG DyLight 680 (Thermo, 35518; 1:10,000) or anti-rabbit IgG DyLight 800 4× PEG (Thermo, SA5-35571; 1:10,000). After washing, membranes were imaged by a near-infrared imaging system (Odyssey LI-COR) with Image Studio software (Li-COR, v5.2; Supplementary Figs. 1 and 2 ). Densitometric analysis of immunoblotting images was performed using Fiji software. BN-PAGE gels were transferred to polyvinylidene fluoride (PVDF) membranes using a Mini Trans-Blot Cell (Bio-Rad). After wet transfer, polyvinylidene fluoride (PVDF) membranes were destained in methanol, blocked and incubated with antibodies against NDUFS1 (Abcam, ab169540; 1:1,000) and ATP5A1 (Abcam, ab14748; 1:1,000) for complex I and complex V immunodetection, respectively (Supplementary Fig. 2a ). Transcript levels RNA from early oocytes and spin-crashed yolk-free late-stage oocyte lysates was extracted using TRI reagent (Sigma, T9424) followed by RNeasy and Oligotex mRNA column (Qiagen) purification, following the manufacturer’s instructions. cDNA was synthesized with a First Strand cDNA synthesis kit (Thermo, K1612). Quantitative real-time PCR was performed using SYBR Green I Master (Roche, 04887352001) in a LightCycler 480 with LightCycler software v1.5.1 (Roche); with the following pairs of primers: ndufs1 forward: 5′-GGTGCGGTATGATGATGTGG-3′, reverse: 5′-ACAGCTTTCACACACTTGGC-3′; ndufs5 forward: 5′-GTCCGAAAGTTGTGCAGTCA-3′, reverse: 5′-CGGATCTGCCCAATTCCATG-3′; ndufv2 forward: 5′-GCATACAATGGAGCAGGTGG-3′, reverse: 5′-CATCCATGCTGTCTCTGTGC-3′; mt-nd3 forward: 5′-ATTTGATCCTCTGGGCTCTG-3′, reverse: 5′-AGCGCAATCTCTAGGTCAAA-3′; mt-nd5 forward: 5′-GGTCATCCACGATCAAATCCA-3′, reverse: 5′-ACCGAAACGATAATAGCTGCC-3′; gapdh forward: 5′-AGTTATCCCTGAGCTGAACG-3′, reverse: 5′-CTGATGCAGTCTTAATGGCG-3′; mt-rnr2 forward: 5′-ACTACCCGAAACTAAGCGAG-3′, reverse: 5′-ATCTTCCCACTCTTTTGCCA-3′. Nuclear-DNA-encoded genes were normalized to gapdh levels and mitochondrial-DNA-encoded genes were normalized to mt-rnr2 . Measurement of FMN and glutathione Samples were prepared using the automated MicroLab STAR system from Hamilton Company in the presence of recovery standard for quality control by Metabolon. After protein precipitation in methanol, metabolites were extracted and analysed by ultrahigh-performance liquid chromatography with tandem mass spectrometry by negative ionization. Raw data were extracted, peak-identified and processed for quality control using Metabolon’s hardware and software. Immunostaining paraffin ovary sections Human and frog ovaries were fixed in 4% PFA in PBS overnight at 4 °C, washed, embedded in paraffin blocks and cut into 5 µM sections. After deparaffinization, antigen retrieval was performed by heating the slides for 15 min in 10 mM sodium citrate at pH 6. Sections were blocked and permeabilized in 3% BSA, 0.05% Tween-20 and 0.05% Triton X-100 for 1 h at room temperature. Sections were incubated overnight at 4 °C in the presence of primary antibodies (1:100): anti-ATP5A1 (Abcam, ab14748) and anti-HSPE1 (Thermo, PA5-30428); then 2 h at room temperature with secondary antibodies (1:500). Antibodies and dyes used were as follows: goat anti-rabbit Alexa488 or Alexa555 (1:500, Thermo, A-11008, A-21428), goat anti-mouse Alexa647 (Thermo, A21236) and Hoechst dye (1:500, Abcam, ab145597). A droplet of mounting medium (Agilent, S302380) was added onto the section before imaging using the LAS X software (Leica, v3.5.5.19976) in a Leica TCS SP8 microscope equipped with 40× (NA 1.30, Leica 506358) and 63× (NA 1.40, Leica 506350) objectives. Statistics and reproducibility Sample sizes were chosen based on published studies to ensure reliable statistical testing and to account for variability among outbred populations. Experimental limitations were also taken into account, such as the number of primordial oocytes that could be obtained from human ovaries. All experiments were performed on isolated oocytes or tissues. Sample randomization was performed by two means. First, all outbred frogs used in this study were chosen by blinded animal facility personnel without our knowledge. Second, all isolated oocytes or tissue samples were first grouped together and then randomly distributed to different experimental groups. Blinding during data collection was not required as standard experimental procedures were applied for different groups, such as western blots and immunohistochemistry. Blinding during data analysis was performed in oocyte survival experiments by involving multiple lab members for analysing blinded datasets. Blinding for the analysis of other experiments was not required since the different experimental groups were analysed using the same parameters. All data are expressed as mean ± s.e.m. A simple linear regression was performed to fit a model between the mitochondrial protein abundances of primordial follicle and ovarian somatic cell samples (Fig. 3d,e ). Unpaired two-tailed Student’s t -test was used in all other analysis, P values are specified in figure legends, and those <0.05 were considered significant. Multiple t -tests were used in Figs. 1d , 4c and 5b,c and Extended Data Figs. 2c,d , 3k,l and 6b , and were corrected by the Šidák–Bonferroni method using GraphPad Prism. In Xenopus proteomics experiments, q values were calculated as adjusted P values and significance was considered for q value < 0.05 for comparing protein levels. A fold-change heatmap was generated using JMP (version 13.2) software. For Extended Data Fig. 6f , we excised the indicated bands in Extended Data Fig. 6e from one of three gels represented in Fig. 4a ; gel-identification MS was performed once. MS Sample preparation For isobaric-tag-based quantification for Xenopus , mitochondrial extracts from early (stage I) oocytes, late (stage VI) oocytes, gastrocnemius muscle, heart, liver and white adipose tissues were processed in two parallel experiments: stage I, stage VI and muscle in triplicates; and stage I, heart, liver and white adipose tissue in duplicates. Samples were quantified and 100 μg of each sample was processed with slight modifications from ref. 46 . In brief, methanol-precipitated proteins were dissolved in 6 M guanidine hydrochloride (GuaCl). Samples were then digested with LysC (20 ng µl −1 ) in 2 M GuaCl overnight at room temperature. The next morning, samples were further diluted to 0.5 M GuaCl and digested with trypsin (10 ng µl −1 ) and further LysC (20 ng µl −1 ) for 8 h at 37 °C. Later, samples were speed-vacuumed, and the resulting pellet was resuspended in 200 mM EPPS pH 8.0. Ten-microlitre volumes of tandem mass tag (TMT) stock solutions (20 µg µl −1 in acetonitrile) were added to 50 μl of samples, and samples were incubated 3 h at room temperature. The TMT reaction was quenched with a 0.5% final concentration of hydroxylamine. The samples were combined in one tube, acidified by 10% phosphoric acid, and subjected to a MacroSpin C18 solid-phase extraction (The Nest Group) to desalt and isolate peptides. TMT mixes were fractionated using basic pH reversed-phase fractionation in an Agilent 1200 system. Fractions were desalted with a MicroSpin C18 column (The Nest Group) and dried by vacuum centrifugation 50 . For label-free proteomics for human oocytes, human primordial follicles and ovarian somatic cells were collected from two individuals who underwent ovarian surgery. Samples were dissolved in 6 M GuaCl pH 8.5, diluted to 2 M GuaCl and digested with LysC (10 ng µl −1 ) overnight. Samples were further diluted down to 0.5 M GuaCl and digested with LysC (10 ng µl −1 ) and trypsin (5 ng µl −1 ) for 8 h at 37 °C. Samples were acidified by 5% formic acid and desalted with home-made C18 columns. For detection of complex I and II subunits from BN-PAGE gels, gel bands were destained, reduced with dithiothreitol, alkylated with iodoacetamide and dehydrated with acetonitrile for trypsin digestion. After digestion, peptide mix was acidified with formic acid before analysis through liquid chromatography with MS/MS. Chromatographic and MS analysis TMT and label-free samples were analysed using a Orbitrap Eclipse mass spectrometer (Thermo) coupled to an EASY-nLC 1200 (Thermo). Peptides were separated on a 50-cm C18 column (Thermo) with a gradient from 4% to 32% acetonitrile in 90 min. Data acquisition for TMT samples was performed using a Real Time Search MS3 method 51 . The scan sequence began with an MS1 spectrum in the Orbitrap. In each cycle of data-dependent acquisition analysis, following each survey scan, the most intense ions were selected for fragmentation. Fragment ion spectra were produced through collision-induced dissociation at a normalized collision energy of 35% and they were acquired in the ion trap mass analyser. MS2 spectra were searched in real time with data acquisition using the PHROG database 52 with added mitochondrially encoded proteins. Identified MS2 spectra triggered the submission of MS3 spectra that were collected using the multinotch MS3-based TMT method 53 . Label-free samples were acquired in data-dependent acquisition mode and full MS scans were acquired in the Orbitrap. In each cycle of data-dependent acquisition analysis, the most intense ions were selected for fragmentation. Fragment ion spectra were produced through high-energy collision dissociation at a normalized collision energy of 28%, and they were acquired in the ion trap mass analyser. Gel bands were analysed using a LTQ-Orbitrap Velos Pro mass spectrometer (Thermo) coupled to an EASY-nLC 1000 (Thermo). Peptides were separated on a 25-cm C18 column (Nikkyo Technos) with a gradient from 7% to 35% acetonitrile in 60 min. The acquisition was performed in data-dependent acquisition mode and full MS scans were acquired in the Orbitrap. In each cycle, the top 20 most intense ions were selected for fragmentation. Fragment ion spectra were produced through collision-induced dissociation at a normalized collision energy of 35%, and they were acquired in the ion trap mass analyser. Digested bovine serum albumin was analysed between each sample and QCloud (ref. 54 ) was used to control instrument performance. Data analysis Acquired spectra were analysed using the Proteome Discoverer software suite (v2.3, Thermo) and the Mascot search engine (v2.6, Matrix Science 55 ). Label-free data were searched against the SwissProt Human database. Data from the gel bands were searched against a custom PHROG database 52 that includes 13 further entries that correspond to mitochondrially encoded proteins for the Xenopus samples and the SwissProt mouse database for the mouse samples. TMT data were searched against the same custom 'PHROG' database. False discovery rate in peptide identification was set to a maximum of 5%. Peptide quantification data for the gel bands and the label-free experiments were retrieved from the 'Precursor ion area detector' node. The obtained values were used to calculate an estimation of protein amount with the top3 area, which is the average peak area of the three most abundant peptides for a given protein. For the TMT data, peptides were quantified using the reporter ion intensities in MS3. Reporter ion intensities were adjusted to correct for the isotopic impurities of the different TMT reagents according to the manufacturer's specifications. For final analysis, values were transferred to Excel. For all experiments, identified proteins were selected as mitochondrial if they were found in MitoCarta 3.0 (ref. 56 ). MS3 spectra with abundance less than 100 or proteins with fewer than 2 unique peptides were excluded from the analysis. Each TMT channel was normalized to total mitochondrial protein abundance. A total of 926 mitochondrial proteins were identified (and 807 quantified) in 3 biological replicates from wild-type outbred animals, representing 80% of known mitochondrial proteins (Supplementary Table 1 and Extended Data Fig. 3b ). Although the mitochondrial proteome in diverse cell types could be quite different 57 , we found comparable levels of mitochondrial housekeeping proteins (such as the import complexes TIMMs and TOMMs) across different maturity stages (Extended Data Fig. 3c and Supplementary Table 1 ), enabling us to compare and contrast changes in other pathways. For human somatic cell samples, we analysed three dilutions: the 1× reference had a similar level of protein loading to that of the primordial follicle sample (0.55 µg total protein); a twofold dilution (0.25 µg total protein); and a fivefold dilution (0.1 µg total protein). In scatter plots (Fig. 3d,e ), we estimated differences in mitochondrial complex I protein abundance using the twofold somatic cell dilution, a conservative approach that compared primordial follicle samples (0.55 µg total protein) to somatic cells half their loading concentration (0.25 µg total protein), nevertheless observing similar levels of the mitochondrial import machinery subunits TOMMs and TIMMs. The fivefold-dilution somatic cell sample was useful for establishing detection limits; indeed, many complex I subunits absent in oocytes were detected with high confidence even at this dilution. In the heatmap (Extended Data Fig. 5 ), we considered normalizing our data using the mitochondrial loading controls citrate synthase and COX4I1 to estimate differences in protein abundance. The abundance of COX4I1 fell within the linear range of our proteomic methodology ( R 2 = 0.99), in contrast to that for citrate synthase ( R 2 = 0.89) whose higher abundance led to measurement saturation at higher concentrations. Therefore, COX4I1 was chosen to normalize protein abundances in the heatmap representation. We identified 454 mitochondrial proteins (Supplementary Table 3 ; 298 and 397 proteins were quantified for early oocyte and somatic cell samples, respectively), representing 40% of all known mitochondrial proteins. Here too, levels of the mitochondrial import proteins TIMMs and TOMMs were similar between oocytes and ovarian somatic cells (Fig. 3d,e ), demonstrating an equivalent mitochondrial abundance that facilitated comparison of protein levels between different cell types. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this paper. Data availability Isobaric-tag-based quantification data shown in Fig. 3 , Extended Data Fig. 3 and Supplementary Tables 1 and 2 are available through PRIDE (ref. 58 ) with the identifiers PXD025366 and PXD030576 . Label-free data shown in Fig. 3 , Extended Data Fig. 5 and Table 3 are available through PRIDE (ref. 58 ) with the identifier PXD025369 . Data for the gel band identification in Extended Data Fig. 6 and Supplementary Table 4 are available through PRIDE (ref. 58 ) with the identifier PXD025371 . Source data are provided with this paper.

Immature human egg cells skip a fundamental metabolic reaction thought to be essential for generating energy, according to the findings of a study by researchers at the Center for Genomic Regulation (CRG) published today in the journal Nature. By altering their metabolic activity, the cells avoid creating reactive oxygen species, harmful molecules that can accumulate, damage DNA and cause cell death. The findings explain how human egg cells remain dormant in ovaries for up to 50 years without losing their reproductive capacity. "Humans are born with all the supply of egg cells they have in life. As humans are also the longest-lived terrestrial mammal, egg cells have to maintain pristine conditions while avoiding decades of wear-and-tear. We show this problem is solved by skipping a fundamental metabolic reaction that is also the main source of damage for the cell. As a long-term maintenance strategy, its like putting batteries on standby mode. This represents a brand new paradigm never before seen in animal cells," says Dr. Aida Rodriguez, postdoctoral researcher at the CRG and first author of the study. Human eggs are first formed in the ovaries during fetal development, undergoing different stages of maturation. During the early stages of this process, immature egg cells known as oocytes are put into cellular arrest, remaining dormant for up to 50 years in the ovaries. Like all other eukaryotic cells, oocytes have mitochondria—the batteries of the cell—which they use to generate energy for their needs during this period of dormancy. Using a combination of live imaging, proteomic and biochemistry techniques, the authors of the study found that mitochondria in both human and Xenopus oocytes use alternative metabolic pathways to generate energy never before seen in other animal cell types. A complex protein and enzyme known as complex I is the usual "gatekeeper" that initiates the reactions required to generate energy in mitochondria. This protein is fundamental, working in the cells that constitute living organisms ranging from yeast to blue whales. However, the researchers found that complex I is virtually absent in oocytes. The only other type of cell known to survive with depleted complex I levels are all the cells that make up the parasitic plant mistletoe. According to the authors of the study, the research explains why some women with mitochondrial conditions linked to complex I, such as Leber's Hereditary Optic Neuropathy, do not experience reduced fertility compared to women with conditions affecting other mitochondrial respiratory complexes. The findings could also lead to new strategies that help preserve the ovarian reserves of patients undergoing cancer treatment. "Complex I inhibitors have previously been proposed as a cancer treatment. If these inhibitors show promise in future studies, they could potentially target cancerous cells while sparing oocytes," explains Dr. Elvan Böke, senior author of the study and Group Leader in the Cell & Developmental Biology program at the CRG. Oocytes are vastly different to other types of cells because they have to balance longevity with function. The researchers plan to continue this line of research and uncover the energy source oocytes use during their long dormancy in the absence of complex I, with one of the aims being to understand the effect of nutrition on female fertility. "One in four cases of female infertility are unexplained—pointing to a huge gap of knowledge in our understanding of female reproduction. Our ambition is to discover the strategies (such as the lack of complex I) oocytes employ to stay healthy for many years in order to find out why these strategies eventually fail with advanced age," concludes Dr. Böke.

10.1038/s41586-022-04979-5

Medicine

'Digital twins'—An aid to tailor medication to individual patients

Danuta R. Gawel et al, A validated single-cell-based strategy to identify diagnostic and therapeutic targets in complex diseases, Genome Medicine (2019). DOI: 10.1186/s13073-019-0657-3 Journal information: Genome Medicine

http://dx.doi.org/10.1186/s13073-019-0657-3

https://medicalxpress.com/news/2019-07-digital-twinsan-aid-tailor-medication.html

Abstract Background Genomic medicine has paved the way for identifying biomarkers and therapeutically actionable targets for complex diseases, but is complicated by the involvement of thousands of variably expressed genes across multiple cell types. Single-cell RNA-sequencing study (scRNA-seq) allows the characterization of such complex changes in whole organs. Methods The study is based on applying network tools to organize and analyze scRNA-seq data from a mouse model of arthritis and human rheumatoid arthritis, in order to find diagnostic biomarkers and therapeutic targets. Diagnostic validation studies were performed using expression profiling data and potential protein biomarkers from prospective clinical studies of 13 diseases. A candidate drug was examined by a treatment study of a mouse model of arthritis, using phenotypic, immunohistochemical, and cellular analyses as read-outs. Results We performed the first systematic analysis of pathways, potential biomarkers, and drug targets in scRNA-seq data from a complex disease, starting with inflamed joints and lymph nodes from a mouse model of arthritis. We found the involvement of hundreds of pathways, biomarkers, and drug targets that differed greatly between cell types. Analyses of scRNA-seq and GWAS data from human rheumatoid arthritis (RA) supported a similar dispersion of pathogenic mechanisms in different cell types. Thus, systems-level approaches to prioritize biomarkers and drugs are needed. Here, we present a prioritization strategy that is based on constructing network models of disease-associated cell types and interactions using scRNA-seq data from our mouse model of arthritis, as well as human RA, which we term multicellular disease models (MCDMs). We find that the network centrality of MCDM cell types correlates with the enrichment of genes harboring genetic variants associated with RA and thus could potentially be used to prioritize cell types and genes for diagnostics and therapeutics. We validated this hypothesis in a large-scale study of patients with 13 different autoimmune, allergic, infectious, malignant, endocrine, metabolic, and cardiovascular diseases, as well as a therapeutic study of the mouse arthritis model. Conclusions Overall, our results support that our strategy has the potential to help prioritize diagnostic and therapeutic targets in human disease. Background One of the most important problems in health care today is that many patients do not respond to drug treatment. According to an FDA report, this affects 38–75% of patients with common diseases [ 1 ]. This problem is closely linked to increasing costs and difficulties in drug development [ 2 ]. One important driver of this high rate of failure is suggested by genome-wide association studies (GWAS), which identify increasing numbers of genetic variants that may affect highly diverse pathways and cell types in the same disease [ 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 ]. While genomic medicine has paved the way for addressing this diversity [ 15 ], the scale of the problem is indicated by single-cell RNA-sequencing (scRNA-seq) studies, which have shown altered expression of thousands of genes across many different cell types [ 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 ] . While such studies have resulted in the identification of potential novel disease mechanisms, no single-cell type, pathway, or gene has been shown to have a key regulatory role in any disease. Instead, the dispersion of multiple causal mechanisms across multiple cell types is supported by several other studies [ 6 , 8 , 9 , 24 ]. An extreme consequence of such complexity could be that a prohibitive number of drugs may be needed for effective treatment of each disease. To address this problem, we would ideally need to (1) characterize all disease-associated cell types and pathways, followed by (2) prioritization of the relatively most important. To our knowledge, neither of these two challenges has been systematically addressed. One reason is that many cell types may not be accessible in patients, and another reason lack of methods to prioritize between the cell types and pathways [ 24 ]. Here, we hypothesized that a solution to systematically investigate multicellular pathogenesis and its diagnostic and therapeutic implications could be to use scRNA-seq data to construct models of disease-associated cell types, their expression profiles, and putative interactions. We will henceforth refer to such models as multicellular disease models (MCDMs). The importance of interactions in an MCDM lies in that they link the cell types into networks. As a simplified example, if the interactions were unidirectional, they could be traced to find upstream cell types and mechanisms for therapeutic targeting. However, biological interactions are often more complex. We therefore hypothesized that network tools could be used to prioritize cell types, mechanisms, and potential drug targets. In support, methods from network science have been applied to analyze genome-wide data from different diseases [ 25 , 26 ]. We and others have used such methods to identify biomarkers and therapeutic targets based on bulk expression profiling data of individual cell types [ 12 , 27 , 28 ], as well as to develop a mathematical framework to rank network nodes [ 29 ]. A core concept is that the most interconnected nodes in a network tend to be most important. Indeed, a large body of evidence supports that such analyses can be formalized and used to find crucial nodes in a wide range of systems, ranging from proteins essential for cell survival to relevant web pages in a Google search [ 30 , 31 ]. Because many cell types are not accessible from patients, we started with a mouse disease model. We focused on a mouse model of antigen-induced arthritis (AIA), because it allows potential analysis of all cells in the target organ, joints, and adjacent lymph nodes. We used our recently developed method for translational scRNA-seq [ 32 ]. The resulting MCDMs and complementary analyses of patients with RA and 174 other diseases supported multicellular pathogenesis of great complexity. Our analyses indicate that network analyses of the MCDMs can help to prioritize cell types and genes for diagnostics and therapeutics. General applicability of our strategy was supported by prospective diagnostic studies of 151 patients with 13 autoimmune, allergic, infectious, malignant, endocrine, metabolic, and cardiovascular diseases, as well as 53 age- and sex-matched controls. The therapeutic potential of the strategy was supported by network-based analyses of these diseases, as well as a study of the mouse model of arthritis. Taken together, our results support that our strategy may have the potential to prioritize therapeutic and diagnostic targets in complex diseases. Methods Study design In summary, this study describes a scalable, step-wise strategy to construct MCDMs and exploit them for diagnostics and therapeutics. The strategy was validated by both clinical and experimental studies. The strategy is based on applying network tools to organize and analyze scRNA-seq data from a mouse model of arthritis and human rheumatoid arthritis. Diagnostic validation studies were performed using expression profiling data and potential protein biomarkers from prospective clinical studies of 13 diseases. A candidate drug was examined by a treatment study of a mouse model of arthritis, using phenotypic, immunohistochemical, and cellular analyses as read-outs. Mouse model of rheumatoid arthritis In order to construct high-resolution multicellular disease models (MCDMs), we first performed scRNA-seq analysis in a mouse model of antigen-induced arthritis (AIA). This model is suitable for generating MCDMs in an in vivo setting because most disease-associated cell types and subsets can be potentially identified in the inflamed joint and adjacent lymph nodes. Arthritis was triggered in six anesthetized female 129/SvE mice (aged 8 to 20 weeks) on day 21 by intra-articular injection of 30 μg methylated bovine serum albumin (mBSA) in 20 μL, in the left knee joint after subcutaneous pre-sensitization with 100 μg mBSA in incomplete Freund’s adjuvant and 200 μg mBSA in complete Freund’s adjuvant two (day 7) and three (day 1) weeks earlier, respectively. The right knee joint was injected with 20 μL phosphate-buffered saline (PBS) as negative control. In some experiments, mBSA was injected in both joints to allow assessment of arthritis and scRNA-seq data from joint cells from the same arthritic animal. One week after intra-articular injection of mBSA (day 28), mice were sacrificed, and joints were either used for immunohistochemistry or scRNA-seq. For assessment of the degree of arthritis, joints were routinely fixed in 4% paraformaldehyde (PFA), decalcified in formic acid/sodium citrate buffer, embedded in paraffin, and cut into 4-μm-thick sagittal sections before hematoxylin and eosin (H&E) staining, as described in [ 33 ]. Specimens were examined in a blinded manner for pannus formation, cartilage and subchondral bone destruction, and synovial hypertrophy on an arbitrary scale, 0–3, as described in [ 33 ]. Arthritis frequency (score one or higher) and arthritis severity (median score) were calculated and compared to non-arthritic controls using the Mann-Whitney U test. All experimental procedures were performed according to the guidelines provided by the Swedish Animal Welfare Act and approved by the Ethical Committee for research on animals in Stockholm, Sweden (N271-14). For scRNA-seq experiments (see below), joints and lymph nodes from naïve (control) and arthritic mice draining the sites used for mBSA injection (axillary and popliteal) were isolated and single-cell suspensions were prepared by triturating the joint and lymph nodes and passing them gently through a 70-μm cell strainer. Red blood cells were lysed by adding RBC lysis solution (Sigma-Aldrich) according to the manufacturer’s instructions. Single-cell RNA-seq wet lab protocol All scRNA-seq experiments were performed using the Seq-Well technique [ 32 ]. Briefly, single-cell suspensions prepared from cultured cells or tissue samples using standard techniques were counted and co-loaded with barcoded and functionalized oligo-dT beads (Chemgenes, Wilmington, MA, USA; cat. no. MACOSKO-2011-10) on microwell arrays synthesized as described in [ 32 ]. For each sample, 20,000 live cells were loaded per array, and libraries from three samples were pooled together for sequencing, resulting in a coverage of 6.6 reads per base. The microwell arrays were then covered with previously plasma-treated polycarbonate membranes, and the membranes were allowed to seal to the bead and cell co-loaded microwell arrays at 37 °C for 30 min. Next, cell lysis and hybridization were performed, followed by bead removal, reverse transcription, and whole transcriptome amplification. Libraries were prepared for each sample using the Nextera XT DNA Library Preparation Kit (Illumina, San Diego, CA, USA; cat. no. FC-131-1096) according to the manufacturer’s instructions. Libraries were sequenced using the NextSeq 500/550 system, and sequencing results were analyzed as described below. Validation of the single-cell RNA-seq analytical process To verify our scRNA-seq setup, we mixed the two colorectal cancer (CRC) cell lines SW480 and HT29 before application to the single-cell array and sequenced them altogether according to the procedures described above. These two CRC cell lines (SW480 and HT29) were a kind gift from Xiaofeng Sun (Linkoping University). We showed that the cells from our previously mixed cell lines were correctly assigned to their corresponding cell lines, verifying our single-cell sequencing and clustering methods (Additional file 1 : Figure S1). Quality check and clustering was performed as described below, using 26 colon cancer cell lines profiled with microarrays (GSE10843). Here, a low cutoff of 4000 unique transcripts per cell was added as a criterion to Seurat (Additional file 1 : Figure S2). In total, 233 cells passed the quality criteria and were separated into two main clusters: SW480 and HT29, as expected. However, a subcluster of SW480, with a profile resembling that of the SW620 CRC cell line, was also identified (Additional file 1 : Figure S1). Single-cell RNA-seq data processing The single-cell data was processed into digital gene expression matrices following James Nemesh, McCarrol’s lab Drop-seq Core Computational Protocol (version 1.0.1, ) using bcl2fastq Conversion and Picard software. The indexed reference for alignment of the reads was generated from GRCh38 (April 2017, Ensembl) for human data (validation of the wet lab and cell type identification protocols; see the “ Methods ” section) and GRCm38 (June 2017, Ensembl) for mouse data using STAR software. Only primary alignments towards the reference genome were considered during downstream analyses, according to the mapping quality using STAR software. The quality of cells was assessed by having a minimum of 10,000 reads, 400 transcripts, and less than 20% mitochondrial genes per cell. Outliers were then removed based on an overestimation of transcripts count, due to the risk of duplicates in the library resulting in two or more cells sharing a cell barcode. This resulted in a total of 7086 and 1333 cells for the joint and lymph node data, respectively. The single-cell data was then normalized using Seurat [ 34 ] for further analysis. To reduce the noise within the data, K -nearest neighbor smoothing was applied for each tissue matrix separately, using a minimum k of 5 or, if more than 5000 cells were captured, ~ 0.1% of the total number of cells [ 35 ]. Cell type identification To cluster the cells and define the cell types, reference component analysis (RCA) was performed [ 18 ]. Each cell was projected against reference bulk expression profiling data, which were generated or derived from public databases, as described for each dataset below. The RCA references were prepared as described in the original paper [ 18 ]. Briefly, all genes with log10 (fold change) expression values greater or equal than 1 in any sample, relative to the median across all samples, were included. For each cell, we also saved the Pearson correlation p value from the RCA algorithm. Next, the reference component features were calculated, and data were clustered in a stepwise procedure. First, cells were clustered as described above. Second, those cells with non-significant p values ( p > 0.05) were removed. Cells within each cluster that were significantly correlated (Benjamini-Hochberg adjusted p < 0.05) with their RCA-predicted cell type were identified accordingly, and cells with non-significant correlations were labeled as undetermined. To construct the reference for the cell type identification in mouse experiments, we used the data from the study GSE10246. The resulting reference contained 5030 genes and 31 cell types/states (Additional file 2 ). t -SNE plots of joint and lymph node cells were created using MATLAB function tsne() , with perplexity parameter set to 40, based on the distance matrix obtained from RCA. Cells were colored according to the clusters identified with RCA. For clustering of the CRC cell lines, a reference was constructed using microarray data from GSE10843 and the same parameters as for the mouse reference. The resulting reference contained 2303 genes and 26 different cell lines (Additional file 2 ). Identification of differentially expressed genes For single-cell data, differentially expressed genes (DEGs) were identified between each cell type and all other cell types, within each tissue separately, using Monocle (version 2.6.1) [ 36 , 37 ]. A negative binomial distribution was used to define the dataset with a lowest detection limit of 0.5. Genes detected in at least three cells within a group were considered as expressed. Genes were considered as significantly differentially expressed if the q value < 0.05. For microarray data, DEGs were identified using the LIMMA R package. Pathway, biomarkers, and drug target enrichment analyses Identification of pathways, biomarkers, and therapeutic targets was done with Ingenuity Pathway Analysis software [ 38 ]. Network construction and centrality analysis Our work included the different analysis of two kinds of cell-cell interaction networks, where one is based on potential spatial interactions and the other on predicted molecular interactions (see details below). The former was used as a proxy for molecular interactions, either because the cell types were not accessible from human patients or to prioritize cell types and tissues for scRNA-seq studies in animal models or human patients. We have made the resulting models of each 175 diseases available for such prioritization. The spatial interaction network models were created by empirical knowledge of which cells could potentially interact in the body. Second, using scRNA-seq data, we derived a refined network models, MCDMs, in which interactions were based on possible regulatory molecular interactions. These were inferred using Ingenuity upstream regulation analysis of the differentially expressed genes in each cell type. We validated the inferred interactions by another, recently described method, which is based on ligand-receptor interactions. However, both cases represent networks of possible physical interactions between cell types. The spatial network represented an undirected average network consisting of more nodes (45 cell types), while the latter resulted in a more sophisticated weighted and directed network of fewer nodes (e.g., six cell types for the joint MCDM from mouse AIA). There are many different methods to measure centrality. Given that we have no information about the kind of paths in the MCDM that will provide the best functional representation of the underlying chain of gene-regulatory processes, and that testing many different methods would involve risk of over-parametrization, we assumed a maximum entropy principle and used the less biased theoretical-information approach that for navigation is random walk centrality [ 39 ]. This metric is based on the average length of an average random walker moving at the network considering the weights, which has also been used by others in molecular biology. As our work aimed for using centrality as a generalizable concept, we also tested other metrics, such as the subgraph closeness centrality that is based on that the flow is concentrated to the closest paths. We found that although less significant similar results held for this metric as well. In support of centrality analyses to prioritize cell types, we found tumor cells to be most central in scRNA-seq data from colorectal cancer and significant correlations between centrality in scRNA-seq data from both mouse AIA and human RA. Construction of MCDMs We constructed the MCDMs using scRNA-seq data from colorectal cancer, mouse AIA (sick or healthy lymph nodes and joints), and human RA synovium. The MCDMs showed genome-wide mRNA expression of each cell type as well as potential types and directions of intercellular interactions. For MCDM construction, we started by identifying cell-type-specific genes, i.e., DEGs in one cell type compared with all others, using the methods described above. Using those gene lists, MCDMs were constructed. First, the Ingenuity Pathway Analysis (IPA) software was queried for prediction of the upstream regulators of cell-type-specific DEGs for each cell type separately. Here, we focused on upstream regulators that were secreted or membrane-bound. Next, we searched for predicted upstream regulators among the DEGs of other cell types. If such an upstream regulator was found, an interaction was assumed between the cell types. To systematically validate the MCDM cellular interactions derived from Ingenuity, we used the novel CellPhoneDB [ 40 ] framework. CellPhoneDB is a publicly available curated repository of ligands, receptors, and their interactions, which integrates a statistical tool for the inference of cell-cell communication networks from human single-cell transcriptomic data. Specifically, cell-type interactions between ligands and receptors from mouse RA and healthy joint MCDMs, and mouse RA lymph node MCDM were analyzed with CellPhoneDB, using the default parameters (the ligands and receptors included should be expressed in at least 10% of the cells for each cluster, and the cluster labels of every cell were randomly permuted 1000 times). As CellPhoneDB is developed for human scRNA-seq data, mouse genes were mapped to human orthologs using the BioMart database [ 41 ]. In total, 6203 (82.2%) and 4808 (87.4%) mouse genes from RA and healthy joint, respectively, could be mapped to humans. An interaction between two cell types was considered significant if the CellPhoneDB analysis predicted any interaction between the cell types with a significance score of p < 0.05. Cell centrality analysis Cell centrality was determined by the random walk centrality and subgraph centrality [ 39 ]. For the scRNA-seq MCDM, directions were derived from the IPA upstream regulatory analysis, and weights were added to the interactions to include biological information for the computation of the coefficient. The weights were based on the number of cells of each cell type and its number of predicted upstream regulators (as described in the “Construction of MCDMs” section). The assumption behind these weights is that the chances of interactions between cell types are likely to increase with the number of cells and upstream regulators. Specifically, the weight for each interaction was derived by multiplying the number of cells by the number of predicted upstream regulators for two interacting cell types. Finally, the centrality of each cell was determined by the subgraph centrality using the normalized weighted adjacency matrix [ 42 ]. In order to validate that centralities were not biased by our choice of the Ingenuity database, we recomputed MCDMs and centrality measures using publicly available ligand-receptor interactions in mice [ 43 ] instead of Ingenuity. Centrality analysis of MCDMs from colorectal tumors The hypothesis behind these analyses was that, since tumor cells have a causal role, they should also be more central than the surrounding immune and stromal cells. To test this hypothesis, processed FPKM (fragments per kilobase of transcript per million reads mapped) values for a scRNA-seq experiment of colon cancer (GSE81861) were downloaded from Gene Expression Omnibus [ 18 ]. T cells, B cells, and undefined cells were re-clustered using RCA, generating a novel reference expression profiling compendium. This consisted of 49 microarrays profiling 11 samples: B cell, CD4+, CD8+, monocytes, natural killer cells, naïve T cells, PBMC, Th1, Th17, Th2, and T regulatory (Treg) cells. Microarrays were processed as described below. RCA reference was prepared as described above, i.e., following instructions in [ 18 ]. Briefly, we included all genes with log10 (fold change) expression values above one in any sample, relative to the median across all samples. For each cell, we saved the correlation p value from the RCA algorithm. Next, we checked if all cells within each cluster were significantly correlated (BH adjusted p < 0.05) with the RCA-predicted cell type from reference data. All cells that did not fulfill this requirement were labeled as undetermined. In total, 579 cells were analyzed (Additional file 2 ). The DEGs were identified using Monocle and a truncated normal distribution ( tobit ) owing to its FPKM format, and the lowest detection limit was set at 0.1. The estimateSizeFactors() function was used to normalize for differences in mRNAs recovered across cells, and genes that were expressed in at least three cells were considered as present within a group. A tumor MCDM was constructed and centrality analyses were performed as described above (Additional file 1 : Figure S3). Generation of an expression profiling reference compendium of immune cells This compendium was used as a reference for classifying cell types in the scRNA-seq data from colorectal cancer and for deconvolution analyses of expression profiling data from CD4+ T cells. PBMCs were isolated from human peripheral blood using Lymphoprep (Axis-Shield Density Gradient Media, Oslo, Norway; cat. no. 1114545) according to the manufacturer’s instructions. Total RNA was extracted from one million PBMCs for microarray analysis. CD4+ T cells were isolated from two thirds of the remaining PBMCs using the CD4+ T Cell Isolation Kit, human (Miltenyi Biotec, Bergisch Gladbach, Germany; cat. no. 130-096-533) and LS Separation Columns (Miltenyi Biotec, Bergisch Gladbach, Germany; cat. no. 130-042-401) according to the manufacturer’s instructions. The negative fraction contained the CD4+ T cells. Total RNA was extracted from one million CD4+ T cells for microarray analysis. Remaining of the CD4+ T cells were incubated with anti-human CD4-FITC (Miltenyi Biotec, Bergisch Gladbach, Germany; cat. no. 130-092-358), anti-human CD127-PE (Becton, Dickinson, Franklin Lakes, NJ, USA; cat. no. 561028), anti-human CD183 (CXCR3)-PerCP/Cy5.5 (Biolegend, San Diego, CA, USA; cat. no. 353713), anti-human CD196 (CCR6)-PE/Cy7 (Biolegend, San Diego, CA, USA; cat. no. 353417), anti-human CD45RA-APC (Biolegend, San Diego, CA, USA; cat. no. 304111), anti-human CD25-APC/Cy7 (Biolegend, San Diego, CA, USA; cat. no. 302613), and anti-human CD194 (CCR4)-PE/Dazzle (Biolegend, San Diego, CA, USA; cat. no. 359419) for fluorescence-activated cell sorting (FACS) of naïve CD4+ T cells (only for counting, CD4+CD45RA+), Th1 (CD4+CXCR3+CCR6-CCR4-), Th2 (CD4+CXCR3-CCR6-CCR4+), Th17 (CD4+CXCR3-CCR6+CCR4+), and Treg (CD4+CD127lowCD25hi) cells. The remaining third of the PBMCs was used to isolate naïve CD4+ T cells using the Naïve CD4+ T Cell Isolation Kit II, human (Miltenyi Biotec, Bergisch Gladbach, Germany; cat. no. 130-094-131) and LS Separation Columns (Miltenyi Biotec, Bergisch Gladbach, Germany; cat. no. 130-042-401) according to the manufacturer’s instructions. The negative fraction contained the naïve CD4+ T cells (NT cells). Total RNA was extracted from NT cells for microarray analysis. The positive fraction from NT magnetic isolation was pooled with the positive fraction from CD4+ T cell magnetic isolation and incubated with anti-human CD3-Pacific Blue (Biolegend, San Diego, CA, USA; cat. no. 300418), anti-human CD4-FITC (Miltenyi Biotec, Bergisch Gladbach, Germany; cat. no. 130-092-358), anti-human CD56 (NCAM)-PE (Biolegend, San Diego, CA, USA; cat. no. 318305), anti-human CD19-PerCP/Cy5.5 (Biolegend, San Diego, CA, USA; cat. no. 302229), anti-human CD14-PE/Cy7 (Biolegend, San Diego, CA, USA; cat. no. 325617), and anti-human CD8-APC (Becton, Dickinson, Franklin Lakes, NJ, USA; cat. no. 555369) antibodies for FACS of CD4+ (only for counting, CD3+CD4+) and CD8+ (CD3+CD8+) T cells, natural killer cells (CD56+), B cells (CD19+), and monocytes (SSClowCD14+). Total RNA was isolated from 11 cell types (PBMCs, B cells, CD4+ and CD8+ T cells, monocytes, natural killer cells, naïve T cells, Th1, Th17, Th2, and Treg cells using the AllPrep DNA/RNA Micro kit (Qiagen, Hilden, Germany; cat. no. 80284) according to the manufacturer’s instructions and used for microarray analysis. Cell centrality correlation with enrichment of genes harboring RA-associated genetic variants Pearson correlation was calculated between subgraph centrality score and −log( p value) of the enrichment of genes harboring RA-associated genetic variants among the DEGs in each cell type. Genes harboring genetic variants associated with RA were downloaded from DisGeNet (February 2017), one of the largest publicly available collections of genes and genetic variants associated with human diseases ( ) [ 44 ]. We removed two long non-coding RNA, 21 gene symbols beginning with LOC, and three microRNAs, leaving 207 genes. Since the RA-gene associations were identified in human samples and RA cells were derived from a mouse model, we searched for mouse orthologs for all human RA-associated genes. The list of human and mouse orthologs was downloaded from Ensembl Compara in June 2017 (Ensembl version 89). For 169 out of 207 human RA associated genes, we identified mouse orthologs. Enrichment of human RA-associated genes in mouse MCDMs Since RA-gene associations were identified in human samples and RA cells were derived from a mouse model, we used mouse orthologs for all genes harboring genetic variants associated with RA. The 169 mouse orthologs were used for enrichment (Additional file 2 ). As a background, we used all mouse genes annotated in the NCBI database on June 16, 2017. Enrichment results are reported in (Additional file 3 : Tables S1 and S2). Identification of disease-associated genes and cell types by meta-analysis of genome-wide association studies and cell-type-specific epigenetic markers We first identified diseases analyzed with genome-wide association studies (GWAS) by downloading GWAS data compiled by the National Human Genome Research Institute (NHGRI). First, we manually classified 180 traits as diseases (Additional file 2 ) and excluded the remaining traits from the analysis. The identified diseases belonged to 20 out of 21 disease chapters listed in ICD-10-CM (International Classification of Diseases, Tenth Revision, Clinical Modification, Additional file 2 , Additional file 3 : Table S3): infectious (I); neoplasms (II); diseases of the blood and blood-forming organs including immune mechanisms (III); endocrine, nutritional, and metabolic diseases (IV); mental and behavioral disorders (V); diseases of the nervous system (VI); diseases of the eye (VII); diseases of the ear (VIII); diseases of the circulatory system (IX); diseases of the respiratory system (X); diseases of the digestive system (XI); diseases of the skin (XII); diseases of the musculoskeletal system and connective tissue (XIII); diseases of the genitourinary system (XIV); pregnancy, childbirth and the puerperium (XV); conditions originating in the perinatal period (XVI); congenital malformations, deformations, and chromosomal abnormalities (XVII); symptoms, signs, and abnormal clinical and laboratory findings, not elsewhere classified (XVIII); injury, poisoning, and certain other consequences of external causes (XIX); and factors influencing health status and contact with health services (XXI). One hundred seventy-five of the diseases belonged to all 17 disease-associated chapters, while five diseases belonged to chapters XVIII, XIX, and XXI. The number of diseases in each ICD-10-CM chapter is illustrated in (Additional file 1 : Figure S4). Next, we downloaded single nucleotide polymorphisms (SNPs) associated with these diseases from DisGeNet (February 2017) [ 44 ], which integrates 46,589 unique SNPs from GWAS, expert-curated repositories, and scientific literature. In total, 9880 out of 46,589 SNPs (21.2%) were associated with the given 180 diseases (Additional file 2 ). Among these, 8316 SNPs were mapped to 4475 unique genes (3518 with Entrez identifier). In order to identify cell types significantly associated with GWAS diseases, we selected cell types with cell-type-specific epigenetic markers significantly enriched for SNPs associated with each disease (Additional file 2 ). These cell types and their disease associations were compiled into a compendium, which will henceforth be referred to as CellComp (Additional file 1 : Figure S5). First, we downloaded the processed BED files from ENCODE ( ) for each of the cell types corresponding to healthy cells, which in total contained 45 cell types. We focused on nine epigenetic markers that are present in most cell types (Additional file 2 ). These cell types included those of the nervous, immune, and circulatory systems, as well as stromal and tissue-specific cell types. GWAS disease SNPs overlapping with epigenetic markers were used to calculate a disease-cell type p value for each marker-disease-cell type triplet, using the Fisher exact test. Specifically, each disease was defined by the SNPs from DisGeNet (see above) in combination with all linkage disequilibrium (LD) SNPs associated with the disease SNP set, obtaining 175 out of the 180 diseases considered. LD SNPs were retrieved from SNAP ( ) using the R package “rsnps” with default parameters of R 2 = 0.8, within 500 base pairs for each SNP. For each cell type and disease pair, we calculated an overlap Fisher exact p value. For each marker, we then formed a disease-disease similarity score based on similarities in their epigenetic associations by Pearson correlation of −log p values of the disease-epigenetic profiles, which resulted in a p value for the disease-disease associations for each disease pair and each marker. We then computed the direct genetic overlaps of all diseases and found them to be highly significant, although the marker with the lowest p value was H3K36me3 (Additional file 3 : Table S4). For robustness analysis, we also computed a binomial enrichment test by counting the numbers of significant disease-cell associations (Fisher’s exact test, p < 0.05), which also showed significant enrichment for many markers. We performed the analysis by constructing a single epigenetic disease association score for each of the 45 cell types and 175 diseases by combining each of the p values from the markers using the Fisher method (also known as Fisher’s combined probability test method). Using this score, we found a significant association (Fisher’s exact test, p < 0.05) of each disease with a median range of 20 (0–45) cell types (Additional file 3 : Table S4). The disease-cell associations are shown in a cluster diagram (Additional file 1 : Figure S6). Pathway analysis of genes harboring disease-associated genetic variants In order to identify the potentially most disease-relevant cell types in the MCDMs, we performed pathway analysis of the genes harboring genetic variants associated with the GWAS diseases, using the IPA software (Additional file 1 : Figure S7). We found that the most significant pathways were immune-related, including the top scoring Th1 and Th2 activation pathway ( p = 3.22 × 10 −34 ). To examine if this result was caused by an overrepresentation of immune-related GWAS diseases, a medical doctor manually classified the diseases as either primarily immune or non-immune (Additional file 2 ). Then, we repeated the analyses for all primary non-immune-related diseases. This also resulted in the identification of the Th1 and Th2 activation pathway as one of the top scoring pathways ( p = 3.33 × 10 −14 ). Construction of a cell type-disease network To get an overview of the disease-associated cell types, we constructed a network where cell types were depicted as nodes with sizes proportional to the number of diseases they were associated with. The nodes were linked based on manual curation of potential spatial interactions in the body. For example, bronchial epithelial cells can spatially interact with T lymphocytes but not with uroepithelial cells (Additional file 2 ). In the resulting network, immune cells were most interconnected and also appeared to be associated with more diseases than less connected cells. Indeed, we found a significant positive correlation between the degree of the cell type and the number of diseases a cell type was associated with (from the e pigenetic association score , Pearson r = 0.31, p = 0.038). Moreover, we classified all cell types into cell categories, namely immune cells, epithelial cells, muscle cells, neural cells, hepatocytes, fibroblasts, and osteoblasts (Additional file 2 ). For each cell category, we calculated a general cell-class-disease association p value by Fisher combining p values for all diseases and all cell types in cell category. Small p values from the chi-square distribution were numerically approximated through a normal distribution approximation followed by Taylor series expansion of the cumulative probability distribution [ 35 ]. Public profiling data of CD14+, CD4+, and B cells from patients with rheumatoid arthritis For the construction of RA modules based on expression profiles in CD14+, CD4+ T cells, and B cells, we downloaded public microarray experiments from GEO. We analyzed the datasets GSE57386 (CD14+), GSE56649 (CD4+ T cells), and GSE4588 (B cells). Module construction and classification with elastic net were performed as described below. Prospective clinical expression profiling studies of CD4+ T cells from patients with 13 different diseases We conducted prospective clinical studies to validate the importance of CD4+ T cells in 13 diseases from the following ICD-10-CM chapters: neoplasms (breast cancer, chronic lymphocytic leukemia); endocrine, nutritional, and metabolic diseases (type I diabetes, obesity); diseases of the circulatory system (atherosclerosis); diseases of the respiratory system (acute tonsillitis, influenza, seasonal allergic rhinitis, asthma); diseases of the digestive system (Crohn’s disease, ulcerative colitis); and diseases of the skin and subcutaneous tissue (atopic eczema, psoriatic disease). Study participants were recruited by clinical specialists based on diagnostic criteria defined by organizations representing each specialist’s discipline. Age- and gender-matched healthy controls ( n = 151 and 53, respectively) were recruited in the Southeast region of Sweden from outpatient clinics at the University Hospital, Linköping; Ryhov County Hospital, Jönköping, a primary healthcare center in Jönköping; and a medical specialist unit for children in Värnamo. Study participants represented both urban and rural populations with an age range of 8–94 years. Patients with type I diabetes and obesity had an age range of 8–18 years. Eleven patients had more than one diagnosis and are included in the reported patient numbers in the following description. For the bioinformatic analyses, when comparing patients with different diagnoses, patients suffering from both diseases in question were excluded (for example, when classifying patients with atherosclerosis versus influenza, patients having both of those diseases were excluded from this specific calculation). ICD-10-CM chapter II: neoplasms Breast cancer Patients with breast cancer were recruited at first diagnosis at an outpatient clinic based on clinical examination (palpation), radiological analyses (mammography and ultrasonography), and pathologist’s evaluation of biopsy material from mastectomy and sentinel nodes. Blood sampling was performed before surgery, and all included patients had invasive ductal or lobular cancers. First, eight patients were recruited (median age [range], 73.5 [67–82] years). For a second validation study, an independent group of 24 patients (median age [range], 61.5 [35–88] years) was recruited, based on the same inclusion criteria. All recruited patients were women. Chronic lymphocytic leukemia (CLL) Patients ( n = 8; three women; median age [range], 69.5 [51–80] years) with untreated CLL were recruited from two hematological outpatient clinics. ICD-10-CM chapter IV: endocrine, nutritional, and metabolic diseases Type I diabetes mellitus Children and adolescents who met the criteria defined by the International Society for Pediatric and Adolescent Diabetes [ 8 ] were recruited, i.e., fast-plasma-glucose level above 7.0 mmol/L at two occasions, alternative non-fasting plasma glucose above 11.1 mmol/L, and symptoms of hyperglycemia. Patients with type I diabetes or those who received insulin treatment for more than 4 weeks were excluded ( n = 8, two females; median age [range], 12.5 [11–16] years). Obesity Children who fulfilled international criteria for overweight or obesity were included based on standards for anthropometric measuring and those with diabetes mellitus were excluded [ 9 ]. Age- and gender-correlated body mass index (BMI) was calculated and defined as weight (kg) divided by stature square (m 2 ). The median BMI was 28.0 (23.0–39.5). In total, 17 patients were recruited, including children (seven females; median age [range], 14.0 [8–60] years). ICD-10-CM chapter IX: diseases of the circulatory system Atherosclerosis Patients were recruited by the same surgeon based on standard criteria [ 45 ] at an outpatient clinic, at least 3 months after coronary artery bypass graft surgery. In total, 12 patients were recruited (two females; median age [range], 71 [49–80] years). The patients were on continuous medication with statins. Patients with diabetes mellitus were excluded. ICD-10-CM chapter X: diseases of the respiratory system Acute tonsillitis Patients ( n = 6, six females; median age [range], 37.0 [26–46] years) with clinical signs of acute tonsillitis were recruited, and the diagnosis was confirmed through a rapid antigen diagnostic test or throat culture before the administration of antibiotics ( n = 6). Influenza Patients with influenza A ( n = 9) and influenza B ( n = 1) were included in the study. Influenza diagnosis was verified by PCR analysis on nasopharyngeal secretions using the Xpert Flu/RSV XC assay (Cepheid, Sunnyvale, CA) according to the manufacturer’s instructions. A total number of 10 patients were recruited (four females; median age [range], 63.0 [23–97] years). Blood samples were drawn while the patients were still symptomatic. Most patients had not started any antiviral therapy at the time of sampling, but some had received one dose of oseltamivir. Seasonal allergic rhinitis In total, 13 patients with seasonal allergic rhinitis were recruited (11 females; median age [range], 38.0 [19–53] years) based on clinical history for at least two pollen seasons, and positive skin prick tests or radioallergosorbent tests (RAST) for birch or grass. Samples were obtained during the pollen season after at least 1 day of symptoms and before treatment. Asthma Patients were recruited based on standard criteria, i.e., at least 2-year history of recurrent wheezing and baseline bronchodilator reversibility of ≥ 12%. All patients were treated with inhaled glucocorticoids and bronchodilators as required. In total, 17 patients were recruited (six females; median age [range], 49.0 [16–74] years). ICD-10-CM chapter XI: diseases of the digestive system All patients with the inflammatory bowel diseases (IBDs) UC and CD were recruited at a gastroenterology outpatient clinic by the same gastroenterologist, based on clinical evaluation, endoscopy, and/or MRI, as well as characteristic histopathological findings and exclusion of differential diagnosis. Ulcerative colitis In total, 10 patients with UC (five females; median 515 age [range], 51.5 [20–69] years were recruited. Crohn´s disease In total, 11 patients with CD (nine female; median age [range], 516 50.0 [31–76] years) were enrolled in the study. All patients were in remission. None of the study subjects had received any systemic immunosuppressive medication three months prior to study entry. ICD-10-CM chapter XII: diseases of the skin and subcutaneous tissue All atopic eczema and psoriasis patients were diagnosed by the same dermatologist (OS), based on standard criteria, medical history, and/or histopathological findings, at a dermatology outpatient clinic. Atopic eczema In total, nine patients (three females; median age [range], 42.0 [12–76] years) were recruited. Psoriasis In total, 11 patients (six females; median age [range], 48.0 [20–71] years) with mild to severe plaque-type psoriasis were recruited. Atopic eczema patients had active eczema for at least 1 week, and a diagnosis for at least 2 years. Psoriasis patients were diagnosed for at least 1 year, and assessment of disease severity was performed using the Psoriasis Area and Severity Index (PASI). Median PASI was 8.1 (range, 4.2–16.2). None of the study subjects had received any systemic immunosuppressive medication or phototherapy 3 months prior to study entry. Isolation of peripheral CD4+ T cells Briefly, PBMCs were prepared from fresh blood samples from the patients of 13 diseases and healthy controls, as previously described [ 12 ], using Lymphoprep (Axis-Shield PoC) according to the manufacturer’s protocol. Total CD4+ T cells were enriched from PBMCs by FACS. Human IgG (Sigma-Aldrich, St Louis, MO, USA) at a final concentration of 200 μg/mL was used to block cells prior to staining. Mouse anti-human CD4-FITC (BD Pharmingen San Diego, CA, USA), Mouse antihuman CD3-Pacific Blue (Biolegend San Diego, CA, USA), and all matched isotype controls were purchased. Cell sorting was performed on a FACS Aria flow cytometer (BD Biosciences, San Diego, CA, USA), and the data was analyzed by FlowJo 7.6 (Tree Star, Inc., San Carlos, CA). After sorting, the purity of total CD4+ T cells was more than 98%. Preparation of RNA for expression profiling Total RNA was extracted using the AllPrep DNA/RNA Micro kit (Qiagen, Hilden, Germany; cat. no. 80284) according to the manufacturer’s instructions. RNA concentration and integrity were evaluated using the Agilent RNA 6000 Nano Kit (Agilent Technologies, Santa Clara, CA, USA; cat. no. 5067-1511) on an Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA). Microarrays were then further computationally processed as described in One-Color Microarray-Based Gene Expression Analysis Low Input Quick Amp Labeling protocol (Agilent Technologies, Santa Clara, CA, USA). Microarray data processing All gene expression microarrays were processed as described above using the LIMMA R package. All probes with an expression below 1.2 times the background signal were removed. To test whether the sex or age of the patients had any confounding effects on the CD4+ T cell microarray data for the 13 diseases, a principal component analysis (PCA) was performed. The results showed no clear differences in any of the components (Additional file 1 : Figure S8). To test for possible confounding effects within T cell subsets, we performed deconvolution analysis of expression profiles from the CD4+ T cells using profiles from Th1, Th2, Th17, and Treg cells. Those profiles were derived from the above-described reference expression profiling compendium of human immune cells. Deconvolution of bulk CD4+ T cell data with sorted T cell subsets using CIBERSORT We tested whether T cell subtypes differed significantly between sexes, ages, and diseases. For this purpose, we performed in vivo sorting of nine different immune cell types, where four were tested in this study, namely Th1, Th2, Th17, and Treg cells, followed by microarray analysis. Next, we applied CIBERSORT [ 46 ] with default parameters using our own reference transcriptomics data for each of the patients from our 13 diseases and the controls. This showed a high overlap of the different age groups, sexes, and diseases (Additional file 1 : Figure S9). Human protein interactome STRING (v.10) [ 47 ] was used to construct the human protein-protein interaction network (PPIn) as a representation of the human protein interactome. All the interactions with a confidence score greater than 0.7 were considered, for both direct (physical) and indirect (functional) interactions. The resulted network consisted of 11,228 vertices (proteins) and 212,419 edges (interactions). Module construction The module construction was based on the integration of expression data for diseases and controls, and the PPIn. Given a disease d i , its associated module M i was defined by the set of genes with highly correlated expression patterns, forming cliques into the PPIn and enriched for DEGs. Given the modules M 1 , ... , M 13 , the shared interaction neighborhood [ 9 ] was defined as the union of modules, comprising all the genes from all the disease modules. The modules consisted of median 392 genes [201–735]. A full list of individual module genes is provided in Additional file 2 . Top canonical pathways enriched in module genes are reported in Additional file 2 . Diagnostic potential of CD4+ T cell expression profiles Classification of patients and controls was performed using elastic net function lassoglm() on MATLAB in the Statistics Toolbox, choosing lambda ( λ) from the minimum deviance of leave-one-out cross-validation starting from all the measured genes of the platform. Prediction p values of case versus controls were based on the leave-one-out estimates. We calculated the area under the precision recall curve using the perfcurve() MATLAB function. The p values were calculated using the two-sided Wilcoxon rank-sum testing on the classifier discriminant function outputs. For each of the 13 diseases, expression of disease-specific module genes separated patients from controls with high accuracy (median AUC 0.98, range 0.82–1; median p < 2.8 × 10 −5 , range 8.5 × 10 −8 to 7.8 × 10 −4 ). For example, the AUC for breast cancer was 1 ( p = 1 × 10 −5 ). An independent validation study of the module in 24 breast cancer patients and 14 healthy controls yielded an AUC of 0.82; p = 1.7 × 10 −3 , which was significantly higher than that for random genes (one-sided, permutation test, p < 3.7 × 10 −258 ). We also found that the respective module union genes separated patients with different diagnoses from each other (median AUC 0.98, range 0.27–1; median p < 1.0 × 10 −3 , range 1.32 × 10 −5 to 0.69; Additional file 2 ). Box plot was created using MATLAB boxplot() function with default settings. Outliers were defined by the algorithm underlying boxplot() function, i.e., points were assumed to be outliers if they were greater than q 3 + w × ( q 3 − q 1 ) or less than q 1 − w × ( q 3 − q 1 ), where q 1 is the 25th percentile, q 3 is the 75th percentile, and w corresponds to ± 2.7 σ and 99.3% coverage according to the function description. Classification robustness was confirmed via 20 additional classifiers Classification robustness was confirmed via 20 additional classifiers, namely, Coarse KNN, Cosine KNN, Fine KNN, Cubic KNN, Weighted KNN, Medium KNN, Complex Tree, Medium Tree, Simple Tree, Linear Discriminant, Logistic Regression, SVM Coarse Gaussian, SVM Cubic, SVM Fine Gaussian, SVM Linear, SVM Medium Gaussian, SVM Quadratic, Ensemble Subspace KNN, Ensemble Bagged Trees, and Ensemble Subspace Discriminant, all implemented with MATLAB Classification Learner App. AUC and p values were calculated as described above. Training and fivefold cross-validation were repeated 100 times. Average AUC and p values are reported in Additional file 2 . Prioritization of genes with the highest predictive value for classification To rank genes based on their predictive value, randomized elastic net was performed. Randomized elastic net was implemented as a modification of randomized lasso [ 48 ] where the lasso technique was replaced with elastic net as follows: for selected λ , and α = 0.5, data were permuted by adding random penalty factors from the interval [1/ α ,1] for each predictor; model coefficients were estimated (elastic net); 10,000 permutations were performed; and predictors with non-zero coefficients in at least one of the 10,000 permutations were selected (Additional file 2 ). Measurement of the proteins in CD and UC We measured the levels of CXCL1, CXCL8, CXCL11, CCL20, TNF-α, and IL-1β (see above) in the serum of 15 UC patients, 11 CD patients, and 20 healthy controls using the U-PLEX Biomarker Group 1 (hu) assays, SECTOR (1PL) (Meso Scale Discovery, Rockville, MD, USA; cat. no. K15067 L-1) according to the manufacturer’s instructions. The U-PLEX technology ( ) can measure a maximum of 10 proteins at the same time and requires only 25 μl from each sample. The biomarker group used (cat. No. K15067 L-1) was custom designed and only measured the six proteins of interest stated above. Drug target analysis of the shared interaction neighborhood In order to test the therapeutic relevance of the SIN, we downloaded all drugs that had at least one known target in human cells ( n = 1790; out of which 1408 were approved) from approved and investigational drugs from DrugBank (version 5.0.3). We then tested whether the SIN was enriched for these drugs and found the SIN genes to be significantly enriched for these drug targets ( n = 302; Fisher exact test OR 2.92; p = 2 × 10 −53 ). An important therapeutic implication is that drugs targeting the SIN can be potentially used for more than one disease. Identification of drugs suitable for targeting the SIN In order to identify drugs suitable for targeting the SIN, we computationally predicted which of the 1790 drugs would mainly target disease genes in the SIN. The predictions were based on prioritization of drugs in network proximity to the disease genes within SIN [ 49 ]. Briefly, we calculated the distance between direct drug targets ( T ) and disease genes ( G ) within the SIN ( M ) on the PPIn [ 47 ] using the shortest path distance measure ( d): {d}_d\left(G,T\right)=\frac{1}{\left\Vert T\right\Vert }{\sum}_{t\in T}{\min}_{g\in G}d\left(g,t\right). {d}_d\left(G,T\right)=\frac{1}{\left\Vert T\right\Vert }{\sum}_{t\in T}{\min}_{g\in G}d\left(g,t\right). Disease genes were defined as DEGs between patients and controls, as well as genes harboring genetic variants associated with each disease. To assess the significance of the distance between drugs and disease-associated genes in the SIN ( dd ), we created a reference distribution corresponding to the expected distances using randomly selected groups of drug targets and disease genes. We performed 1000 degree-preserving randomizations using a binning approach that grouped nodes with a certain degree interval together, such that there were at least 300 nodes in a bin. Next, the average d and standard deviation of reference d distribution were used to convert observed distance to a normalized distance: z\left(G,T\right)=\frac{d_d\left(G,T\right)-{\mu}_{d_d\left(G,T\right)}}{\sigma_{d_d\left(G,T\right)}}. z\left(G,T\right)=\frac{d_d\left(G,T\right)-{\mu}_{d_d\left(G,T\right)}}{\sigma_{d_d\left(G,T\right)}}. Subsequently, z -scores were transformed to p values using MATLAB ztest() function, left-sided test (Additional file 2 ). To further validate these results, we also modified the approach from Guney et al. [ 49 ] by calculating the distance between direct drug targets ( T ) and disease genes ( G ), and between disease genes and the SIN ( M ) on the PPIn using the minimum shortest path distance measure d {d}_{dm}\left(G,M,T\right)=\frac{1}{\left\Vert T\right\Vert }{\sum}_{t\in T}{\min}_{g\in G}d\left(g,t\right)\cdotp \left(1+\frac{1}{\left\Vert G\right\Vert }{\sum}_{g\in G}{\min}_{m\in M}d\left(m,g\right)\right). {d}_{dm}\left(G,M,T\right)=\frac{1}{\left\Vert T\right\Vert }{\sum}_{t\in T}{\min}_{g\in G}d\left(g,t\right)\cdotp \left(1+\frac{1}{\left\Vert G\right\Vert }{\sum}_{g\in G}{\min}_{m\in M}d\left(m,g\right)\right). This modification takes into consideration disease genes (defined as described above) that are not a part of the SIN. To assess the significance of the distance between drug, disease genes, and the SIN ( ddm ), we performed 1000 degree-preserving randomizations using binning approach, as described above. Next, the average d and SD of reference d distribution was used to convert observed distance to a normalized distance: {z}_m\left(G,M,T\right)=\frac{d_{dm}\left(G,M,T\right)-{\mu}_{d_dm\left(G,M,T\right)}}{\sigma_{d_{dm}\left(G,M,T\right)}} {z}_m\left(G,M,T\right)=\frac{d_{dm}\left(G,M,T\right)-{\mu}_{d_dm\left(G,M,T\right)}}{\sigma_{d_{dm}\left(G,M,T\right)}} and z -scores were then transformed to p values as described above (Additional file 2 ). We identified five such drugs, which targeted nine out of 13 diseases. We validated one of these drugs, the peroxisome proliferator-activated receptor (PPAR)-alpha agonist bezafibrate, in a T cell-dependent mouse model of rheumatoid arthritis described above, which was not among the 13 diseases that were analyzed to construct the SIN. Treatment study of bezafibrate in a mouse model of RA To test if bezafibrate could dampen inflammation, we administered bezafibrate to 8- to 20-week-old female 129/SvE mice subjected to the antigen-induced RA model, as described above. Bezafibrate (Sigma-Aldrich, B7273) was dissolved overnight in dimethyl sulfoxide (DMSO; 0.1 g bezafibrate/mL) and further diluted in PBS before treatment. Three different treatment protocols were used to test the effect of bezafibrate on arthritis development: (1) systemically at pre-sensitization days 1 and 7 (8 mg bezafibrate/kg included in the immunization solution containing mBSA + Freund’s adjuvant, n = 5), (2) systemically by injections after triggering of arthritis (4 mg bezafibrate/kg in a total volume of 500 μL PBS, intraperitoneally on days 21, 24, and 26, n = 4), and (3) locally by a single intra-articular injection (0.6 mg bezafibrate/kg included in the mBSA solution used to trigger arthritis, n = 5). For all treatments, control mice subjected to antigen-induced arthritis (AIA, n = 5) were injected in the same manner with the same volume of PBS/DMSO used for bezafibrate delivery (max 0.1% DMSO). Systemic delivery of bezafibrate after triggering of arthritis prevented the development of arthritis (Additional file 1 : Figure S10). Next, we examined if bezafibrate delivery would have local effects. This was done by local treatment with a single intra-articular injection of bezafibrate. We also examined the effects of bezafibrate delivery at antigen sensitization: neither treatment locally or at sensitization had any ameliorating effects on arthritis (Additional file 1 : Figure S10). To validate the effects of bezafibrate on T cell proliferation, we proceeded with a proliferation assay, as described below. CD4+ T cell proliferation assay Spleen and lymph nodes draining sites of injection (axillary and popliteal) were isolated, and single-cell suspensions (splenocytes and lymph nodes combined) were prepared by passing the spleen and lymph nodes gently through a 70-μm cell strainer. Red blood cells were lysed by adding RBC lysing solution (Sigma-Aldrich) according to the manufacturer’s instructions. Carboxyfluorescein diacetate succinimidyl ester (CFDA-SE, Sigma-Aldrich) at 5 μM was added to cells and incubated for 5 min in the dark at room temperature (RT, 25 °C). This was followed by washing the stained cells five times with FACS buffer (PBS + 1% FBS). Stained cells (2 × 10 6 /mL in a total volume of 200 μL) from mice subjected to AIA, mice subjected to AIA, and treated with Bezafibrate as described above and from naïve control mice were stimulated with mBSA (50 μg/mL) and cultured for 72 h at 37 °C in 5% CO 2 and 95% humidity. After 72 h, cells were harvested and analyzed for diminished carboxyfluorescein succinimidyl ester (CFSE) stain by FACS. CFSE-stained, non-stimulated cells from a naïve mouse were used to define non-proliferating cells [ 50 ]. Proliferation assay To determine the effect of bezafibrate on T cell proliferation in the RA model, spleen, axillary, and popliteal lymph node cells from naïve, bezafibrate-treated, and non-treated controls subjected to AIA were stained with 5 μM CFDA-SE (Sigma-Aldrich) by incubation for 5 min in the dark at RT. Stained cells were washed five times with FACS buffer (PBS + 1% FBS). Stained cells (2 × 10 6 /mL in a total volume of 200 μL) were stimulated with mBSA (50 μg/mL) and cultured at 37 °C in 5% CO 2 and 95% humidity. After 72 h, cells were harvested and analyzed for diminished CFSE-stain by FACS; see gating strategy (Additional file 1 : Figure S11) CFSE-stained cells from a naïve mouse were used to define non-proliferating cells [ 50 ]. Assessment of antigen recall responses of CD4+ T helper cells among spleen and lymph node cells showed that the systemic intraperitoneal treatment with bezafibrate that protected from arthritis also inhibited proliferation of CD4+ T helper cells, which was not the case for bezafibrate treatment locally or at sensitization (Additional file 1 : Figure S10). Thus, the protective effect of bezafibrate on arthritis development is contingent on its ability to inhibit T helper cell proliferation. Results scRNA-seq study of a mouse model of arthritis shows wide dispersion of pathogenic mechanisms in multiple cell types We performed scRNA-seq analyses of a mouse model of RA, antigen-induced arthritis (AIA). In this model, arthritis is triggered by intra-articular injection of the antigen mBSA in mice previously sensitized with mBSA (Fig. 1 a). Histologically, the arthritic tissue resembled that found in human RA, with infiltration of inflammatory cells into the synovium, cartilage/bone destruction, and hyperplasia of the synovial lining (Fig. 1 b). scRNA-seq was performed on whole arthritic joints, as well as on joint lymph nodes (Fig. 1 c). In total, we recovered 8420 single cells after filtering on a minimum of 10,000 reads and 400 transcripts per cell. Cell type classification was performed using Reference Component Analysis, RCA (see the “ Methods ” section) [ 18 ]. This method uses bulk expression profiles of known cell types as references to classify single-cell profiles based on genome-wide transcriptional similarity. We first tested RCA by analyzing if it correctly classified in-house scRNA-seq data from two cancer cell lines using bulk profiling data from 26 cell lines as a reference and found that this was the case (see the “ Methods ” section, Additional file 1 : Figure S1). We next used RCA to classify cell types in our mouse scRNA-seq data. We identified nine cell types in arthritic mice and seven in healthy controls (Fig. 1 d, Additional file 1 : Figures S12 and S13). The cell types were dendritic cells, CD4+T lymphocytes, T regulatory cells (Treg), B lymphocytes, macrophages, granulocytes, common myeloid progenitor cells (promyeloids), adipocytes, and osteoblasts. These cell types are similar to those primarily or secondarily involved in human RA [ 23 , 51 , 52 ] and are partially similar to cell types identified by scRNA-seq analysis of synovium from patients with rheumatoid arthritis (RA) [ 23 , 52 ]. In order to identify and prioritize mechanisms for therapeutic targeting, we first identified differentially expressed genes (DEGs) using Monocle. Similar to the scRNA-seq study of human RA [ 23 ], DEGs were calculated for each cell type compared to all other cell types in each tissue separately (see the “ Methods ” section). We performed pathway analysis of the DEGs in AIA using the Ingenuity Pathway Analysis (IPA). The most significant pathways enriched among differentially expressed genes were found in T cells and related to T cell activation and differentiation, e.g., Cd28 signaling in T helper cells ( p = 2.51 × 10 −12 and Th1 differentiation ( p = 5.75 × 10 −9 ). These pathways included genes with key roles for activation and differentiation, such as Il2 , Itk , and Ifngr1 . Fig. 1 scRNA-seq analysis of a mouse model of antigen-induced arthritis (AIA). a An overview of the AIA mouse model. b Representative joint images from naïve mice and arthritic joints after hematoxylin and eosin (H&E) staining. B, bone marrow; S, synovial cavity; C, cartilage. Arrows indicate (1) infiltration of inflammatory cells to the synovium, (2) cartilage/bone destruction, and (3) hyperplasia of the synovial lining. c A schematic overview of seq-well scRNA-seq and cell type identification using reference component analysis (RCA). d t-SNE plot of 7086 healthy and RA joint cells ( n = 4 healthy mice samples and 5 sick mice samples), and 1333 healthy and AIA lymph nodes cells ( n = 4 healthy mice samples and 5 sick mice samples), colored by RCA clusters Full size image Intuitively, therapeutic targeting of genes in the most significant pathways would appear ideal. Indeed, drugs that inhibit Th1- or Th17-like responses have been developed to treat RA [ 53 ]. However, those drugs have shown variable efficacy [ 54 ]. One reason could be the involvement of multiple other pathways in T cells and other cell types, which are not targeted. Indeed, the other most significant T cell pathways were highly diverse, e.g., calcium-induced T lymphocyte apoptosis ( p = 1.48 × 10 −10 ), Nfat activation ( p = 2.13 × 10 −9 ), Cdc42 signaling ( p = 5.37 × 10 −8 ), Nur77 signaling in T lymphocytes ( p = 3.47 × 10 −7 ), and production of nitric oxide and reactive oxygen species ( p = 3.55 × 10 −6 ). A similar, and non-overlapping, diversity was found in granulocytes, e.g., virus entry via phagocytosis ( p = 1.29 × 10 −9 ), Mtor signaling ( p = 1.29 × 10 −9 ), integrin signaling ( p = 6.91 × 10 −8 ), leukocyte extravasation ( p = 5.12 × 10 −6 ), caveolar-mediated endocytosis ( p = 1.12 × 10 −5 ), and Vegf signaling ( p = 3.02 × 10 −5 ). Further analyses of all differentially expressed genes in all the cell types showed a great variety of pathways, therapeutic targets, and biomarkers: 285 pathways, 263 drugs, and 873 biomarkers were significantly associated with any of the cell types ( p < 0.05). The median number of pathways per cell type was 46 (0–205), and the median number of cell types associated with each pathway was 2 (1–8), (Additional file 2 ). This diversity suggested that specific therapeutic targeting of the most significant pathway in one cell type would not suffice because of multiple other pathways in the same or other cell types. Instead, an impractical number of drugs targeting multiple pathways might be needed. Indeed, the number of drugs predicted to target significant pathways in each cell type was 55.5 (0–144), and the number of cell types predicted to be targeted by each drug was 1 (1–8). Only one drug, sirolimus, targeted most of the identified cell types. This is a potent immunosuppressant, which has been tried in refractory cases of RA, but has significant side effects [ 55 ]. We repeated the above analyses in cell types identified by scRNA-seq of human synovium from patients with RA [ 23 ]. Similar to the AIA mice there was a great diversity of pathways, which were dispersed across multiple cell types. There were also highly significant pathway overlaps between the same cell types in human and mouse data (Additional file 1 : Figure S14, Additional file 2 ). Effective drug targeting of such complex changes is a formidable challenge, which may explain why many patients with autoimmune diseases do not respond to treatment. This highlights the need for novel, systems-level approaches to prioritize cell types and pathways for therapeutic targeting. Here, we examined if network principles could be applied to aid in cell, target, and drug prioritization. Thus, we constructed network models of the cell types in the lymph nodes and joints from arthritic mice—henceforth referred to as multicellular disease models (MCDMs). Interactions were inferred by connecting whole networks of differentially expressed genes in each cell type with their predicted regulators in all other identified cell types. These predictions were based on significantly enriched interactions in the IPA program (see the “ Methods ” section, Additional file 2 ) [ 38 ]. For example, Il1b was predicted as a possible upstream regulator of DEGs in granulocytes. We found that Il1b was differentially expressed in promyeloids. This led to the identification of a potential interaction between these two cell types (Fig. 2 a–c; Additional file 2 ). The predicted interactions were supported by similar results using another recently described method (Additional file 2 ) [ 40 ]. Fig. 2 Multicellular disease models (MCDMs) from a mouse model of AIA. MCDMs were constructed based on scRNA-seq data by connecting differentially expressed genes in each cell type with predicted upstream regulators in all other cell types. Cell type size corresponds to centrality score. Numbers indicated by the nodes denote the number of identified cells of specific type (for example in RA joint, we have identified 4258 granulocytes). a An MCDM of lymph nodes from arthritic mice. b An MCDM from arthritic joints. c Multicellular model of a healthy mouse joint (lymph node model is not shown because there was only one predicted interaction). Gene names of predicted upstream regulators are indicated on arrows. Treg, T regulatory cells. d Correlation between centrality score of cell types and enrichment of genes harboring genetic variants identified by GWAS and expert curated repositories among differentially expressed genes (the genes were derived from DisGeNet and the analysis based on the mouse orthologues of the human genes) Full size image In the lymph node, only two cell types, T and B lymphocytes, had predicted interactions (Fig. 2 a). By contrast, in the joints, all cell types were connected with each other in a multi-directional manner, mainly by cytokines and chemokines (Fig. 2 b). A possible explanation for more interactions in the joints could be a larger number of cells than in lymph nodes ( n = 7086 and n = 1333, respectively). However, we found no significant correlation between the number of cells of each cell type and the number of outgoing edges/interactions (Pearson, sick joint r = 0.23, p = 0.66; healthy joint r = 0.61, p = 0.39). Therefore, a more likely explanation is structural differences between lymph nodes and the whole joint. Different cell types in lymph nodes may potentially interact less because they are more localized in dedicated tissue compartments than in joint fluid or tissue. Visual inspection of the resulting MCDM networks revealed no obvious key regulatory cell type, such as a “hub” that had many more interactions than the others, nor an upstream cell type that regulated the others in a linear chain (Fig. 2 a, b). Similar results were obtained for MCDMs derived from scRNA-seq from human RA (Additional file 2 ), as well as when cell types were connected based on another ligand-receptor-based method to infer interactions (see the “ Methods ” section, Additional file 2 ) [ 40 ]. This suggested that pathogenic mechanisms were dispersed across multiple cell types. In support of multicellular pathogenesis, the differentially expressed genes in most cell types in scRNA-seq data from both mouse arthritis and human RA were significantly enriched for genes harboring genetic variants associated with RA. Those genes were derived from the DisGeNet database [ 44 ], and the mouse analyses based on the mouse orthologs of those genes (see the “ Methods ” section, Additional file 2 ). Below, we present data supporting that multicellular pathogenesis is a general characteristic of complex diseases and that network analysis in combination with converging biomedical data can be used to prioritize the most relevant cell types and genes for diagnostics and therapeutics. Analyses of MCDM network properties and enrichment of RA-associated genetic variants in mouse arthritis supported that the relatively most important cell types can be prioritized Because of the complex, multi-directional interactions in the MCDMs, we examined if network centrality could help to prioritize the relatively most important cell types. A detailed explanation of centrality is given in the “Methods” section [ 39 , 42 ]. Briefly, centrality is a measure of interconnectivity, and our assumption was that the most central cell types would be relatively most important for pathogenesis. We used random walk centrality as a metric for centrality. For robustness, we also included subgraph centrality analysis, and a complementary annotation source to Ingenuity ([ 43 ], see the “ Methods ” section) which for each case showed similar results as the random walk centrality (Additional file 4 ). As a positive control, we started by analysis of scRNA-seq from human colorectal tumors [ 18 ]. We hypothesized that tumor cells would be more central than surrounding stromal and immune cells. To characterize surrounding cell types, we used RCA and generated a reference expression profiling compendium (see the “ Methods ” section). We constructed an MCDM and found that tumor cells were most central, followed by immune and stromal cells (Additional file 1 : Figure S3). We continued by analyses of centrality in the mouse MCDM. The relevance of centrality as an indication of the pathogenic importance of the MCDM cell types was supported by a significant correlation between centrality and the degree of enrichment of genes associated with RA by genetic variants among differentially expressed genes in each cell type (Pearson r = 0.85, p = 0.03) (Fig. 2 d). An MCDM of human RA supported multicellular pathogenesis and centrality to prioritize cell types In order to test the translational value of centrality, we constructed an MCDM based on scRNA-seq data from human synovium from RA patients [ 23 ]. We found a significant correlation between centrality and the degree of enrichment of genes harboring genetic variants associated with RA (Pearson r = 0.57, p = 0.04). A possible explanation for the lower correlation coefficient in the human synovium MCDM compared to AIA could be that the human MCDM lacked a predominant cell type in RA, namely granulocytes. By contrast, this cell type category was among the most central and also most enriched for genes harboring RA-associated genetic variants in mouse arthritis. This emphasizes the importance of performing scRNA-seq analysis on whole organs to identify pathogenic cell types. However, such analysis may be complicated by not knowing all organs and cell types involved in many diseases. In order to obtain an estimate of cell types and organs involved in human RA, we analyzed cell-type-specific epigenetic markers that were enriched for genetic variants associated with RA (henceforth referred to as GWAS-enriched epigenetic markers; see the “ Methods ” section; Additional files 2 and 5 ). The epigenetic markers were identified in the Encyclopedia of DNA Elements [ 56 ]. These markers included both activating and repressive elements identified in 45 primary human cell types. This resulted in the identification of 24 cell types and subsets that could be ordered based on their significance of association (Fig. 3 a, Additional file 2 ). The cell types belonged to two main categories, immune cells and local stroma or parenchymal cells. Fig. 3 Network models of disease-associated cell types. a 24 cell types and subsets that were significantly enriched for GWAS-enriched epigenetic markers associated with RA. Cell type size corresponds to association −ln ( p value). b Network model of cell types associated with human rheumatoid arthritis (RA). Nodes correspond to cell types, node size corresponds to significance of association (−log10 RA GWAS-epigenetic marker enrichment p value). Cell types with potential spatial interactions are linked, and cell type position depends on the centrality score as indicated by the rings in the background. c Bar plot of cell type classes ordered by significance of association with 175 human diseases (Fisher combined GWAS-enriched epigenetic markers – disease association p value calculated for each cell type class). d Network model of cell types associated with 175 diseases, based on the same parameters as in b (for details see results ) Full size image As expected, immune cell types were most significantly associated. The latter category pointed to organs known, or thought to be involved in RA, namely the joints, lungs, heart, skin, and liver. Some of these are difficult to study in human patients, which led us to ask if centrality could be used to further prioritize cell types and thereby organs. To explore the centrality of these categories, we would ideally need the expression profiles of each cell type to infer molecular interactions and construct a cellular network. Since many of the cell types are not possible, or difficult to obtain from patients, we instead used potential spatial interactions as a simple and binary proxy for molecular interactions. For example, a T cell can spatially interact with both bone and kidney cells, but the latter cannot interact with each other. These predicted interactions were inferred based on manual curation (Additional file 2 ). Using these interactions, we constructed a cellular network model of human RA (Fig. 3 b). In this model, cell type size corresponded to relative pathogenic importance, as defined by significance of GWAS-enriched epigenetic markers, and position in the network to centrality. The model indicated that centrality could potentially be used to prioritize cell types. Similar to the mouse and human scRNA-seq MCDMs, the central cell types were mainly from the immune system, while the peripheral ones were local parenchymal or stromal cell types from different tissues (Fig. 3 b). Returning to the question if centrality could help to prioritize between local cell types, osteoblasts were actually less central in the model than, for example, epithelial cells and fibroblasts from the lungs. While this may seem unexpected given the importance of joint involvement in RA, this agrees with immune reactions in the lungs being proposed to have a primary pathogenic role [ 51 ]. Taken together, these findings supported the dispersion of pathogenic mechanisms in multiple immune and local tissue cell types in RA. A network model of 175 diseases supported multicellular pathogenesis and centrality to prioritize cell types Our analyses of AIA and RA led us to ask if multicellular pathogenesis is a general characteristic of human diseases and if the most important cell types can be identified in each disease. To address the first question, about multicellular pathogenesis, we identified cell types that were significantly associated with 175 diseases that had been analyzed with GWAS. This was done using GWAS-enriched epigenetic markers, as described above for RA (Additional files 2 and 5 ). We found that the diseases were associated with a median of 20 (0–45) cell types, which could be potentially ranked in order of relative importance, based on significance of association (Fig. 3 c). This ranking showed that immune cells were more significant than local stroma and parenchymal cells, and was supported by pathway analysis of genes harboring genetic variant associated with the 175 diseases, which showed that the Th1 and Th2 activation pathway was most significant ( p = 3.22 × 10 −34 , Additional file 5 ), followed by other immune-related or general pathways. Even after removal of immune diseases, the Th1 and Th2 activation pathway remained significant ( p = 3.3 × 10 −14 ; see the “ Methods ” section). Next, we examined if centrality analysis could be applied to prioritize the most important cell types in human diseases. To formally test if there was an association between cellular centrality and disease risk, we constructed a single, multicellular network model of the 175 diseases analyzed with GWAS (henceforth referred to as GWAS diseases). We used the same construction principles as for the model of RA (Fig. 3 d, see the “ Methods ” section). In support of centrality as an explanation for increased disease risk, we found a significant correlation between centrality and GWAS-enriched epigenetic markers (Pearson r = 0.51, p = 3.5 × 10 −4 ). In summary, the above analyses supported multicellular pathogenesis as a general disease characteristic, and the potential to prioritize cell types using GWAS-enriched epigenetic markers and centrality. Thus, cellular network models of individual diseases, like the one for RA, may help to prioritize organs or cell types to construct scRNA-seq-based MCDMs from human patients. To facilitate such studies, we have provided network models of each of the 175 diseases, as well as the underlying data (Additional files 2 and 5 ). High cell type interconnectivity implies diagnostic potential of central and clinically accessible cell types Both the MCDM of RA and the network model of 175 diseases supported multicellular pathogenesis. While this complicates therapeutic targeting, the centrality analyses indicated a potential diagnostic advantage: because of its interconnectivity with other cell types, any single, central cell type could serve as a diagnostic sensor of all other disease-associated cell types in an MCDM. We examined this possibility in public microarray data from some central cell types in RA patients, namely CD4+ T cells, B cells, and CD14+ cells. To prioritize between the large number of differentially expressed genes between RA and controls in each cell type, we identified so called modules, i.e., genes that co-localized on the protein-protein interaction (PPI) network (Fig. 4 a) [ 2 ]. We found that module genes separated patients and controls with high accuracy (area under the curve, AUC CD4+ T cells = 1.0, p = 5.4 × 10 −4 , module size = 43; AUC CD14+ = 0.74, p = 3.4 × 10 −3 , module size = 8; AUC B cells = 0.88, p = 4.9 × 10 −3 , module size = 35). Fig. 4 Diagnostic potential of CD4+ T cells based on clinical profiling studies of 13 diseases. a Toy model of a disease module. Disease-associated genes (red) are mapped on proteins (blue) in the human protein-protein interaction network. Disease-associated genes that co-localize form a module. b Overview of the module-based analyses. First step is the identification of disease modules for each of the 13 diseases profiled in the prospective microarray study of CD4+ T cells. For each disease module, genes separate patients from healthy controls. For pairwise comparison of the diseases, genes in the union of two respective modules separate patients with different diseases; for example, genes in influenza and asthma modules separate patients with influenza from patients suffering from asthma with AUC of 0.99, p = 3.3 × 10 −5 , as shown in c . c Heatmap presenting area under the curve (AUC) values of 13 disease classifications based on the module intersections genes, using elastic net. d An independent validation study of classification accuracy of breast cancer patients ( n = 24) and healthy subjects ( n = 14) based on previously preselected biomarkers (genes) measured in CD4+ T cells. Classification was performed with elastic net, preserving same lambda ( λ ) value as estimated for the original study. e – j Potential diagnostic classification of IBD patients based on six secreted plasma proteins identified in the intersection of ulcerative colitis (UC) and Crohn’s disease (CD) modules. These proteins could separate patients from healthy controls (HCs). e CXCL11; f CCL25; g CXCL1; h CXCL8; i IL1B; j TNF. k Crohn’s disease and ulcerative colitis patients’ classification based on normalized protein levels of CXCL1 and CXCL8. UC, ulcerative colitis; CD, Crohn disease; HC, healthy controls. Star denotes p value < 0.05. d – k The bars in the boxes represent median and 25th and 75th percentiles, while whiskers extend to ± 2.7 σ (see the “ Methods ” section) Full size image This led us to examine if a central, functionally relevant and clinically accessible cell type could be used diagnostically in clinical studies of human patients with multiple diseases. We focused on expression profiling of the CD4+ T cell because of its centrality, accessibility in peripheral blood, and the pathway analyses of the 175 diseases, described above. Prospective clinical studies of 13 diseases demonstrate diagnostic potential of CD4+ T cells Collection of T cell expression profiling data across multiple diseases is a considerable challenge since it requires involving specialists from primary, secondary, and tertiary care, laboratory facilities to define inclusion/exclusion criteria, as well as a standardized protocol for T cell isolation and analysis. To our knowledge, such a study has not been previously undertaken. Here, we performed such a study of 13 different diseases to evaluate the diagnostic potential of peripheral, total CD4+ T cells, using a highly standardized protocol. The results were validated by independent studies. We analyzed autoimmune, allergic, infectious, malignant, endocrine, metabolic, and cardiovascular diseases (see the “ Methods ” section), and age/gender-matched controls. The diseases were Crohn’s disease, ulcerative colitis, psoriasis, seasonal allergic rhinitis, asthma, atopic eczema, acute tonsillitis, influenza, breast cancer, chronic lymphocytic leukemia, type 1 diabetes, obesity, and atherosclerosis. In order to prioritize between differentially expressed genes in each disease, we constructed modules (Fig. 4 b). For all but one of the 13 diseases, expression of module genes separated patients from controls with high accuracy (Fig. 4 c, marked with red box, median of area under the curve, AUC, was 0.98, range 0.82–1; median p < 2.8 × 10 −5 , range 8.5 × 10 −8 to 7.8 × 10 −4 ; Additional file 2 ). One disease was separated with less high accuracy, namely obesity (AUC = 0.82). While obesity is increasingly recognized as an inflammatory disease [ 57 ], a likely explanation is that it has a greater metabolic component than the other investigated diseases. The classifications were not dependent on the classification method since similar results were obtained using 20 different methods (Additional file 2 ). As an example, the AUC for breast cancer was 1, p = 10 −4 . An independent validation study in 24 breast cancer patients and 14 healthy subjects yielded an AUC of 0.82; p = 1.7 × 10 −3 (Fig. 4 d), which was significantly higher compared to random genes ( p < 3.7 × 10 −258 ). We also found that the disease module genes separated patients with different diagnoses from each other (Fig. 4 c, median AUC 0.98, range 0.27–1; median p < 1.0 × 10 −3 , range 1.32 × 10 −5 to 0.69; Additional file 2 ). As an example, the inflammatory bowel diseases ulcerative colitis (UC) and Crohn’s disease (CD), which may be difficult to separate in clinical practice [ 58 ], were classified with a cross-validated AUC of 0.82 ( p < 0.03) . Taken together, our analyses of 13 different diseases support the diagnostic potential of CD4+ T cells. However, since expression profiling of T cells is complex in clinical settings, we searched for module intersection genes that encoded secreted proteins, which could more readily be measured diagnostically in sera. We identified six such proteins in UC, CD, and healthy controls ( n = 15, 11, and 20, respectively). All six proteins were differentially expressed in patients versus controls (non-parametric Wilcoxon test had 10 −8 < p < 4.5 × 10 −3 , Fig. 4 e–j). Using random elastic net, we ranked those module intersection genes by their predictive value in discriminating patients with CD from patients with UC. We proceeded and aimed for a combination of biomarkers to separate UC and CD. For this purpose, we applied our previously described strategy [ 59 ] that avoids any additional parameter inferences. Therefore, we expect the results to be reproducible for a new validation cohort. We normalized each protein to have unit variance and zero mean in the healthy controls and tested whether the sum of the normalized levels separated UC and CD. We found that summing the two proteins with highest individual predictive values (CXCL1 and CXCL8, Additional file 2 ) separated UC and CD with an AUC of 0.81 (double-sided p = 9.47 × 10 −3 , Fig. 4 k). We emphasize that this simple approach is likely translatable to new studies, as long as control samples exist. For all disease pairs of all 13 diseases, we therefore provide respective gene lists rank-ordered by their predictive value to find highly predictive combinations of protein biomarkers, similarly to what we described for UC and CD (Additional file 2 ). Pleiotropic mechanisms in T cells are highly enriched for genetic variants and drug targets Next, we analyzed the therapeutic potential of T cells. As described above, this was complicated by the involvement of multiple pathways and drug targets in T cells. Therefore, we needed a complementary principle to prioritize disease-associated genes in this cell type. We hypothesized that, since T cells were associated with multiple diseases, there could be overlapping, pleiotropic, disease mechanisms. If so, those mechanisms could have a relatively greater pathogenic importance and therefore be prioritized for therapeutic targeting. If this could be shown, a general implication could be improved drug prioritization based on analysis of pleiotropic mechanisms in central cell types. Indeed, modules from the 13 diseases partially overlapped on the PPI network, and their union formed what henceforth will be referred to as a shared interaction neighborhood (SIN; see the “ Methods ” section). The pathogenic and therapeutic importance of the SIN was supported by highly significant enrichment of genes harboring genetic variants associated with the GWAS diseases ( n = 261 genes, odds ratio (OR) = 2.83, p = 1.5 × 10 −37 ), as well as drug targets ( n = 302, OR = 2.92, p = 2 × 10 −53 ). Identification and validation of drugs targeting the SIN Because the above analyses supported a general pathogenic and therapeutic importance of the SIN, we hypothesized that it could be exploited to identify drugs for that could be effective in many diseases, in which CD4+ T cells had a central role. We computationally tested this hypothesis in the 13 diseases analyzed above, as well as by a therapeutic study of AIA. To find drugs that optimally targeted SIN genes, we used a recently described network-based method [ 49 ]. Briefly, we computationally predicted which of the 1790 drugs in DrugBank would mainly target genes that were in network proximity with the SIN genes. We identified five such drugs, which targeted nine of the 13 diseases (Additional file 2 ). We tested one of these drugs in the mouse model of AIA. The drug, bezafibrate, is a PPARa agonist used to treat hyperlipidemia and to prevent cardiovascular disease [ 60 ]. To our knowledge, bezafibrate treatment has not been described in RA. However, positive effects of PPARg agonists have been described [ 61 , 62 ]. In our study of the AIA mice, histological specimens from arthritic joints that had or had not been treated with bezafibrate were stained with hematoxylin and eosin and examined in a blinded manner for an arthritis score [ 33 ]. This ranged from 0–3, where 0 = no signs of inflammation, 1 = mild inflammation, with proliferation of the synovial lining layer, and 2 and 3 = different degrees of influx of inflammatory cells, as described [ 33 ]. We found that bezafibrate treatment significantly decreased the arthritis score ( p < 0.05; arthritis score in mock-treated control mice median of 1, range 0.5–2.0; bezafibrate-treated mice median of 0, range 0.0–0.5; Fig. 5 a, b). Since we have previously shown influx of neutrophils and lymphocytes in untreated arthritis [ 63 ], and no inflammatory cells were found in the joints following bezafibrate treatment, our study is the first to show this effect in a mouse model of arthritis. Because the lymphocyte response was induced by a specific antigen, bovine serum albumin (BSA), we could test if this response was affected by bezafibrate. Indeed, we found significantly decreased antigen recall responses in CD4+ T cells in a proliferation assay following bezafibrate treatment ( p = 0.032; number of proliferating cells in mock-treated control mice median of 1195, range 562–1599; bezafibrate-treated mice median of 259.5, range 107–809; Fig. 5 c). Fig. 5 Bezafibrate protects against antigen-induced arthritis (AIA). Female mice with mBSA-induced arthritis were intraperitoneally (i.p.) treated with bezafibrate ( n = 4) or mock (AIA control, n = 5). a Arthritis severity was scored based on histopathology day 28 in the two groups (H&E staining, vertical bars indicate median, differences between groups evaluated using the Mann-Whitney U test, * p < 0.05). b Representative H&E joint image from the bezafibrate-treated mice. c Antigen recall response of CD4+ helper T cells among spleen and lymph node cells isolated from mock- (AIA control, n = 5) or bezafibrate-treated ( n = 4) mice; vertical bars indicate mean ± SEM, differences between groups evaluated using the two-sided Mann-Whitney U test * p < 0.05) Full size image Discussion Understanding of pathogenic mechanisms and identification of drug targets in complex diseases are daunting challenges, because of the involvement of thousands of genes in many different cell types. Ideally, it would be possible to identify and target a single key regulatory cell type and mechanism in each disease. Traditionally, such targets are sought based on empirical or screening approaches. However, unbiased genome-wide approaches like GWAS and scRNA-seq studies have indicated the dispersion of multiple pathogenic mechanisms across many cell types [ 13 , 64 , 65 ]. This may explain the difficulties in drug discovery and why many patients do not respond to treatment. Despite this, systematic characterization and prioritization of disease-associated cell types and mechanisms for diagnostics and therapeutics remain unresolved challenges. One obvious approach would be to identify and target the most significant pathway. In our scRNA-seq study of AIA, we found that pathways involved in T cell differentiation were most significant. This is consistent with the current understanding of RA pathogenesis and has resulted in drugs targeting such pathways. However, the effects have been variable. A likely explanation was suggested by our systematic analyses of the scRNA-seq data, which revealed a large number of other pathways and therapeutic targets in each cell type, as well as limited overlap between the cell types. This indicates the need for systems-level approaches to organize and prioritize cell types and mechanisms for diagnostic and therapeutic purposes. Here, our results support that network-based principles can be applied for both. We organized scRNA-seq data from AIA and human RA into MCDMs. Instead of any unique cell type or mechanism having an obvious key regulatory role, most cell types in the MCDMs interacted, forming multidirectional networks, in which multiple cell types potentially contributed to pathogenesis. Although it is possible that one cell type and mechanism had a key role, our analysis of genetic and epigenetic data supported that pathogenic mechanisms were dispersed in multiple cell types. This led us to examine if multicellular pathogenesis was a general disease characteristic. Indeed, our analyses of GWAS-enriched epigenetic markers in 175 diseases showed that a median of 20 cell types was associated with each disease. However, those analyses were based on only 45 out of an unknown total number of cell types in the human body. Therefore, the number of disease-associated cell types is likely to be much higher. This complexity emphasized the need for strategies to prioritize the most important disease-associated cell types. Our subsequent analyses supported that such prioritization is feasible, based on centrality analyses of our scRNA-seg-based MCDMs and network models of 175 diseases. We have provided individual models of all the 175 diseases (Additional file 5 ), as well as the underlying data (Additional file 2 ), to help prioritization of cell types and tissues for scRNA-seq-based MCDM construction. Such prioritization is important because most complex diseases involve multiple organs, some of which may not be known to be affected in the diseases. For example, the network model of human RA included 24 cell types and cell categories, from different organs. One of these was bronchial epithelium, which is not clinically associated with RA, but recently proposed to have causal role in this disease [ 51 ]. While dispersion of pathogenic mechanisms in multiple cell types complicates drug discovery, high interconnectivity between cell types may have an unexpected diagnostic advantage: any central cell type can potentially serve as a sensor of all other disease-associated cell types. This was supported by our analyses of three different immune cells in peripheral blood from RA patients. To our knowledge, this potential advantage has not been previously explored. This would be complicated by the need to perform coordinated clinical studies of multiple diseases according to standardized protocols. In this work, we did perform such studies. We first identified T cells as a suitable diagnostic and therapeutic candidate cell type both because of its centrality and its clinical accessibility. The diagnostic potential of T cells was supported by prospective clinical studies of patients with 13 highly diverse diseases. Expression profiles of this cell type could be used to separate the different diseases from healthy controls and each other, with high accuracy. An independent validation study showed the potential of T cell profiling to diagnose breast cancer. We also found that the profiles could be used to infer a limited number of protein biomarkers for two diseases with partially similar phenotypes, UC and CD. We propose that the data and methods can be used to identify diagnostic proteins for any of the 13 diseases, or newly generated data from other cell types that have been prioritized based on scRNA-seq-derived MCDMs. For this purpose, we have included gene lists rank-ordered by the predictive value that can be used to prioritize biomarkers for further studies. Moreover, based on the example of UC and CD, we describe a method to identify possible combinations of biomarkers from those lists, for better classification accuracy. While high interconnectivity between cell types was diagnostically advantageous, it complicated prioritization between the many cell types, mechanisms, and drugs. Here, we focused on T cells because of the centrality and GWAS analyses. However, this approach was also challenging due to the involvement of multiple pathways and predicted drug targets in T cells. We hypothesized that, since T cells were associated with multiple diseases, this pleiotropy could be a sign of pathogenic importance. If so, the same pleiotropic mechanisms could potentially be exploited for therapeutic targeting of multiple diseases. This was supported by overlap, the SIN, between modules from the 13 diseases. The pathogenic and therapeutic importance of the SIN was shown by highly significant enrichment of genes harboring disease-associated genetic variants, as well as therapeutic targets. Using network tools, we computationally predicted five out of 1790 drugs that optimally targeted the SIN. We validated one, bezafibrate, in the AIA model. An implication of this study is that drug prioritization may be improved by analysis of pleiotropic mechanisms in central cell types. This is a finding of considerable potential importance that, to our knowledge, is novel and merits further studies. Limitations of the study include that the bioinformatics analyses of gene interactions, pathways, biomarkers, and drug targets were based on a manually curated aggregate of multiple data sources, which may be confounded by, for example, cell type- or tissue-specific variations. However, we repeated the analyses using independent ligand-receptor data with similar results. The therapeutic implications of our mouse study should be interpreted with caution, since it was based on a standardized protocol and inbred mice of the same age and sex. Indeed, the treatment effects of a related drug were less pronounced in human patients with RA [ 61 , 62 ]. A likely explanation could be the greater complexity and variability of human RA, as indicated by the pathway analyses of scRNA-seq data from human RA synovium. Thus, combinations of drugs targeting multiple pathways may be required. This highlights the need for future studies aiming at detailed characterization of dispersion of pathogenic mechanisms in MCDM cell types, as well as individual variations. A potential clinical implication is that, for severe diseases that require costly or risky medications, scRNA-based MCDMs may provide a framework to tailor treatments for individual patients, similar to how we now take high-resolution imaging for granted [ 24 ]. Another important implication is suggested by recent advances in digital medicine, where different computational methods, such as artificial intelligence, have been applied for automated diagnostics of medical images [ 66 ]. Given the molecular complexity of common diseases, successful implementation of digital medicine will require integration of high-resolution molecular data with routine data, such as medical images. We and others have developed network methods for integration of heterogeneous large-scale data, which may prove useful in this context [ 2 , 29 ]. Conclusions Our findings support that MCDMs and network principles may have the potential to prioritize cell types and mechanisms for biomarker and drug discovery. All the presented data and methods have been made available to facilitate such studies. Availability of data and materials The single-cell raw data generated in this study is publicly available on SRA database ( ) with accession PRJNA504425. The code for module construction is available at . All codes for analyzing the MCDMs are available at . Change history 28 April 2020 An amendment to this paper has been published and can be accessed via the original article. Abbreviations AIA: Antigen-induced arthritis AUC: Area under the curve BMI: Body mass index BSA: Bovine serum albumin CD: Crohn’s disease CFDA-SE: Carboxyfluorescein diacetate succinimidyl ester CFSE: Carboxyfluorescein succinimidyl ester CLL: Chronic lymphocytic leukemia CRC: Colorectal cancer DEGs: Differentially expressed genes DMSO: Dimethyl sulfoxide ENCODE: Encyclopedia of DNA Elements GWAS: Genome-wide association studies H&E: Hematoxylin and eosin HCs: Healthy controls i.p.: Intraperitoneally IBDs: Inflammatory bowel diseases IPA: Ingenuity Pathway Analysis mBSA: Methylated bovine serum albumin MCDMs: Multicellular disease models NHGRI: National Human Genome Research Institute OS: Dermatologist PASI: Psoriasis Area and Severity Index PBS: Phosphate-buffered saline PCA: Principal component analysis PFA: Paraformaldehyde PPAR: Peroxisome proliferator-activated receptor PPI: Protein-protein interaction PPIn: Protein-protein interaction network promyeloids: Myeloid progenitor cells RA: Rheumatoid arthritis RAST: Radioallergosorbent tests RCA: Reference component analysis RT: Room temperature scRNA-seq: Single-cell RNA-sequencing study SIN: Shared interaction neighborhood SNPs: Single nucleotide polymorphisms Treg: T regulatory cells UC: Ulcerative colitis

Advanced computer models of diseases can be used to improve diagnosis and treatment. The goal is to develop the models to "digital twins" of individual patients. Those twins may help to computationally identify and try the best medication, before actually treating a patient. The models are the result of an international study, published in the open access journal Genome Medicine. One of the greatest problems in medical care is that medication is ineffective in 40-70% of patients with common diseases. One important reason is that diseases are seldom caused by a single, easily treatable "fault". Instead, most diseases depend on altered interactions between thousands of genes in many different cell types. Another reason is that those interactions may differ between patients with the same diagnosis. There is a wide gap between this complexity and modern health care. An international research team aimed to bridge this gap by constructing computational disease models of the altered gene interactions across many cell types. "Our aim is to develop those models into 'digital twins' of individual patients' diseases in order to tailor medication to each patient. Ideally, each twin will be computationally matched with and treated with thousands of drugs, before actually selecting the best drug to treat the patient", says Dr. Mikael Benson, professor at Linköping University, Sweden, who led the study. The researchers started by developing methods to construct digital twins, using a mouse model of human rheumatoid arthritis. They used a technique, single-cell RNA sequencing, to determine all gene activity in each of thousands of individual cells from the sick mouse joints. In order to construct computer models of all the data, the researchers used network analyses. "Networks can be used to describe and analyse most complex systems", says Dr. Benson. "A simple example is a soccer team, in which the players are connected into a network based on their passes. The player that exchanges passes with most other players may be most important". Similar principles were applied to construct the mouse "twins", as well as to identify the most important cell type. That cell type was computationally matched with thousands of drugs. Finally, the researchers showed that the "best" drug could be used to treat and cure the sick mice. The study also demonstrated that it may be possible to use the computer models to diagnose disease in humans. The researchers focused on the same cell type that was used for drug identification. This cell type, T cells, plays an important role in the immune defence, and serves as a fingerprint of the whole digital twin. The researchers analysed T cells from patients with thirteen diseases, including autoimmune diseases, cardiovascular diseases and various types of cancer. The diagnostic fingerprints could be used not only to distinguish patients from healthy people, but also to distinguish most of the diseases from each other. "Since T cells function as a sort of spy satellite, which is continuously surveying the body to discover and combat disease as early as possible, it may be possible to use this cell type for the early diagnosis of many different diseases", says Mikael Benson.

10.1186/s13073-019-0657-3

Nano

PEDOT:PSS: Improving thermoelectric materials that convert heat to electricity and vice-versa

Engineered doping of organic semiconductors for enhanced thermoelectric efficiency, DOI: 10.1038/nmat3635 Journal information: Nature Materials

http://dx.doi.org/10.1038/nmat3635

https://phys.org/news/2013-05-pedotpss-thermoelectric-materials-electricity-vice-versa.html

Abstract Significant improvements to the thermoelectric figure of merit ZT have emerged in recent years, primarily due to the engineering of material composition and nanostructure in inorganic semiconductors 1 (ISCs). However, many present high- ZT materials are based on low-abundance elements that pose challenges for scale-up, as they entail high material costs in addition to brittleness and difficulty in large-area deposition. Here we demonstrate a strategy to improve ZT in conductive polymers and other organic semiconductors (OSCs) for which the base elements are earth-abundant. By minimizing total dopant volume, we show that all three parameters constituting ZT vary in a manner so that ZT increases; this stands in sharp contrast to ISCs, for which these parameters have trade-offs. Reducing dopant volume is found to be as important as optimizing carrier concentration when maximizing ZT in OSCs. Implementing this strategy with the dopant poly(styrenesulphonate) in poly(3,4-ethylenedioxythiophene), we achieve Z T = 0.42 at room temperature. Main Thermoelectric devices directly convert heat to electricity and vice versa without moving parts or working fluids, making them reliable and compact compared with conventional heat engines. OSCs offer numerous advantages over ISCs for thermoelectric applications, such as low cost, large-area deposition, high toughness and elasticity, material abundance and low weight. Furthermore, OSCs have low thermal conductivity ( κ ), which increases their energy conversion efficiency as defined by the thermoelectric figure-of-merit Z T = S 2 σ T / κ , where S is the Seebeck coefficient, σ is the electrical conductivity and T is the absolute temperature. As the material parameters constituting ZT do not suffer from the same trade-offs in OSCs as they do in ISCs (for example, OSCs do not typically obey the Wiedemann–Franz law, as the correlation between σ and κ is weak), they offer new routes to optimization of thermoelectric efficiency that remain for the most part unexplored. Doping is key to maximizing the thermoelectric power factor ( S 2 σ ), because it determines free-carrier concentration (and hence S ; ref. 2 ) and affects carrier mobility ( μ ). For OSCs, which traditionally suffer from low thermoelectric power factor, the effect of dopants on mobility is especially large, because dopants in van der Waals bonded solids can not only modify the conformation of conducting host molecules and thereby alter their carrier transport properties, but also typically increase the tunnelling distance between these molecules and hence greatly reduce the rate of thermally activated hopping. An important consequence of the weak van der Waals bonding characteristic of OSCs is that dopants often exhibit a very small ionization fraction ( b ) over a wide range of dopant concentrations 3 . The presence of a large number of non-ionized dopants can significantly reduce carrier mobility and consequently the thermoelectric power factor. This effect on S 2 σ is illustrated in Fig. 1 , where the free-carrier concentration ( n ) normalized by the host material’s total density of states ( N 0 ) is given on the horizontal axis and the ratio of total dopant volume to total host volume is given on the vertical axis. This total volume ratio is expressed as the product of the subunit (for example, monomer) volume ratio ( r )and the subunit concentration ratio ( χ ). Essentially these plots (described in more detail below) represent S 2 σ versus carrier concentration and mobility, because μ decreases exponentially with increasing r χ (ref. 4 ), and hence capture the principal properties governing S and σ = n q μ (where q is the unit charge). As in ISCs, S and σ have opposite dependences on n (for example, S ~ln( n / N 0 )), leading to a particular value of n that maximizes S 2 σ . However, the exponential dependence of μ on r χ leads to a variation in μ over typical doping concentrations in OSCs that is much larger than the variation in μ over typical doping concentrations in ISCs. Therefore, as Fig. 1 makes clear, minimizing r χ is just as important as optimizing n when maximizing S 2 σ in OSCs. Figure 1: The doping (or dedoping) trajectory required for efficient maximization of thermoelectric power factor in an OSC. a – c , Calculated 4 dependence of S 2 σ on the normalized carrier concentration ( n / N 0 ) and the ratio of total dopant volume to total host volume ( r χ ), for different degrees of carrier localization ( α / l ). S 2 σ is normalized by its maximum value as illustrated by the colour bar. d , Three-dimensional plot of normalized S 2 σ showing the steep ascent of the dedoping trajectory along both n / N 0 and r χ axes in the direction of maximum S 2 σ . e , Measured dedoping trajectory for DMSO-mixed (circles) and EG-mixed (squares) PEDOT:PSS. χ is determined by XPS data and n / N 0 is derived by fitting measured S to a numerical calculation 4 . The forbidden area occurs for n / N 0 > χ when r = 1.3. Arrows in magnified image for data indicate the direction of the dedoping process. Full size image We experimentally demonstrate the key importance of minimizing r χ for achieving high thermoelectric performance in an OSC by removing non-ionized dopant species from poly(3,4-ethylenedioxythiophene) (PEDOT) doped by poly(styrenesulphonate) (PSS). PEDOT:PSS is promising as an organic-based thermoelectric material owing to its stability in air 5 and potential for very high σ (measured over 3,000 S cm −1 ; ref. 6 ). As for other OSCs, it suffers from a small ionization fraction for the PSS dopant in pristine PEDOT:PSS (typically near 0.3 (refs 7 , 8 )) and requires a large amount of PSS to create a high carrier concentration; PEDOT:PSS mass ratios of 1:2.5 or 1:6 are typically used for high-conductivity samples 7 . Measurements of S 2 σ in PEDOT:PSS samples have yielded very low values (of the order of 1–10 μW m −1 K −2 ) due to low values of S at these high carrier concentrations 8 , 9 , 10 ; attempts to control PSS concentration and hence free-carrier concentration have proved difficult owing to its large molecular weight (long chain length) 11 . The hydrophilic nature of PSS and the hydrophobic nature of PEDOT offer a means for selective dedoping of PSS through the use of hydrophilic solvents such as ethylene glycol 5 (EG). Samples were first prepared by mixing PEDOT:PSS (Clevios PH1000 from H. C. Starck; 1:2.5 mass ratio) solutions with 5 vol% of EG or dimethylsulphoxide (DMSO) to enhance the electrical conductivity 7 and spin-coated on a glass substrate (see Methods and Supplementary Information for more details regarding sample preparation). After thermal annealing, the EG treatment was applied by immediately immersing the samples in an EG bath (kept at 60 °C for EG-mixed PEDOT:PSS and at room temperature for DMSO-mixed PEDOT:PSS) for a certain length of time to induce a desired amount of physical dedoping. Pairs of gold electrodes separated by various distances ( L = 60, 80, 100 and 120 μm) were deposited by electron-beam evaporation on each sample to define electrical measurement spacings of various lengths. S and σ were then derived differentially by linear fittings of thermal voltage and resistance (determined by a standard four-point probe method) versus spacing length, to remove contact effects and therefore yield more accurate results than single-spacing measurement techniques 2 , 8 , 9 (see Methods and Supplementary Information for more details regarding electrical and thermal property measurements). The electrical conductivities of the pristine spin-coated films (before dedoping) were 639±21 S cm −1 for EG-mixed PEDOT:PSS and 620±37 S cm −1 for DMSO-mixed PEDOT:PSS. The Seebeck coefficients of the pristine films were 27.3±1.9 μV K −1 (EG-mixed) and 33.4±2.2 μ V K −1 (DMSO-mixed). The abundance of PSS in freshly spin-coated PEDOT:PSS films leads to phase separation in which charge transport occurs by tunnelling between conductive nanoscale PEDOT-rich islands that are each surrounded by a PSS-rich shell 11 , 12 . As shown in Fig. 2a , EG treatment leads to a significant reduction in film thickness, mainly due to a decrease in PSS concentration 5 . This selective removal of PSS was confirmed by X-ray photoelectron spectroscopy (XPS), as the intensity of the sulphur atom (S(2p)) from the sulphonate group in PSS (166–170 eV) was observed to decrease with respect to the intensity of the S(2p) from the thiophene group (162–166 eV) in PEDOT with longer dedoping time ( Fig. 2b ). At 60 and 120 min of dedoping, the S(2p) signals for PSS and PEDOT have the same intensity, indicating that the concentrations of PEDOT and PSS monomers have become equal and stabilized within the sample. As the hydrophilic PSS molecules become covered by hydrophobic PEDOT molecules, selective removal of PSS by the hydrophilic EG solvent becomes less and less efficient. Furthermore, after approximately 60 min the film thickness stops decreasing (as shown in Fig. 2a ) and the Seebeck coefficient levels off (as shown in Fig. 3a ), indicating that the carrier concentration has become constant. These observations indicate that the dedoping process comes to an end after approximately 60 min. Figure 2: Selective removal of PSS by EG treatment. a , Measured thickness of PEDOT:PSS film at various EG treatment times (that is, various PSS dedoping levels). b , S(2p) XPS spectra of EG-mixed and DMSO-mixed PEDOT:PSS films before PSS dedoping and after 60 and 120 min of dedoping. Full size image Figure 3: Thermoelectric properties of PEDOT:PSS at various dedoping times. a – e , Seebeck coefficients ( a ), electrical conductivities ( b ), thermoelectric power factors ( c ), vertical (cross-plane) thermal conductivities ( d ) and thermoelectric figure-of-merit ( e ) at T = 297 K in EG-mixed and DMSO-mixed PEDOT:PSS measured during the EG treatment (dedoping) process. Lateral (in-plane; in the direction of measured S 2 σ ) thermal conductivities were used to derive ZT . Error bars for S and σ were defined as the standard deviation of measured data from the linear fit line, and error bars for were determined by the standard deviations of measured temperature rise at different frequencies and the standard deviations of measured thicknesses (see Supplementary Information ). f , σ (normalized by maximum value for b = 1) versus S for different magnitudes of b (assuming r = 1.3). Inset: derived b during the dedoping process. Full size image The presence of dopants with non-zero volume ( r χ > 0) leads to a decrease in the total carrier density of states ( N t ) relative to its undoped value ( N 0 ), which can be calculated using N t = N 0 /(1+ r χ ). This smaller number of states per volume results in an increased tunnelling distance for hopping 4 and hence an exponential decrease in μ . This decrease in μ further depends on the carrier localization length ( α ) with respect to the distance between PEDOT monomers ( l ), because α and the tunnelling distance together define the overlap of carrier wavefunctions between initial and final states that determines the tunnelling probability. The effect of r χ on S 2 σ is plotted in Fig. 1 for different α / l , based on numerical calculations 4 for a Gaussian density of states width of 0.1 eV (typical for OSCs at room temperature 13 ). To track the dedoping process for EG-treated PEDOT:PSS, χ was measured at various dedoping times by calculating the ratio of PSS and PEDOT S(2p) intensities. r was approximated as 1.3 based on the ratio of PSS monomer and PEDOT monomer molecular weights (182 and 140, respectively). A normalized localization length of α / l = 2.1 was used for both EG-mixed and DMSO-mixed PEDOT:PSS (ref. 4 ), because α has been shown to be primarily governed by the host material and not significantly affected by dopant type or chemical additive 4 , 14 . The above properties were then used in a numerical calculation of Seebeck coefficient 4 to derive n / N 0 at the corresponding dedoping times. As shown in Fig. 1e , the observed dedoping trajectories in PEDOT:PSS aim straight at the point of maximum S 2 σ by not only reducing n / N 0 but also by reducing r χ , leading to significant enhancements in S 2 σ ( Figs 1d and 3d ). As shown in Fig. 1e , the dedoping trajectory reaches n / N 0 = 0.13 and χ = 0.96; at this point the DMSO-mixed samples reaches S 2 σ = 469 μW m −1 K −2 , which is approximately half of the calculated maximum possible value for PEDOT:PSS (~1,100 μW m −1 K −2 ). d σ /d S is typically negative in ISCs, because changes in μ are small relative to changes in n , and S and σ have opposite dependences on n . In both EG-mixed and DMSO-mixed PEDOT:PSS, however, S and σ are observed to simultaneously increase during dedoping ( Fig. 3a,b ), indicating that the mobility enhancement due to a decrease in tunnelling distance overwhelms the reduction in carrier concentration. As shown in Fig. 3f , this unique trend of positive d σ /d S occurs in doped OSCs at high dopant concentrations and depends on the dopant ionization fraction b . Values of b in PEDOT:PSS were derived using b = 1/ χ × n / N t and were found to be almost constant ( b ≈0.3) over the range of dedoping levels tested ( Fig. 3f , inset), indicating that the EG-treatment dedoping process removed approximately 7 non-ionized PSS molecules for every 3 ionized PSS molecules. Cross-plane (vertical) thermal conductivity in PEDOT:PSS at various dedoping times was determined by a differential 3 ω method 15 that compared the thermal response to a line heater in samples with PEDOT:PSS layers to that of samples without PEDOT:PSS layers. As shown in Fig. 3c , as dedoping progressed, decreased from 0.30 to 0.22 W m −1 K −1 in DMSO-mixed PEDOT:PSS and from 0.32 to 0.23 W m −1 K −1 in EG-mixed PEDOT:PSS. As PSS molecules are much larger than PEDOT molecules and contain a much greater number of covalent bonds, the removal of PSS and consequent increase in the average van der Waals character of bonds within the sample probably explains this thermal conductivity reduction. While carrier transport ( S 2 σ ) is measured in the in-plane (lateral) direction, measurement of thermal conductivity in this direction ( κ ∥ ) is challenging for low-thermal-conductivity films less than 100 nm thick. To perform in-plane thermal measurements, thick films were prepared by multiple spin-coatings with the same spin conditions used for thin films. As the film anisotropy is expected to be induced mainly by spin-coating 11 , 12 and the spin-coating parameters were the same for thick and thin samples, κ ∥ is expected to be similar for both thick and thin films of the same dedoping. By comparing the thermal responses of the thick film to line heaters of wide width (50 μm film thickness) and narrow width (2 μm≈film thickness), both the thermal anisotropy factor and the cross-plane value were measured and used to derive κ ∥ (ref. 16 ). For pristine (non-dedoped) thick films, κ ∥ was found to be 0.42±0.07 W m −1 K −1 for DMSO-mixed PEDOT:PSS and 0.52±0.11 W m −1 K −1 for EG-mixed PEDOT:PSS. On the basis of the expectation that dopants have an isotropic effect on thermal transport, the measured thermal anisotropy factor ( for DMSO-mixed PEDOT:PSS and 1.62±0.35 for EG-mixed PEDOT:PSS) was assumed to be constant with dedoping and was used to convert to κ ∥ at various dedoping levels. Using in-plane values for both thermoelectric power factor and thermal conductivity, we derive a maximum ZT value of 0.42 for DMSO-mixed PEDOT:PSS at room temperature and a maximum value of 0.28 for EG-mixed PEDOT:PSS ( Fig. 3e ), the former being the highest value yet reported among OSCs. This high thermoelectric figure-of-merit indicates the importance of minimizing dopant volume ( r χ ) when engineering the thermoelectric properties of an organic semiconductor, which has been shown to simultaneously improve all of the properties constituting ZT . Methods PEDOT:PSS mixed with 5 vol% of EG or DMSO was spin-cast on a pre-cleaned glass substrate for electrical measurements ( S and σ ) and on a pre-cleaned silicon substrate (capped with 100 nm SiO 2 ) for thermal measurements ( κ ). Spin-coating conditions were kept the same for both electrical and thermal samples to ensure the same anisotropy in the resulting films. After spin-coating, samples were thermally annealed at 130 °C for 15 min, and were immersed in EG solvent for various times to induce different levels of dedoping. All of the above steps for sample preparation were performed in nitrogen gas; the resultant PEDOT:PSS films were stable in air. Electrical ( S and σ ) and thermal ( κ ) measurements were made in ambient (air) conditions. S and σ were determined differentially, with several electrode spacings used to remove contact effects. Both the electrical resistance and the thermal voltage measured between the electrodes increased linearly with spacing length ( Supplementary Fig. S3 ). Thermal conductivity was determined by a differential technique in which the temperature difference across the PEDOT:PSS layer was isolated by comparing samples with and without the PEDOT:PSS layer. The derived temperature difference across the PEDOT:PSS layer was constant over a range of frequencies tested. More information regarding both sample preparation and sample property measurements can be found in the Supplementary Information .

Thermoelectric materials can be used to turn waste heat into electricity or to provide refrigeration without any liquid coolants, and a research team from the University of Michigan has found a way to nearly double the efficiency of a particular class of them that's made with organic semiconductors. Organic semiconductors are carbon-rich compounds that are relatively cheap, abundant, lightweight and tough. But they haven't traditionally been considered candidate thermoelectric materials because they have been inefficient in carrying out the essential heat-to-electricity conversion process. Today's most efficient thermoelectric materials are made of relatively rare inorganic semiconductors such as bismuth, tellurium and selenium that are expensive, brittle and often toxic. Still, they manage to convert heat into electricity more than four times as efficiently as the organic semiconductors created to date. This greater efficiency is reflected in a metric known by researchers as the thermoelectric "figure of merit." This metric is approximately 1 near room temperature for state-of-the-art inorganic thermoelectric materials, but only 0.25 for organic semiconductors. U-M researchers improved upon the state-of-the-art in organic semiconductors by nearly 70 percent, achieving a figure-of-merit of 0.42 in a compound known as PEDOT:PSS. "That's about half as efficient as current inorganic semiconductors," said project leader Kevin Pipe, an associate professor of mechanical engineering as well as electrical engineering and computer science. Pipe is a co-author of a paper on the research published in Nature Materials on May 5, 2013. PEDOT:PSS is a mixture of two polymers: the conjugated polymer PEDOT and the polyelectrolyte PSS. It has previously been used as a transparent electrode for devices such as organic LEDs and solar cells, as well as an antistatic agent for materials such as photographic films. One of the ways scientists and engineers increase a material's capacity for conducting electricity is to add impurities to it in a process known as doping. When these added ingredients, called dopants, bond to the host material, they give it an electrical carrier. Each of these additional carriers enhances the material's electrical conductivity. In PEDOT doped by PSS, however, only small fraction of the PSS molecules actually bond to the host PEDOT; the rest of the PSS molecules do not become ionized and are inactive. The researchers found that these excess PSS molecules dramatically inhibit both the electrical conductivity and thermoelectric performance of the material. "The trouble is that the inactive PSS molecules push the PEDOT molecules further apart, making it harder for electrons to jump between PEDOT molecules," Pipe said. "While ionized PSS molecules improve electrical conductivity, non-ionized PSS molecules reduce it." To improve its thermoelectric efficiency, the researchers restructured the material at the nanoscale. Pipe and his team figured out how to use certain solvents to remove some of these non-ionized PSS dopant molecules from the mixture, leading to large increases in both the electrical conductivity and the thermoelectric energy conversion efficiency. This particular organic thermoelectric material would be effective at temperatures up to about 250 degrees Fahrenheit. "Eventually this technology could allow us to create a flexible sheet—-think of Saran Wrap—-that can be rolled out or wrapped around a hot object to generate electricity or provide cooling," Pipe said.

10.1038/nmat3635

Biology

Fish and ships: Vessel traffic reduces communication ranges for Atlantic cod, haddock

Jenni A. Stanley et al, Underwater sound from vessel traffic reduces the effective communication range in Atlantic cod and haddock, Scientific Reports (2017). DOI: 10.1038/s41598-017-14743-9 Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-017-14743-9

https://phys.org/news/2017-12-fish-ships-vessel-traffic-ranges.html

Abstract Stellwagen Bank National Marine Sanctuary is located in Massachusetts Bay off the densely populated northeast coast of the United States; subsequently, the marine inhabitants of the area are exposed to elevated levels of anthropogenic underwater sound, particularly due to commercial shipping. The current study investigated the alteration of estimated effective communication spaces at three spawning locations for populations of the commercially and ecologically important fishes, Atlantic cod ( Gadus morhua ) and haddock ( Melanogrammus aeglefinus ). Both the ambient sound pressure levels and the estimated effective vocalization radii, estimated through spherical spreading models, fluctuated dramatically during the three-month recording periods. Increases in sound pressure level appeared to be largely driven by large vessel activity, and accordingly exhibited a significant positive correlation with the number of Automatic Identification System tracked vessels at the two of the three sites. The near constant high levels of low frequency sound and consequential reduction in the communication space observed at these recording sites during times of high vocalization activity raises significant concerns that communication between conspecifics may be compromised during critical biological periods. This study takes the first steps in evaluating these animals’ communication spaces and alteration of these spaces due to anthropogenic underwater sound. Introduction Sound is an efficient way to communicate in the marine environment, and animal inhabitants and people alike have developed ways to exploit this fact. Many organisms occupying the oceans actively use and produce sound. Marine mammals use sound as a primary method for communicating underwater over large distances, over shorter spatial scales fishes do the same. Marine invertebrates produce sound both actively for behavioural display purposes as well as passively due to feeding or movement. Features of ambient sound are a result of the characteristics of all of the contributing sound sources, including those composed of biological sounds such as animals vocalising (biotic), physical sounds, such as wind and water movement (geophysical or abiotic) and anthropogenic sounds such as shipping or construction (anthropogenic) 1 . Many marine organisms utilize ambient sound to navigate, choose their settlement or residence location, and to modify their daily behaviour, e.g., breeding, feeding and socializing 2 , 3 , 4 . Due to these reasons, ambient underwater sound is an important feature of marine habitats. Anthropogenic sound in certain ocean regions has increased considerably in recent decades due to various human activities such as resource acquisition, global shipping, construction, sonar, and recreational boating 1 . As ocean sound increases, so does the concern for its effects on populations of acoustic signallers, making this a topic of significant scientific research focus. Effects of anthropogenic sound exposure can be seen in the physiology and behaviour of a range of marine organisms, from invertebrates 5 , 6 to marine mammals 7 , with studies on these effects to date largely focusing on high-amplitude sources. Sound exposure can cause temporary hearing loss and threshold shifts 8 , reduction in temporal resolution ability 9 , damage and hair cell death in the inner ear 10 , 11 , and stress responses 12 . However, few studies have addressed the effects of lower-level and chronic sound exposures 13 , 14 . Fishes represent over half of all vertebrate species, and more than 800 species from greater than 100 families are known to produce sound 15 . Not surprisingly these families have evolved a large diversity of sonic organs and sound producing mechanisms 16 . This variety of mechanisms has led to the production of diverse vocalizations and acoustic characteristics between species and populations. Fish vocalizations are an important component of the marine soundscape 17 , 18 and they provide valuable information regarding the behaviour of the signaller in a variety of different contexts, such as general interactions, territorial displays, feeding, contact vocalization, and courtship interactions 19 . Fishes exhibit an array of reproductive and social behaviours and the majority of species live fairly close to the coast or in fresh water environments, consequently they are exposed to the various human activities which produce sound 20 . In addition to the increasing amount of anthropogenic sound in the marine environment, these sounds often have prominent frequencies which fall within the frequency range of acoustic communication signals, therefore, having the potential to reduce communication efficiency. One of the most widespread, yet poorly understood means in which fishes could be affected by chronic, lower-level anthropogenic sound, such as vessel sound, is through the disruption of acoustic communication by masking 21 , 22 . In this situation, the receiver experiences an increase in the threshold of detection or discrimination of the signal which could potentially lead to complete or partial loss of received signal, misinterpretation of the signal, and/or subsequently changes in the response 21 , 23 . Although there is a growing body of literature on how signallers may avoid masking from anthropogenic sound, much of the research has been conducted on terrestrial organisms and marine mammals. To date, there have been very few documented studies on the potential of anthropogenic sound to mask, disrupt or reduce acoustic communication in fishes, and fewer still on the means of avoiding masking in the presence of extraneous sound 24 , 25 . Assessing the impacts of anthropogenic sound on the ecology of fishes is not the only concern. Fish provide livelihoods to hundreds of millions of people, and is a primary source of protein for >1 billion people worldwide with growth expected for more than 9 billion by 2050 26 , which is a difficult target without strict and ongoing management. The central topographical feature within the Gerry E. Studds Stellwagen Bank National Marine Sanctuary (SBNMS), Stellwagen Bank itself, has supported high catch rates of Atlantic cod ( Gadus morhua ) and haddock ( Melanogrammus aeglefinus ) for centuries, and includes past knowledge of predicable spawning areas for cod within the sanctuary and greater Massachusetts Bay. Gulf of Maine cod stock contains genetically distinct spring- and winter-spawning subpopulations, and recent studies have highlighted waters both inshore and within the sanctuary as supporting seasonal spawning activity 27 . The spawning components of the Gulf of Maine cod stock are overfished, with the population at a historic low of about 82% less (winter stock) and 77% less (spring stock) than the same populations a decade ago 28 . The Gulf of Maine haddock stock is currently considered stable, and fishing quotas have recently been dramatically increased for this species due to an increase in stock size and to compensate for tighter controls on ground fish like cod 29 . The sanctuary and greater Massachusetts Bay waters also support the spawning activity of haddock, with major spawning locations on Stellwagen Bank occurring from January to May, usually peaking in February to April 30 . In addition to supporting these biologically important habitats, the sanctuary experiences high anthropogenic activity and subsequently increased levels of ambient underwater sound, particularly due to commercial shipping with a Traffic Separation Scheme running through its centre. The purpose of this investigation was to examine the ambient soundscape (up to 1000 Hz) at three sites, two within and one inshore of SBNMS, which have been documented to support spawning activity for the Gulf of Maine cod and haddock stocks. These data were then used to calculate the estimated effective vocalization radius for each species in these areas during spawning time periods. These results take the first steps in assessing these animals’ communication spaces and the alteration of these spaces due to varying levels of background sound. Results Vocalization characteristics Vocalizations from Atlantic cod and haddock were present during the three-month recording period at each of the study sites. Atlantic cod grunts and haddock knocks during these recording periods were consistent with previously reported spawning vocalizations, with frequency and time-based measurements matching previous studies 31 , 32 . In the Atlantic cod spawning sites, ‘grunts’ were present for 100% (spring) and 83% (winter) of the days within the three-month sampling period. In the haddock spawning site ‘knocks’ and variations of the knock were present for 62% of the days within the three-month period (January 88.5%, February 75% and March 50%). Atlantic cod “grunts” (n = 40) had a mean peak fundamental frequency (f 1 ) of 53 Hz (range = 41–69 Hz), mean duration of 232 ms (159–541 ms) and a mean number of pulses of 9.2 pulses per grunt (range = 7–11). Haddock “knocks” consisted of several arrangements including short slow knocks, short fast knock, long slow knocks, and long fast knocks (Fig. 1a ). Haddock “knocks” (n = 40) had a mean peak frequency of 258 Hz (range = 184–356 Hz), mean sound duration of 6.3 s (range = 389 ms – 36 s), and mean number of knocks of 15.4 (range = 3–132) (Fig. 1b ). Figure 1 Panels showing acoustic characteristics of a spawning vocalization of a representative ( a ) an Atlantic cod grunt recorded within the three recording sites, and ( b ) Haddock knocks. Top panels: spectrogram of vocalizations, indicating frequency range. Middle panels: wave form of vocalization. Bottom panels: normalized power spectral density (PSD). Sounds were band pass filtered between 30 and 500 Hz for Atlantic cod and 100–1000 Hz for Haddock during the boxed pulses for the normalized PSD. Spectrograms and power spectra were computed using a 512-point fast Fourier transform (FFT), Hann-window, 80% overlap. Fish images with full permission by Scandinavian Fishing Year Book. Full size image Ambient sound levels Ambient sound levels at the three study sites ranged from 84.7 to 139.9 dB re 1 µPa in the full spectrum band (10–1000 Hz) and 78.8 to 137.7 dB re 1 µPa in the combined octave bands (1,2 & 3,4) (Table 1 , Fig. 2 ). Both full spectrum and combined octave band (matched with species vocalization) ambient sound pressure levels differed significantly among sites over the three-month sampling period (full spectrum: Kruskal-Wallis; H = 12128, P < 0.001, combined octave band: H = 13518, P < 0.001). Table 1 Summary statistics of full spectrum (10–1000 Hz) and combined octave band sound pressure levels, AIS data, and estimated effective vocalization radius from each of the three recording sites over the three-month recording period. Full size table Figure 2 Multiplot showing daily means of full spectrum and combined band octave sound pressure levels (SPL RMS ), and estimated effective vocalization radius (m). ( a ) Atlantic cod spring spawning site, ( b ) Atlantic cod winter spawning site, break in graph indicates a period of no acoustic data, ( c ) haddock winter spawning site. Combined band octave levels for both the winter and spring Atlantic cod spawning include bands 1 & 2, and for the haddock winter spawning site include bands 3 & 4 to best match vocalization frequency (see methods for f bw ). Full size image The Atlantic cod winter spawning site had both the highest mean full spectrum sound level (111.1 dB re 1 µPa) and combined octave band levels (103.6 dB re 1 µPa) of all the recording sites over the three-month spawning period, as well as the highest 10 th and 90 th percentiles respectively (Full spectrum; 105.1 and 117.3 dB re 1 µPa, Combined bands; 97.7 and 111.2 dB re 1 µPa) and maximum sound level of 139.9 dB re 1 µPa which was due to a large vessel transiting over the site (Table 1 ). The Atlantic cod spring spawning site had both the lowest mean full spectrum sound level and combined octave band levels (99.7 and 89 dB re 1 µPa respectively) of the three recording sites, as well as the lowest 10 th and 90 th percentiles respectively (Full spectrum; 94.1 and 105.5 dB re 1 µPa, Combined bands; 82.6 and 95.1 dB re 1 µPa). The haddock winter spawning site had intermediate mean full spectrum sound levels (105.6 dB re 1 µPa) and combined octave band levels (99.5 dB re 1 µPa) over the three-month spawning period, as well as intermediate percentiles (Figs 2 & 3 ). Figure 3 Example of visual representation of 1-hour vessel passage and haddock vocalizations at haddock winter spawning site. ( a ) spectrogram of 1-hour vessel passage, ( b ) Full spectrum sound level over 20–1000 Hz frequency range, ( c ) Power spectra of 20 sec length of recording when vessel is at its closest approach to hydrophone with a full spectrum sound level of 113.5 dB re 1 µPa in the 10–100 Hz frequency range (>90 th percentile), ( d ) Power spectra of 20 sec recording when vessel left immediate vicinity of hydrophone with a full spectrum sound level of 101.2 dB re 1 µPa in the 10–1000 Hz frequency range (50 th percentile). In figure a) indicates time section for plot c and ◽ indicates time section for plot d. Colour bar units are dB re 1 µPa 2 Hz −1 . FFT: 1024, Hann window, 80% overlap. Full size image The pairwise multiple comparison demonstrated the Atlantic cod winter site had the highest full spectrum sound pressure levels (Dunn’s; vs. haddock winter Q = 54, P < 0.001, vs. Atlantic cod spring Q = 111, P < 0.001), followed by the haddock winter site (vs. Spring cod Q = 58, P < 0.001), and the Atlantic cod spring site had the lowest sound levels over the three-month period. The differences in combined octave band sound levels among sites followed the same pattern as the full spectrum levels, with the Atlantic cod winter site having the highest sound levels (Dunn’s; vs. haddock winter Q = 36, P < 0.001, vs. Atlantic cod spring Q = 114, P < 0.001), followed by the haddock winter site (vs. Atlantic cod spring Q = 79, P < 0.001) and finally the Atlantic cod spring site (Table 1 , Fig. 3 ). Effective vocalization radius estimation As the estimated effective vocalization radius was calculated by integrating the varying levels of ambient sound over the duration of the sampling period, these ambient levels influenced the effective vocalization radius greatly. The estimated effective vocalization radius at the three study sites ranged from 1.2 to 21.6 m (Table 1 , Fig. 2 ), with significant differences among the three sites (Kruskal-Wallis; H = 273, P < 0.001). The Atlantic cod spring site had the greatest mean effective vocalization radius of 15.3 m. The vocalization radius was less than 11.3 m 10% of the sampling period, or seven out of 74 days, and 19 m or less for 90% of the sampling period, or 67 of the 74 days. The winter Atlantic cod site had a much lower estimated effective vocalization radius than its spring counterpart with a mean of 2.7 m. The radius was 2.1 m or less for eight days, 10% of the sampling period, and was no greater than 3.4 m, 90% or 75 of the 84 sampling days. For the haddock winter spawning site, two source levels (SL) were used to estimate the effective vocalization radius, a low SL and a high SL (for rationale see methods section). Of all four vocalization estimates, the lower estimate for haddock winter (L) had the smallest mean estimated vocalization range of 1.8 m. The estimated vocalization radius was 1.5 m or less for eight of the 82 sampling days (10%) and 2.2 m or less for 74 of the 82 days. Automatic Identification System Vessel Tracking and relationship with ambient sound levels To further understand the contribution vessel sound had on the ambient soundscape at the recording site, the relationship between the daily numbers of AIS tracked vessels within the 10 nm radius and the daily combined octave band sound pressure levels were tested for a correlation (Fig. 4 ). Figure 4 Maps showing ( a ) locations of recording sites within Massachusetts Bay and Stellwagen Bank National Marine Sanctuary in relation to the adjacent northeast coast of the United States, ( b ) AIS vessel tracks over the three-month recording period for both the Atlantic cod winter spawning site and the haddock winter spawning site within a 10 nm radius. Polygons marks the boundaries of Stellwagen Bank National Marine Sanctuary. location of the Spring Cod Conservation Zone, the site of the Atlantic cod spring spawning recording location. location of the Atlantic cod winter spawning recording location. location of the haddock winter spawning recording location. Boston traffic separation scheme. Maps created in ArcMAP 10.3.1 . Full size image There was a significant difference in daily number of AIS tracked vessels within a 10 nm radius between the Atlantic cod winter, Atlantic cod spring and haddock winter spawning sites (Kruskal-Wallis; H = 184.9, P = < 0.001). The Atlantic cod spring spawning site had the greatest number of AIS tracked vessels with a mean of 14.5 vessels per day, compared to means of 7 and 2.7 per day at the Atlantic cod winter and haddock winter spawning sites (Table 1 , Figs 4 & 5 ). Figure 5 Bar graphs showing daily means of the estimated effective vocalization radius (m) and the daily number of AIS tracked vessels within a 10 nm radius of the recording site. ( a ) Atlantic cod spring spawning site, ( b ) Atlantic cod winter spawning site and ( c ) haddock winter spawning site, with H and L indicating effective vocalization radius calculated with high and low haddock source levels respectively. The breaks in graph indicates a period of no AIS data. Note different axis scales. Full size image For the Atlantic cod winter spawning site, there was a statistically significant positive correlation between the daily number of AIS tracked vessels and the daily combined octave band sound levels; when the number of daily AIS vessels increased, the combined octave band sound levels increased (Pearson correlation; r (79) = 0.543, p < 0.0001). Since the effective vocalisation radius was calculated using the ambient sound levels, there was also a significant negative relationship between the number of AIS vessels and the daily estimated effective vocalization radius ( r (79) = −0.544, p < 0.0001). A similar relationship occurred at the haddock winter spawning site: there was a statistically significant positive correlation between the daily number of AIS vessels and the daily combined octave band sound levels ( r (78) = 0.509, p < 0.0001), and consequently a statistically significant negative relationship between the daily number of AIS vessels and the daily estimated effective vocalization radius (high source level: r (78) = −0.495, p < 0.0001, low source level: r (78) = −0.495, p < 0.0001). However, the Atlantic cod spring spawning site showed no significant relationship between the daily number of AIS vessels and the daily combined octave band sound levels ( r (72) = −0.129, p = 0.279) or the daily estimated effective vocalization radius ( r (72) = −0.124, p < 0.301). Discussion Rising levels of anthropogenic underwater sound is of mounting concern in all marine environments. While high intensity sources hold much of research and management attention, more moderate sounds of much longer duration, like those produced by commercial shipping vessels, dominate background noise conditions over much larger areas and thus have the potential to effect greater numbers of marine animals. The results from the present study illustrated that ambient sound across the Atlantic cod and haddock spawning sites varied significantly, and as a result so too did the estimated effective vocalisation radius. These spaces were extremely reduced in the presence of sound produced by large vessels and at times the vocalisations of fin whales. Both the “grunt” and “knock” vocalizations emitted by Atlantic cod and haddock occupy the same frequency range as many underwater anthropogenic sound sources 33 , with the peak of acoustic energy in the 50–260 Hz frequency band 34 , 35 . Field and laboratory measurements have shown that this bandwidth contains the range of the greatest acoustic sensitivity in both species 36 , 37 . The mean effective vocalization radii for spawning Gulf of Maine cod were estimated to be 2.7 m in winter and 15.3 m in spring spawning locations. Similarly, over the three-month winter spawning sample, Gulf of Maine haddock mean effective vocalization radii were between 1.8 m (low SL) to 3.5 m (high SL). The difference in effective vocalization radius between cod spawning locations appeared to be largely driven by the presence of large vessel activity in the surrounding environment, e.g., in the Atlantic cod winter spawning site the effective vocalization radius was as low as 1.3 m when there was a total of 13 AIS tracked vessels within a 10 nm radius of the recording site, and as high as 4.4 m in the presence of two AIS tracked vessels. There is no baseline information available on the distances cod and haddock have evolved to use acoustic signals . It would be informative to examine masking under a range of different conditions in which they spawn, including other populations or sites with lower vessel traffic. Unfortunately, there are very few locations known where these populations spawn that are not heavily impacted by humans and heavily targeted as a fishery resource. Atlantic cod and haddock are known to exhibit complex “lekk” spawning behaviour, whereby males arrive to spawning grounds first and form dense aggregations over a small area and compete for dominance and females, via courtship displays, acoustic communication and aggression towards rivals. Females visit the aggregation, select a dominant male, initiate a spawning event and return to previous locations 38 , 39 . Therefore, females are not in continuous contact with males during the spawning season, and an attraction cue is essential for courtship 40 . Vertical and horizontal separation between males and females in spawning locations have been reported for several populations of Atlantic cod in the wild 40 , 41 . Haddock vocalization behaviour in the spawning season indicates that acoustic signals may also be used as a medium range signal to mediate the migration or attraction to spawning locations in transient populations, and not only over short distances (0–10 m) 42 , 43 , 44 . Male haddock have often been observed to repeat long knocking vocalizations for hours at a time, often in solitary display with no other fish close by 35 , 45 . This behaviour indicates that the male is occupying a home range or territory where it is exhibiting unambiguous sexual readiness to females 35 . It is also hypothesized that the chorus of large aggregations of male Atlantic cod at spawning locations may serve as a long-range signal attracting females to the area 46 . If the signal or chorus is undetected or misinterpreted due to masking it could lead to the mistiming or unsuccessful location of spawning aggregations, which is critical to the survival of these managed populations. Mounting evidence suggests that acoustic communication can affect the survival and reproductive success of fishes, including direct evidence for Atlantic cod 47 . The Gadidae family contain several vocal species, where the sounds produced are species specific and usually relatively simple. However, haddock, produces a variety of knock sounds which are used in a diverse range of behavioural contexts 31 , 38 . Evidence suggests that haddock vocalizations serve to not only get male and female fish together in a specific part of the ocean but also play a key role in synchronizing the reproductive behaviour in males and females 35 . Unlike haddock who have a wide acoustic repertoire, Atlantic cod are thought to be less versatile vocalists during courtship, they produce single “grunts” which are believed to function as both an agonistic display but also to be especially significant as a reproductive advertisement and used during spawning 48 . If anthropogenic sound reduces the efficiency of the vocalizations utilized by these species, this interference could potentially impact their reproductive success and survival through the incorrect assessment of the quality of potential mates or competitors, reduction in the ability to attract mates and/or the mistiming of gamete release. Although the behavioural effects of masking are often difficult to measure, other quantifiable effects of anthropogenic sound on the reproductive and developmental physiology of Atlantic cod have been documented 14 , 49 . Sierra-Flores et al ., (2015) demonstrated that a daily randomized 60-minute exposure to a linear sweep (100–1000 Hz @ 132 dB re 1 μPa) over a two-week period resulted in a significant reduction in total egg production and fertilization rates, reducing the total number of viable embryos by over 50% compared to a control. Effects of anthropogenic sound are also not limited to the adult population, Nedelec et al . (2015) revealed that for newly hatched Atlantic cod two days of exposure to both regular and random noise from ships reduced growth, and led to faster yolk sac utilization. After 16 days, fish exposed to regular ship noise had reduced body width-length ratios and were easier to catch in predator avoidance experiments 14 . Several studies conducted in the field on marine teleosts (including cod and haddock) have confirmed that masking of a signal can occur under relatively quiet background sound conditions 36 , 50 . These studies demonstrated that hearing thresholds increased with decreasing frequency separation between the signal and the masking sound band 51 . As with many other organisms, fishes evolved in environments with varying levels of background ambient sound, they have regularly encountered loud sources of naturally occurring biotic and geophysical sounds including wind, rain, the action of waves at the surface of the water. There are also several examples that illustrate both hetero- and conspecific sounds have the potential to overlap in both the frequency and time domain and therefore have the potential to mask communication 52 , 53 , 54 . Furthermore, several solutions, “masking releases”, to ensure the audibility of a signal over the background noise have been observed in other taxa and have been suggested could be occurring in fishes. The simplest way to avoid the impacts of a potential threat is to avoid it; however, when applying this to underwater sound it is certainly not always possible. This is especially true if a species is dependent on a certain area for critical resources, or with sources whose sounds dominated certain biologically relevant frequencies and have long-distance propagation properties, such as the low frequency sound produced by large vessels 55 . These limitations apply to both populations in this study and particularly for spawning components of Gulf of Maine cod in Massachusetts Bay, as they are known to exhibit extremely site fidelity returning to the same spawning locations year after year 56 . There is also evidence for directional masking release which could allow for the detection and discrimination of a signal in the presence of another. Although the particle velocity component of the sound field was not measured in the current study, it may prove to be important when addressing short-range communication and masking in these species. In some cases, the sound pressure and particle velocity components of the sound field are directly related, however, in relatively shallow waters this is not the case, and deviations between these two components can be especially high in the near field (close to the sound source). Although one’s ability to detect vocalisations wasn’t addressed in data of the current study, this component may be a factor resulting in vocalisations being detectable at different ranges than those calculated by pressure based equations 57 . Additionally, the particle velocity component encountered in the nearfield may give listeners information on the directionality of the sound sources when the signal of interest and the competing sound are received from differing directions. This may allow the listener to utilize binaural timing differences to enhance discrimination 58 . This has been briefly investigated in a few marine mammal species 59 , 60 , and could be possible in fishes, however, the extent of the masking release is unknown 37 . Temporal avoidance is another potential solution used by signallers in the presence of high background sound. Here a signaller could take advantage of intrinsic gaps or fluctuations in the competing signals. However, this strategy is not always possible in species using acoustic communication during critical time periods due to daylight, tidal or moon phase synchrony, and especially in the case of locations where the low frequency sound from large commercial vessels is the competing signal. In this case there is often no predicable time of the day where vessel noise is not present, or in some locations where distant low-frequency vessel sound is a constant. Temporal adjustments to the communication signal itself, such as increasing the duration of a brief signal has also shown to considerably enhance its detectably 61 . In the case of longer acoustic signals, an increase in duration or redundancy by repetition can subsequently increase the probability that part of the signal will be received during a period of quieter background sound 62 , 63 . This mechanism seems plausible for some fish species, especially those who exhibit complex or repetitive vocalization structure, such as haddock. Conversely, this may not be possible for species with simple or singular vocalization structure, such as Atlantic cod. With high interspecific diversity, temporal patterns of sounds are thought to be the most crucial factors in carrying information during acoustic communication in many teleost fishes, with receivers extracting information from pulse number, duration and repetition rate 64 , 65 . Consequently, population wide adaptations in temporal signal patterns could certainly occur across evolutionary time, but seem unlikely to occur over the current time scale of increasing background sound. There has been less than a handful of studies that have shown fish species capable of altering their vocalization signals in the presence of anthropogenic sound, and those that have, Cyprinella vesta and Opsanus tau , have been found to alter the power of their call (Lombard effect) while also changing behavioural traits such as distance between signalers 66 , 67 . In the current study, ambient sound levels fluctuated greatly within recording sites, occurring on both an hourly and daily time scale (85–143dB and 79–138dB re 1 µPa full spectrum and combined octave bands respectively) . In support with previous studies, large vessel activity around the recording sites appeared to be a predominant contributor to increases in the ambient sound levels, especially in the acoustic bands occupied by the vocalizations, and regularly increased the ambient background sound by at least 10 dB (re 1 µPa in the 20–1000 Hz range) 68 , 69 . The Atlantic cod winter spawning site had the second highest AIS-monitored traffic, it also had over double the acoustic power (≥3 dB) more than 50% of the time in both full spectrum and combined band octave sound levels compared to the least trafficked recording site. Accordingly, there was a significant positive correlation between the combined octave band sound levels and the number of AIS vessels present within a 10 nm radius at both Atlantic cod and haddock winter spawning sites. There were also increases in the ambient sound levels, and consequently a temporary reduction in the correlation, due to the vocalizations of baleen whales, particularly fin whales who produces both 20 and 40 Hz pulses 70 , and to a lesser degree fishes, however, their influence on local acoustic conditions at the sites was less pervasive and their presence did not affect the correlation. Conversely, there was no significant correlation between the two factors at the Atlantic cod spring spawning site. There are several possible reasons for this; the number of AIS tracked vessels was larger at this site, this was predominately because the data was collected five years subsequent to the other sites. In that time, it is estimated there was at least a 30% increase in the number of vessels with AIS installed in the region (US Coast Guard, 2017). The number of smaller sized components of traffic installed with AIS transmitters was increasing, consequently a higher proportion of the total vessels present in the recording area were registered, but with significant differences in the amount and type of sound they emitted relative to the large ocean-going commercial traffic highly represented at the other two sites 33 , 71 . When examining the size class of AIS tracked vessels at this site there were higher numbers of vessels <40 m in length compared to the other two sites. The total number of AIS vessels quantified at the spring spawning location had subsequently less impact on the combined octave band sound levels at this site. The site’s coastal proximity, proximity to many recreational marinas and to the Boston vessel separation scheme, and sampling season relative to the other sites also likely differentiate its vessel activity, and the sound signature of that activity, from the two more offshore locations. For these reasons, AIS data is not a perfect representation of sound that could potential mask other biological signals in all environments, and care must be taken when using AIS data to infer masking potential. The results of the current study have taken the first steps in assessing the direct influence of anthropogenic sound on the communication spaces of two ecologically and commercial important fish species at three locations highly influenced by human activities . It highlights the ever-increasing need to better understand the role anthropogenic sound has in the disruption of intraspecific acoustic communication. Future research should focus on examining the extent to which specific species can compete with anthropogenic noise through adaptation or adjustment of their acoustic signals, address the abilities to receive and interpret signals in the presence of another, increase the accuracy of vocalization propagation and detection models by incorporating oceanographic, and bathymetric variables, as well as updating species-specific source level measurements in the field. Further consideration on the use of multisensory cues and how they supplement each other is also needed to support our understanding of the behavioural and physiological effects of prolonged exposure to low frequency sound. This research also highlights the need to gain a better understanding of the spatial and temporal use of unique habitats that are predictably used for critical life history events in declining populations. Identifying and better understanding these consequences at lower trophic levels is important to advancing the management of shared acoustic space. Methods Instrumentation The acoustic recordings were made using Marine Autonomous Recording Units (MARUs; Cornell University Bioacoustics Research Program 72 ). At all recording sites, the units continuously sampled at a rate of 2000 Hz with a flat frequency response sensitivity (±1.0 dB) of ~151.2 dB re 1 μPa between 10 and 2000 Hz (HTI-94-SSQ, High Tech Inc., Gulfport, MS, USA). Deployments Between January 2006 – February 2007 and April – June 2011, several marine autonomous recording units (MARUs; Cornell University Bioacoustics Research Program (Clark et al ., 2002)) were deployed within the Sanctuary to calculate the spatial and temporal variability of soundscapes and to detect vocally active marine species. From these deployments, sites for the current study were determined using prior knowledge of spawning areas and spawning dates for the two Gadoid species; Atlantic cod ( Gadus morhua ) and haddock ( Melanogrammus aeglefinus ). Recordings were also examined for times of high vocal activity attributable to spawning behaviour (Stanley, Van Parijs & Hatch; in prep). Three-months of recordings each from three different sites were chosen to represent spring and winter spawning periods for Atlantic cod and a winter spawning period for haddock. The Atlantic cod spring spawning site, 42°31′5.58″N, 70°41′43.26″W, occurred within the Spring Cod Conservation Zone (SCCZ), in 51.4 m of water in northern Massachusetts Bay, 5 km south of Gloucester, USA (Fig. 4 ). The SCCZ is a seasonal fisheries closure area established to attempt to provide protection for a historic and once predictable coastal cod spawning aggregation ( ). The substrate at this site was predominately fine-grained sediment, occasionally broken up by cobble and boulder deposits and larger bedrock structures 73 . Recordings were utilized from 15 April – 27 June 2011 during the spring spawning season 74 . The Atlantic cod winter spawning site, 42°26′29.69″N, 70°33′29.59″W, occurred within the SBNMS, in 49.3 m of water with a gravel dominated substrate. Recordings from 01 November 2006 – 31 January 2007 were utilized to represent the winter spawning season (Fig. 4 ) 75 . The haddock winter spawning site, 42°28′11.30″N, 70°14′32.82″W, also occurred within the SBNMS at a depth of 66.4 m. The substrate at this recording site was largely dominated by gravel, but also had areas of sand and cobble and/or boulder areas. Recordings from the 6 th January – 28 th March 2006 were used to best represent the winter spawning period 76 . Acoustic analysis Vocalization characterization Using previously used sound classification parameters 31 , 32 , 34 , 35 , the acoustic data from the three study sites were hand browsed for haddock in Raven Pro 1.5 (The Cornell Lab of Ornithology, NY, USA) and run through an Atlantic cod detection algorithm 77 to ensure vocalizations from the two species were present during the selected recording periods. Forty randomly selected vocalizations from each species were selected during their spawning period, and summary statistics were taken including peak fundamental frequency, sound duration and number of pulses (cod), and peak frequency, sound duration, and number of knocks (haddock). Each day was examined and daily percentage presence of vocalizations for the specific species was calculated for the 3-month period at each site. Ambient noise analysis The ambient sound was measured over the entire three-month spawning period at each recording site. Using purpose-written MATLAB scripts, sound pressure levels for the full spectrum (10–1000 Hz with a 2 kHz sample rate) were calculated at 1 s resolution at each of the three sites, and daily metrics were also calculated for comparison (SPL; RMS, median, 10 & 90 th percentiles). The precise bandwidths for the auditory filters of the species of interest are unknown, but have been described as being slightly larger than other vertebrates 78 . Thus, filters were approximated using octave filter banks. This method is considered more suitable to gauge the audibility of a signal in the presence of ambient noise. Using MATLAB scripts modified from octbank.m by Christophe Couvreur, octave band analyses were conducted at 1 s resolution to characterize the bands with centre frequencies (ƒ c ) at 31.5 (ƒ bw 22 – 44 Hz) (band 1), 63 (ƒ bw 44 – 88 Hz) (band 2), 125 (ƒ bw 88 – 177 Hz) (band 3), and 250 Hz (ƒ bw 177 – 355 Hz) (band 4), additionally daily metrics were calculated for comparison (SPL; RMS, median, 10 & 90 th percentiles) over the all three-month periods at each site. Bands 1 & 2 were selected for Atlantic cod recording sites and bands 3 & 4 for the haddock recording sites as these bands best matched the frequency distribution of the vocalization types for each species. Daily sound pressure levels were calculated in the combined bands 1 & 2 (band 1,2) and 3 & 4 (band 3,4) for Atlantic cod and haddock respectively, for use in the effective vocalization radius calculation (ANL). All acoustic analysis was carried out in MATLAB R2015b (Mathworks Inc., USA). Estimated effective vocalization radius Using the modified sonar equation from Clark et al . (2009), adapted for the use with fishes 78 , 79 , the estimated effective vocalization radius was calculated for each day during the recording periods. This gave an estimated radius in which a single Atlantic cod and/or haddock vocalization could theoretically propagate under the ambient noise levels encountered over the three-month recording period. For the purpose of this study we assumed; (1) signal detection was limited by ambient noise, (2) vocalization source level did not vary in response to varying ambient noise levels i.e., Lombard effect, (3) fish hearing had equal omnidirectional sensitivity, and (4) the sound source propagates approximately omnidirectionally. $${\rm{SE}}={\rm{SL}}\,-\,{\rm{TLsp}}\,-\,{\rm{MSL}}\,-\,{\rm{DT}}$$ (1) SE is signal excess, when SE = 0 it is routinely defined in respect to sonar systems as the 50% probability of signal detection 80 ; SL is the source level of the fish vocalization at 1 m from the source – 127 dB re 1 µPa @ 1 m for Atlantic cod 46 , and as there is no published research on the source level of haddock vocalizations two levels were used from unpublished research findings, 119.2 and 125 dB re 1 µPa @ 1 m for haddock low and haddock high respectively (Hawkins; pers. comm.); MSL is the mean sound level for the site, calculated as the mean daily combined octave band level of sound (SL RMS ) (band 12 or 34 for cod or haddock respectively) for each day in the three-month recording period for each site; TL sp is the simplified spherical spreading transmission loss, calculated as 20 log[r (m)] 81 , the spherical spreading transmission loss model was used due the relatively low source levels of the vocalizations and the water depth at the site. The vocalisations are estimated to propagate over a shorter distance than the depth range of the water, therefore are assumed to propagate in a spherical manner. DT is the detection threshold, defined as the difference between the signal and the sound at a threshold where the signal can still be perceived by the recipient. There are no precise data for the detection threshold in fishes, therefore the current study used a detection threshold of 15 dB, which is considered conservative and attempts to incorporate the understanding of the masked detection thresholds of Atlantic cod and haddock 37 , 51 , 79 . Effective vocalization radius ( r , eq. 3 ) was derived from eq. 2 when SE = 0. $${\rm{TLsp}}=20\,{\rm{log}}\,{\rm{r}}\,$$ (2) TL sp will give r when SE = 0, and therefore: $$r=10({\rm{SL}}-{\rm{MSL}}-{\rm{DT}}/20)$$ (3) Automatic Identification System and Vessel Tracking The sanctuary partnered with the US Coast Guard to gather early data from implementation of the Automatic Identification System (AIS) in Massachusetts Bay, providing high-resolution information since 2004 on the distribution and density of large commercial traffic through the Traffic Separation Scheme that transits the sanctuary accessing the Port of Boston. The relationships between the daily number of AIS vessels within a 10 nm radius of the Atlantic cod and Haddock spring and winter spawning study sites, the daily sound pressure levels and the daily estimated effective vocalization radii were investigated to further understand and identify the acoustics drivers at each site. Following the methods of Hatch et al . 82 , AIS data collected during the study period were extracted and reformatted using AIS Miner (U. S. Coast Guard Research and Development Centre, 2005) and custom software written in Python V2.5.1. (Python Software Foundation, 2007) added to the NOAA data package 83 . The daily number of AIS tracked vessels within a 10 nm radius around the recording sites over the 3-month sampling period was calculated, excluding vessels with a ground speed of zero. This spatial extent was chosen following the rationale of Hatch et al . 2008, with an extended vessel radius. This radius would roughly estimate the area within which a ship with a source level of ≥180 dB re 1 µPa would ensonify the recording site at levels >116 dB re 1 µPa, therefore including sources positioned at a greater distance from the recording site while still rising above ambient sound levels at the sites (Table 1 ). The theoretical source level of 180 dB re 1 µPa was used because a large proportion of commercial shipping vessels are in the range of 170–190 dB 69 , 84 . Statistical analysis For statistical tests, including detecting significant differences in ambient sound levels in the full spectrum and combined octave bands, estimating effective vocalization radii, and estimating the number of AIS vessels during respective spawning periods among sites, non-parametric statistical methods were used to test for differences among sites as the data had unequal variance among treatments and some data had a non-normal distribution 85 . To compare differences among sites, the Kruskal-Wallis test was used. If such tests provided significant results, a Dunn’s pairwise multiple comparison was used to isolate differences among individual sites. A Pearson Correlation test was performed to assess the relationship between the number of AIS vessels present per day and the combined octave band levels. All analyses were performed using SigmaPlot 13 (Systat Software Inc) and SPSS Statistics (IBM) Software. Data availability The datasets generated during and analysed during the current study are readily available from the corresponding author on request.

NOAA scientists studying sounds made by Atlantic cod and haddock at spawning sites in the Gulf of Maine have found that vessel traffic noise is reducing the distance over which these animals can communicate with each other. As a result, daily behavior, feeding, mating, and socializing during critical biological periods for these commercially and ecologically important fish may be altered, according to a study published in Scientific Reports. Three sites in Massachusetts Bay, two inside Stellwagen Bank National Marine Sanctuary (SBNMS) and one inshore south of Cape Ann, were monitored for three months by researchers at the Northeast Fisheries Science Center (NEFSC) laboratory in Woods Hole, Mass. and at the sanctuary offices in Scituate, Mass. Vocalizations, such as Atlantic cod grunts and haddock knocks, were recorded by bottom-mounted instruments at each site during spawning in winter and spring. "We looked at the hourly variation in ambient sound pressure levels and then estimated effective vocalization ranges at all three sites known to support spawning activity for Gulf of Maine cod and haddock stocks," said Jenni Stanley, a marine research scientist in the passive acoustics group at the NEFSC and SBNMS and lead author of the study. "Both fluctuated dramatically during the study. The sound levels appear to be largely driven by large vessel activity, and we found a signification positive correlation with the number of Automatic Identification System (AIS) tracked vessels at two of the three sites." AIS is an automatic tracking system, used on ships and by vessel traffic services. It provides information on a vessel, such as its unique identification number, position, course and speed, which can be displayed on a shipboard radar or electronic chart display. Ambient sounds - those in the surrounding environment - include animals vocalizing, physical sounds such as wind and water movement or geological activity, and human-produced sound from ships and marine construction. Many marine animals use ambient sound to navigate, to choose where to settle, or to modify their daily behaviors including breeding, feeding and socializing. Cod grunts were present for 100 percent of the spring days and 83 percent of the winter days. Haddock knocks were present for 62 percent of the winter days within the three-month sampling period. However, ambient sound levels differed widely at the three sites, both on an hourly and daily time scale. The Atlantic cod winter spawning site, nearest the Boston shipping lanes, had the highest sound levels, while the Atlantic cod spring spawning site inshore south of Gloucester, Mass. had the lowest. Sound levels in the haddock winter spawning site, further offshore in the sanctuary, were in the middle of the range detected in the study. Study data were also used to calculate the estimated distance a fish vocalization would be heard at each of the spawning sites. The effective radius ranged widely, from roughly 4 to 70 feet, and was largely dependent on the number of tracked vessels within a 10 nautical mile radius of the recording sites. Lower-level, chronic exposure to increased ambient sound from human activities is one of the most widespread, yet poorly understood, factors that could be changing fish behavior. If they cannot hear as well as they need to, then sound signals from other fish can be lost, compromised, or misinterpreted in ways that can cause a change in behavior. Since Atlantic cod, for example, vocalize to attract mates and listen for predators, not hearing those signals could potentially reduce reproductive success and survival. "Anthropogenic sound in certain ocean regions has increased considerably in recent decades due to various human activities such as global shipping, construction, sonar, and recreational boating," Stanley said. "As ocean sound increases, so does the concern for its effects on populations of acoustic signalers, which range from invertebrates to marine mammals. We don't know if or to what extent specific species can adapt or adjust their acoustic signals to compete in this environment." In addition to Stanley, other researchers involved in the study were Sofie Van Parijs at the NEFSC's Woods Hole Laboratory and Leila Hatch at Stellwagen Bank National Marine Sanctuary.

10.1038/s41598-017-14743-9

Physics

Magnetism fosters unusual electronic order in quantum material

Robert Birgeneau, Magnetism and charge density wave order in kagome FeGe, Nature Physics (2023). DOI: 10.1038/s41567-023-01985-w. www.nature.com/articles/s41567-023-01985-w Journal information: Nature Physics

https://dx.doi.org/10.1038/s41567-023-01985-w

https://phys.org/news/2023-03-magnetism-fosters-unusual-electronic-quantum.html

Abstract Electron correlations often lead to emergent orders in quantum materials, and one example is the kagome lattice materials where topological states exist in the presence of strong correlations between electrons. This arises from the features of the electronic band structure that are associated with the kagome lattice geometry: flat bands induced by destructive interference of the electronic wavefunctions, topological Dirac crossings and a pair of van Hove singularities. Various correlated electronic phases have been discovered in kagome lattice materials, including magnetism, charge density waves, nematicity and superconductivity. Recently, a charge density wave was discovered in the magnetic kagome FeGe, providing a platform for understanding the interplay between charge order and magnetism in kagome materials. Here we observe all three electronic signatures of the kagome lattice in FeGe using angle-resolved photoemission spectroscopy. The presence of van Hove singularities near the Fermi level is driven by the underlying magnetic exchange splitting. Furthermore, we show spectral evidence for the charge density wave as gaps near the Fermi level. Our observations point to the magnetic interaction-driven band modification resulting in the formation of the charge density wave and indicate an intertwined connection between the emergent magnetism and charge order in this moderately correlated kagome metal. Main In quantum materials where the energy scale of electron–electron correlations is comparable to that of the electronic kinetic energy, a range of emergent electronic phases are found 1 . Well-known systems that exhibit such rich phase diagrams include the unconventional superconductor families of cuprates 2 and iron-based superconductors 3 , 4 , where magnetic order, charge density wave (CDW) order, nematicity and superconductivity have been found in close proximity due to the entwinement of a multitude of degrees of freedom with similar energy scales, including spin, charge and orbitals. Besides engendering large quantum fluctuations that may give rise to novel phases, the intertwinement of multiple orders allows tunability of the ground state via the coupled order parameters. It is expected that, when quantum topology is realized in such a strongly correlated regime, unprecedented exotic phases are to be discovered. Recently, kagome lattices have been extensively investigated for the rich emergent physics associated with its lattice geometry in the presence of topology 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 . Among the insulators, quantum spin liquid states may exist due to the magnetic frustration 20 . Analogously, the geometric frustration for an electronic state could also cause localization as a way to quench the kinetic energy, thereby leading to a regime with narrow electronic bandwidth where electron correlation effects could be strong enough to induce emergent phases 5 , 6 . Equally importantly, the point group of the kagome lattice is that of graphene, hence the kagome lattice also shares many properties of the Dirac fermions. Therefore, kagome lattices have been intensely explored as a model system to realize topology in the presence of strong electron correlations. Among the known kagome metals, some exhibit magnetic order with ferromagnetically ordered sheets that are either ferromagnetically or antiferromagnetically stacked 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , with ordering temperatures often exceeding room temperature. Another class of recently discovered kagome materials, AV 3 Sb 5 (A = K, Rb, Cs) 17 , 18 , 19 , hosts CDW order near 100 K that may be associated with time-reversal symmetry-breaking 21 , 22 . The interpretation on the origin of these orders ranges from those initiated from a strongly interacting picture to those based on electronic instabilities from band structure 23 , 24 , 25 , 26 , 27 , 28 . For a two-dimensional (2D) kagome lattice, three key features have been identified in the electronic structure (Fig. 1d ): (1) a flat band derived from the destructive interference, (2) a Dirac crossing at the Brillouin zone (BZ) corner (K point in the inset of Fig. 1d ) and (3) a pair of Van Hove singularities (VHSs) at the BZ boundary (M point). It has been proposed that the large density of states (DOS) from the kagome flat bands could induce ferromagnetism 5 . Alternatively, at special fillings where the VHS is at the Fermi level ( E F ), interaction between the saddle points could also induce CDW order 6 , 25 , 26 , 27 . The large energy separation between the flat band and the VHSs (Fig. 1d,e ) may be one reason why CDW order and magnetic order have not been commonly observed simultaneously within one system. Efforts at tuning the ordered phases include tuning the energy location of either the flat bands or VHSs relative to the chemical potential via chemical doping 29 , hydrostatic pressure 30 and uniaxial strain 31 , all of which have demonstrated potential yet so far have resulted in decoupled order parameters of either magnetism or CDW order. Fig. 1: Crystal structure and electronic structure. a , The crystal structure of hexagonal FeGe with the unit cell shown by solid lines. Fe atoms are labelled in red and Ge atoms in grey. A-type AFM order is illustrated by red arrows denoting spin ordering of Fe atoms. b , A top view of the crystal structure to visualize the Fe kagome layer. c , A summary of the transitions as a function of temperature. d , A simplified schematic summary of AFM-induced band splitting plotted in energy ( E ) versus momentum. In the PM state, the kagome flat band appears at E F , with a Dirac crossing and a pair of VHSs well below E F . In the AFM phase, the ordered moment within each ferromagnetic layer induces exchange splitting of the spin-minority and spin-majority bands, resulting in an upshift of the blue VHS towards E F . At the onset of CDW order, a gap ( Δ CDW ) opens at the VHS band near E F . The temperature evolution (horizontal direction) is aligned with the temperature axis in c . e , DFT calculated band structures for the PM state, as well as those for the spin-minority and spin-majority bands within an FM layer in the AFM state. Colour indicates the spin projection ( s z ). Flat band (FB) regions are highlighted with coloured shades. f , DOS calculations for the PM and AFM phases. Source data Full size image Very recently, it was discovered that CDW order appears deep in a magnetically ordered kagome metal (FeGe) 32 , 33 , providing the opportunity to explore the possible connection between magnetism and CDW order on a kagome lattice. Hexagonal FeGe (refs. 34 , 35 ) (Fig. 1a,b ), which is isostructural to FeSn and CoSn, consists of stacks of Fe kagome planes with both in-plane and inter-plane Ge atoms. A sequence of phase transitions has been found (Fig. 1c ). First, an A-type antiferromagnetic (AFM) order appears below a Néel temperature \(T_{\mathrm{N}} \approx 410\) K, with moments aligned ferromagnetically along the c axis within each plane and anti-aligned between layers (Fig. 1a ). At 100 K, where CDW order takes place ( T CDW ), two types of CDW order were found with Q CDW1 = (0.5, 0, 0)/(0, 0.5, 0), that is, completely in-plane, and Q CDW2 = (0.5, 0, 0.5)/(0, 0.5, 0.5), which also has an out-of-plane component, similar to that found in the AV 3 Sb 5 systems 17 . Finally, at a lower temperature of \(T_{\mathrm{canting}} \approx 60 \, {\mathrm{K}}\) , the magnetic moments cant from the c direction to give a c axis double cone (canted) AFM structure. In addition, evidence of strong coupling between the CDW order and AFM order has been reported in the form of an enhanced magnetic moment at the onset of the CDW order 32 . Here, we use a combination of high-resolution angle-resolved photoemission spectroscopy (ARPES), inelastic neutron scattering (INS) and density functional theory (DFT) to study the formation of the CDW order from the magnetic phase to gain insights into the potential contributors to the origin of CDW order in magnetic FeGe. In particular, in the AFM state, we observe saddle points in the vicinity of E F , as well as kagome flat bands and Dirac cones (DCs). With the guidance of DFT calculations, we identify an orbital-selective contribution of VHS to a potential nesting condition near E F , enabled by the spin splitting of the bands within each ferromagnetic (FM) kagome layer. At T CDW , gaps open on the VHS bands across the Q CDW vectors. In addition, we find that the electronic states on the VHS bands couple to an optical phonon mode. We therefore identify the key ingredients for the formation of CDW in kagome FeGe to be the presence of orbital-selective VHS near E F induced by the spin splitting of the ordered moment as well as electron–phonon coupling. The observations taken together suggest an intimate connection between the CDW order and magnetism on the kagome lattice and lay the groundwork for theoretical efforts incorporating electron correlations to identify CDW order formation in a magnetic kagome lattice. We begin with an overview of the electronic structure of FeGe from DFT calculations. Consistent with previous first-principle calculations 36 , 37 , 38 , the DOS calculated for the paramagnetic (PM) phase exhibit a large peak at E F , which splits into two in the AFM phase (Fig. 1f ). By directly comparing with the band structures calculated in each phase (Fig. 1e ), the peaks in the DOS are indeed associated with the kagome flat bands. Moreover, although the magnetic structure in FeGe is AFM, each kagome layer is FM with a well-defined spin-majority and spin-minority contribution. From Fig. 1e , we confirm that the flat bands above E F in the AFM phase correspond to the spin-minority bands while those below E F correspond to the spin-majority bands within each FM layer. Hence, one can largely interpret the AFM band structures as an interaction-driven magnetic splitting of the kagome bands within each FM layer (Fig. 1d ). Interestingly, the consequence of such splitting is that the VHSs that were relatively far from E F in the PM state are brought to the vicinity of E F in the AFM phase in the spin-minority sector, as shown by the marked VHS1 and VHS2 in Fig. 1e . Although DFT cannot capture all the details of the experimental observations because this is a moderately correlated system, we will show in the following that this qualitative understanding is consistent with our experimental observations. Next, we gauge the strength of the electron correlations in FeGe by comparing the measured DOS with DFT. From our angle-integrated photoemission spectrum, we can identify key spectral features that match with those from the calculated DOS in the AFM phase (Supplementary Fig. S1 ), which allows us to extract an overall renormalization factor of 1.6, close to that in some of the iron-based superconductors 4 . With such a renormalization factor, the DFT calculated band structure for the AFM state can provide a reasonable match with that of the measured dispersions covered in a large energy range (Fig. 2c ). In particular, we can identify a broad band near −0.75 eV that matches the location of the spin-majority kagome flat bands. Next, we examine the key signatures of the kagome electronic structure near E F in the AFM phase. From a detailed photon energy measurement, we identify 69 eV photons to cross the k z = 0 plane while 47 and 102 eV photons to cross the k z = π plane (Supplementary Fig. S3 ). Since the photoemission process is subject to a substantial k z broadening effect 39 , we discuss the following using the 2D projected notation of the BZ (Fig. 2a ). We find two types of terminations among the samples measured, corresponding to the Ge and Fe kagome terminations, similar to the case of FeSn (ref. 9 ). From a detailed analysis of the two terminations (Supplementary Figs. S2 , S3 , S4 and S6 ), we can identify two sets of VHSs and DCs that are termination independent and therefore intrinsic to the bulk kagome layers. The first set (marked in blue as VHS1 and DC1) consists of linear bands that cross at −0.6 eV at the BZ corner, \({\bar {\rm {K}}}\) points (Fig. 2d ). The DC structures can be better visualized in a stack of constant energy contour plots in Fig. 2e , where circular contours shrink down to the Dirac point and expand out again. Furthermore, the linear dispersions from DC1 rise to cross E F , forming a VHS on the BZ edge, \({\bar {\rm {M}}}\) point (Fig. 2d,e ). The saddle point nature of this VHS can be better visualized from a series of cuts across the \({\bar {\rm {M}}}\) point (Fig. 2f ), where the band top of the hole-like band parallel to \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) has a minimum energy at the \({\bar {\rm {M}}}\) point, indicative of the electron-like nature in the orthogonal direction. The Fermi momenta ( k F ) of this VHS1 (denoted by blue dots) form circular Fermi pockets (Fig. 2e ). Besides this set of DC and VHS, a second set (labelled in dark green) consists of a DC2 at −0.2 eV that is connected to VHS2 slightly below E F with a much smaller band velocity compared with that of VHS1. In addition, we also identify a DC3 at −0.06 eV. For the hexagonal lattice symmetry, the five Fe 3 d orbitals are split into three groups: \(d_{x^2 - y^2}\) / d xy , d xz / d yz and \(d_{z^2}\) . Guided by orbital-projected DFT calculations and consideration of both in-plane and out-of-plane band velocities (Supplementary Figs. S4 and S5 ), we identify DC1/VHS1 to be of \(d_{x^2 - y^2}\) / d xy origin, DC2/VHS2 to be of d xz / d yz origin and DC3 to be of \(d_{z^2}\) origin. A summary of the identified VHSs and DCs is shown in Fig. 2b . We note that, while common features can be identified between DFT and measured dispersions, the precise locations of the VHSs are not well captured by DFT even with the overall renormalization, suggesting non-negligible correlation effects at play. Fig. 2: Key signatures of kagome band structure. a , 3D and 2D projected BZ notations with polarization vectors indicated. b , Schematic of DCs and associated VHSs. c , Large energy range dispersions measured compared with that from DFT calculation after renormalization by a factor of 1.6. The lower kagome flat bands are highlighted similarly to in Fig. 1e . d , Spectral image measured along \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) using LH 47 eV photons. e , Constant energy contours measured with 102 eV photons and LH polarization. Dashed blue lines indicate a DC at the \({\bar {\rm {K}}}\) point and the corresponding VHS at \({\bar {\rm {M}}}\) . f , Stack of cuts taken along k x at k y = k L − 0.3 Å −1 , k L − 0.2 Å −1 , k L − 0.1 Å −1 , k L and k L + 0.1 Å −1 , as indicated in e as solid white lines. Thick red lines at the top of d and fourth figure in f indicate they are the high symmetry cuts, as illustrated in a . The Fermi crossings are marked by blue dots, also marked on the corresponding cuts in e . All data are taken at 11 K. Full size image If the qualitative understanding of the spin splitting from DFT calculations is correct, the VHSs near E F observed in the AFM state must be the spin-minority set that is shifted up as the magnetic moment orders with lowering temperature. To check for this, we examine the temperature evolution of VHS1, which as a mostly in-plane orbital can be tracked more accurately as it is less affected by k z broadening compared with VHS2. Figure 3e shows the spectral images of the VHS1 at 11 K and 260 K. The VHS point visibly shifts in energy as temperature is varied, as also seen in the extracted momentum distribution curves (MDCs) from −0.04 eV as a function of temperature (Fig. 3f ). The peak from the VHS band is observed to shift away from the M point as temperature is lowered, consistent with the shifting-up behaviour of the VHS (Fig. 3h ). We can further extract the energy location of the VHS. In Fig. 3a , we identify the bands leading to VHS1. Since the VHS point is slightly above E F , to track its energy position, we first fit the band dispersion at each temperature from an MDC analysis. From the MDC, we identify a main peak next to another feature in the form of a smaller peak. We fit the MDC line profile to two Lorentzian functions on a Gaussian background. The location of the VHS band can therefore be extracted as a function of binding energy (Fig. 3d ). We then fit for the location of the band top using a parabolic function of the dispersion at each temperature. We note that only the dispersion lower than −30 meV is used in the parabolic fit to avoid any complication due to the potential opening of a CDW gap or dispersion kink. The extracted location of VHS1 is plotted as a function of temperature in Fig. 3h , showing an upward shift as temperature is lowered towards T CDW . Additional evidence for the band shift is seen in the peak shift from both energy distribution curves (EDCs) and MDCs as a function of temperature (Supplementary Fig. S7 ). The trend of the shift is consistent with the ordering of the magnetic moment ( m ), as can be seen by overlaying in Fig. 3g the m 2 measured from neutron diffraction 32 . Fig. 3: Temperature evolution of VHS. a , Spectral images across the \({\bar {\rm {M}}}\) point along \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) , taken with LH-polarized 69 eV photons. b , A stack of MDCs (from E F to −0.1 eV) from a . The red line in the stack is the MDC curve along the dashed red line in a . c , The MDC from the red dashed line in a (identical to the red curve in b ). We fit the curve with two Lorentzians (red and blue), one background Gaussian profile (light green) and a constant background (yellow) curve. d , Fitting results from the MDC stack. Red circles are peak positions of VHS1-forming band dispersions (red curve in b ). We use a parabolic fit to extrapolate locations of VHS1 above E F . e , Spectral images across the \({\bar {\rm {M}}}\) point along \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) , showing the temperature evolution of VHS1. f , MDCs taken at the dashed grey line in e , measured as a function of temperature. g , The ordered magnetic moment measured by neutron scattering, reproduced from ref. 20 (green). h , Extracted VHS1 locations (blue) and the fitted peaks in MDC taken from f (grey). Error bars are from the s.d. resulting from the fitting process. Source data Full size image As the VHSs are shifted to the vicinity of E F by the magnetic order, we therefore examine the potential role of the VHSs in the formation of the CDW order. At 140 K above the onset of the CDW order, VHS1 is above E F while VHS2 is below E F . They form distinct Fermi surfaces. In particular, the Fermi surface from VHS1 ( xy / x 2 – y 2 ) can be better observed under an in-plane polarization (linear vertical, LV) while VHS2 ( xz / yz ) can be better observed in the linear horizontal (LH) polarization that contains an out-of-plane component (Fig. 4a ). Noticeably, the VHS1 band forms a circular Fermi surface, while that of the VHS2 band resembles more that of the triangular Fermi surface predicted by kagome models at VHS filling 6 . We can further confirm their contribution to the potential nesting condition using DFT-based Lindhard susceptibility calculations (Supplementary Fig. S11 ). While DFT does not accurately reproduce the Fermi surface, the chemical potential can be tuned to identify energies where such Fermi surface contours are revealed. A susceptibility calculation at such energy that reproduces the VHS2 fermiology indeed produces a peak at the 2 × 2 CDW wavevector, while that done for the VHS1 fermiology does not (Supplementary Fig. S11 ). This suggests that the VHS contribution to the CDW formation may be orbital dependent in FeGe, as has also been reported for CsV 3 Sb 5 (ref. 40 ). Fig. 4: Observation of CDW gap. a , Fermi surface taken on kagome termination at 140 K with 69 eV photons, with the left half under LH polarization and the right half under LV polarization. b , The same as a but taken at 13 K. Blue and green lines/curves in a and b delineate the Fermi surfaces formed by VHS1 and VHS2, respectively, serving as guide to the eyes. c , d , \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) cut (cut 1) from a ( c ) and b ( d ). e , f , Cut 2 from a ( e ) and b ( f ), showing CDW gap opening on the VHS2 band. Green dashed lines in c – f indicate VHS2 bands. g , EDC from cut 1 marked by the arrow, showing the appearance of two features at 13 K. h , EDC from cut 2 marked by the arrow, showing the gap opening. i , Spectral image taken with LH-polarized 47 eV photons across the \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) high symmetry direction at 140 K > T CDW (left) and 70 K < T CDW (right). The data have been divided by the Fermi–Dirac function convolved with the energy resolution to reveal signal above E F . The Fermi momenta of four band dispersions are indicated ( k 1 , k 2 , k 3 and k VHS ), where k VHS corresponds to the VHS crossing. j , EDC at k VHS , showing a gap below T CDW , as well as EDCs for other band crossings as labelled in i , showing no gap opening. k , Temperature evolution of the symmetrized EDC taken at k VHS . l , Integrated spectral weight across the CDW gap energy between [−0.02, 0.02] eV at k VHS as a function of temperature (red circles). The red line is the smoothed average of the data points, showing a suppression onset at T CDW . Source data Full size image Furthermore, we can identify participation of the VHS bands in the CDW ordering by comparing the band dispersions across T CDW . From a comparison of the Fermi surface at 140 and 13 K (Fig. 4a,b ), we see that the VHS2 undergoes the largest change as evident in the increased curvature of its Fermi pocket. This can also be seen by comparing two cuts across T CDW . On the cut along \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) (Fig. 4c,d ), even though matrix elements suppress the spectral weight of VHS2, we can still identify the presence of k z broadened VHS2 via a hump in the EDC (Fig. 4g ). In the CDW phase, an additional peak appears close to E F in the EDC (Fig. 4g ), consistent with the splitting of the band (Supplementary Fig. S9 ). Such band splitting at the M point is reminiscent to that observed in CsV 3 Sb 5 (ref. 40 ) and is interpreted to be the result of folding and gap opening at the M point. On a parallel cut away from \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) , gap opening can also be observed on VHS2 with a typical bend-back behaviour of the dispersion (Fig. 4e,f,h and Supplementary Fig. S9 ). Besides VHS2, we also observe spectral gap opening on VHS1. Spectral images across the \({\bar {\rm {K}}}\) – \({\bar {\rm {M}}}\) – \({\bar {\rm {K}}}\) cut are shown for 140 K > T CDW (Fig. 4i , left side) and 70 K < T CDW (Fig. 4i , right side). The spectra have been divided by the Fermi–Dirac function convolved with the instrument resolution to also uncover spectral weight above E F . Noticeably, a spectral weight suppression is observed on VHS1, not on any other bands crossing E F . This is further confirmed by the EDC for the VHS1 at k VHS , where a spectral weight suppression below T CDW indicates the gap opening (Fig. 4j and Supplementary Fig. S8 ). The observable lower edge of the gap below E F is −20 meV, consistent with the observed full gap size of 50 meV by scanning tunnelling microscopy 33 . In contrast, the EDCs for the other band crossings away from the BZ edge, L points in the CDW ordered phase show ungapped peaks at E F in the CDW phase, indicating that the CDW gap is momentum dependent. Figure 4k shows a continuous temperature evolution of the symmetrized EDC taken at k VHS , where a gap appears below T CDW . The integrated spectral weight within the gap energy of 20 meV of E F is shown in Fig. 4l , where the smoothed average of the scattered data points (Fig. 4l , red line) shows a trend consistent with the onset of the depletion of the spectral weight at T CDW = 100 K (Supplementary Fig. S8 ). Lastly, we report on the observation of electron–phonon coupling in FeGe. Dispersion kinks have been reported in unconventional superconductors such as the cuprates and iron-based superconductors 2 , 3 , 39 , 41 , 42 . The origin of such dispersion kinks has been proposed to indicate electron–boson coupling, to either phonons or magnons. On the VHS1 band where a CDW gap is observed, a kink structure is observed in the electronic dispersion (Fig. 5a ). We can use MDC analysis to extract the real and imaginary parts of the self-energy from the peak position and the full-width at half-maximum of the MDC fitting, respectively (Fig. 5b–d ), showing the coupling of the electrons to a bosonic mode at around 30 meV below E F . We can then extract an electron–boson coupling constant from the slope of the real part of the self-energy as well as from the ratio of the renormalized band velocity ( \(v_{\mathrm{F}}\) ) and the bare-band velocity ( \(v_{{\mathrm{F}}^0}\) ), both of which give a consistent coupling constant λ of approximately 0.5–0.6 (refs. 43 , 44 ). Notably, this kink structure has little resolvable temperature dependence from 11 K ( T < T canting ) to 70 K ( T canting < T < T CDW ) to 140 K ( T CDW < T ). Fig. 5: Electron–boson coupling in FeGe. a , VHS1-forming band dispersion (same one as in Fig. 3 ). White dots are fitted band dispersions from MDC analysis. On the right, MDC fitting results at 11 K ( T < T canting ), 70 K ( T canting < T < T CDW ) and 140 K ( T CDW < T ) are overlaid. b , A stack of MDCs directly visualizing the kink structure. c , The imaginary part of the self-energy derived from width analysis of MDC fittings. Σ denotes self-energy. We introduce one electron–phonon (e–ph) coupling mode together with background (b.g.) signals considering moderate electron–electron (e–e) interaction (Fermi-liquid-like behaviour). d , The real part of the self-energy analysis of peak position from MDC fittings. We overlay the Kramers–Kronig transformed electron–boson coupling mode from the imaginary part of the self-energy (solid blue curve in c ). A bosonic mode is observed at 30 meV. e , INS measurement at 120 K of the in-plane phonon spectra at L = [−3.1, −2.9] with DFT calculated phonon spectra overlaid. f , INS measurement of the in-plane magnon spectra with L = [−2.6, 0]. The colour bar of f is same for e . g , INS measurement geometry illustrating the relationship between neutron momentum transfer Q , phonon wavevector q and vibrational direction. h , Measured phonon spectra at the Γ (0, 0, 3) point (cut 1), showing no temperature dependence. i , Measured phonon spectra at the M points (cut 2, averaged between (1, −0.5, 3) and (−1, 0.5, 3)), showing hardening of the optical phonon mode across T CDW . In h and i , blue, green and red dots indicate 8 K, 70 K, 120 K data, respectively. Solid lines are fitting with a Lorentzian and a linear background. Error bars represent one s.d. The fitted energy at (8 K, 70 K, 120 K) is (32.74 ± 0.13 meV, 33.94 ± 0.10 meV , 33.57 ± 0.15 meV). The colour scheme for h is same for i . j , Identified phonon mode ( A 2u ) for the optical branch outlined in e . Source data Full size image To identify the origin of this bosonic mode, we carried out INS measurements. As presented in Fig. 5f , the observed spin waves stemming from the AFM zone centre and in-plane Γ point cross the M point at 70 meV, far above the interaction point of 30 meV. Therefore, a direct electron–magnon interaction at the M point is ruled out. In contrast, the phonon spectra do reveal optical modes near 33–38 meV around the M point (Fig. 5e ). More interestingly, a hardening of this optical phonon mode at 33 meV is observed across the T CDW from 120 K to 70 K, appearing at the M point and extending to the K points (Fig. 5i and Supplementary Fig. S10b ). Such hardening is absent at the Γ points (Fig. 5h ). This optical phonon mode, therefore, is likely the mode responsible for the dispersion kink observed in ARPES measurements. We note that the size of the hardening is about 1 meV, which is too small to be resolved in the temperature-dependent kink analysis (Fig. 5a ). To identify the phonon mode, we carried out phonon calculations using DFT, which match excellently with the measured phonon spectra (Fig. 5e ). The optical phonon is determined to be an A 2u mode involving the movement of the Fe atoms from the kagome plane as well as movement of the inter-layer Ge atoms (Fig. 5j ). In addition, we observe no acoustic phonon softening from 8 K to 70 K (Supplementary Fig. S10a ), contrary to the behaviour in the A 2u optical mode across the same temperature region. The electron–phonon coupling here in FeGe is similar in many aspects to that in the AV 3 Sb 5 system, where a dispersion kink at a similar energy scale of 36 meV has also been reported 42 . In addition, a hardening of the B 3u longitudinal optical phonon mode around 10 meV has been observed in AV 3 Sb 5 across its T CDW (ref. 43 ), while no acoustic phonon softening is observed 42 , 44 . Having presented all the data, we now discuss the implications. We have experimentally identified potential key ingredients for the formation of the CDW order in magnetic FeGe, namely the orbital-dependent VHSs that are brought to the vicinity of E F due to the magnetic splitting of the bands. The contribution of the VHSs to the formation of the CDW order is evident in the opening of gaps at E F in the CDW phase while other bands remain gapless. In addition, we identify electron–phonon coupling around the M point of the BZ that is connected by the CDW wavevector. We can first compare the phenomenology of the CDW order in FeGe with that in non-magnetic kagome AV 3 Sb 5 (refs. 17 , 18 , 19 , 30 , 31 , 32 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 ). In both CDW ordered phases, VHSs are observed to be in the vicinity of E F , albeit accidental in AV 3 Sb 5 and magnetism assisted in FeGe. The VHSs are observed to have orbital-dependent contributions gauged from their fermiology. In both cases, the xz / yz orbitals are proposed to be dominant 45 . The CDW gap at E F is observed at the BZ edge, which is connected by wavevectors that correspond to the diffraction peaks seen by X-rays and neutrons. Electron–phonon coupling is present in both systems in the absence of acoustic phonon softening at the CDW wavevectors 42 , 43 , 44 . These similarities point to a potential universality of the phenomenology of CDW among the kagome metals. However, the microscopic origin of the CDW order in kagome system is still unresolved. In the weakly correlated AV 3 Sb 5 system, nesting of VHSs has been proposed based on 2D models. In FeGe, however, the moderate electron correlations deem a purely nesting-driven scenario unlikely. This is reflected in the larger degree of mismatch between DFT and measured dispersions in FeGe as well as the difficulty to stabilize a CDW order from DFT alone. If nesting is indeed the dominant driver, then it may be expected that other A-type AFM kagome systems, such as FeSn (ref. 9 ), could also develop 2 × 2 CDW order when the spin splitting shifts the VHSs at M into the proximity of E F . Since FeGe is the only kagome lattice magnet with CDW discovered so far, it is therefore likely that correlation effects, perhaps even orbital-dependent correlation effects, must be taken into consideration in addition to the VHS presence near E F to understand the origin of the CDW order in magnetic kagome systems. Regardless of the ultimate microscopic theoretical description of the CDW on stacked kagome lattices, our work reveals an intimate interaction of the CDW order and magnetism in a moderately correlated kagome metal and provides experimental groundwork towards a potential universal understanding of CDW order in kagome systems. Methods Single crystals were synthesized using chemical vapour transport 32 . ARPES measurements were carried out at BL 5-2 of the Stanford Synchrotron Radiation Lightsource and the MAESTRO beamline (7.0.2) of the Advanced Light Source, with a DA30 electron analyser and an R4000 electron analyser with deflector mode, respectively. FeGe single crystals were cleaved in situ in ultra-high vacuum with a base pressure better than 5 × 10− 11 torr. Samples were cleaved in the (001) orientation. Two types of terminations were found, one corresponding to the kagome layer and another to the Ge layer. Both bulk and surface states were identified on both types of terminations. A detailed discussion on the terminations is provided in the Supplementary Information . The data shown in the main text come from the Ge termination unless specifically stated otherwise, but all the features discussed are those that are found in both terminations, hence being representative of the bulk kagome lattice. The energy and angular resolutions used were better than 20 meV and 0.1°, respectively. Measurements shown were taken with a beam spot smaller than 50 μm × 50 μm. All gap measurements were done with correction of the Fermi level by a measurement of polycrystalline gold that was electrically connected to the sample. INS experiments were carried out on the ARCS (BL-18) spectrometer at the Spallation Neutron Source at Oak Ridge National Laboratory 51 on 0.9 g of co-aligned single-crystal samples. The definition of momentum transfer Q in reciprocal space in units of Å −1 is given by Q = H a ∗ + K b ∗ + L c ∗ , where H , K and L are Miller indices. The direction vectors a , b and c are specified with direction as shown in Fig. 1a,b . The reciprocal lattice vectors a ∗ , b ∗ and c ∗ are calculated using a ∗ = 2π( b × c )/[ a ⋅ ( b × c )], b ∗ = 2π( c × a )/[ a ⋅ ( b × c )] and c ∗ = 2π( a × b )/[ a ⋅ ( b × c )]. The scattering plane is [ H , H , L ]. We measured the neutron spectrum at 120 K, 70 K and 8 K by scanning the angle about the vertical axis by 140° in 1° steps with incident energies of 100 meV (Fig. 5f ) and 45 meV (Fig. 5e ). Cuts and slices from the Q , E data set are extracted using Mantid 52 and Horace 53 . Magnon and phonon spectrum are plotted using low- Q ( L = [−2.6, 0]) and high- Q ( L = [−3.1, −2.9]) data, respectively. The vibration direction of phonons observed by INS is always parallel to the momentum transfer Q (Fig. 5g ), so the phonon mode shown in Fig. 5e is mostly the A 2u mode vibrating along the c axis. Constant- Q cuts of the measured phonon spectra are extracted at the Γ (Fig. 5h ) and M points (Fig. 5i ) and fitted using a Lorentzian and a linear background. DFT calculations are performed with the Vienna ab initio simulation package 54 , 55 . The exchange–correlation interaction between electrons is mimicked with the generalized gradient approximation as parameterized by Perdew et al. 56 . A cutoff energy of 350 eV and a k mesh of 12 × 12 × 8 are used throughout. Experimental lattice constants are employed, and the atomic positions are fully relaxed until the remaining force is less than 1 meV Å −1 . Both the paramagnetic and antiferromagnetic structures are calculated. All calculations are performed considering the spin–orbital coupling, except for the structure relaxation and phonon dispersion calculation. The phonon dispersion is calculated by the finite displacement method as implemented in the PHONONPY software 57 . To calculate the Fermi surface and electron Lindhard susceptibility, we first fitted the DFT band structure with Wannier orbitals (Fe d and Ge p ) as implemented in Wannier90 software 58 and then calculate the 3D Fermi surface in the full BZ based on the obtained tight-binding Hamiltonian with a k mesh of 150 × 150 × 80. The electron susceptibility is calculated on the basis of the Fermi surface by employing the method described in ref. 59 . Data availability Source data are provided with this paper. All other data that support the findings of this study are available from the corresponding authors upon reasonable request. Code availability The band structure and phonon calculations used in this study are available from the corresponding authors upon reasonable request.

Physicists were surprised by the 2022 discovery that electrons in magnetic iron-germanium crystals could spontaneously and collectively organize their charges into a pattern featuring a standing wave. Magnetism also arises from the collective self-organization of electron spins into ordered patterns, and those patterns rarely coexist with the patterns that produce the standing wave of electrons physicists call a charge density wave. In a study published this week in Nature Physics, Rice University physicists Ming Yi and Pengcheng Dai, and many of their collaborators from the 2022 study, present an array of experimental evidence that shows their charge density wave discovery was rarer still, a case where the magnetic and electronic orders don't simply coexist but are directly linked. "We found magnetism subtly modifies the landscape of electron energy states in the material in a way that both promotes and prepares for the formation of the charge density wave," said Yi, a co-corresponding author of the study. The study was co-authored by more than a dozen researchers from Rice; Oak Ridge National Laboratory (ORNL); SLAC National Accelerator Laboratory; Lawrence Berkeley National Laboratory (LBNL); the University of Washington; the University of California, Berkeley; Israel's Weizmann Institute of Science; and China's Southern University of Science and Technology. The iron-germanium materials are kagome lattice crystals, a much-studied family of materials featuring 2D arrangements of atoms reminiscent of the weave pattern in traditional Japanese kagome baskets, which features equilateral triangles that touch at the corners. "Kagome materials have taken the quantum materials world by storm recently," Yi said. "The cool thing about this structure is that the geometry imposes interesting quantum constraints on the way the electrons are allowed to zoom around, somewhat analogous to how traffic roundabouts affect the flow of traffic and sometimes bring it to a stop." By nature, electrons avoid one another. One way they do this is to order their magnetic states—spins that point either up or down—in the opposite direction of their neighbors' spins. Dai, a co-corresponding study author, said, "When put onto kagome lattices, electrons can also appear in a state where they are stuck and cannot go anywhere due to quantum interference effects." When electrons cannot move, the triangular arrangement produces a situation where each has three neighbors, and there is no way for electrons to collectively order all neighboring spins in opposite directions. The inherent frustration of electrons in Kagome lattice materials has long been recognized. Yi said the lattice restricts electrons in ways that "can have a direct impact on the observable properties of the material," and the team was able to use that "to probe deeper into the origins of the intertwinement of the magnetism and charge density wave" in iron-germanium. They did so using a combination of inelastic neutron scattering experiments, which were performed at ORNL, and angle-resolved photoemission spectroscopy experiments that were performed at LBNL's Advanced Light Source and SLAC's Stanford Synchrotron Radiation Lightsource, as well in Yi's lab at Rice. "These probes allowed us to look at what both the electrons and the lattice were doing as the charge density wave was taking shape," she said. Dai said the findings confirmed the team's hypothesis that charge order and magnetic order are linked in iron-germanium. "This is one of the very few, if not of the only, known example of a kagome material where magnetism forms first, preparing the way for charges to line up," he said. Yi said the work shows how curiosity and basic research into natural phenomena can eventually lead to applied science. "As physicists, we are always excited when we find materials that spontaneously form an order of some sort," she said. "This means there is a chance for us to learn about the self-organizational abilities of the fundamental particles of quantum materials. Only with that kind of understanding can we one day hope to engineer materials with novel or exotic properties that we can control at will." Dai is the Sam and Helen Worden Professor of Physics and Astronomy. Dai and Yi are each members of the Rice Quantum Initiative and the Rice Center for Quantum Materials (RCQM).

10.1038/s41567-023-01985-w

Earth

Building blocks for life on Earth arrived much later than we thought, billion-year-old rocks show

Mario Fischer-Gödde et al. Ruthenium isotope vestige of Earth's pre-late-veneer mantle preserved in Archaean rocks, Nature (2020). DOI: 10.1038/s41586-020-2069-3 Journal information: Nature

http://dx.doi.org/10.1038/s41586-020-2069-3

https://phys.org/news/2020-03-blocks-life-earth-thought-billion-year-old.html

Abstract The accretion of volatile-rich material from the outer Solar System represents a crucial prerequisite for Earth to develop oceans and become a habitable planet 1 , 2 , 3 , 4 . However, the timing of this accretion remains controversial 5 , 6 , 7 , 8 . It has been proposed that volatile elements were added to Earth by the late accretion of a late veneer consisting of carbonaceous-chondrite-like material after core formation had ceased 6 , 9 , 10 . This view could not be reconciled with the ruthenium (Ru) isotope composition of carbonaceous chondrites 5 , 11 , which is distinct from that of the modern mantle 12 , or of any known meteorite group 5 . As a possible solution, Earth’s pre-late-veneer mantle could already have contained a fraction of Ru that was not fully extracted by core formation 13 . The presence of such pre-late-veneer Ru can only be established if its isotope composition is distinct from that of the modern mantle. Here we report the first high-precision, mass-independent Ru isotope compositions for Eoarchaean ultramafic rocks from southwest Greenland, which display a relative 100 Ru excess of 22 parts per million compared with the modern mantle value. This 100 Ru excess indicates that the source of the Eoarchaean rocks already contained a substantial fraction of Ru before the accretion of the late veneer. By 3.7 billion years ago, the mantle beneath southwest Greenland had not yet fully equilibrated with late accreted material. Otherwise, no Ru isotopic difference relative to the modern mantle would be observed. If constraints from other highly siderophile elements besides Ru are also considered 14 , the composition of the modern mantle can only be reconciled if the late veneer contained substantial amounts of carbonaceous-chondrite-like materials with their characteristic 100 Ru deficits. These data therefore relax previous constraints on the late veneer and are consistent with volatile-rich material from the outer Solar System being delivered to Earth during late accretion. Main Ruthenium is a highly siderophile element (HSE) and is therefore expected to be sequestered in the metallic core during Earth’s differentiation. Contrary to this prediction, the abundances of Ru and other HSEs in the modern mantle are higher than expected compared with metal–silicate equilibrium conditions 15 , 16 . This observation is most commonly explained by HSE replenishment of the mantle through the addition of a late veneer after core formation. Relative abundances of HSEs that are close to chondritic compositions in the mantle suggest that the late veneer must have consisted of primitive meteoritic material 17 , 18 , amounting to ~0.5% of Earth’s mass 18 . The chemical composition of the late veneer and its origin are a longstanding matter of debate, especially in the context of how and when Earth accreted its water and volatiles 3 , 6 , 9 , 10 . Previous studies debated whether significant amounts of volatile-rich carbonaceous-chondrite-like material were added by the late veneer during the final stages of Earth’s accretion 6 , 9 , 10 or had already been incorporated during earlier stages of Earth’s growth 3 , 5 , 7 , 8 , 11 . Mass-independent ruthenium isotopic variations among meteorites and Earth have provided evidence that the late veneer was derived from reduced and volatile-poor inner Solar System materials most similar to enstatite chondrites 5 , 11 , 12 , 19 . This is in contrast to constraints from the relative abundances of volatile elements such as selenium (Se), tellurium (Te) and sulfur (S) and the Se isotope composition in the silicate Earth that were used to argue for a CM or CI carbonaceous-chondrite-like late veneer composition 2 , 9 , 10 . Owing to its distinct Ru isotope composition, volatile-rich carbonaceous-chondrite-like material from the outer Solar System was excluded as possible late-veneer source material 5 , 11 , and thus the late veneer seemed unlikely to be the primary source of water and volatiles on Earth 5 , 11 . It should be noted, however, that this conclusion depends on the premise that the Ru in Earth’s mantle originates solely from the late accreted materials that were added after cessation of core formation 11 , 15 , 16 , 18 . If Earth’s pre-late-veneer mantle retained a significant fraction of Ru during metal–silicate differentiation 13 , 20 , as recently suggested, this conclusion would be invalid. Investigating Ru isotope signatures in the putative remnants of pre-late-veneer mantle would thus not only provide insights into the timescales and efficiencies of mixing the late veneer into Earth’s mantle, but also introduce constraints on the composition of the material that was added as a late veneer. To our knowledge, no unambiguous isotopic evidence for the preservation of pre-late-veneer mantle on Earth existed until now. For instance, resolvable excesses in 182 W reported for 3.8 billion-year-old (Gyr-old) Archaean rocks from Isua (Greenland) and Acasta (Canada) in conjunction with relatively low HSE abundances observed in 3.5–3.2-Gyr-old Archaean komatiites from the Pilbara Craton (Australia) and the Barberton greenstone belt (South Africa) were interpreted to reflect sluggish mixing of the late veneer into the early Archaean mantle 21 , 22 . However, it was later suggested that the mantle sources of the 3.8–3.7-Gyr-old Isua supracrustal belt (ISB) rocks, including 3.8-Gyr-old Eoarchaean peridotites from the Narssaq ultramafic body (NUB) and the south of the Isua supracrustal belt (SOISB), already had HSE abundances at about 60–100% of the modern mantle value 14 , 23 . This suggests that the late veneer was to a large extent mixed into the ambient mantle by ~3.8 billion years ago (Ga). To reconcile 182 W excesses with the presence of modern-mantle-like HSE abundances, it was proposed that a small amount of core material could have been entrained into proto-Earth’s mantle as a consequence of the Moon-forming giant impact 20 , 24 . However, 182 W anomalies could also be generated by early mantle differentiation processes during approximately the first 50 Myr of the Solar System 25 , 26 , 27 , 28 , 29 or by core–mantle interactions in the sources of mantle plumes 30 . In summary, 182 W and HSE concentration data alone fail to provide an unambiguous test of whether pre-late-veneer mantle domains were preserved. Here we explore the potential use of mass-independent Ru isotope variations in terrestrial rocks as a tool to investigate whether pre-late-veneer isotope signatures can be found in the Archaean mantle. While the Ru isotope composition of the modern mantle is well constrained 12 , this is not the case for the Archaean mantle. To address this issue, we determined the Ru isotope composition for a set of ultramafic rocks from different Eoarchaean and Palaeoproterozoic terranes (Extended Data Table 1 ; see Methods for details). We focus on the 100 Ru/ 101 Ru and 102 Ru/ 101 Ru ratios to constrain the Ru isotope compositions of the mantle sources of these rocks because these isotope ratios are measured at the highest precision and also show the largest variability among meteoritic materials 5 , 19 , 31 . The results are reported as ε unit (0.01%) deviations of mass bias-corrected 100 Ru/ 101 Ru and 102 Ru/ 101 Ru ratios from a terrestrial standard. Exotic composition of Archaean mantle We report Ru isotope data for samples from four different cratons. The Ru isotope compositions obtained for ultramafic samples from the Pilbara Craton (3.5–3.2 Gyr old), the Superior Province (Abitibi greenstone belt, 2.7 Gyr old) and the Kaapvaal Craton (Bushveld Complex, 2.05 Gyr old) are indistinguishable from the Ru solution standard (Fig. 1 ), indicating that their Ru isotope compositions reflect that of the modern terrestrial mantle. By contrast, Eoarchaean 3.8–3.7-Gyr-old ultramafic rocks from the North Atlantic Craton, originating from various localities of the Itsaq gneiss complex (IGC) in southwest Greenland (the NUB, SOISB, ISB and the Ujaragssuit Nunât layered intrusion) exhibit a uniform and well-resolved excess in ε 100 Ru of +0.22 ± 0.04 (95% confidence interval, Fig. 1 ) combined with a smaller excess in ε 102 Ru of +0.09 ± 0.02 (95% confidence interval, Fig. 2a ). Chromitites from the younger 3.0-Gyr-old Seqi ultramafic complex in southwest Greenland show the same excesses in ε 100 Ru and ε 102 Ru. The combined ε 100 Ru and ε 102 Ru excesses in these rocks represent mass-independent isotope anomalies of nucleosynthetic origin and indicate that the Ru in the southwest Greenland mantle source is enriched in nuclides produced by the slow neutron capture process (s-process) of nucleosynthesis compared with the modern mantle (Fig. 2a ). The isotope excesses cannot be explained by mass-independent fractionation effects or by inherited fissiogenic Ru nuclides (see Methods and Extended Data for details about the accuracy of the Ru isotope data). Fig. 1: ε 100 Ru data for Archaean and Palaeoproterozoic rocks, the modern mantle and chondrites. The individual results for all analysed samples (Extended Data Table 1 ) are shown with the composition of the modern mantle 12 . The uncertainties for individual data points reflect the external uncertainty of the method (2 s.d. for samples measured n < 4 times) or 95% confidence intervals of replicate analyses of a given sample (if n ≥ 4). The mean values for 3.8–3.7-Gyr-old Eoarchaean samples from the IGC in southwest Greenland and chromitite samples from the Bushveld complex are shown as solid vertical black lines. The darker grey and blue areas represent the respective 95% confidence intervals; the light grey and blue areas limited by dashed lines indicate the 2 s.d. uncertainty of the mean values. The uncertainty for the modern mantle composition is 2 s.d. (ref. 12 ). Numbers on the right of the data points refer to the sample identifiers given in Extended Data Table 1 . Source data Full size image Fig. 2: Ru isotope plot illustrating compositional differences between enstatite, ordinary, average carbonaceous, CI and CM carbonaceous chondrites, the modern mantle and the Eoarchaean mantle. a , The dashed line represents a mixing line between the modern mantle composition (ε 100 Ru = 0) and an s-process component defined by Ru isotope data for pre-solar silicon carbide grains 34 . The compositions of enstatite chondrites (EC, ε 100 Ru = –0.08 ± 0.04, 95% confidence interval); ordinary chondrites (OC, ε 100 Ru = –0.29 ± 0.03, 95% confidence interval) 5 , CI chondrites (CI, ε 100 Ru = –0.24 ± 0.13, 2 s.d.) 5 ; CM chondrites (CM, ε 100 Ru = –0.69 ± 0.38, 95% confidence interval) 5 , and average carbonaceous chondrites (average CC, ε 100 Ru = –0.90 ± 0.12, 95% confidence interval) 5 are shown for comparison. The uncertainties for CI chondrites reflect a single measurement and are thus shown with the external uncertainty of the method (2 s.d. as stated in ref. 5 ). Uncertainties for the modern and the Eoarchaean mantle composition are the same as stated in Fig. 1 . Note that the uncertainty for the modern oceanic mantle composition from the literature is shown as 2 s.d. (ref. 12 ). b , Heliocentric zoning of ε 100 Ru anomalies 5 .The presence of an s-process-enriched reservoir that contributed to Earth’s growth is inferred from the Ru isotope composition obtained for the Eoarchaean mantle of southwest Greenland (Fig. 1 ). Chondrite groups formed at increasing heliocentric distances exhibit more negative ε 100 Ru because they are more depleted in s-process Ru relative to Earth’s modern mantle 5 . The ε 100 Ru uncertainty for carbonaceous chondrites in b is shown as 2 s.d. to account for the significant within-group variation of their ε 100 Ru values (image adapted from ref. 5 , Springer Nature). Source data Full size image The s-process-enriched composition inferred for the Archaean southwest Greenland mantle is an unexpected finding because the Ru isotope compositions reported for all meteorites are deficient in s-process Ru and exhibit negative ε 100 Ru and ε 102 Ru anomalies relative to Earth’s modern mantle 5 , 19 , 31 . The southwest Greenland data provide unambiguous evidence for s-process-enriched building material that contributed to the early stages of Earth’s growth. Owing to the observed heliocentric zoning of ε 100 Ru anomalies among meteorites 5 , we speculate that this reservoir was most probably located in the innermost region of the Solar System, within 1 astronomical unit(Fig. 2b ). Pre-late-veneer Ru isotopic signature The 100 Ru excess provides unequivocal evidence that the mantle source of the Greenland rocks did not receive the full complement of late veneer material 21 . Furthermore, it also requires that Ru (and possibly other HSEs) was not completely stripped from the mantle during the latest stages of core formation 13 . Otherwise, no Ru isotope anomaly would be observed. The uniform and ubiquitous presence of the ε 100 Ru anomaly in various 3.8–3.7-Gyr-old ultramafic rock types from different Eoarchaean terranes in Greenland (Isuakasia, Færingehaven) suggests that a larger mantle domain is lacking a full late veneer component 20 , 21 . The presence of the ε 100 Ru anomaly in the younger Mesoarchaean chromitites from Seqi (minimum age of 3.0 Gyr, Akia terrane) also indicates that even 700 Myr later, the southwest Greenland mantle had not fully equilibrated with the late veneer. Such a prolonged timescale for mixing-in of the late veneer component is consistent with significant HSE depletions observed in Archaean mafic rocks from the Pilbara and Kaapvaal cratons 22 , which previously had been explained by sluggish inmixing of late veneer material. As outlined above, the ε 100 Ru excess identified in Eoarchaean ultramafic rocks from southwest Greenland indicates that Ru was not completely sequestered in the core, most probably because some late accretionary component had been delivered during the waning stages of core formation 14 , 23 , 32 . Depending on the composition of this early late veneer material, the 100 Ru excess measured in the Greenland rocks would then represent a minimum estimate for the ε 100 Ru excess of the pure pre-late-veneer mantle. The nature of the early component that supplied the 100 Ru excess and was already mixed into the Greenland mantle before 3.8 Ga, probably inner Solar System material (Fig. 2b ), can be further constrained by osmium isotope systematics. This is because the initial osmium isotopic compositions of chromitite and peridotite samples from the IGC overlap the 187 Os/ 188 Os composition of chondrites at 3.8 Ga (refs. 14 , 23 , 32 , 33 ) (Extended Data Table 2 ). Assuming that the positive ε 100 Ru anomaly and the chondritic Os signature were both imparted by this component, it is unlikely that it is represented by any known chondritic meteorites because these all exhibit negative ε 100 Ru values and chondritic Os isotope compositions. Importantly, owing to its positive ε 100 Ru value, this material cannot derive from a carbonaceous-chondrite-like Moon-forming impactor 8 because carbonaceous chondrites also exhibit the most negative ε 100 Ru values among all known chondrite groups 5 (Fig. 2 ). Carbonaceous-chondrite-like late veneer Regardless of the precise nature and origin of the early accreted component, the 100 Ru excess inferred for the Eoarchaean southwest Greenland mantle source could only be balanced by the addition of chondritic materials with negative ε 100 Ru to yield the composition of the modern mantle (ε 100 Ru = 0). This mixing relationship is further illustrated in Fig. 3 , where possible ε 100 Ru compositions for the pre-late-veneer mantle are calculated by subtracting enstatite, ordinary or carbonaceous-chondrite-like materials from the Ru isotopic composition of the modern mantle. The model is based on a recently proposed inefficient core formation scenario where about 20% of the Ru (~1.4 ng g −1 ) in the modern mantle derives from the pre-late-veneer stage 13 . Assuming a minimum late accretion contribution of 60% for the ≥3.8-Gyr-old Itsaq mantle source 14 , only the addition of a late veneer consisting of carbonaceous chondrites could account for ε 100 Ru ≈ 0, as observed for the modern mantle 12 (Fig. 3 ). The required proportion of late accreted material would amount to a maximum estimate of 0.3% of Earth’s mass of average carbonaceous chondrite or CM carbonaceous chondrite material, consistent with a recent estimate based on Se isotopes 10 . A late veneer consisting of CM-like material is also supported by the abundances of volatile elements in the silicate Earth 2 , 9 , 10 . This conclusion remains robust, even if a larger fraction of Ru was present in the pre-late-veneer mantle or the late accretion component in the mantle by 3.8 Ga was >60%, but in these cases a lower carbonaceous chondrite mass fraction would be sufficient. Ordinary chondrites would only become viable late veneer materials if the Greenland mantle contained a significantly lower late veneer contribution by 3.8 Ga (<50%). A late veneer consisting of carbonaceous chondrites is consistent with the relative abundances of S–Se–Te and the Se isotopic composition of the modern mantle 2 , 9 , 10 , but the addition of a late veneer composed of ordinary chondrites cannot be reconciled with these constraints, because the relative abundances of S–Se–Te and the Se isotope composition of ordinary chondrites are distinct from those of Earth’s mantle 9 , 10 . If a major part of the late veneer consisted of core fragments from differentiated impactors 24 , one potential caveat would be that this material cannot readily account for chondritic S–Se–Te and broadly chondritic HSE relative abundances in Earth’s mantle 2 , 9 , 16 , 18 . Fig. 3: Model estimates for the amount of late veneer added to the Eoarchaean mantle based on Ru isotope compositions. a – c , The model illustrates the effect of subtracting variable mass fractions of late veneer from the Ru isotope composition of the modern mantle. The composition of the modern mantle was fixed at ε 100 Ru = 0, as indicated by the composition obtained for samples from the Bushveld complex (Extended Data Table 1 ), which is indistinguishable from previously reported data for the modern oceanic mantle composition 12 . Dotted lines indicate the respective mass fractions of different chondritic late-veneer compositions subtracted from the modern mantle composition. Solid black lines show the minimum and maximum ε 100 Ru values for different chondrite classes in a : enstatite chondrites (EC, ε 100 Ru = –0.08 ± 0.04, 95% confidence interval) and ordinary chondrites (OC, ε 100 Ru = –0.29 ± 0.03, 95% confidence) 5 ; b : carbonaceous chondrites (CC, ε 100 Ru = –0.90 ± 0.61, 2 s.d.; ± 0.12, 95% confidence interval) 5 and CI chondrites (CI, ε 100 Ru = –0.24 ± 0.13, 2 s.d.) 5 ; and c : CM chondrites (CM, ε 100 Ru = –0.69 ± 0.38, 95% confidence interval) 5 . The uncertainties for carbonaceous chondrites in b are given as 2 s.d. to account for significant within-group variation in their ε 100 Ru values (the dashed line indicates 95% confidence interval uncertainty). The 2 s.d. uncertainty for CI chondrites reflects the external uncertainty of the method (as stated in ref. 5 ). The amount of subtracted late veneer material for each respective chondrite composition was adjusted to match a presumed Ru concentration in the pre-late-veneer mantle of ~1.4 ng g −1 , corresponding to ~20% of the Ru contained in the modern mantle. The yellow boxes indicate the composition of the Eoarchaean mantle inferred from the mean value of all analysed 3.8–3.7-Gyr-old samples originating from various localities of the IGC (ε 100 Ru = +0.22 ± 0.04, average value shown as solid black line, dashed lines indicate 95% confidence interval uncertainty, Extended Data Table 1 ). The solid vertical dashed line indicates the minimum late veneer contribution inferred for the mantle source of peridotite samples from the NUB and the SOISB based on previously reported 187 Os/ 188 Os data and HSE concentrations 14 . The parameters used for the mixing model are given in Extended Data Table 3 . Source data Full size image Collectively, our data imply that the distinct 100 Ru isotope excess in the Eoarchaean southwest Greenland mantle source is best explained by late mixing of a carbonaceous-chondrite-like late veneer fraction into Earth’s mantle. Thus, contrary to previous Ru isotope constraints on the late veneer 5 , 11 , these data imply that significant amounts of volatile-rich outer Solar System materials including water and volatiles were added with the late veneer. This revised view also agrees with other constraints, such as those independently obtained from the relative abundances and isotope compositions of Earth’s volatile elements 1 , 2 , 9 , 10 , which also indicate that the major share of Earth’s volatiles was inherited from a carbonaceous chondrite source 1 , 2 , 3 , 4 . Finally, our data demonstrate that investigating the Ru isotope composition of terrestrial rocks represents a powerful analytical tool for identifying primordial mantle heterogeneities arising from incomplete equilibration of the ambient mantle with Earth’s late-stage building blocks. Methods Samples The samples analysed in this study comprise ultramafic rocks from four different cratons. The North Atlantic Craton is represented by peridotite and chromitite samples of from various localities of the Eoarchaean IGC and the Mesoarchaean Seqi ultramafic complex in southwest Greenland. The sample set is complemented by chromitites from the Kaapvaal Craton (Bushveld Complex, South Africa) and the Pilbara Craton (Australia) and a komatiite reference sample from the Superior Province (Abitibi greenstone belt, Canada). Where available, the Ru concentration and osmium isotope data for the samples analysed in this study are given in Extended Data Table 2 . There were no data available for some samples, so Ru and Os data obtained for similar samples from the same location are listed. The IGC in southwest Greenland represents one of the few localities where remnants of Eoarchaean mantle are preserved. The IGC comprises two Eoarchaean crustal terranes (Isuakasia and Færingehavn), where possible mantle rocks are exposed as ultramafic lenses in the 3.8–3.7-Gyr-old Isua supracrustal belt and ultramafic bodies in the 3.8-Gyr-old SOISB, both located in the Isuakasia terrane 35 , 36 . In the Færingehavn terrane, such rocks are exposed in the 3.8-Gyr-old NUB 37 . The peridotite and chromitite samples investigated in this study were selected to cover all of these different localities. Samples 10-9 and 10-11 from the NUB are massive, coarse-grained peridotites. Olivine is the dominant phase in these rocks. They also contain orthopyroxene, amphibole, spinel and magnetite 14 . The chemical compositions of these samples, including concentration data for highly siderophile elements and 187 Os/ 188 Os data, were reported in a previous study 14 . Sample 10-27 is a harzburgite from the SOISB. The mineral assemblage of this rock is comparable to other harzburgites collected from the same locality 14 . These rocks are typically spinel-peridotites with harzburgitic mineral assemblages composed mainly of olivine and variable amounts of orthopyroxene, amphibole and opaque phases 14 . A minimum age of 3.8 Gyr was estimated for the analysed SOISB and NUB peridotites on the basis of field relationships with surrounding 3.8-Gyr-old tonalitic gneisses and crosscutting dykes 14 , 36 , 37 . Two of the investigated chromitites (194856, 194857) were collected from the Ujargssuit Nunât layered intrusion 38 . For chromitites from this locality, Pt–Os model ages as old as 4.36 Gyr were reported 39 . Samples 194882B and 19488C are chromitites from a locality close to the inland ice that most probably belongs to the same sequence as the chromitites from the Ujargssuit Nunât layered intrusion. The dunite sample 194907 was collected from an antigorite lens located within the northeastern part of the 3.7-Gyr-old Isua supracrustal belt, which has previously been referred to as Dunite Lens B 40 , 41 . The ISB dunites also contain orthopyroxene and spinel, and very minor amounts of clinopyroxene that has mostly been altered 40 . Two of the analysed chromitites (186466, 186479) derive from the Seqi ultramafic complex. The major and trace element compositions, including concentration data for platinum group elements of these samples, were reported in a previous study 42 . The Seqi ultramafic complex represents a peridotite enclave hosted by tonalitic orthogneiss within the 3.0-Gyr-old Akia terrane. A minimum age for the ultramafic body is constrained by 2.98-Gyr-old crosscutting granitoid sheets 42 , although unpublished Re–Os isotope data show a consistent 3.1 Gyr mantle depletion age for the Seqi ultramafic complex. The highly refractory peridotites and chromitites are interpreted as representing the remnant of a fragmented layered complex or a magma conduit. The ultramafic rocks formed from a magnesian-rich, near-anhydrous magma as olivine dominated cumulates with high modal contents of chromite 42 . Their parental magma was generated by high degrees of partial melting of a mantle source that probably represents the precursor of the regional sub-continental lithospheric mantle. The chromitite sample Pil 16-61 was collected from the Warawoona Group located within the Pilbara Craton. The chromitite may be as old as the associated Mount Ada basalt unit (3.5 Gyr) or it may be part of the younger 3.2-Gyr-old Dalton Suite sill complex 43 . The investigated komatiite rock (OKUM) from the 2.7-Gyr-old Abitibi greenstone belt (Canada) is a commercially available rock reference sample provided by the International Association of Geoanalysts. Two chromitites (UG2, LG6) from the 2.05-Gyr-old Bushveld complex (South Africa) were used as a reference sample to validate the analytical method and to assess the precision and the accuracy of the Ru isotope measurements. The Ru isotope composition of UG-2 was previously determined employing a different digestion method (alkaline fusion) 5 . Ruthenium separation and purification The required amount of sample material to yield sufficient Ru for a high-precision measurement was estimated on the basis of previously reported Ru concentrations (10-9, 10-11, UG-2, LG-6, OKUM, 186466, 186479) 42 , 44 , 45 or Ru concentrations reported for samples of similar composition from the same locality (10-27, 194907) 23 , 46 . When information was not available (for example, for samples 194856, 194857and Pil16-61), the Ru concentrations were determined from a 1 g powder test portion digested in a high-pressure asher in reverse aqua regia (5 ml concentrated HNO 3 and 2.5 ml concentrated HCl). Before quantification of Ru for these samples, the digestion solution was dried down, converted twice with 5 ml of 6 M HCl, taken up in 0.2 M HCl and loaded on a cation column to remove matrix elements as described below. Ruthenium concentrations were determined in the eluted Ru fractions by external calibration using a quadrupole inductively coupled plasma mass spectrometer (ThermoScientific iCap). We note that the concentrations determined by this procedure may underestimate the actual concentration of samples because some Ru may have been lost as a volatile tetroxide (RuO 4 ) when the aqua regia solutions were dried down. These concentrations are therefore considered to be only approximate values. In a similar manner, the Ru contents of two chromitites (194882B, 194884C) were estimated from a 1% sample aliquot taken after NiS digestion and cation column chemistry as described below. However, we note that these estimates represent only approximate values too because the Ru yield of the NiS procedure is <100%. For the NiS procedure, powder aliquots of 5–10 g were digested using a NiS fire assay technique 47 . For chromitite samples with high Ru concentrations (UG-2, LG-6, 194856, 194857, Pil16-61), one NiS digestion with 5 g of sample powder was needed to yield sufficient amounts of Ru. Multiple NiS digestions with 10 g of sample powder had to be prepared for ultramafic samples with lower Ru concentrations (harzburgites, dunites, komatiites and some chromitites). The total number of NiS digestions and the amount of sample material used for each respective NiS bead are given in Extended Data Table 2 . Appropriate amounts of Ni, S, borax and Na 2 CO 3 were added to each 5–10 g sample portion and thoroughly mixed. The mixture was fluxed in a muffle furnace for 75 min at 1,000 °C. After cooling, the NiS beads were physically removed from the quenched silicate melt. For the majority of samples, the NiS procedure resulted in about one to three beads of about 1cm in diameter that could readily be recovered from the quenched silicate. The Ru yield of the NiS procedure was determined on the basis of sample powders with known Ru concentrations (UG-2, OKUM, 10-9, 10-11, 186466, 186479). The Ru yield for these samples usually varied from 60–95%. However, in case of three replicate UG-2 digestions the Ru yields were only of the order of 10–20%. The lower yields resulted from incomplete homogenization and subsequent inefficient extraction of NiS beads from the quenched silicate. The NiS digestions for these samples produced finely dispersed millimetre- to micrometre-sized spherules within the quenched silicate. Careful homogenization of the NiS sample–flux mixtures before digestion helped to avoid this problem. The NiS beads were crushed in an agate mortar and transferred into 60 ml Savillex beakers to which 30 ml of concentrated HCl was added. The solutions were evaporated to near dryness on a hotplate at 100 °C. This step was repeated with another 30 ml of concentrated HCl and 20 ml of 1 M HCl. Ruthenium was separated from the dissolved NiS beads using cation exchange chromatography 48 . Each dissolved NiS bead from a single fire assay digestion was split over three cation columns filled with 10 ml AG50 X8 resin, respectively. The resin was equilibrated with 20 ml of 0.2 M HCl. Ruthenium and other platinum group elements were loaded and eluted in 14 ml of 0.2 M HCl. The eluted Ru fractions from each sample were recombined and a small aliquot (1%) was taken to determine the amount of Ru and remaining matrix elements (mainly Ni). If significant amounts of matrix elements passed through the column (if Ni/Ru > 1), the combined fractions of samples were passed for a second time over a single 10 ml cation column. The Ru yields from the cation column were usually >95%. The eluted sample solutions were dried down on a hotplate, recombined and Ru was further purified using a macrodistillation unit as described elsewhere 48 . After the distillation, the purified Ru fractions were dried down on a hotplate and dissolved in 0.5 ml of 0.28 M HNO 3 , from which a small aliquot was prepared as a predilution to determine the Ru yield and to check for potential interfering elements. The distillation yields were usually between 40 and 80%. The total Ru yield of the analytical procedure, including NiS digestion, column chemistry and distillation, is typically 30–70%, estimated from samples with known Ru concentrations (UG-2, OKUM, 10-9, 10-11, 186466, 186479). The total yield of the three UG-2 digestions was only 6–21%. The yields of the distillation for these samples were 50–80%, so the low total Ru yields are caused by inefficient extraction of NiS beads, as described above. However, neither the total Ru yield of the entire analytical procedure nor the respective yields from the NiS digestion or the Ru distillation have any effect on the accuracy of the Ru isotope data (Extended Data Fig. 1 ). The procedural blank for a single NiS digestion, including column chemistry and distillation, varied between 185 and 435 pg ( n = 3). The blank contribution was ≪ 1% for the majority of samples and <2% for OKUM and 194907 given that ≥30 ng of Ru were processed for each respective NiS digestion. Mass spectrometry The Ru isotope measurements were performed using a ThermoScientific Neptune Plus multicollector inductively coupled plasma mass spectrometer in the Institut für Geologie und Mineralogie at the University of Cologne. For the measurements, the Ru fractions were further diluted in 0.28 M HNO 3 to yield Ru solutions of 100 ng ml −1 . The diluted solutions were checked for the presence of interfering elements (Zr, Ni) that could affect the accuracy of the isotope data and cannot be monitored online during the measurements. The sample solutions were introduced into the mass spectrometer at an uptake rate of ~50 μl min −1 using an ESI microflow PFA nebulizer attached to a Cetac Aridus II desolvator. The isotope measurements were conducted with total ion beam intensities between 8 × 10 −11 and 2 × 10 −10 A, obtained for 100 ng ml −1 Ru sample and standard solutions using conventional Ni H-cones. The set-up was optimized to yield oxide production rates ≪ 1% (CeO/Ce). The measurements were conducted in static mode and the seven stable Ru isotopes ( 96 Ru, 98 Ru, 99 Ru, 100 Ru, 101 Ru, 102 Ru and 104 Ru) as well as 97 Mo and 105 Pd were monitored simultaneously. Each Ru isotope analysis consisted of an on-peak baseline on a solution blank (40 integrations of 4.2 s) followed by 100 integrations of 8.4 s for each sample or standard solution and typically consumed about 90 ng Ru. Each sample analysis was bracketed by measurements of an in-house Ru standard solution (Alfa Aesar Ru). The data were internally normalized to 99 Ru/ 101 Ru = 0.7450754 (ref. 31 ) using the exponential law to corrected for mass-dependent isotope fractionation. The isotope data are reported as ε i Ru = ([( i Ru/ 101 Ru Sample )/( i Ru/ 101 Ru Standard )] – 1) × 10 4 , calculated relative to the bracketing standard of each analytical session. The accuracy and precision of the Ru isotopic measurements were evaluated by replicate digestions and multiple analyses of the UG-2 chromitite (Bushveld), which was used as a reference sample. The Ru isotope data obtained for UG-2 in this study agree well with previously reported data 5 , where a different digestion method (alkaline fusion) was used for sample decomposition. This demonstrates that the isotope data obtained by the NiS method yield accurate results. The external reproducibilities (2 s.d.) obtained for a total number of 103 individual measurements from 8 replicate UG-2 digestions are ±0.43 ε 96 Ru, ±0.49 ε 98 Ru, ±0.12 ε 100 Ru, ±0.16 ε 102 Ru and ±0.30 ε 104 Ru. Correction for mass-dependent isotope fractionation The exponential law is one of the most commonly used methods to correct for natural and instrumental mass-dependent isotope fractionation. One potential caveat in using this correction for Ru isotope measurements could be that the distillation technique used to purify the Ru could induce an isotope fraction that would not follow the exponential law. This could cause apparent isotopic anomalies for a given sample as a consequence of inaccurate mass fractionation correction. The exponential law assumes that the logarithmic fractionation β = ln( r/R ) of a given isotopic ratio is expressed as a function of the mass log difference Δ(ln M ) = ln( M 2 / M 1 ). Considering two isotopic ratios ( r = 99 Ru/ 101 Ru and r’ = 100 Ru/ 101 Ru) the exponential law predicts that mass fractionation produces a linear array in a ln( r/r ’) plot 49 . This is illustrated in Extended Data Fig. 2 for the measured raw ratios of 99 Ru/ 101 Ru and 100 Ru/ 101 Ru. The ratios in this figure are not corrected for mass fractionation and are normalized to a reference ratio ( R and R ′, respectively) 31 . If the mass fractionation experienced by the samples is accurately described by the exponential law, the ratios should fall on a linear array with a slope of ~0.5. Two distinct mass fractionation lines can be observed in the plot for different sessions. The slopes for both groups of sessions are indistinguishable within error and are in very good agreement with the slope predicted by the exponential law. Most importantly, the samples purified by distillation fall on the same respective mass fractionation line as their associated Alfa Aesar bracketing standards. This clearly demonstrates that the Ru distillation does not induce any non-exponential mass-fractionation effects for the samples in comparison with the bracketing standard. This observation is also independent from the Ru yield of the samples and does not change if other Ru isotope ratios are considered. Thus, the Ru isotope anomalies obtained for the southwest Greenland samples cannot reflect inaccurate mass fractionation correction. The shift observed for samples and associated standards plotting on a distinct mass fractionation array in Extended Data Figure 2 was caused by maintenance in May 2019 during which a Faraday cup was replaced. However, because the data are reported as relative deviations in parts per 10 4 from the Alfa Aesar bracketing standard, and because samples and bracketing standards are shifted by the same magnitude, this does not affect the accuracy of the data. This is also confirmed by replicate digestions of sample 10-9 that were analysed in both groups of sessions. The ε 100 Ru values for this sample are indistinguishable within analytical uncertainty. Another argument against non-exponential mass-fractionation effects is that nucleosynthetic Ru isotope anomalies caused by variable contributions of s-process Ru nuclides would not lead to any ε 104 Ru anomalies in the 99 Ru/ 101 Ru normalization scheme. As the ε 104 Ru values for all analysed samples fall within the external reproducibility of the method (±0.30 for ε 104 Ru), this demonstrates that sample distillation does not cause non-exponential mass fractionation effects. Isobaric interferences The accuracy of the Ru isotopic measurements could be compromised by isobaric interferences from Mo, Pd, Zr and Ni argide species, or potential effects relating to remaining S in the analysed sample solutions. While interferences from Mo and Pd are simultaneously monitored and corrected for during the measurements 48 , isobars from Zr and Ni argides are not. Owing to the design of the collector block and limited availability of collectors, Zr and Ni could not be simultaneously monitored during the measurements. However, Zr is very effectively separated from Ru by cation exchange chemistry. Hence, all analysed sample solutions (except for one digestion of sample 10-27) had Zr intensities indistinguishable from the background of the 0.28 M HNO 3 and the Ru standard solution. Only one analysed sample (10-27) had a slightly elevated Zr/Ru ratio of 0.0008 and, hence, its ε 96 Ru value is slightly elevated due to an isobaric interference from 96 Zr that could not be corrected. The Zr contained in the one analysed sample solution most probably reflects a random contamination from the laboratory equipment used during sample preparation that was not observed for other samples. In the case of Ni we noticed during the initial stage of the project that a few processed reference samples still contained considerable amounts of Ni after the cation chemistry. For these samples, even after further purification of Ru by distillation, smaller amounts of Ni (between 1 and 10 ng ml −1 ) were observed in the sample solutions to be analysed. During the isotopic measurements Ni readily forms argide species in the plasma that interfere with Ru isobars 50 . To assess potential effects from Ni argide species on the measured Ru isotope data, a 100 ng ml −1 Ru standard solution was doped with varying amounts of Ni to yield concentrations between 0.2 pg ml −1 and 50 ng ml −1 . The results of this test show that the measured Ru isotope compositions for 100 ng ml −1 Ru solutions are not affected for samples with Ni/Ru ratios <0.01 (Extended Data Fig. 3a, b ). For sample solutions with higher amounts of Ni, positive anomalies are observed, which are most pronounced for 98 Ru and to a lesser extent for 100 Ru. Other Ru isotope masses ( 102 Ru and 104 Ru), owing to the lower abundance of the higher mass Ni isotopes, are not significantly affected by Ni argide species. To avoid any interferences from Ni argides during the isotopic measurements, the final dilutions of all samples analysed in this study were carefully checked for their Ni contents before the analysis. The intensity of Ni, monitored by scanning the mass of 58 Ni, in the finally diluted sample solutions was indistinguishable from the background intensity observed for the Ru solution standard and for 0.28 M HNO 3 (10–30 mV on 58 Ni). These negligible amounts of Ni are insignificant and have no effect on the measured data. The minimal Ni background originates from the Ni cones of the experimental set-up. To eliminate any potential effects of S in the analysed solutions on the isotopic measurements, the S from the crushed NiS beads was almost completely removed by evaporation as H 2 S gas during dissolution of the beads with concentrated HCl. After further purification of Ru by column chemistry and distillation, the S contents in the final sample solutions were <25 ng ml −1 for all analysed samples. Tests with S-doped Alfa Aesar Ru standard solutions showed that even if a 100 ng ml −1 Ru standard solution contains large excesses of S (S/Ru = 5), the accuracy of the Ru isotope measurements is not compromised (Extended Data Fig. 3c, d ). Nuclear field shift Previous studies have shown that mass-independent Ru isotope anomalies could be caused by nuclear field shift-induced fractionation effects 51 . In meteorites and their components, such effects could be a primary feature resulting from evaporation/condensation processes. However, experimental studies have shown that mass-independent effects could also be generated in the laboratory during sample preparation 51 . In this section, we explore the potential effects of nuclear field shift-induced fractionation of Ru isotopes. These fractionations can be predicted on the basis of differences in the mean-squared nuclear charge radii between nuclides of a given element. The resulting effects on the measured Ru isotopic composition can be calculated in ε units using the following equation 51 : $${\varepsilon }_{{m}_{i}}=\,\left({\rm{\delta }}{r}_{{m}_{1},{m}_{i}}^{2}-\,\frac{{m}_{2}({m}_{i}-{m}_{1})}{{m}_{i}({m}_{2}-{m}_{1})}{\rm{\delta }}{r}_{{m}_{1},{m}_{2}}^{2}\right)\,\alpha $$ where \({m}_{1}\) and \({m}_{2}\) are the atomic masses of the two isotopes of an element chosen for internal normalization, \({m}_{i}\) refers to the atomic mass of another isotope indexed with variable \(i\) , \({\rm{\delta }}{r}^{2}\,\) denotes the difference in the mean-squared nuclear charge radii of the respective isotope pair and α is an adjustable parameter that determines the magnitude of mass-independent fractionation, which is a function of temperature T as 1/ T . In plots of ε 102 Ru and ε 96 Ru versus ε 100 Ru (Extended Data Fig. 4 ), the slopes calculated for the nuclear field shift fractionation are clearly distinct from the slope predicted by a variation in s-process Ru nuclides. The Ru isotopic composition obtained for Eoarchaean southwest Greenland rocks does not plot on the calculated slope for nuclear field shift but instead plots on the s-process mixing line. As such, the anomalies identified in the southwest Greenland rocks cannot be explained by nuclear field shift-induced fractionation and therefore reflect isotope anomalies of nucleosynthetic origin. Fissiogenic Ru The spontaneous fission of uranium has been shown to produce Ru nuclides with relative abundances that are distinct from naturally occurring Ru (ref. 52 ). Fissiogenic Ru primarily consists of 99 Ru (33.4%), 101 Ru (28.9%), 102 Ru (24.7%) and 104 Ru (12.4%) 53 . The presence of an inherited fission-produced fraction of Ru in a rock sample would induce a characteristic isotope anomaly pattern that would be distinct from anomalies of nucleosynthetic or nuclear field shift origin. Because 96 Ru and 100 Ru are not a significant fission product 53 , the presence of an inherited fraction of fissiogenic Ru in a rock sample would cause negative ε 96 Ru and ε 100 Ru anomalies. These are not observed in any of the analysed samples. On the other hand, a deficit of such an inherited fissiogenic Ru component would yield positive ε 96 Ru and ε 100 Ru anomalies, which are also not observed. This is shown in Extended Data Fig. 4b , where samples with an excess or a deficit of such a fissiogenic Ru components would fall on a mixing line with a distinct slope. Hence, the isotopic composition of the samples with a positive ε 100 Ru anomaly cannot be explained by either an excess or a deficit of fissiogenic Ru nuclides. Data availability The data that support the findings of this study are available from the EarthChem library ( ). Source data for Figs. 1 – 3 and Extended Data Fig. 1 – 4 are provided with the paper.

Ancient rocks from Greenland have shown that the elements necessary for the evolution of life did not come to Earth until very late in the planet's formation—much later than previously thought. An international team of geologists—led by the University of Cologne and involving UNSW scientists—have published important new findings about the origin of oceans and life on Earth: they have found evidence that a large proportion of the elements that are essential to the formation of oceans and life—such as water, carbon and nitrogen—only came to Earth very late in its history. Many scientists previously believed that these elements had already been there at the beginning of our planet's formation. However, the geological investigations published in Nature today have shown that most of the water in fact only came to Earth when its formation was almost complete. Volatile elements such as water originate from asteroids, the planetary building blocks that formed in the outer solar system. There has been a lot of discussion and controversy in the scientific community around when precisely these building blocks came to Earth. Dr. Mario Fischer-Gödde from the Institute of Geology and Mineralogy at the University of Cologne, who led the work, says we are now able to narrow down the timeframe more precisely. "The rocks we analyzed are the oldest preserved mantle rocks. They allow us to see into the early history of the Earth as if through a window. "We compared the composition of the oldest, approximately 3.8 billion-year-old, mantle rocks from the Archean Eon with the composition of the asteroids from which they formed, and with the composition of the Earth's mantle today." To understand the temporal process, the researchers determined the isotope abundances of a very rare platinum metal called ruthenium, which the Archean mantle of the Earth contained. Like a genetic fingerprint, the rare platinum metal is an indicator for the late growth phase of the Earth. "Platinum metals like ruthenium have an extremely high tendency to combine with iron. Therefore, when the Earth formed, ruthenium must have been completely discharged into the Earth's metallic core," says Professor Fischer-Gödde. Professor Martin Van Kranendonk, the UNSW scientist who was part of the research, says the reason why this is of such interest relates directly to understanding the origins of life on Earth, how we humans came to be, and in fact, to whether we might be alone, or have neighbours in the universe. "This is because the results show that Earth did not really become a habitable planet until relatively late in its accretionary history," he says. "If you combine this with the evidence for very ancient life on Earth, it reveals that life got started on our planet surprisingly quickly, within only a few hundred million years. Now this might sound like a lot of time, and it is, but it is far different from what we used to think, that life took half a billion, or even a billion years to get started. "And this gives hope for finding life on other planets that had a shorter geological history and period of 'warm and wet' conditions than Earth, because if life could get started quickly here, then perhaps it got started quickly elsewhere." Professor Dr. Carsten Münker, also at the University of Cologne, added: "The fact that we are still finding traces of rare platinum metals in the Earth's mantle means that we can assume they were only added after the formation of the core was completed—they were certainly the result of later collisions of the Earth with asteroids or smaller planetesimals." Scientists refer to the very late building blocks of Earth, which arrived through these collisions, as the 'late veneer." "Our findings suggest that water and other volatile elements such as carbon and nitrogen did indeed arrive on Earth very late in the 'late veneer' phase," Professor Fischer-Gödde says. The new findings are the result of collaboration among scientists from Germany, Denmark, England, Australia and Japan. The scientists are planning further field trips to India, northwestern Australia, and Greenland to investigate more rock samples.

10.1038/s41586-020-2069-3

Biology

Chimpanzees are 'indifferent' when it comes to altruism

Nature Communications, nature.com/articles/doi:10.1038/ncomms13915 Journal information: Nature Communications

http://nature.com/articles/doi:10.1038/ncomms13915

https://phys.org/news/2016-12-chimpanzees-indifferent-altruism.html

Abstract An important debate centres around the nature of prosociality in nonhuman primates. Chimpanzees help other individuals in some experimental settings, yet they do not readily share food. One solution to this paradox is that they are motivated to help others provided there are no competing interests. However, benefits to recipients could arise as by-products of testing. Here we report two studies that separate by-product from intended helping in chimpanzees using a GO/NO-GO paradigm. Actors in one group could help a recipient by releasing a food box, but the same action for another group prevented a recipient from being able to get food. We find no evidence for helping—chimpanzees engaged in the test regardless of the effects on their partners. Illusory prosocial behaviour could arise as a by-product of task design. Introduction The evolution of behaviour directed at improving the welfare of unrelated individuals, especially when these behaviours are costly for the actor, is an evolutionary puzzle. Although the main conundrum comes from behaviours that decrease an actor’s fitness (biological altruism), low-cost behaviours that benefit others are also an anomaly. Typically, non-kin targeted helping is explained by mutualism, reciprocity, policing and reputation which, over the long run, are likely to increase the actor’s fitness 1 . Equally puzzling are the motivations underlying prosocial acts. An act is considered prosocial if it is intended to benefit others 2 . This distinguishes genuine acts of prosociality such as helping from self-serving alternatives where benefits to others arise as incidental by-products. To understand the nature and evolutionary origins of prosocial behaviour, recent work has probed social behaviours to determine their possible underlying psychological mechanisms. More specifically, researchers have focused on whether prosocial acts are motivated to foster the welfare of others or instead have a more self-serving nature. Owing in part to their complex social life, nonhuman primates, and chimpanzees in particular, have been the focus of intense scrutiny with regard to the psychological mechanisms underlying prosocial acts. A number of candidates for prosocial behaviour have been observed in the wild. Chimpanzees risk injury by going on border patrols, come to the aid of each other in conflicts, affiliate with victims of conflicts, groom each other and share food 3 , 4 , 5 , 6 . Even more spectacular examples of potential prosocial behaviour have also been observed, such as adoption of orphans and anecdotal accounts of rescues 7 , 8 , 9 . However, it is difficult to infer intentions from observations alone, particularly anecdotes 10 . Chimpanzees and other nonhuman primates can receive immediate or delayed benefits from their actions, calling into question whether they have the goal of improving the welfare of others. Grooming can be reciprocated and can be immediately beneficial for the groomer who experiences reduced stress as well as benefits from the parasites they eat 11 , 12 , 13 , 14 . Consoled individuals do not appear to experience reduced stress but consolers benefit by receiving less redirected aggression 15 , 16 . The benefits of border patrols and coalitions, like cooperative hunting, are shared, making these mutualistic interactions. Adoptions typically come long after the orphans lost their mothers 8 . Food transfers tend to be done in response to harassment and begging (manipulation) although active transfers do occur 3 , 17 , 18 . To address the underlying motivations of apparently prosocial acts, experiments are done in captivity. The experiments fall into two types, sharing and helping. In sharing experiments, subjects typically choose between an outcome that benefits both themselves and a partner or themselves only (mutualistic preference tasks), or between an outcome that benefits a conspecific or no one (altruistic preference tasks). In tasks that involve a food delivery apparatus, chimpanzees rarely transfer food to each other even when there is no cost to doing so 19 , 20 , 21 , 22 , 23 , 24 . A token exchange paradigm involving interactions with human experimenters showed stronger evidence for prosocial preferences 25 , but overall, evidence for food transfers in chimpanzees is weak. Cooperatively breeding monkeys transfer food more often and there has been some evidence from other nonhuman primates, but results are mixed 19 , 26 , 27 , 28 , 29 , 30 , 31 . The strongest evidence for prosocial behaviour comes from experiments of instrumental aid (helping) in chimpanzees. Chimpanzees will hand objects to experimenters who reach for them and will do so in the absence of immediate food rewards, something that capuchin monkeys do not do 32 , 33 , 34 . They also help conspecifics by choosing the correct tool to access a reward 35 , releasing a door to allow conspecifics to pass through 33 and releasing food and tokens for partners 36 . The implication is that sharing and helping are fundamentally different prosocial behaviours, with the former failing to elicit prosociality in experiments due to competitive motivations for food 37 , 38 , 39 . The ability to infer goals in others—something that chimpanzees, at least, have been shown to do in other situations 40 , 41 —and the motivation to help others achieve those goals, have been suggested to have evolved far earlier than our Homo lineage 37 , 38 , 39 . An alternative interpretation is that there is no difference between the motivation to help and share in chimpanzees and that apparent prosocial behaviour might arise as a by-product of self-regarding preferences. An important feature of helping studies is that subjects might engage with the task due to stimulus enhancement (the socially influenced attraction to environmental features). Stimulus enhancement is an important mechanism for social learning in chimpanzees; observing another individual perform an action on an object can motivate the observer to act on it as well, but without imitating the actions. The same motivation that underlies social learning in chimpanzees—namely attraction to a test due to the actions of other individuals—might also drive apparent prosociality. As well, prior reward histories might also influence the results of the task since in helping studies, subjects first perform the task alone to experience the outcomes themselves; carryover effects could create expectations of rewards 42 . To address the stimulus enhancement hypothesis, we conducted two experiments on a captive group of chimpanzees that had previously shown evidence of helping behaviours 33 , 36 . Chimpanzees were assigned to one of two groups. In both of these, the subjects (actors) had no access to a food box but the box was accessible to a conspecific (recipient). Actors could release a wooden peg. In the GO group, releasing the peg unlocked the food box, allowing food to be shaken out. Doing so in the NO-GO group had the opposite effect—it locked the food box, preventing the recipient from getting food. If chimpanzees are helpful, actors in the GO group were expected to release the peg more often than the NO-GO group in both experiments. Alternatively, if chimpanzees are spiteful—or competitively motivated by food—they would show the reverse pattern 43 , 44 . A lack of a difference between the behaviour of chimpanzees in the GO and NO-GO groups would lend support to the stimulus enhancement hypothesis and thus call into question the nature of prosocial behaviour in chimpanzees. Actors were first tested without prior experience with the food box (experiment 1). This is an important innovation; all prior apparatus-based experiments on nonhuman primate prosociality have trained subjects on the contingencies nonsocially before introducing a test partner 20 , 21 , 22 , 23 , 24 , 26 , 27 , 28 , 29 , 30 , 31 , 34 , 35 , 36 . Subjects in these studies are therefore reinforced for some choices more than others, and these learning effects may carry over into the tests. Experiment 1 allowed us to determine what the chimpanzees learned solely through observing the effects of their actions on conspecifics. Later, in experiment 2, we provided them with training as in all other tests of prosociality by giving them access to the food box prior to testing; in this way, they learned the consequences of their actions through personal experience. The key finding is that there was no difference between the two groups of chimpanzees. They were just as likely to release a peg to prevent access to food (NO-GO group) as they were to provide access (GO group). Response rates in both groups declined over time in experiment 1. After personal experience with the apparatuses, chimpanzees only released the peg when doing so resulted in them getting food for themselves. Stimulus enhancement can fully account for apparent prosociality in chimpanzees; prior evidence for prosocial behaviour may have been by-products of experimental designs, producing an illusion of helping in our closest living relatives. Results Experiment 1 In experiment 1, six chimpanzees (3 males, 3 females—mean age 13 years) were randomly assigned to be actors in the GO group and seven (4 males, 3 females—mean age 13 years) to the NO-GO group (see Supplementary Table 1 for further details). To minimize the effects of personal relationships, three male chimpanzees of a similar age range were chosen to be recipients. None of the chimpanzees were genetically related. Recipients were individually given experience getting food out of the food boxes before being paired with the actors. In addition to the peg connected to the food box ( Fig. 1 ), a distractor rope and peg were placed in the actor’s room (peg room); actors were also given a towel soaked in fruit juice. The purpose of the distractors, which were identical to those used in previous helping studies 33 , 36 , was to lower the rate of peg releasing below ceiling levels. Each pair was tested across four sessions of four trials each (see Methods for details). Figure 1: Experimental apparatuses and setup. ( a ) Chimpanzees in the GO group (blue food box on left) could release a peg, allowing food to be shaken out, whereas those in the NO-GO group (green food box on right) would prevent food from being shaken out. ( b ) In the test conditions, recipients would sit in front of the food box (left of image) and actors would face them across a keeper’s corridor (right); the peg was attached to the mesh of the actor’s room (right). Full size image Actors were no more likely to release the peg to help the recipient in the GO group as they were to release it to block recipients’ access to food in the NO-GO group (exact Mann-Whitney U test, N GO =6, N NO-GO =7, U =17.00, P =0.628; Fig. 2 and Supplementary Movies 1–3 ). Chimpanzees in both groups initially released the peg at high rates (83% in GO and 86% in NO-GO), but quickly declined to do so in subsequent trials (Spearman’s rho: GO r s =−0.822, N =16, P <0.001; NO-GO r s =−0.703, N =16, P =0.002). Both groups showed a typical extinction curve consistent with lack of reinforcement, along with ‘spikes’ at the start of each session (spontaneous recovery of stimulus attraction). In sum, chimpanzees in experiment 1 did not behave in a manner consistent with prosocially motivated helping: actors did not appear to intend the social outcomes that resulted from their actions and any social consequences did not appear to be intrinsically rewarding. Figure 2: Results of experiment 1. Across four sessions, actors in both the GO group (blue, open circles) and NO-GO group (green, closed circles) showed a decline in the percentage of trials in which they released the peg. At the start of each session (dashed line) there is a spike in performance consistent with spontaneous recovery. Full size image To determine what the actors understood of the effects of their choices in experiment 1, they were given a post-test knowledge probe (one session of four trials). Actors started in the peg room. Actors could access the food box (that is, the food room) via a raceway that connected both rooms. This control followed experiment 1, rather than preceded it, to test for any unintended learning effects on part of the actors (namely being personally rewarded for certain actions). During this knowledge probe, actors resumed peg releases, even in the NO-GO group for which this action was not personally beneficial (25% go, 61% NO-GO). While there was no significant difference between the two groups (exact Mann-Whitney U test, N GO =6, N NO–GO =7, U =8.00, P =0.073), there was a trend for more releases by the NO-GO group, for whom releasing was not even personally beneficial—actors in experiment 1 seemed to have failed to understand the affordances of the apparatuses by observing the effects on a conspecific. Experiment 2 Actors were then given training sessions in which they individually learned about the respective food boxes’ affordances. This is the approach taken by almost all previous studies on sharing and helping 20 , 21 , 22 , 23 , 24 , 26 , 27 , 28 , 29 , 30 , 31 , 34 , 35 , 36 . As in the previous knowledge probe, actors could move through the raceway from the peg room to the food room to access the food box. They were required to either release the peg before shaking the food box (GO group) or to inhibit releasing before shaking (NO-GO group). The criterion for success was to get food out of the food box in at least three consecutive trials. All but two actors (two chimpanzees in the NO-GO group) passed and were then given another knowledge probe. The training clearly improved understanding: actors in the GO group now released the peg 96% of the time and those in the NO-GO group never did so (0%; Fig. 3b ). Following this, actors were again placed into the peg room and were again either paired with recipients in the food room (test trials—three sessions of four trials each) or with partners in an unconnected neighbouring room (testing for mere social presence effects: social control—three sessions of four trials each). Release rates in these test and control trials of experiment 2 were even lower than in experiment 1 ( Fig. 3a ). In the test trials, there was no difference in peg release between the GO and NO-GO group (exact Mann-Whitney U test, N GO =6, N NO-GO =5, U =13.5, P =0.792). Furthermore, there was no difference in peg release between the test and social control for these two groups (GO, Wilcoxon exact test: z =8, N =6, P =0.892; NO-GO group, Wilcoxon exact test: z =3, N =5, P =1; Fig. 3b ). Following experiment 2, actors were given post-test knowledge probes, where they again demonstrated a clear understanding of the contingencies of the test: in the GO group actors always released the peg while the actors of the NO-GO group never did. Figure 3: Results of experiment 2. ( a ) In the test conditions, chimpanzees in the GO group (blue) and NO-GO group (green) released the peg at low rates across the three sessions of four trials. ( b ) Release rates for actors in the GO group were at ceiling in the pre-test (probe 1) and post-test (probe 2) knowledge probes, whereas NO-GO group actors never released the peg (mean±s.e.m.). Release rates—for both groups—were very low in both the test and social control. Full size image The results from these two experiments—in conjunction with the data from the knowledge probes—demonstrate that chimpanzees did not act to produce benefits for others in a helping context. Chimpanzees did not take into account the social consequences of their actions, even after having learned personally about the outcomes of their actions. Any benefits or harm to conspecifics that arose did so as incidental by-products of a personal interest in the stimulus, leading to the peg release actions. Actors showed an initial interest in the task (possibly even independent of social effects) which was quickly extinguished due to lack of reinforcement. Chimpanzees did not show any evidence for being motivated to influence outcomes that benefitted or harmed others. Discussion The instrumental helping experiments presented here showed that chimpanzees will actively ‘help’ conspecifics by performing a low-cost action that allows them access to food. However, they are just as likely to perform the same action when the effect is to prevent conspecifics from getting food. Chimpanzees, then, are no more prosocial than they are spiteful. Furthermore, regardless of outcomes for their partners, chimpanzees show a rapid decline in engaging with the task. Personal experience with the task only increases performance if the actors benefit personally from their actions. Chimpanzees do demonstrate an understanding of the consequences of their actions, but are indifferent to any effects on others. These findings reconcile studies in which chimpanzees did not show signs of prosociality with those that did, and highlight the similarities between helping and sharing. In studies of active food sharing, chimpanzees failed to show a preference for prosocial outcomes when given a choice between outcomes that benefitted a conspecific from those that did not 19 , 20 , 21 , 22 , 30 . Some studies did report evidence for food sharing 24 , 25 , 31 . However, prosocial choices occurred at a fairly low rate, raising questions about how prosocial the subjects were. More critically, each of these studies is open to alternative explanations. In House et al . 24 , chimpanzees only showed weakly prosocial choices in a GO paradigm in which there was only one piece of food that could be delivered despite showing no prosocial preferences in other conditions. In a series of experiments by Claidiere et al . 31 chimpanzees were faced with paired choices, one of which was mutually beneficial and one which was purely selfish; results were inconsistent and subjects failed knowledge controls. Perhaps the strongest positive evidence for sharing in an experiment comes from Horner et al . 25 However, in this token exchange study, in which subjects preferred to exchange coloured tokens with an experimenter for wrapped food for both themselves and a partner, the results can be explained by a conditioned preference for the sound of food being unwrapped 45 ; furthermore, there were no controls for task comprehension. Instrumental helping studies in chimpanzees have shown more consistent evidence for prosociality, but here as well, alternative explanations have not been ruled out. First, the distinction between helping and sharing is not as clear-cut as has been suggested 37 , 38 , 39 . All instrumental helping studies with conspecific recipients involved delivering food or the means to get food 33 , 35 , 36 , 46 , 47 , 48 , blurring the distinction between sharing and helping. In the one direct test comparing prosociality rates when food or non-food items were delivered, there was no difference, showing that food delivery does not specifically impede prosocial behaviour 36 . The most highly cited evidence for helping in chimpanzee comes from experiments in which subjects hand objects back to human experimenters 32 , 33 . However, the chimpanzees had prior experience with handing objects back to their caregivers and were reinforced intermittently for this behaviour. A variable reinforcement schedule such as this produces persistent responding and it is not surprising that it generalized to a similar context in testing. Even though the behaviour of handing objects to humans was not intentionally trained for those studies, the prior learning history of the subjects has to be taken into consideration. When training is an explicit part of the experimental procedure, caution is needed in interpreting the results. In one study, chimpanzees that had been trained in symbolic use transferred food to their partners 49 . However, explicit training through standard shaping and training produced similar results in pigeons 50 , highlighting the importance of flexible behaviours in response to novel circumstances 35 . Other, more recent, helping studies 35 , 36 , 46 , 48 , 51 can be explained by social tool use (giving a tool to a partner to get food for oneself), responding to solicitation (begging), or task persistence whenever food was visible, calling into question prosocial motivations. One suggestion for the inconsistent evidence for prosociality in chimpanzees is that they are too competitive and that other species might be better models. Bonobos, which are as closely related to humans as are chimpanzees, have been suggested as having a more peaceful temperament 52 . Evidence for prosocial preferences has come from studies in which they open doors for conspecifics allowing them to co-feed 53 , 54 . However, when given prosocial preference tasks as used in chimpanzees, there has been no evidence for active food sharing in bonobos 19 , 55 . In the only test of instrumental helping in bonobos, there was no evidence of prosociality, with only the relatively solitary orangutans handing tools to distressed partners 47 . Other non- Pan ape species have shown no signs of sharing 19 , 30 , 56 , evidence in Old World primates has been mixed, with self-regarding preferences for social contact as possible explanations for apparent prosociality 29 , 30 , 57 , 58 , 59 . Among New World primates, capuchin monkeys have not consistently exhibited prosocial sharing 19 , 31 , 60 , 61 , 62 although see 63 , 64 . They will ‘help’ humans by passing objects to them in exchange for food 34 , 65 , although they do not transfer objects to help conspecifics get food 66 . The strongest evidence for prosociality comes from more distantly related New World monkeys that provide alloparental care for offspring, leading to the suggestion that cooperative breeding is a key driver for the evolution of primate prosociality 27 , 28 , 30 , 67 , 68 . However, these results have not been consistently replicated 26 , 69 and various experimental details such as test order effects and partner presence cast some doubt on the evidence for sharing in these primates 70 (though see ref. 71 ). As yet, there have been no tests of instrumental helping in callitrichines. Almost all prior food delivery (sharing) and instrumental helping studies have involved training in which actors first learn through personal experiences the results of their actions. While helpful in demonstrating that the animals have learned the task contingencies, this can create an expectation for getting rewarded in the test context. While it is important to demonstrate task understanding, this could be done after testing, or between tests of naïve then trained subjects as done here (ABA design). By giving chimpanzees the opportunity to observe the consequences of their actions on others before giving them personal experience mitigates the food expectation while still demonstrating task comprehension. Chimpanzees are able to learn through others by observing them 72 , so there is no reason that they could not learn to help, or hinder, others solely on the basis of observed outcomes. The difference between studies that find evidence for prosociality in chimpanzees and other primates and those that do not can be attributed to design features of the experiments. Experimental setups which contain stimuli that are sufficiently interesting for chimpanzees and other nonhuman primates initially elicit actions. Once a novel feature of the environment ceases to be engaging or of personal value, interest in it diminishes, without any consideration for how this affects other individuals. If prior studies of chimpanzee prosocial behaviour that simply presented a single choice (do something or do nothing) had been designed such that the outcome of an action was harmful rather than beneficial to a partner, then chimpanzees might have been considered to be spiteful. The key strength of the approach taken in our study is that the same action (release a peg) under the same stimuli (shaking food box) had two opposing effects. The GO/NO-GO approach is a more powerful method for teasing apart other-regarding motivations from self-interest, in contrast to all prior helping studies that only use a GO design. Stimulus enhancement is one important determinant of chimpanzees’ apparent prosociality in experimental settings. By engaging with an interesting or novel feature of the environment, particularly when doing so has a history of providing rewards, chimpanzees can benefit others as incidental by-products. Social benefits can also arise as by-products if the subject is trying to gain social contact, play with the partner, signal dominance, respond to harassment and so on 73 . Self-interest, rather than concern for the welfare of others, could explain putative prosocial behaviour in chimpanzees. It might have been the case that the stakes were too low in the experiments reported here and that in other contexts—for example, where peg release might free a conspecific from confinement—motivated aid might be elicited. But a general point highlighted by our studies is that prosocial motivations cannot be elucidated from prosocial actions without first controlling for intrinsic interest in experimental setup (for example, by using a GO/NO-GO method). Future studies on helping, sharing, comforting and informing will have to directly address the motivational substrate. Thus, even though chimpanzees have elsewhere shown to recognize something of the goals of others 74 , they appear to lack the motivation to see those goals realized. Studies on chimpanzees and other nonhuman animals can shed light on the origins of our own prosocial behaviour and its importance for large-scale nonkin cooperation. Methods Experiment 1 We tested 13 chimpanzees with three recipients (16 chimpanzees overall, see below). These animals were rescued from illegal wildlife trade and kept at the Ngamba Island Chimpanzee Sanctuary, a forested island in Uganda. The research was approved and reviewed by the local ethics committee of CSWCT (Chimpanzee Sanctuary and Wildlife Conservation Trust), the organization running the Chimpanzee Sanctuary in Uganda, as well as UWA (Ugandan Wildlife Authorities) and UNCST (Ugandan National Council for Science and Technology). The chimpanzees live freely on the island, and come in at night to the sleeping rooms where they were tested in the morning. Participation was voluntary and after testing the subjects re-joined the rest of the group. They were not food or water deprived. Subjects had previously participated in studies on cooperation and helping 33 , 36 , 75 , 76 , 77 , 78 , 79 , but the current setup involved a novel apparatus. For further details on the subjects, see Supplementary Table 1 . These studies took place in the summer of 2011. One subject (recipient) could interact with a Plexiglas box containing food (food box). A second subject (actor) could not get food from the box in the test conditions, but could release a wooden peg connected to the box. For one group of actors, releasing the peg freed the box, allowing the recipient to get food out by shaking it (GO group). For the other group of actors, doing so locked the previously functional (that is, food-delivering) box in place so that the food could no longer be extracted (NO-GO group). During tests, actor and recipient were in separate rooms of the sleeping area across a 2 m wide corridor used by experimenters and keepers: actors were in the room with peg access (peg room) and recipients were in the room with food box access (food room). Actors and recipients could not interact physically, but could see and hear each other, as well as the entire apparatus (that is, food box and attached peg). The peg and food rooms were bridged by an overhead ‘raceway’ that was used during knowledge probes, but which was closed during test and control trials. The food boxes for both groups of subjects (GO and NO-GO) were Plexiglas boxes fixed to the outside of the mesh of the recipient’s room. The lower end of each food box was set 60 cm above the ground. Directly below each food box, a metal hopper channelled the food into the food room; the experimenter could also drop food down the hopper directly ( Supplementary Fig. 1 ). Hopper and food box were both placed directly under the overhead raceway. The food boxes could be directly accessed by the experimenters, but not by the subjects. Recipients could get food (shelled peanuts) only indirectly, namely by shaking the boxes using a chain attached to the bottom that led through a hole in the Plexiglas food box into the food room. The use of a chain and shaking boxes was designed to be noisy so as to attract the attention of the actors, as need for help may have to be signalled through instrumental actions or communication 36 . A series of trays inside the boxes limited the rate at which peanuts cascaded down to the opening at the bottom, necessitating repeated shaking—thus getting a few peanuts per shake. The GO apparatus had a strong, inflexible cord running across the corridor to the peg attached to bars of the peg room. This static cord prevented the GO box from being shaken. However, once the peg was released, a small rubber cord attached to a metal angle allowed the food box to be shaken repeatedly, dispensing food in the food room. The NO-GO apparatus, on the other hand, had a strong rubber cord running across the corridor to the peg attached to the mesh of the peg room. This cord allowed the food box to be shaken repeatedly, causing food to come out—as long as the peg maintained tension on the rubber cord. Once the peg was released, the NO-GO food box fell flush to mesh of the food room, and could no longer be shaken (it lacked the small rubber cord of the GO apparatus). The GO apparatus was marked with blue tape and the NO-GO with green tape to facilitate coding. There is no reason to believe that these colours had any influence on the chimpanzees’ behaviours. There was no demonstration or familiarization phase for the actors prior to testing. This was done to avoid any unintended learning effects that could have come about from actors getting the food themselves. Actors had to learn about the consequences of their actions during the test trials, but they could easily see across the corridor, a distance that had been used in a prior helping study 36 . During testing, actors and recipients were brought into their respective rooms (actors into the peg room; recipients into the food room). The doors and the raceway were closed so that none could access the others’ room or other parts of the sleeping area during the test. The recipient that each actor started with was counterbalanced across actors. Following this, recipients were always exchanged every two trials, in a fixed order (after Asega came Baluku, followed by Mawa, then Asega, and so on). We kept this order across all studies. At the start of each trial, actors were given distractor items (towels soaked in fruit juice) to prevent a ceiling effect for peg releases; also, they were given a 6 m long rope (that served as distractor to reduce random pulling behaviour); all of these were also done in prior helping studies 33 , 36 . Both actors and recipients were also distracted with single peanuts at the start of each trial to keep them in position while the peg was placed and the food box baited. The soaked towel and individual peanuts also served to maintain motivation, particularly for the actors who received no food rewards during the test. When the actors were in position away from the apparatus, the experimenter showed them a handful of about 45 peanuts (a small handful) and then walked over to bait the food box by pouring the peanuts into the top shelf. He then placed the peg into the mesh of the actor’s room and then signalled the start of the trial. All this while, the actor as well as the recipient were kept away from the apparatus by a human helper each who provided the actor with single peanuts. The helper aimed to ensure that the actor would observe the baiting of the food box, but would also ensure that the actor would not leave position prematurely. Before each trial, both helpers stopped providing single peanuts to actor and recipient, respectively, then the actor was given the juice-soaked towel, and then both helpers and the experimenter left the testing area: this was the start of a trial. Each trial lasted 60 s regardless of outcome. After the 60 s, the experimenter blocked the apparatus so that no more food could be released by the recipients. For each actor there were four sessions with four trials in each session. To maintain the recipients’ motivation we provided motivational trials: if they did not receive any food in three successive experimental trials, they were given a motivational trial with a 50% probability of getting food. If after this the recipients again did not receive food in another experimental trial, they again received a motivational trial. This continued until recipients received food in an experimental trial. To ensure that these motivational trials did not interfere with the actors’ motivations and knowledge of the apparatus, actors were moved out of sight before motivational trials commenced. After completing the 16 trials of the experimental phase, actors were given post-test knowledge probes to determine whether they learned about the effects of their actions on the apparatus through observation. They were given one session of four trials. Actors started in the peg room; there was no partner in the food room. Once in position, the door to the overhead raceway was opened, allowing them access to the food box by traversing over to the food room. The raceway access was closed after 60 s had passed and subjects were given an additional 60 s to gather the food from the apparatus. This protocol was followed whether subjects released the peg or not for both GO and NO-GO groups, that is, subjects would not necessarily get food, and they could remain in the peg room if they failed to cross. At the end of the trial (120 s in total), the experimenter locked the food box and the actor was moved back to the peg room unless already there (unless all four trails were finished, upon which the actor was let go to join the conspecifics in the outdoor area). All trials were videotaped with Sony digital cameras. The primary measure of whether the peg was released or not by the actor was coded live. All reported tests are two-tailed. Twenty percent of the trials were coded for reliability by an assistant blind to the study’s design and purpose. Reliability for whether the peg was released by the actor was excellent (Cohen’s kappa 0.95). Experiment 2 Eleven chimpanzees from experiment 1 participated as actors in experiment 2; two from the NO-GO group failed to pass the training phase (Nkumwa, a female; and Kisembo, a male). The GO group consisted of six individuals (three males) and the NO-GO group had five (three males). The same three chimpanzees were used as recipients ( Supplementary Table 1 ). The setup and apparatus were the same as in experiment 1. One additional room was used for a social control condition; here, instead of being in the food room, the recipient was in a room adjacent to the peg room (that is, a room without access to either peg or food box). In this way, the recipient was still present (that is, the control was still social), but unable to interact with the setup. The overall procedure was the same as in experiment 1. The key difference was the addition of familiarization/training trials in which the actors directly experienced the consequences of their actions on the apparatus during the training phase. Actors were also given knowledge probes before and after testing to ascertain that they had learned—and remembered—how to use the apparatus to their personal benefit. Furthermore, actors were also tested in a social control condition in which the recipient was present and visible, but in a third room in which they could not interact with the apparatus (see above). The number of sessions was reduced to three, rather than four, due to testing time constraints. Still, each session contained four trials (see Supplementary Table 2 ). Actors were given the same test apparatus (GO or NO-GO) that they had interacted with in experiment 1. There were three steps in the training phase, all of which had to be passed before actors could advance to the testing phase. In step 1, actors were in the food room. The apparatus was in the same functional state for both the GO and NO-GO groups (configured so that shaking it would release food). The experimenter baited the apparatus with approximately 45 peanuts. One more peanut was dropped down the hopper so that no initial shaking would be required to get this one. The purpose of this was to attract the chimpanzees to the food box. Once the chimpanzee arrived at the food box, they were given a 60 s trial. They had up to ten trials in a session—with a maximum of two sessions—to reach criterion level of performance of shaking the food box at least five times in each of three successive trials. All subjects—with the exception of one male (Kisembo)—passed this criterion, and moved on to step 2. In step 2, actors of both groups started in the food room as before and had 120 s to shake the food box to get food (it was no longer necessary to drop peanuts down a hopper to lure them to the food box). The experimenter then placed the peg into the mesh of the peg room. After the actor had shaken the food box (five shakes), the overhead raceway was opened, allowing actors the choice to leave the food room and go to the peg room to release the peg or not. Actors in the NO-GO group thus had to inhibit moving over and releasing the peg (since doing so would have locked the box into position and prevented the food from being accessible). However, the opening of the raceway allowed subjects from the GO group to make their food box functional: the GO group actors could move to the peg room, release the peg there and then go back to the food box to shake food out. As in step 1, there was a criterion level of performance. Actors had to be able to shake the food box in its functional state at least five times in three successive trials within a maximum of two sessions of ten trials; for the GO group, this meant releasing the peg prior to shaking, for the NO-GO group, this meant inhibiting releasing of the peg. All 12 remaining actors passed criterion and moved to step 3 of training. In step 3, having learned to release the peg (GO group) and to inhibit releasing it (NO-GO group), actors now started the trial in the peg room. The trial began with the raceway to the recipient’s room locked. Actors were distracted by slowly handing them peanuts (as in the test phase of experiment 1). As in the test of experiment 1 they were then shown a handful of peanuts, which was placed into the food box. The peg was then attached to the mesh of the peg room where the actor was and the raceway was then opened. GO actors had to release the peg before passing over the raceway to the food room. NO-GO actors had to instead inhibit releasing the peg, then to cross the raceway to be able to get food out of the food boxes. The same criterion as in step 2 applied, and again subjects had a maximum of two sessions of ten trials to reach criterion. Eleven actors reached criterion and thus passed the final step of training. Only one actor, Nkumwa in the NO-GO group, failed and was thus excluded from further testing. The pre-test knowledge probe was the same as the post-test knowledge probe of experiment 1. Actors had access to the food box in the recipient’s room via a raceway. There was a single session of four trials. In the test, actors were given the same apparatus they had used in experiment 1 as well as the training phase of experiment 2. They were paired with recipients as in experiment 1. In addition to the test condition with the recipient in the food room (as in experiment 1), a social control condition had the recipient sitting in the room adjacent to the peg room, fully visible to the actor. The actor was free to release the peg in the social control condition, but since the food room was empty, releasing the peg had no effect on the movement of the food box or consequences for a conspecific. The test and social control trials were presented in a blocked design counterbalanced across actors. That is, actors were given three sessions of each condition (either social control or test) before switching to the other condition. There were four trials in each session. Following the test phase, actors were given four post-test knowledge probe trials, as described above for the pre-test knowledge probe, to determine whether they remembered their prior training. Data coding and analyses were conducted as in experiment 1. Data availability All summary data are included in the manuscript and supplement; requests for more detailed data collected for this study are available from the corresponding author on request. Additional information How to cite this article: Tennie, C. et al . The nature of prosociality in chimpanzees. Nat. Commun. 7, 13915 doi: 10.1038/ncomms13915 (2016). Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

New research into chimpanzees suggests that, when it comes to altruistically helping a fellow chimpanzee, they are 'indifferent'. The paper, published in Nature Communications, found no evidence that chimpanzees had a tendency to help others - or conversely to be spiteful - when there was no anticipated benefit to themselves. In two experiments, chimpanzees could determine whether or not a peer received access to food, and they displayed no preference for either providing access or denying access with their actions. The team, from the universities of Birmingham, Manchester and St Andrews, and the Max Planck Institute in Leipzig, argue that previous experiments that imply 'prosocial' or helpful behaviour in non-human primates were likely a by-product of task design that nudged chimpanzees into participation. Chimpanzees at the Ngamba Island Chimpanzee Sanctuary, a forested island in Uganda, participated in two studies. None of the chimpanzees were known to be closely related to each other. In the first experiment, 13 chimpanzees were split into two groups and tested with three recipient chimpanzees. They were able to see and hear one another and the equipment, including a food box containing shelled peanuts that could only be reached from the recipients' room. Recipients in the room could shake the food boxes. Sometimes shaking the box had no effect because it was locked in position by a peg in the room of the chimpanzee being tested. Other times, shaking the box allowed the recipients to get the peanuts inside. The test subjects could not get the food, but they could see what the recipients were doing. In one group (the "go group"), releasing the peg would unlock the food box and allow the recipients to get the peanuts. The second group of chimpanzees (the "no go group") saw the recipients who could get the food, but releasing the peg would lock the box and restrict the access to food. As in previous studies, the chimpanzees released the peg at high rates but this quickly declined when there was no reward for them. Importantly, there was no difference in peg release between the two groups - the chimpanzees were no more likely to release the peg to help the recipient in the go group as they were to block access to food in the no-go group. In the second experiment, 11 of the same chimpanzees were given training sessions in which they individually learned the impact of their actions on the recipients' food boxes by being able to go to the room containing the food box. This additional knowledge did not lead to a rise in peg release where food would be given to another chimpanzee, nor did it decrease peg releases when this would prevent others from eating. Dr Claudio Tennie, from the University of Birmingham, explained, "The results of these two experiments combined demonstrate that the chimpanzees did not act in a manner that would produce benefits for others in a task where there was no perceived benefit to themselves." "Indeed, given that the participants were just as likely to prevent access to food as they were to permit access, chimpanzees are no more altruistic than they are spiteful. Even after they demonstrated a clear understanding of the consequences of their actions, they remained indifferent to any effects these actions may have on others." Dr. Keith Jensen, from the University of Manchester, added, "The evolution of social behaviour, and what drives individuals to act altruistically, is an important and active area of debate. Chimpanzees appear indifferent to others when it comes to giving food to others in experimental settings, yet puzzlingly, they have been shown to help humans and other chimpanzees. There has been an appealing suggestion that the roots of human altruism extend down at least as far as our common ancestor with chimpanzees. However, the results of this study challenges that view. "Helping" might have formerly arisen as a by-product of interesting tasks." "If true, this would mean that prosocial behaviour has developed late in evolution, after our split with the other apes."

nature.com/articles/doi:10.1038/ncomms13915

Biology

Harnessing functions of microbiota to combat fungal pathogens

Sunde Xu et al, Fusarium fruiting body microbiome member Pantoea agglomerans inhibits fungal pathogenesis by targeting lipid rafts, Nature Microbiology (2022). DOI: 10.1038/s41564-022-01131-x Journal information: Nature Microbiology

https://dx.doi.org/10.1038/s41564-022-01131-x

https://phys.org/news/2022-05-harnessing-functions-microbiota-combat-fungal.html

Abstract Plant-pathogenic fungi form intimate interactions with their associated bacterial microbiota during their entire life cycle. However, little is known about the structure, functions and interaction mechanisms of bacterial communities associated with fungal fruiting bodies (perithecia). Here we examined the bacterial microbiome of perithecia formed by Fusarium graminearum , the major pathogenic fungus causing Fusarium head blight in cereals. A total of 111 shared bacterial taxa were identified in the microbiome of 65 perithecium samples collected from 13 geographic locations. Within a representative culture collection, 113 isolates exhibited antagonistic activity against F. graminearum , with Pantoea agglomerans ZJU23 being the most efficient in reducing fungal growth and infectivity. Herbicolin A was identified as the key antifungal compound secreted by ZJU23. Genetic and chemical approaches led to the discovery of its biosynthetic gene cluster. Herbicolin A showed potent in vitro and in planta efficacy towards various fungal pathogens and fungicide-resistant isolates, and exerted a fungus-specific mode of action by directly binding and disrupting ergosterol-containing lipid rafts. Furthermore, herbicolin A exhibited substantially higher activity (between 5- and 141-fold higher) against the human opportunistic fungal pathogens Aspergillus fumigatus and Candida albicans in comparison with the clinically used fungicides amphotericin B and fluconazole. Its mode of action, which is distinct from that of other antifungal drugs, and its efficacy make herbicolin A a promising antifungal drug to combat devastating fungal pathogens, both in agricultural and clinical settings. Main The microbiome plays a crucial role in host development, fitness and health 1 , 2 , 3 , 4 . Numerous microbiome studies have provided insights into the composition of various microbial communities and their dynamic changes in response to exogenic factors 2 , 5 , 6 , 7 , 8 . However, autogenic factors in microbiome assembly and function are mostly overlooked, especially in terms of the importance of microbe–microbe interactions 1 . Detailed knowledge of physical and molecular interactions between different microbiome members has a high potential to aid the development of new strategies for targeted manipulations of the microbiome to improve host health 3 , 4 , 9 , 10 . In addition, the underlying interactions could also be considered in a broader sense for microbial communities of pathogenic host organisms, for example phytopathogenic fungi. Interactions between bacteria and fungi are crucial for modulating microbial communities in terms of biodiversity, ecosystem functions and host health when associated with multicellular organisms 11 . Molecular mechanisms governing inter-kingdom communication are reflected by numerous biological interactions, from antagonism to mutualism 12 . In agroecosystems, investigations of bacterial–fungal interactions are mainly focused on the suppression of pathogens 4 , 12 , 13 . Bacteria generally interfere with fungal pathogens by competition for resources and secretion of lytic enzymes or bioactive compounds (antibiosis) 14 . In most cases, the detailed modes of action of many antagonistic bacterial–fungal interactions and the involved bioactive compounds remain elusive 15 . Sometimes the underlying mechanisms have been only partially deciphered. For example, herbicolin A, which can be biosynthesized by certain Pantoea agglomerans isolates, was identified about 40 years ago as the bacterium’s most important bioactive compound against fungi and Mollicutes 16 , 17 , 18 , 19 , 20 , 21 . However, the structure of the biosynthetic gene cluster for herbicolin A as well as its mode of action remain elusive. Most phytopathogenic fungi can produce sexual structures such as perithecia that allow them to survive on infested plant debris before the next infection cycle 22 . Fusarium graminearum (Fg), the major pathogen causing Fusarium head blight (FHB) in wheat ( Triticum aestivum ), uses this strategy with high efficiency 23 . FHB epidemics not only affect global wheat production, but also contaminate grain with mycotoxins 24 . Interfering with perithecium formation to decrease the initial inoculum is a potential strategy for FHB management. Current literature mostly focuses on changes of the plant microbiome as a response to disease or contrasts healthy with diseased states 4 , 25 , 26 , 27 . The initial origin of disease-responsive microorganisms remains in most cases elusive; they might be associated with the plant host or with the pathogen itself. Similar to soil, plants and animals, fungal structures are also known to provide microhabitats for adapted bacterial communities; they can also be involved in destructive interactions with their hosts. Perithecia provide a highly specific ecological niche for bacteria that we hypothesized could play a pivotal role in disease establishment. Preliminary assessments of bacterial–fungal interactions in such microhabitats indicated antagonistic activity via emission of volatile bioactive metabolites 28 , 29 . At present, there is a lack of large-scale assessments of fungal fruiting body microbiomes and mechanistic insights into interactions between prevalent bacteria and their hosts. In this study, bacterial communities associated with Fg perithecia were profiled. Phylogenetically distinct bacteria were always detectable in perithecia from separated geographic locations in China. In parallel, representatives of 30 core microbiome members were cultured and screened for antagonistic activity. This resulted in the isolation of one highly active antagonist ( Pantoea agglomerans ZJU23) with pronounced antagonistic activity against FHB. We showed that herbicolin A (HA) secreted by ZJU23 is responsible for suppression of the phytopathogen. HA binds ergosterol and disrupts the integrity of ergosterol-containing fungal lipid rafts suppressing fungal growth, perithecium formation and virulence of Fg. Furthermore, we showed that HA has a wide antifungal spectrum including human opportunistic fungal pathogens, indicating a potential for its use in controlling fungal diseases in agricultural and clinical settings. Results Insights into perithecium-associated bacterial communities Using scanning electron microscopy, morphologically distinct bacteria were found to be embedded in or attached to the surface of Fg perithecia that were collected from geographically distant fields (Fig. 1a and Extended Data Fig. 1a ). For a detailed analysis of bacterial communities, we sampled perithecia, rice stubbles and rhizosphere soil from 13 sites within major wheat-growing regions (Fig. 1b and Supplementary Data 1 ). The bacterial community was profiled on the basis of bacterial 16S ribosomal RNA (rRNA) gene fragment sequences clustered into operational taxonomic units (OTUs; 97% sequence similarity) (Extended Data Fig. 1b ). We obtained 49,457,306 high-quality reads from 195 samples (ranging from 5,933 to 547,412 reads, with an average of 253,627 reads per sample), phylogenetically representing 4,912 OTUs (mean: 1,787 OTUs per perithecium sample; 976 OTUs per rice stubble sample; 3,179 OTUs per rhizosphere soil sample; Supplementary Data 2 ). Rarefaction curves indicated that the sequencing depth of the current study was sufficient to capture bacterial diversity (Extended Data Fig. 2a ). Fig. 1: Assessment of perithecium-associated bacterial communities. a , Scanning electron microscopy observation of bacteria on perithecia collected from the field. The experiment was repeated three times with similar results. b , Locations of 13 sampling sites located in two Chinese provinces. AH, Anhui Province; JS, Jiangsu Province. c – d , Richness ( c ) and Shannon ( d ) index of the bacterial communities from perithecia, rice stubbles and soil. Different letters indicate significantly different groups ( P < 0.001, ANOVA, Tukey's HSD). The horizontal lines within boxes represent medians. The tops and bottoms of boxes represent the 75th and 25th percentiles, respectively. The upper and lower whiskers extend to data no more than 1.5× the interquartile range from the upper edge and lower edge of the box, respectively. e , Principal coordinates analysis based on Bray–Curtis distance showing that the microbiota of perithecia is distinct from that of soil and rice stubbles ( P = 0.001, PERMANOVA with ADONIS test). f , Stacked bar plot combined with Bray–Curtis distance-based dendrogram showing the averaged bacterial community composition for each compartment at each sampling site. The stacked bar plot includes the phylum and proteobacterial class levels as indicated. g , The core microbiome of perithecium-associated bacteria. The inner ring represents the OTUs that were reproducibly detected in the perithecia and infected rice stubbles with relative abundance greater than 0.1%. The relative abundances of distinct perithecium and stubble OTUs are shown as blue (perithecium) and green (stubble) heatmaps. Replicates for each compartment, n = 65 biologically independent samples. Source data Full size image Assessments of α-diversity revealed a significant difference in richness and Shannon index among perithecium, rice stubble and rhizosphere soil samples (Fig. 1c,d ; P < 0.05, ANOVA, Tukey's HSD). Detailed examination of the bacterial composition of these samples showed that the perithecium microbiota exhibited higher diversity than those found in stubble, but lower diversity than those in rhizosphere soil (Fig. 1c,d ). Unconstrained principal coordinate analysis (PCoA), based on a Bray–Curtis distance matrix, revealed that the perithecium communities formed a distinct cluster and indicated that they are more similar to the stubble communities than to the rhizosphere soil communities (Fig. 1e ; P = 0.001, PERMANOVA with Adonis test). Moreover, constrained PCoA (CPCoA) plotting with complementary analyses indicated that the clusters explained 23.3% of the variance within the three compartments (Extended Data Fig. 2b ; P = 0.001, permutation test with anova.cca). Further analysis based on average Bray–Curtis distances of taxonomic composition at the phylum level across 13 sites revealed that bacterial communities from perithecia, stubble and rhizosphere soil clearly clustered according to the respective compartment (Fig. 1f ). Different OTUs visualized in a Manhattan plot showed similar distribution patterns (Extended Data Fig. 2c and Supplementary Data 2 ). Predominant OTUs (relative abundance >0.1%, 111 OTUs) were defined as the core microbiome of perithecia and account for 2.3% (111/4,912) of the total OTUs, and 80.4% and 74.4% relative abundance in the perithecium and stubble microbiomes, respectively. These OTUs were classified into 13 bacterial orders and 32 families, belonging to the phyla Proteobacteria, Actinobacteria, Bacteroidetes and Firmicutes (Fig. 1g ). Microbial source tracking analysis revealed that perithecium-associated bacteria mainly originated from the stubble microbiome (mean ± s.e.m. fractional contribution, 92.85 ± 0.81%; Extended Data Fig. 2d and Supplementary Data 3 ). The overall results indicated that the microbiome of perithecia was distinguishable from those of the other analysed compartments. ZJU23 isolate reduces FHB and mycotoxin contamination To test whether there are beneficial bacteria in the Fg perithecium microbiome for biological control of FHB, we isolated and identified bacteria from 65 independently pooled perithecium samples collected from the aforementioned fields 30 (Extended Data Fig. 1c ). A total of 2,036 bacterial isolates were recovered from perithecia, which shared >97% 16S rRNA gene similarity to 30 OTUs in the core perithecium microbiome (relative abundance >0.1%, a total of 111 OTUs) (Fig. 2a and Supplementary Data 2 ). Among them, 113 isolates representing 13 core OTUs demonstrated various degrees of inhibitory activity against fungal growth in dual-culture assays (Fig. 2a and Supplementary Data 4 ). Notably, a bacterial isolate (termed ZJU23) showed the strongest inhibitory activity against Fg during co-cultivation, producing radial inhibition zones ≥20 mm (Fig. 2b ). ZJU23 was matched to OTU_10 in the core perithecium microbiome (Supplementary Data 2 ), which was identified as Pantoea agglomerans on the basis of phylogenetic analysis and genome sequencing. The ZJU23 genome comprises one chromosome and three plasmids (Supplementary Fig. 1 ). Fig. 2: Isolation and identification of the F. graminearum antagonist ZJU23. a , Information of core perithecium microbiome and cultivable bacteria recovered from Fg perithecia. The maximum-likelihood phylogenetic tree of the bacterial genera tree was constructed on the basis of full-length 16S rDNA sequences. The inhibition zones of representative strains and the relative abundance of corresponding OTUs in the perithecium microbiome are included in the bar plots on the right. b , The antagonistic activity of ZJU23 against Fg PH-1. c , Inhibition efficiency of ZJU23 against the formation of Fg perithecia on carrot agar medium (left panel) and corresponding perithecium numbers in each treatment (right panel). Data are presented as mean ± s.e.m. of n = 7 biologically independent samples. Different letters indicate significantly different groups ( P < 0.05, ANOVA, Tukey's HSD). The non-antagonistic P. agglomerans Pa58 was used as a control. d , e , Inhibition efficiency ( e ) of ZJU23 against deoxynivalenol (DON) toxisome formation and DON production. DON-toxisome formation labelled with Tri1-GFP in each treatment was observed at 60 h under toxin-inductive conditions using a confocal microscope ( d ). The DON production in each treatment was quantified at 7 days post-inoculation (dpi). Data are presented as mean ± s.e.m. of n = 5 biologically independent samples. Different letters indicate significantly different groups ( P < 0.05, ANOVA, Tukey's HSD). The experiment was repeated three times with similar results ( e ). f , g , Biocontrol efficacy of ZJU23 against FHB on wheat anthers and seedlings in a growth chamber ( f ). The representative infected anthers are shown at 5 dpi. The lengths of lesions on wheat coleoptiles at 7 dpi ( g , right) were measured and statistically analysed ( n = 7 biologically independent wheat coleoptiles, mean ± s.e.m., P < 0.05, ANOVA Tukey's HSD). The experiment was repeated three times ( g , left) with similar results. h , i , The disease index of FHB after treatment with ZJU23 ( h ) and efficiencies of ZJU23 in controlling DON production ( i ) in field trials conducted in three continuous years from 2017 to 2019. Water was used for the untreated control. Data are presented as mean ± s.e.m. of n = 3 biologically independent samples, *** P < 0.001, ** P < 0.005, two-sided Student’s t -test with false discovery rates (FDR) correction. Source data Full size image ZJU23 treatment significantly suppressed perithecium formation, with 80–90% reduction both on carrot ( Daucus carota ) agar plates (Fig. 2c ) and rice straws (Extended Data Fig. 3a ), in comparison with untreated control and non-antagonistic P. agglomerans strain Pa58 treatment. Pa58 isolated from the perithecium microbiome showed similar growth and colonization patterns to ZJU23 and was used as a control to exclude the possibility that nutrient competition is involved in the observed effects (Supplementary Fig. 2 ). Deoxynivalenol (DON) synthesized in fungal DON-toxisomes is a critical virulence factor of Fg 24 , 31 . We found that ZJU23 substantially suppressed the formation of fungal DON-toxisomes as indicated by Tri1-GFP in the Fg hyphe grown in the DON induction medium (Fig. 2d ). Comparative assessments confirmed that DON production was significantly decreased when Fg was co-cultured with ZJU23 (Fig. 2e ). To determine whether ZJU23 inhibited growth of Fg in planta, we conducted biocontrol experiments both in a growth chamber and in the field (see Supplementary Methods for details). Treatment with ZJU23 by foliar spray almost completely suppressed Fg infection on wheat heads (Fig. 2f ), and significantly suppressed disease progression of Fg in germinated wheat seeds (Fig. 2g and Extended Data Fig. 3b ). Moreover, during three consecutive years of field trials, ZJU23 consistently showed a clear biocontrol efficacy of 50–70% against FHB (Fig. 2h ). In addition, ZJU23 substantially suppressed DON production by 50–80% during field trials (Fig. 2i ). Together, these results indicate that ZJU23 is an effective biocontrol agent for the control of FHB and can be used to reduce DON contamination. HA is the antifungal compound produced by ZJU23 To identify antifungal compound(s) produced by ZJU23 and their biosynthetic gene cluster(s), we first performed a mutagenesis screen for the loss of antifungal activity. More than 12,000 himar1 mariner transposon mutants were assessed for antifungal activity, and 80 inactive transposon mutants were identified and subsequently subjected to whole-genome re-sequencing for mapping of the insertion sites. Transposon insertions were mainly distributed in the open reading frames of ORF3911, ORF3913 and ORF3914 and in their neighbouring ORFs (in 41 mutants) (Fig. 3a ). These genes are located in a ~42 kbp DNA region in one orientation that is located on plasmid 1 as illustrated in Fig. 3b , indicating that this region is a potential biosynthetic gene cluster for the antifungal compound in ZJU23. Ten putative proteins were predicted for this gene cluster and assigned A ntifungal c ompound b iosynthetic protein A (AcbA) to AcbJ. To validate whether the AcbA–AcbJ cluster is responsible for antifungal compound biosynthesis, we constructed individual gene mutants and tested their antifungal activity. Bacterial growth was not altered in any of the mutants, when compared with ZJU23 (Supplementary Fig. 3 ). Eight of the mutants (that is, all mutants except those lacking ACBG and ACBJ ) showed no inhibition activity towards Fg (Fig. 3c ), suggesting that the AcbA–AcbJ cluster is responsible for antifungal compound biosynthesis in ZJU23. Fig. 3: Herbicolin A is the major antifungal compound produced by ZJU23. a , Distribution and frequency of transposon disrupted genes in the ZJU23 transposon mutants that lost antifungal activity. Blue dots indicate a potential active compound biosynthetic gene cluster. b , Potential active compound biosynthetic gene cluster and putative assignments of encoded proteins. c , Antifungal activity of individual deletion mutants in the biosynthetic gene cluster. Representative photographs of antifungal activity of ZJU23 or mutants against FG are shown. d , Structure of HA. e , Extracted ion chromatographic (EIC) comparison of HA production by ZJU23 and the indicated mutants. Pure HA was used as a positive control. The calculated mass of HA ([M + H] + at m / z 1300.7378) is indicated by a dashed line. f , Relative growth rate of Fg mycelia grown in liquid medium supplemented with the concentrations of HA indicated on the x axis was calculated from n = 6 biologically independent samples (mean ± s.e.m.). The dashed line indicates the EC 50 concentration. The experiment was repeated three times with similar results. g , Inhibition efficiency of HA (2 μg ml −1 ) against the formation of Fg perithecia on carrot agar medium. The experiment was repeated three times with similar results. h , Inhibition efficiency of HA (2 μg ml −1 ) against conidial germination of Fg. i , j , Inhibition efficiency of HA (8 μg ml −1 ) against DON-toxisome formation and DON production. DON-toxisome formation labeled with Tri1-GFP in each treatment was observed at 36 h under toxin-inductive conditions using a confocal microscope ( i ). The DON production in each treatment was quantified at 7 dpi from n =5 biologically independent samples (mean ± s.e.m., *** P = 9.28 × 10 -7 , two-sided Student’s t -test) ( j ). The experiment was repeated three times with similar results ( j ). k , Proposed biosynthetic pathway of HA. Source data Full size image To identify the antifungal compound synthesized by the AcbA–AcbJ cluster, we compared the metabolic profiles of ZJU23 and four deletion mutants (Δ AcbA , Δ AcbC , Δ AcbD and Δ AcbH ) in the AcbA–AcbJ cluster. Unique peaks in ZJU23 identified using liquid chromatography–mass spectrometry (LC–MS) analysis are expected to correspond to peaks of anti fungal c ompounds (Anti-Cs). Two unique peaks (Anti-C1 and Anti-C2) were specifically found in the supernatant of ZJU23, which were further purified and tested for antifungal activity (Extended Data Fig. 4a ). Anti-C1 with (+)-high resolution mass spectroscopy (HR-MS) (obs. [M+H] + at a mass-to-charge ratio ( m / z )1300.7378) (Extended Data Fig. 4b ) showed higher antifungal activity than Anti-C2. Analysis of 1D ( 1 H, 13 C)/2D (COSY, HSQC, HMBC) NMR spectra revealed that Anti-C1 is a glycosylated lipopeptide containing nine amino acid residues. Further analysis of NOESY correlations (Supplementary Data 5 and Extended Data Fig. 4c–g ) and absolute configurations of amino acids in the macrocycle peptide skeleton using Marfey’s method 32 suggested that Anti-C1 is identical to the known lipopeptide compound HA 16 , 17 (Fig. 3d and Extended Data Fig. 5 ). Anti-C2 also possessed antifungal activity and was further identified as herbicolin B, a deglycosylated intermediate during HA biosynthesis. Herbicolin B is transformed to HA by a glycosyltransferase encoded by AcbI (Extended Data Fig. 6 ). In addition, LC–MS analysis further confirmed that mutants lacking either AcbA to AcbF, AcbH or AcbJ did not produce HA (Fig. 3e ). HA efficiently inhibited mycelial growth in both agar plates and liquid media, and suppressed perithecium formation, conidial germination, DON-toxisome formation and DON production (Fig. 3f–j and Extended Data Fig. 7 ). HA itself has antifungal activity and is not metabolized inside the cell, as indicated by comparing the metabolic profiles of HA-treated protoplast samples with those of the non-treated control using LC–MS (Supplementary Fig. 4 ). Taken together, HA is the major antifungal compound produced by ZJU23, and the AcbA–AcbJ cluster is responsible for HA production. Biosynthesis of HA and origin of AcbA–AcbJ cluster Subsequently, we investigated the biosynthesis process of HA by the AcbA–AcbJ cluster. AcbA, AcbC, AcbD and AcbH are non-ribosomal peptide synthetase (NRPS) proteins. A total of nine modules in these NRPSs are responsible for the activation and incorporation of nine amino acids into the growing peptide chain. In addition, the AcbH 2 module catalyses the cyclization by forming an ester bond between amino acids incorporated by modules AcbA and AcbH 2 (Fig. 3k ). We performed LC–MS analyses of both the wild-type (WT) and mutant samples to detect HA and its predicted intermediates. As predicted, the different intermediates were detectable and highly accumulated in the mutants, compared with the LC–MS profile of the WT (Extended Data Fig. 4a ). Furthermore, the molecular weight of emerged intermediates in each NRPS mutant was matched to the predicted intermediates (Extended Data Fig. 6 ). None of the identified intermediates were glycosylated, which indicated that the glycosylation modification by AcbI was the last step of HA biosynthesis. Taken together, HA is likely the product of the following biochemical pathway: AcbA to AcbC to AcbD to AcbH to AcbI (Fig. 3k ). We next traced the origin of the AcbA–AcbJ cluster. The HA biosynthetic gene cluster was either absent or discontinuous with low identity in 102 P. agglomerans genomes available in public databases and the control strain Pa58 (Extended Data Fig. 8a , and Supplementary Fig. 5 and Data 6 ). Further investigation of homologous sequences revealed that homologues of HA biosynthetic proteins were mainly retrieved from strains belonging to Candidatus Fukatsuia symbiotica, Chromobacterium spp. and Burkholderia spp., with high amino acid sequence identities (>40%) (Extended Data Fig. 8a,b and Supplementary Fig. 6 ). In addition, we noted that homologues in Ca . F. symbiotica had a ~38 kb cluster in a plasmid with 60% amino acid identity to the HA cluster in P. agglomerans (Extended Data Fig. 8c ). This high level of identity was surprising because the two bacteria belong to two distinct taxonomic families (Extended Data Fig. 8d ). This remarkable conservation in microsynteny and sequence homologies between ZJU23 and Ca . F. symbiotica gene clusters, not present in other tested P. agglomerans strains, raised the possibility that the HA cluster in ZJU23 was likely acquired from Ca . F. symbiotica via horizontal gene transfer (HGT) or underwent massive losses independently from the common ancestor of P. agglomerans and Ca . F. symbiotica. Fungal ergosterol biosynthesis is involved in HA resistance One strategy to identify drug targets is to isolate and analyse drug-resistant mutants. Given that HA also has strong inhibition activity towards Saccharomyces cerevisiae BY4741 (EC 50 = 0.3 μg ml −1 ), we isolated HA-resistant mutants in the BY4741 background on rich medium supplemented with 6 μg ml −1 HA. A total of 49 HA-resistant mutants were recovered and subsequently subjected to whole-genome sequencing to map the mutated sites. Each of the resistant mutants had at least one non-synonymous mutation, and 72 mutated sites in 55 genes were identified (Fig. 4a and Supplementary Data 7 ). Notably, 9 genes were found to be directly involved in ergosterol biosynthesis (Fig. 4a ). Gene Ontology (GO) analysis of the genes identified in HA resistance mutants showed a significant enrichment of sterol/ergosterol biosynthetic processes (Fig. 4b ). We further tested the HA susceptibility of available deletion mutants (∆ Erg4 , ∆ Erg6 , ∆ Erg8 and ∆ Erg10 ) in the ergosterol biosynthetic pathway of S. cerevisiae , and validated that mutation of genes involved in yeast ergosterol biosynthesis conferred HA resistance. Furthermore, we examined whether ergosterol biosynthesis is also involved in HA resistance in Fg. FgErg4 and FgSR are critical for ergosterol production and virulence for Fg 33 , 34 . Susceptibility tests indicated that mutants lacking each of these proteins had significantly increased HA resistance (Fig. 4d ), similar to that observed in S. cerevisiae (Fig. 4c ). Together, these data suggest that the ergosterol biosynthetic process is involved in the fungal susceptibility to HA. Fig. 4: Ergosterol biosynthesis is involved in the fungal susceptibility to HA. a , The frequency of mutated genes identified by whole-genome sequencing of 49 HA-resistant mutants. Red dots represent the genes involved in ergosterol biosynthesis. b , GO enrichment analysis of mutated genes in a . The colour and size of the bubble represent the adjusted P value and number of enriched genes in each GO pathway, respectively. c , The susceptibility of ergosterol biosynthesis mutants of S. cerevisiae BY4741 towards HA at 6 μg ml −1 . One HA-resistant mutant in a was used as a control (bottom). Schematic diagram of the ergosterol biosynthetic pathway in fungi. Critical ergosterol biosynthetic enzymes are shown. Broken arrows, multiple enzymatic steps; solid arrows, single enzymatic step (top). d , Left: the susceptibility of ergosterol biosynthesis mutants ∆ FgErg4 and ∆ FgSR in Fg towards HA at 4 μg ml −1 . Right: mycelial growth inhibition of HA against the wild type, ∆ FgErg4 and ∆ FgSR of Fg was calculated from n = 6 biologically independent samples (mean ± s.e.m.). Different letters indicate significantly different groups ( P < 0.05, ANOVA, Tukey's HSD). The experiment was repeated three times with similar results. PDA stands for Potato Dextrose Agar. Source data Full size image HA interacts with ergosterol and disrupts fungal lipid rafts The fungal cell lipid plasma membrane naturally contains regions rich in ergosterol and sphingolipids 35 . As the antifungal activity of HA seemed to be connected with ergosterol (Fig. 4 ), we wondered whether HA directly interacts with ergosterol-containing plasma membranes. To elucidate this, Fg protoplasm was incubated with HA, and the plasma membrane was subsequently extracted. An LC–MS assay was used to quantify HA in the fractionated samples. HA was more abundant in the plasma membrane fraction than in the cytosolic fraction and in the supernatant collected through centrifugation after the incubation (Fig. 5a ). To further identify which fungal plasma membrane component is necessary for the retention of HA, major components of the fungal plasma membrane, including ergosterol, sphingomyelin (SM), phosphatidylethanolamine (PE), 1,2-dipalmitoyl- sn -glycero-3-phosphocholine (DPPC) and 1,2-dioleoyl- sn -glycero-3-phosphocholine (DOPC) were individually coated onto glass discs and air-dried. Subsequently, they were covered with an HA solution. Results indicated that HA was mainly retained by ergosterol in this assay (Fig. 5b ). Fig. 5: HA interacts with ergosterol and targets fungal lipid rafts. a , Distribution of HA in Fg protoplasts. The relative amount of HA in each fraction was calculated from n = 3 biologically independent samples (mean ± s.e.m.). Different letters indicate significantly different groups ( P < 0.05, ANOVA, Tukey's HSD). b , HA retention capability of various lipids. The retained HA was quantified by LC–MS from n = 5 biologically independent samples (mean ± s.e.m.). Different letters indicate significantly different groups ( P < 0.05, ANOVA, Tukey's HSD). c , Effect of HA on artificial lipid rafts. The atomic force microscopy elevation maps (left) and statistical results of the percentage of the elevated phases (right) within the ordered lipid domains at different timepoints are shown. d , Effect of HA on lipid domains of giant unilamellar vesicles (GUVs). GUV containing ergosterol or cholesterol, spiked with 0.1% lipid-conjugated fluorescent dye, were treated with 10 μM HA or left untreated. Top: representative original images from hundreds of GUVs analysed. Scale bar, 8 µm. Bottom: statistical analysis of different types of GUVs from 100 GUVs of n = 3 biologically independent samples (mean ± s.e.m.). Brown, normal GUV; Green, dissolution of GUV. e , Localization of the lipid raft marker FgStoA-GFP in various treatments. The final concentration of HA and amphotericin B was 5 and 8 µg ml −1 , respectively (EC 90 of HA or amphotericin B). The solvent DMSO was used as a control. Scale bar, 5 μm. DIC stands for differential interference contrast. f , Fungal membrane permeability following various treatments. Mycelia of Fg treated with HA (6.5 μM) or amphotericin B (6.5 μM) were stained with FM4-64 dye and observed at the indicated timepoints after treatment. Scale bar, 5 μm. g , Transmission electron micrographs of plasma membrane changes caused by HA treatments in Fg. Scale bar, 0.1 μm. The white arrow indicates a pole. All experiments were repeated three times with similar results. Source data Full size image As HA interacts with ergosterol and associates with ergosterol-enriched cell membranes, we wondered whether HA affects the integrity of lipid rafts in Fg. We first observed the influence of HA on the morphology of artificial lipid rafts using atomic force microscopy (AFM). Although these artificial lipid rafts on a mica support are distinct from those on a real biological membrane, our previous study suggested that the components of the artificial lipid rafts behave similarly to those of biological membranes 36 . HA drastically changed the appearance of reconstituted lipid rafts containing ergosterol, with a clear dissolution of the elevated plains (Fig. 5c ). Furthermore, we created giant unilamellar vesicles (GUVs) of defined lipid compositions with fluorescent DiD (1,1-Dioctadecyl-3,3,3,3-tetramethylindodicarbocyanine) labelling the lipid disorder phase 37 . Upon HA treatment, the phase partitioning of fluorescent DiD in GUVs was rapidly shifted because of phase separation, revealing a dissolution effect (Fig. 5d ). However, phase separation was not observed after HA addition in the GUVs containing cholesterol (Fig. 5d ), indicating that HA specifically disrupted the ergosterol-containing artificial lipid raft. Stomatin protein StoA has been identified as a marker for fungal lipid rafts 38 . We next constructed an Fg strain expressing the StoA orthologue FgStoA-GFP (encoded by FGSG_10909) to indicate Fg lipid rafts, and determined the effect of HA on lipid rafts in vivo. In the solvent or the fungicide amphotericin B treatment, FgStoA-GFP mainly localized along the plasma membrane of mycelia and protoplasm (Fig. 5e ). However, upon HA treatment, the plasma membrane localization of FgStoA-GFP was dramatically diminished, and the fused protein was mainly diffused in the cytoplasm (Fig. 5e ). Furthermore, HA treatment resulted in increased membrane permeability indicated by plasma membrane dye FM4-64 staining (Fig. 5f ). Subsequently, transmission electron microscopy observations indicated that the architecture of cell plasma membranes, but not cell wall, was markedly changed (Fig. 5g and Supplementary Fig. 7 ). In addition, the in silico modelling approach showed that the acyl chain of HA rapidly penetrated the bilayer and was only stable when ergosterol was present in the lipid membranes (Supplementary Fig. 8 , and Videos 1 and 2 ). These results suggested that HA is directly inserted into the Fg plasma membrane by interacting with ergosterol. This interaction echoes observation that yeast mutants with a deficiency in ergosterol biosynthesis were resistant to HA (Fig. 4 ). Collectively, HA was shown to target ergosterol-containing lipid rafts and to disrupt the integrity of the cell plasma membrane in Fg. HA exhibits broad-spectrum antifungal activity As ergosterol is highly conserved in fungal plasma membranes, we decided to examine whether ZJU23 and HA are effective in antagonizing other important phytopathogenic fungi. Both ZJU23 and HA substantially inhibited the mycelial growth of all tested fungal pathogens in vitro (Fig. 6a,b ). The half maximal effective concentration (EC 50 ) of HA in these tested species falls in the same range as that against Fg (Fig. 3f ). HA also substantially inhibited spore germination of Botrytis cinerea (Fig. 6c , left), and significantly reduced the severity of grey mould disease on tomato ( Solanum lycopersicum ) and strawberry ( Fragaria × ananassa ) leaves by 93% and 91%, respectively, at a final concentration of 2 μg ml −1 (Fig. 6d,e ). Furthermore, HA (5 μg ml −1 ) effectively suppressed the appressorium formation of Magnaporthe oryzae (Fig. 6c , right) and decreased the infection lesion area by 86% in barley ( Hordeum vulgare ) leaves (Fig. 6f ). We then examined the impact of HA on the human opportunistic fungal pathogen Candida albicans and found that HA effectively inhibited the growth of C. albicans YW02 with EC 50 = 0.0368 μM, which is more efficient than the clinical fungicides amphotericin B (EC 50 = 0.173 μM) and fluconazole (EC 50 = 5.2 μM) (Fig. 6g ). Furthermore, we found that Aspergillus fumigatus CBS101355 was also more susceptible to HA with EC 50 = 0.38 μM, compared with amphotericin B (EC 50 = 6.5 μM) and fluconazole (EC 50 = 13 μM) (Fig. 6h ). Fig. 6: HA exhibits broad-spectrum antifungal activity. a , Antagonistic activity of ZJU23 against six plant-pathogenic fungi. The radius of the inhibition zone (cm) was measured for each co-culture of ZJU23 with Alternaria alternata ( A. a ), Aspergillus flavus ( A. f ), Botrytis cinerea ( B. c ), Colletotichum gloeosporioides ( C. g ), Magnaporthe oryzae ( M. o ) and Sclerotinia sclerotiorum ( S. s ). Data are presented as mean ± s.e.m., n = 5 biologically independent samples. The experiment was repeated three times with similar results. b , EC 50 of HA against tested plant fungal pathogens. c , Inhibition activity of HA on spore germination of B. cinerea and appressorium formation of M. oryzae . Representative images were taken after 3 h of incubation with HA at 5 µg ml −1 . Scale bar, 10 µm. n = 100 spores. d – f , Effect of HA on diseases caused by B. cinerea and M. oryzae in planta. The inoculated B. cinerea ( d, e ) or M. oryzae ( f ) spore suspension was supplemented with the indicated concentration of HA, and then inoculated onto tomato fruit ( d , top), strawberry leaves ( e , top) or barley leaves ( f , top). The area or length of lesions was measured (bottom) and statistically analysed from n = 10 biologically independent fruits or leaves (mean ± s.e.m.). Different letters indicate significantly different groups ( P < 0.05, ANOVA, Tukey's HSD). DMSO was used as a solvent control. The experiment was repeated three times with similar results. g , h , Growth inhibition rate of HA, amphotericin B and fluconazole at various concentrations in C. albicans YW02 ( g ) and A. fumigatus CBS101355 ( h ) (mean ± s.e.m., n = 3 and 6 biologically independent samples for C. albicans YW02 and A. fumigatus CBS101355, respectively). The experiment was repeated three times with similar results. i , Antifungal activity of HA against fungicide-resistant Fg strains. The final concentration of the fungicides or HA was 5 µg ml −1 . Images were taken after 3 d of incubation. Car, carbendazim; Flud, fludioxonil; Phe, phenamacril; Teb, tebuconazole. Car-R, Flud-R, Phe-R and Teb-R indicate fungicide-resistant strains. The experiment was repeated three times with similar results. Source data Full size image Next, we investigated whether HA was able to inhibit the mycelial growth of fungicide-resistant Fg strains. Four categories of resistant strains were tested, which show resistance to carbendazim, phenamcaril, fludioxonil and tebuconazole, targeting fungal tubulin, myosin I, hybrid histidine kinase and sterol 14α-demethylases, respectively. All tested resistant strains were susceptible to HA, indicating that there is no cross-resistance between HA and other tested fungicides (Fig. 6i ). Moreover, HA showed synergistic effects with two commonly used fungicides for plant fungal disease control, polyoxin B and carbendazim (Extended Data Fig. 9 ). Collectively, these results indicated that HA is a potential effective chemical against fungal diseases in both crops and animals, and might be applicable for the management of fungicide resistance issues in agriculture. Discussion Like animals and plants, fungi are also associated with distinct bacterial communities. However, fungal microbiomes remain poorly understood when compared with those of other hosts 12 , 39 . In the present study, we observed that distinct bacteria are associated with the sexual fruiting bodies of Fusarium graminearum . Microbiome profiling revealed that perithecia from geographically distant locations harboured very similar bacterial communities. Source tracking analyses indicated that the identified bacteria were mainly acquired from adjacent rice stubbles, but with different abundance. This finding indicates that microbiota assembly of fungal fruiting bodies is already initiated during the infectious stage of the pathogen. Previous studies have suggested that the environment and host-specific factors together determine host microbiome assembly 40 , 41 , 42 , 43 , 44 . As perithecia from different field sites showed a similar core bacterial composition, it is evident that Fg has a strong shaping effect on its microbiota. Chemical-guided communication between microbes was recently recognized as an important factor shaping communities 12 , 39 , 45 , 46 . Previous studies indicated that the Fg genome can harbour up to 67 gene clusters involved in secondary metabolism, and some of the biosynthesized compounds showed antimicrobial activity 47 , 48 , 49 . Therefore, we propose that secondary metabolic compounds of Fg perithecia play an important role in shaping the associated microbiome. Recently, distinct fungal compounds have been reported to have a strong effect on the associated bacterial communities 50 , supporting our hypothesis. Our results indicated that a fungal pathogen’s microbiome naturally harbours antagonists in addition to mutualistic colonizers as observed before 29 . Although individual bacteria can have negative impacts on the host fungus, these bacteria probably do not prevail under natural conditions. However, when biotechnologically enriched and re-applied, they may be applicable for plant protection. In this study, P. agglomerans ZJU23 was isolated from Fg perithecia and displayed the highest antagonistic activity towards mycelial growth of Fg. ZJU23 effectively suppressed Fg perithecium formation, mycotoxin biosynthesis and development of Fusarium head blight at high concentrations, while naturally associated with perithecia at lower concentrations (Fig. 2a ). Antagonism within members of the microbiota was described before, but our general understanding of the molecular mechanisms involved in such microbial interaction is still limited 51 , 52 , 53 . Deciphering yet unknown microbial interaction mechanisms will provide the basis for the engineering of beneficial microbiomes that protect host plants and animals 3 , 9 , 54 . A detailed investigation of the antagonistic effects between P. agglomerans ZJU23 and Fg showed that HA is the involved bioactive compound. Although HA was identified around 40 years ago 16 , 17 , 18 , the biosynthetic gene cluster of HA and its mode of action (MOA) towards fungi were hitherto unknown. By combining various approaches, we showed that the AcbA–AcbJ gene cluster is responsible for HA biosynthesis in ZJU23 and unexpectedly discovered that ZJU23 most probably attained the ability to produce HA via horizontal gene transfer. HA consists of a peptide ring with eight amino acids to which a fatty acid chain, a dehydrobutyric acid and a sugar moiety are attached 55 . The MOA of cyclic lipopeptides against fungi are mostly unknown 56 . For example, the MOA of various well-known surfactins and fengycins produced by Bacillus species are still not completely understood; they may interact with phosphatidylethanolamine or the cell wall, by inhibiting DNA synthesis or by interrupting the integrity of mitochondria 57 , 58 , 59 , 60 . Another MOA was observed for Iturin A and syringomycin, which inhibits fungi via a pore-dependent mechanism 61 . Echinocandins secreted by Aspergillus spp. inhibit the synthesis of β-(1,3)-glucan and thus damage the fungal cell wall 62 . Here we found that HA interacts with ergosterol. Furthermore, we established that HA quickly disrupts the integrity of artificial membranes containing ergosterol, but not those containing cholesterol. Complementary microscopy observation showed that HA treatment disrupted lipid rafts, increased membrane permeability, resulted in the formation of abnormal cell membranes and ultimately caused cell death. Mutations in fungal ergosterol biosynthetic pathways resulted in resistance towards HA in S. cerevisiae and Fg. Therefore, our data suggest that the MOA of HA is mainly through disrupting lipid rafts in cell membranes by directly interacting with ergosterol (Extended Data Fig. 10 ). It should be noted that the lipid tail itself is not inhibitory, since various tested compounds with differing lipid tails did not exert antifungal activity (Supplementary Fig. 9 ). In addition, the antifungal activity of HA and ZJU23 was eliminated when the amino acid ring structure in HA was opened by disrupting the ACBH gene. This suggests that the ring structure in HA is essential for its activity. In the present study, we also characterized the Fg perithecium-associated microbiome presumably for the first time in a large-scale approach. This approach allowed us to understand the role of commensal bacteria in the suppression of fungal disease. We determined that the biocontrol agent ZJU23 secretes HA to inhibit Fg by interacting with ergosterol-containing cell membranes. HA has a broad antifungal spectrum and control efficiency towards phytopathogenic fungi. During the past few years, antimicrobial peptides have emerged as potential candidates for developing new antifungal therapies because they are often characterized by negligible host toxicity and low resistance rates 56 . HA inhibited the growth of the human opportunistic pathogen Candida albicans more efficiently than the clinical fungicides amphotericin B and fluconazole. In addition, HA has no visible phytotoxicity towards plant tissues (Supplementary Fig. 10 ). An acute oral toxicity test of pure HA indicated low toxicity in mice (Supplementary Data 8 ). Although HA is a potential candidate for agricultural and clinical applications, further experimentation is required to develop viable application modes of either the pure compound or microbes harbouring the respective biosynthetic cluster. Methods Community profiling based on 16S rRNA gene sequencing The DNA for each sample was extracted with the FastDNA SPIN kit (MP Biomedicals) according to the manufacturer’s instructions. The V5–V7 region of the bacterial 16S rRNA gene was amplified by degenerate PCR primers 799F and 1193R. Two consecutive PCR reactions were carried out according to previous reports 63 . The barcoded amplicons were sequenced on an Illumina NovaSeq6000 platform with 2 × 250 bp reads. The 16S rRNA gene sequences were analysed with EasyAmplicon v1.09 64 , which includes QIIME v1.9.1 65 , QIIME2 v2020.11 66 and USEARCH v11.0.667 67 . All sample metadata are included in Supplementary Data 1 . Sequencing library splitting was conducted with the split_libraries_fastq.py command in QIIME. Then, the paired-end Illumina reads were processed with USEARCH as previously described 63 . Sequences were clustered into OTUs at 97% similarity and the representative sequences were picked with the UPARSE algorithm in USEARCH (-cluster_otus command) 68 . All OTUs were aligned to the SILVA 132 69 database to remove chimera sequences with the UCHIME algorithm 70 in USEARCH (-uchime2_ref command). The taxonomy of the representative sequences was classified with the ‘RDP trainset 16’ database 71 on the basis of the sintax algorithm in USEARCH (-sintax command). OTUs assigned to chloroplasts and mitochondria were removed as host contamination by using the respective script in EasyAmplicon. An OTU table was generated within USEARCH (-otutab command). For alpha and beta diversity, samples were first rarefied at minimal sequences 5,933 by USEARCH (-otutab_norm command). Subsequent diversity analyses were carried out using EasyAmplicon and QIIME2. The core microbiome of perithecia was defined as that containing OTUs with a relative abundance greater than 0.1% in the perithecia and associated stubble microbiome. Cladograms were visualized by GraPhlAn v.0.9.7. Figures were visualized by using the ggplot2 v.2.3.2 package in R v.4.0.3 and ImageGP 72 ( ). Clustering trees were visualized by ggtree v2.2.4 in R. The Bayesian source tracking model was used for identifying the source of perithecium-associated microbiota with Sourcetracker 73 . All related tables are included in Supplementary Data 2 . For the reproducibility of analyses with custom scripts, all information is available at . Samples of each site were used for sequencing, with five technical repeats. Bacterial isolation, identification and assessment of antifungal activity Perithecia were vortexed in phosphate-buffered saline (PBS) buffer (E607016, Sangon Biotech) for 5 min and then ground with sterile mortar and pestle. Homogenized perithecium samples were then allowed to settle for 15 min and the supernatants were serially diluted, the resulting supernatants were plated and cultivated on 9 cm Petri dishes in 1:10 (v/v) tryptic soy medium, Reasoner’s 2A agar medium and Luria-Bertani (LB) medium (details in Supplementary Methods ) for 3 d at 30 °C 30 . Antagonistic activity of bacterial isolates towards Fg PH-1 was assessed with a dual-culture assay in Waksman’s agar 74 . All isolates were tested in triplicates, and their inhibition zones were measured after 3 d of dual-culture cultivation at 25 °C. The genome of ZJU23 was sequenced using the PacBio RSII sequencing platform. ZJU23 and the 13 other representatives of Pantoea spp. with complete genome sequences were selected to construct a phylogenetic tree with ten housekeeping genes 16Sr DNA, aroE , dnaA , guaA , gyrB , mutL , ppsA , pyrC , recA and rpoB using the neighbour-joining method with Molecular Evolutionary Genetics Analysis version 4.0. Production, purification and characterization of HA To identify antifungal compounds, ZJU23 or its mutants were grown in King's B medium for 3 d at 30 °C, before bacterial cells were removed by centrifugation. The fermentation medium was extracted with Diaion HP-20 resin (13605, Sigma-Aldrich) and then eluted with methanol to obtain crude extracts with an evaporator. The crude sample was dissolved in a methanol solution and then subjected to purification with preparative high performance liquid chromatography (HPLC; C18 column, 5 m, 100.0 mm × 21.2 mm; 10 ml min −1 , 20–100% MeOH/H 2 O in 30 min, followed by 100% MeOH in 10 min) to produce sub-fractions. The unique peaks in the ZJU23 sub-fraction showing antifungal activity were subjected to a semipreparative HPLC (C18 column, 5 m; 250.0 mm × 10.0 mm; 3 ml min −1 ; with 0.1% formic acid in 54% MeOH/H 2 O in 53 min, 65% MeOH/H 2 O in 53 min–70 min) to obtain pure compound. NMR spectra were recorded in methanol-d4 on Bruker AVANCE II 400 MHz; high-resolution mass spectra were obtained on an Agilent 6530 Accurate-Mass Q-TOF LC/MS coupled to an Agilent 1260 HPLC. Isolation of HA-resistant yeast mutants and whole-genome sequencing To select for HA resistance, overnight cultures of a wild-type S. cerevisiae (BY4741) strain were grown in YEPD at 30 °C. Cells were plated on YEPD containing HA (6 μg ml −1 ). Experiments were performed in 10 biological replicates; plates were incubated at 30 °C for 5 d. The whole genomes of mutants and the wild-type strain were sequenced using an Illumina NovaSeq PE150 instrument at Beijing Novogene Bioinformatics Technology. Single nucleotide polymorphism, insertions and deletions (indel), as well as structural variations, were searched with genomic alignments. Construction of transposon mutants and gene deletion mutants in ZJU23 Transposon mutagenesis was conducted to screen mutants without antagonistic activity 75 . The donor Escherichia coli SM10 λpir carrying plasmid pSC123, and the recipient P. agglomerans ZJU23 which was induced to be resistant to rifampicin, were grown in LB medium supplemented with kanamycin and rifampicin overnight, respectively. The supernatant was removed by centrifugation at 5,000 × g , and the pellets were washed with fresh LB medium twice and resuspended in 100 μl medium. The strains were combined and spotted in 100 μl volumes on 0.45 µm millipore filter (F513132, Sangon Biotech) on antibiotic-free LB-agar and incubated overnight at 30 °C. Then the filters were removed, the cells were vortexed in 1 ml LB, diluted and plated onto LB plates containing 50 μg ml −1 kanamycin and 50 μg ml −1 rifampicin. The Petri dishes were incubated at 30 °C for 3 d or until visible colonies developed. The whole genomes of the transformants were sequenced and analysed by Beijing Novogene Bioinformatics Technology. For generation of gene deletion mutants of ZJU23, a PCR-based, λ phage recombinase method was used for targeted deletions. The coding region of the target genes was replaced with the selection marker, conferring resistance to kanamycin. All transformations, selections and confirmations of mutants were conducted using the λ phage recombinase as previously described 76 ( Supplementary Methods ). Primers used for gene deletions are listed in Supplementary Data 9 . HA molecule retention and lipid-binding assay The HA molecule retention assay was performed as described in a previous study 36 . To perform the membrane retention assay, protoplasts of Fg were incubated with homoserine lactones (HSLs) at 30 °C. Then, the cells and supernatants were collected separately by centrifugation (194 × g for 5 min). The protoplasts were treated with buffer 1 (20 mM Tris-HCl, 2 mM EDTA, 1 mM dithiothreitol and 10% glycerol) on ice to break the membrane. Cytosol and membrane fragments were separated by ultracentrifugation (179,000 × g for 24 min at 4 °C). The membrane fragments were dissolved in buffer 2 (buffer 1 with 1% Triton X-100). HA in the fractions was dried by rotary evaporation and resolved with methanol in equivalent volumes and subsequently analysed using the Agilent 6460 LC system (Agilent Technologies). The relative amount of HA was calculated from three technical replicates. The experiment was repeated three times with similar results. The binding ability of HA to various lipids was analysed using the lipid-binding assay 36 . One mg per ml of each lipid species in chloroform was coated on round glass slides by evaporation. Then, the glass slides were incubated with 10 μM HA in dimethyl sulfoxide (DMSO) with PBS at 37 °C. After the incubation, the glass slides were carefully washed with PBS 4–5 times and the lipids were dissolved in DMSO. The amount of HA in each sample was measured by LC–MS. eSM (860061), DOPC (850375), DPPC (850355) (Avanti Polar Lipids) and ergosterol (45480, Sigma-Aldrich) were used. The relative amount of HA was calculated from five technical replicates. The experiment was repeated three times with similar results. Microscopy imaging Green fluorescence signalling of Tri1-GFP-labelled Fg and red fluorescence signalling of FM4-64 were visualized using a Zeiss LSM780 confocal microscope (Carl Zeiss). The survival assay of fungi was performed as previously described with modifications 77 . In brief, after inoculation of F. graminearum _StoA-GFP for 48 h in YEPD liquid medium, various concentrations of HA were added and the mixture was incubated for up to 2 h. Subsequently, 0.2 µl of Live/Dead solution from the Fixable Red Dead cell stain kit (Thermo Fisher) was added to 100 μl of YEPD liquid medium with the F. graminearum _StoA-GFP, followed by incubation for 5 min on a rotator at r.t. The stained mycelia were placed onto microscopic slides and analysed by fluorescence microscopy using the 488 nm and 561 nm lasers. Transmission electron microscopy for visualization of fungal cell membranes was employed as previously described 77 using a Hitachi H7650 transmission electron microscope, operated at 80 Kv (details in Supplementary Methods ). The experiment was repeated three times with similar results. Preparation of giant unilamellar vesicles and permeability analyses Chloroform solutions of lipid mixtures (DOPC/eSM/ergosterol, 2/2/1) doped with 0.5% lipid-conjugated fluorescent dye ( V22887 , Invitrogen) were placed onto a clean indium tin oxide (ITO)-coated glass surface within an area delimited by a rectangular silicone gasket, and the solvent was then removed under vacuum. Sucrose solution (0.1 mol l −1 ) was carefully added to the chamber at the edge of the gasket. The filled chamber was then sealed with another ITO-coated glass and then transferred into an oven at a temperature that was 80 °C above the highest melting temperature of any of the lipids present. At the same time, a 1.4 Vp–p (peak to peak) and 10 Hz sine-wave voltage was applied on the ITO-coated glasses for 60–90 min as previously described 78 , 79 . For lipid domain rearrangement experiments, an analogue of DiIC18 (1,1'-dioctadecyl-3,3,3',3'-tetramethylindocarbocyanine perchlorate) with extended conjugation, DiD-C18, was excluded from the lipid order phase (Lo) into the lipid disorder phase (Ld) for GUVs 37 , which were used to characterize the Ld phase lipid domain. The rearrangements before and after the 10 μM HA treatment for 20 min were recorded under a total internal reflection fluorescence microscope (FV1200, Olympus). The experiment was repeated three times with similar results. AFM A chloroform solution of the lipids (DOPC/ergosterol/eSM, 3/1/6) was mixed in glass vials and dried under vacuum to remove the solvent. After resuspending the dried lipid in 10 mM HEPES buffer (including 150 mM NaCl), tip-sonication was employed to generate a clear solution, enabling the creation of nanosized tiny unilamellar vesicles. In a fluid cell, the vesicle solution (150 μl) was introduced to cleaved mica attached to a glass substrate. The vesicle sample was incubated at 37 °C for 2 min and then at 65 °C for 15 min, followed by rinsing with HEPES buffer (with 150 mM NaCl) at 65 °C to remove excess vesicles. Images were acquired with AFM at room temperature on a JPK Nanowizard 3 scanning microscope in the contact mode using Si 3 N 4 tips (HYDRA2R-100NG-50, APPNANO), with a spring constant of ∼ 0.011 N m −1 , a resonance frequency of 21 kHz in air and a tip radius of <10 nm. AFM scanning was conducted in HEPES buffer with a scanning size of 10 × 10 μm, 256 × 256 pixels and a scanning rate of 1 Hz per line 36 . After obtaining the AFM image with the untreated sample, 10 µmol HA was added into the fluid cell and the sample was statically incubated for 20 min, followed by repetition of the aforementioned protocol. Statistical analyses and reproducibility For determining statistical significance, the P value was calculated using permutational multivariate analysis of variance (PERMANOVA) with ADONIS test, two-sided Student’s t -test or one-way analysis of variance (ANOVA) with Tukey's HSD by SPSS version 24.0 programme. All values are presented as mean ± s.e.m. Experiments were repeated at least three times to confirm reproducibility. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The genome sequence of ZJU23, Pa58 has been deposited in the NCBI BioProject database with accession codes PRJNA707237 and PRJNA795028 . Raw data of amplicon sequencing, genome sequences of herbicolin A-resistant S. cerevisiae mutants and transposon mutants of P. agglomeans ZJU23 are deposited in the Genome Sequence Archive of the Beijing Institute of Genomics (BIG) Data Center with accession numbers CRA003916 , CRA006594 and CRA006602 in bioproject PRJCA003858 ( ). Other data supporting the findings of the present study are available within this article, in Extended Data and the Supplementary Information . Source data are provided with this paper. Code availability Scripts employed in the microbiome analysis are available at .

Fusarium graminearum is a widespread pathogenic fungus that causes Fusarium head blight (FHB) in cereal crops worldwide, especially in wheat. Between 2000 and 2018, more than 4.5 million hectares were annually affected by FHB in China, or around 20% of the total planted area of wheat. This has resulted in annual production losses of more than 3.41 million tons. Researchers led by Prof. Bai Yang from the Institute of Genetics and Developmental Biology (IGDB) of the Chinese Academy of Sciences, Prof. Chen Yun and Prof. Yu Yunlong at Zhejiang University, have recently found that F. graminearum perithecia provided a specific ecological niche for bacteria that could play an important role in disease establishment. Over 2,000 bacterial strains were isolated from the native microbiota of F. graminearum, and subsequently identified and screened for antagonistic activity. The researchers successfully recovered 113 isolates that showed antagonistic activity against the pathogen. One specific bacterial isolate, termed as ZJU23, was identified as Pantoea agglomerans, and had the strongest inhibitory ability against F. graminearum. Subsequently, they found that herbicolin A secreted by ZJU23 was responsible for the observed suppression of F. graminearum. Although herbicolin A was identified about four decades ago, its biosynthetic gene cluster and mode of action against fungi were not known. In this study, the researchers set out to investigate the biosynthetic gene cluster and mode of action against various fungi by combining various approaches, including transposon mutagenesis, liquid chromatography-mass spectrometry, atomic force microscopy and confocal microscopy. By comparing the metabolic profiles of ZJU23 and four deletion mutants of a potential biosynthesis gene cluster, they found that herbicolin A was synthesized by the AcbA-AcbJ cluster. They then proceeded to uncover the mode of action of herbicolin A against various fungi. It is important to note that the modes of action of cyclic lipopeptides against fungi are mostly unknown. It's surprise that herbicolin A was shown to disrupt lipid rafts by interacting with ergosterol, which resulted in the formation of abnormal cell membranes, and, ultimately, caused cell death. In addition, herbicolin A was found to inhibit the growth of Candida albicans and Aspergillus fumigatus, and was more effective than the clinical fungicides Amphotericin B and fluconazole. This provided important evidence for potential applications of herbicolin A beyond agriculture, for example in medicine. Overall, the researchers have deciphered the mechanism of the inhibitory effect of herbicolin A on fungi and identified its biosynthetic gene cluster. This could support future developments for sustainable management of Fusarium head blight worldwide. This work, titled "Fusarium fruiting body microbiome member Pantoea agglomerans inhibits fungal pathogenesis by targeting lipid rafts," was published in Nature Microbiology.

10.1038/s41564-022-01131-x

Physics

Advancing neutron diffraction for accurate structural measurement of light elements at megabar pressures

Bianca Haberl et al, Advancing neutron diffraction for accurate structural measurement of light elements at megabar pressures, Scientific Reports (2023). DOI: 10.1038/s41598-023-31295-3 Journal information: Scientific Reports

https://dx.doi.org/10.1038/s41598-023-31295-3

https://phys.org/news/2023-05-advancing-neutron-diffraction-accurate-elements.html

Abstract Over the last 60 years, the diamond anvil cell (DAC) has emerged as the tool of choice in high pressure science because materials can be studied at megabar pressures using X-ray and spectroscopic probes. In contrast, the pressure range for neutron diffraction has been limited due to low neutron flux even at the strongest sources and the resulting large sample sizes. Here, we introduce a neutron DAC that enables break-out of the previously limited pressure range. Key elements are ball-bearing guides for improved mechanical stability, gem-quality synthetic diamonds with novel anvil support and improved in-seat collimation. We demonstrate a pressure record of 1.15 Mbar and crystallographic analysis at 1 Mbar on the example of nickel. Additionally, insights into the phase behavior of graphite to 0.5 Mbar are described. These technical and analytical developments will further allow structural studies on low-Z materials that are difficult to characterize by X-rays. It has been over 40 years since the ‘sound barrier’ of achieving the pressure of one megabar (= 100 GPa) was broken in a diamond anvil cell 1 and the field of high pressure research has dramatically advanced since. Conditions of the Earth’s core and lower mantle can now be simulated and many geophysical questions have been addressed 2 . Similarly, physical understanding of phase diagrams has immensely advanced and, for example, several new high pressure phases were identified in the ‘simplest’ of all materials, hydrogen (see recent review 3 ). Furthermore, new materials can now be synthesized through high pressure (and high temperature conditions) such as various nitrides 4 , 5 , 6 and, as of particular recent interest, superconducting superhydrides, for example 7 , 8 , 9 . Finally, the field continues to be highly active and multi-megabar pressures were recently achieved with sophisticated toroidal-shaped diamond anvils or double-stage techniques 10 , 11 , 12 . In common to many of these studies is the fact that in situ structure determination under pressure is performed through X-ray diffraction. While such in situ X-ray diffraction is very powerful, there are severe limitations when it comes to low-Z elements. Here, neutron diffraction has evolved as an important tool. Not only are neutrons sensitive to many low-Z elements, neutrons are also capable of distinguishing between different isotopes. As neutrons carry a magnetic moment, they also allow for the detection of magnetic Bragg diffraction. Thus, a number of very important questions in high pressure science can only be addressed by neutron diffraction. For example, for geophysics neutron diffraction can investigate the nature of water within minerals or can provide knowledge of the density and structure of the ices of water, methane, and other light compounds. Neutron diffraction is key for the understanding of phase diagrams of light elements such as hydrogen or carbon. In view of the recently discovered metal superhydrides, neutron diffraction can reveal the exact position of hydrogen in the metal matrix, thus providing important structural information. However, unlike X-ray diffraction, the relatively much lower neutron flux of existing neutron facilities required relatively large sample volumes thus limiting the pressure to a few tens of GPa. Until recently, typical maximum pressures at most user facilities were limited to \(\sim \) 25 GPa when using Paris-Edinburgh cells 13 , although recently breakout studies up to 40 GPa have been reported 14 , 15 . To push for higher pressures, several iterations of neutron diamond anvil cells (DAC) have been developed over time. A major body of work commenced at the Kurchatov Institute in Moscow and was later transferred and further improved in France. There, studies up to 40 GPa on materials such as hydrogen 16 or on magnetic materials have been performed 17 . These studies however, were only able to identify a very small number of reflections, not sufficient for crystallographic analysis and full structure information. Consequently, several efforts world-wide have attempted to further the development of neutron DACs. These efforts focused on high-quality data obtained through single crystal diffraction at the Institute-Laue-Langevin 18 , 19 or the Forschungs-Neutronenquelle Heinz Maier–Leibnitz 20 as well as on higher pressure capabilities using powder diffraction at the Japan Proton Accelerator Research Complex (J-PARC) 21 , the Frank Laboratory of Neutron Physics 22 , as well as at Oak Ridge National Laboratory’s (ORNL) Spallation Neutron Source (SNS) 23 , 24 . Reaching megabar pressures in a neutron diffraction experiment is not simply done by up-scaling conventional DACs as used for X-ray diffraction. Typically, for reaching one megabar in a DAC, culet sizes of not more than \(\sim \) 100 µm are used. For neutron diffraction, where the sample sizes must be much larger, the minimum culet size has been in the millimeter regime. This larger area increases the force requirement more than 10-fold to several metric tons usually supplied by a gas membrane or hydraulic press. The main problem is that such high forces induce significant shear stresses in the diamond cell body and thus in the diamond anvils and also exceeding the strength of the anvil seats resulting in frequent diamond anvil failure. Nonetheless, at the SNS efforts have been underway for over 10 years to develop megabar neutron diffraction. These efforts have centered on the high pressure diffractometer Spallation Neutrons and Pressure (SNAP) 25 . SNAP is a high-flux, medium-resolution instrument equipped with advanced neutron focusing optics and large area detectors that uses white-beam time-of-flight (TOF) neutron diffraction. This combination of white-beam TOF diffraction and large angular coverage presents advantages for high pressure studies, particularly increased Q -coverage, an ability to evaluate texture as well as an ability to handle blocked or heavily attenuated sections of the 2θ-range (since a significant portion of the full diffraction pattern can be measured at each 2θ due to the energy-dispersive nature of white-beam TOF diffraction). This allows for the use of absorbing materials in the pressure cell set-up, specifically gasket materials, without compromising the Q -range, a fact that has highly benefited DAC studies on SNAP 26 , 27 . Over time, several iterations of DACs have been developed on SNAP. Initially, these have used highly elaborate anvil support 23 , which successfully achieved over 90 GPa 23 and obtained high quality crystallographic data on water ice up to 52 GPa 28 . However, the available scattering aperture and thus Q -range severely limited the possible studies to simple cubic systems. Even more detrimental, the design was very expensive and not feasible for a user program. The next upgrade thus focused on very large synthetic diamonds grown by chemical vapor deposition (CVD) in order to reduce stresses in the anvil supports 24 . These enabled full crystallographic analysis and structural refinement to 62 GPa 29 on water ice and were successfully rolled out for a wider user program. More recently, DAC efforts at SNAP have been attempting to take advantage of two key factors, an increased flux of the overall facility (that now operates at 1.4 MW) and significant upgrades to the beamline (an upgraded detector system and a vastly improved neutron focusing guide system). Despite these upgrades however, the maximum pressure in the DACs remained at 65 GPa, a barrier that simply could not be overcome with these large CVD anvils and these cell bodies. Here, we present the development of novel neutron DACs based on smaller gem-quality synthetic diamond anvils, ultra-high precision cell bodies and new anvil supports as well as high-precision in-built boron collimation for reduced backgrounds. These new DACs take advantage of the recent transformative beamline upgrades, since higher flux allows for smaller sample volumes and thus new concepts for DACs. These new designs now enable high quality data obtained from DAC equipped with culet sizes of 700 and 800 µm to one megabar and above. Such neutron diffraction at megabar pressures can be expected to enable new science directions as well as answer pressing questions in geosciences, physics, chemistry and material science. Results Diamond cell designs The diamond anvil cell for megabar neutron diffraction is an opposed-anvil piston-cylinder device with 80 degree aperture tailored for SNS’s dedicated high pressure neutron diffractometer SNAP 25 , 26 . The neutron beam enters along the pressure axis and scatters out radially from this pressure axis. Very high mechanical stability and accuracy in the cell body was achieved by using steel roller bearings between piston and cylinder shown in Fig. 1 a. These roller bearings have essentially eliminated shear stresses that were always present in previous pressure cells, which were most likely the main cause for premature anvil failure. Further, these linear ball bearings have the additional advantage of nearly friction-less operation at low temperatures, eliminating the well-known ‘freezing’ of piston and cylinder in conventional cells. New diamond anvil geometries were developed to accommodate optimal radial and axial support of the conical CVD diamond anvils. The anvils have dimensions of 5 mm in diameter and 4 mm height. The pavilion angle is 40 \(^{\circ }\) and the bottom cone angle is 30 \(^{\circ }\) . The design accommodates for both axial and radial support of the anvil. Whereas conical support alone caused anvil failure due to plastic deformation or cracking of the seat (tungsten carbide or polycrystalline diamond) 23 , 24 , simple axial/flat support lead to splitting of the large diamonds along the axis due to elastic or plastic deformation of the seat material (steel or tungsten carbide). The combination of radial steel binding rings and high strength flat supports (tungsten carbide or steel) has significantly improved the stability of the anvils allowing routine experiments at over 50 GPa. Indeed, anvils with dimensions 5 \(\times \) 4 mm could withstand forces up to 8 metric tons, or about 80 kN, without failure. Figure 1 b,c show anvil and support combining axial and radial support. Culet sizes range from typically 600–900 µm allowing for sample volumes between 0.004 and 0.014 mm \(^{3}\) . Figure 1 Megabar neutron diamond anvil cell. ( a ) A photo of the individual parts of the cell, roller bearings, top piston including seat and anvil, cell body including opposing seat and anvil, and top closing cap. ( b ) A photo of the seat and anvil including the steel binding ring used for additional support. ( c ) A 3D-schematic created in Autodesk Fusion 360 (V2.0.15050, ) of seat and anvil including the collimator built into the seat consisting of amorphous boron with a small amount of epoxy. A close-up photo of the seat in top-view shows the aperture for incoming beam formed by the collimator. The diameter of the hole in the seat is thereby 1 mm diameter. Full size image The seats of the anvils were made from tungsten carbide with an aperture of 23 \(^{\circ }\) (see Fig. 1 ). Since the incoming neutron beam is highly divergent and larger than the sample, the upstream seat requires an in-built collimator that allows for the absorption of all neutrons not illuminating the sample. This collimator is made from a mixture of boron powder (amorphous natural B or isotopically enriched crystalline \(^{10}\) B) and a small amount of epoxy. It is custom drilled for each experiment and allows for low background levels routinely adapted to a given culet size. Following its fabrication, the diamond anvil is optically aligned to the collimator. Typically, dimensions of culet, gasket hole and collimator are 700 µm, 300 µm and 280 µm, respectively. The alignment and effectiveness of the collimators are checked by monitoring diffraction peaks from the rhenium gasket described below. The initially used stainless steel gasket (301, full hard) 23 , 24 limited pressures to about 65 GPa. While Re, and to a lesser extent also W, have a high absorption cross-section for neutrons, this is not an issue for operation on SNS’s SNAP diffractometer. There, the large angular range available through the movable detectors allows for detector placements such that only small slivers of \(2\theta \) -range are attenuated by the gasket. Due to the use of TOF neutron diffraction rather than a monochromatic beam, this still allows for access to the full desired Q -range. The megabar neutron DACs described here thus allow for the necessary high precision in alignment and operation, for the necessary stability in radial anvil support and gasket use as well as for sufficient sample volume as needed for high quality neutron diffraction data. Pressure-load curves A range of samples at various sample sizes and a maximum pressures have been run on the SNAP beamline using these anvil and cell designs. Here we show three different sample materials that cover a range of sample types: The first sample is solid nickel powder loaded without pressure medium, the second sample is water ice loaded as a liquid and the third sample is graphite loaded with gaseous argon used as pressure transmitting medium. In situ diffraction data were collected upon pressurization using a hydraulic DAC press 27 . The resulting pressure-load curves are shown in Fig. 2 . Figure 2 Pressure-load curves obtained from nickel, water ice and graphite. The pressure for the Ni and D \(_2\) O data were obtained from the diffraction data themselves using Ni and D \(_2\) O equations of state 30 , 31 , respectively. For those, pressures were determined through Gaussian fits to the diffraction data (black circles and stars, respectively). Additionally, for short runs, pressures were determined visually from the peak positions (grey circles) in one case. For the graphite loaded in Ar pressure was obtained using the ruby fluorescence method (black diamonds). Refer to the main text and the methods section for further details. In all cases, a fit was added to guide the eye. For Ni and graphite 2nd order polynomial fits were used, while two separate linear fits were to used to comply with a distinct change in slope at \(\sim \) 60 GPa for ice. Full size image For the Ni powder, two different runs using 800 µm and 600 µm culets are shown. Pressure was determined from the accumulated data over the entire collection period and both banks. Following data reduction and background correction, the 111, 220 and 200 Ni reflections were separately fitted with a Gaussian each to extract the lattice parameter a , which was then used for pressure calculation. This was performed for all pressure points obtained with 800 µm culets and all long-exposure pressure points obtained with 600 µm culets. Additionally, pressure was determined visually from the data independently for the 111, 220 and 200 reflections for all data sets. Pressure was then calculated independently for each reflection and the pressure points shown on the pressure-load curve present the average while the uncertainty provided is the corresponding standard deviation. The actual variation due the positional/angular variations and or also pressure gradients may thus be higher. Further details are provided in the supplemental material . For both Ni loadings a rapid, relatively linear pressure increase with load is seen initially, while the curve begins to flatten at higher loads indicating cupping of the anvil culets. The maximum pressure obtained with the 800 µm was 1 Mbar at 5 metric tons. Upon further load increase, the anvils collapsed at 5.13 metric tons. Specifically, \(V/V_0\) = 0.7565 ± 0.0008 was measured experimentally. Using a Vinet EoS with \(B_0\) = 183 GPa and \(B'_0\) = 4.99 30 , this yields a pressure of 100 GPa. Note that all the Ni pressures presented in figures here are based on this EoS. However, it is noteworthy that a more recent EoS 32 using a 3rd-order Birch Murnaghan with \(B_0\) = 201 GPa, \(B_0'\) = 4.4 yields a pressure of 104 GPa for this \(V/V_0\) . The cell with 600 µm culets achieved 1.15 Mbar at 5 metric tons. Specifically, \(V/V_0\) = 0.738 ± 0.003 was measured experimentally. This equates to a pressure of 115 GPa using the Vinet EoS and of 120 GPa using the 3rd-order Birch–Murnaghan EoS quoted above. It should be noted that this cell failed after \(\sim \) 1 h data collection at this load and pressure. Next, a pressure-load curve obtained from water ice (here D \(_2\) O to enhance coherent neutron scattering) is shown. Unlike the solid Ni powder, this was a liquid sample loading. This presents additional challenges in terms of loading and gasket stability upon compression. As for the Ni, pressure was determined from the accumulated diffraction data, sampled over the entire exposure time and both detector areas. A Gaussian fit to the 110 reflection was used to extract the lattice parameter. To calculate pressure, we used \(V_0\) = 12.3 cm \(^3\) mol \(^{-1}\) , \(B_0\) = 23.7 GPa and \(B'_0\) = 4.15 with a Birch–Murnaghan finite-strain EoS 31 . For this particular run on ice, the pressure was increased relatively rapidly up to 65 GPa, the pressure limit in most previous runs. Up to this pressure, short collection periods of 20 min were performed, which were sufficient to determine pressures from the main diffraction peaks and an equation of state. Above 65 GPa, longer data collections were performed. This yielded a maximum pressure of 88.6 GPa for the 640 µm culet used here. An intriguing kink is seen in the pressure-load curve at \(\sim \) 60 GPa, i.e. in the pressure regime where previously catastrophic gasket/anvil-failures consistently occurred. Above and below the kink at \(\sim \) 60 GPa, the pressure-load curve is linear, although at two different slopes. A further, more detailed offline experiment with H \(_2\) O loaded also in a neutron DAC but using ruby as a pressure calibrant, revealed no discontinuous slope change in the pressure load curve. Interestingly however, it did show that the pressure gradient across the sample chamber remained constant at 3–4 GPa upon pressure increase. A manuscript on the details of these findings is in preparation. This gives useful insight into overall pressure gradients. While no significant broadening of diffraction lines was observed upon pressure increase, it should be noted that SNAP is a medium-resolution diffractomter only and thus broadening up to a certain point, may not be detected. Overall, this pressure gradient observed for H \(_2\) O is similar in magnitude to the variation/uncertainty in pressures determined for Ni, i.e. the variation due to different angular positioning ( \(\sim \) 6.5 GPa at 100 GPa, see supplemental ). As in earlier X-ray diffraction studies, such pressure variation/gradients may in future be overcome using hydrostatic pressure media in combination with laser annealing. The final pressure-load curve shown here is obtained from graphite measured (quasi-)hydrostatically with an Ar pressure transmitting medium loaded into the DAC at 2.6 kbar using ORNL’s gas loader 26 . Unlike the Ni and the ice examples, here the sample filled only a fraction the entire chamber to achieve (quasi-)hydrostatic conditions. This experiment demonstrates the capability of obtaining good diffraction patterns on even smaller samples, which is important for future experiments using for example metallic samples loaded in hydrogen. Here, the pressure was obtained using the ruby fluorescence method. A hydrostatic ruby scale 30 was used up to \(\sim \) 20 GPa, at higher pressures the non-hydrostatic ruby scale 33 was used instead. Pressure was measured before and after each data collection whereby pressure drifted upward by 1–3 GPa during long collections. All these measurements are included here. The graphite was loaded to 2.5 metric tons yielding a maximum pressure just above 50 GPa. At this pressure, the gasket started to deform asymmetrically and the cell was decompressed. The anvils were successfully recovered. This is noteworthy as compressed graphite has a tendency to break anvils. Data quality Despite the, in terms of neutron scattering, very small sample volumes, the data quality is sufficient for crystallographic analysis using conventional Rietveld approaches. While achieving megabar pressures for neutron diffraction is a critical breakthrough, it is of equally high importance that not only the positions of the Bragg peaks are determined accurately but also their intensities. This is critical pre-requisite to exploit the ability of neutron diffraction measurements to determine stoichiometry and atomic positions of light atoms such as in D \(_2\) O/H \(_2\) O ices, and metal superhydrides. An illustration of the data quality that can be achieved is demonstrated in Fig. 3 . Figure 3 High pressure neutron diffraction of nickel. Neutron diffraction data obtained from Ni in a cell with ( a ) 800 µm culets and ( b ) 600 µm culets. ( c ) Quantitative Rietveld refinement performed on Ni measured at 100 GPa in the 800 µm culet cell after full background correction. The data set is presented by the measured data (grey), the calculated data (colored) and the difference function beneath. In all plots, Ni is additionally indexed as per Ni’s cubic unit cell, \(Fm\overline{3} m\) . Full size image Ni is chosen here as example element due to its high-symmetry cubic structure, its high coherent neutron scattering cross-section and the fact that it does not exhibit structure phase transitions in the pressure regime probed. The latter allows for reliable pressure determination from measured diffraction data as demonstrated above and benchmarks the accuracy of the data-reduction approach. Fig. 3 a,b show the Ni diffraction data collected in DACs with 800 µm and 600 µm culets, respectively. The data were vanadium-normalized, to account for instrumental artefacts such as the incident flux wavelength dependence and detector efficiencies, (see details in supplemental material ) and have had experimental backgrounds subtracted. Additionally, the data set shown in Fig. 3 c have been corrected for attenuation of the incident beam by Bragg scattering of the upstream diamond attenuation 34 . TOF diffraction is an energy-dispersive technique, whereby the transit time of detected neutrons can be used to resolve their energies. As a consequence, a full diffraction pattern can be extracted from neutrons scattered into even a small solid angle. When combined with large area detectors, this facilitates powerful insight into the sample microstructure, as observations at different angles sample the subsets of differently-oriented crystals relative to the beam (and, in our case, the load axis). Here, however, we have chosen to combine pixels from a wide angular range (spanning 61 \(^{\circ }\) to 126 \(^{\circ }\) ), as the increased total counts improve the statistics of our measurement. There are some considerations that must be made when combining data measured at different angles, which we discuss in the supplementary material . In order to conduct the refinement shown, both the time-of-flight to d -spacing conversion and diffractometer resolution were determined from measurements of a calibrant powder of diamond collected with an identical angular range to the Ni data. This was done using standard calibration processes, developed for the GSAS-II 35 Rietveld software package. The crystal structure of Ni has no refineable atomic positions, so the only crystallographic parameters refined were the lattice parameter and a single isotropic atomic-displacement parameter (ADP) for the Ni nucleus, \(U_{iso}(Ni)\) . The resulting fit, yielding a reduced- \(\chi ^2=1.169\) . The refined lattice parameter was \(a = 3.2123(6)\) Å and \(U_{iso}(Ni)=0.0109 (27)\) Å \(^2\) . Significant further improvement was found by adding a single-coefficient spherical-harmonic model of preferred orientation (PO) to the refinement. This dropped the reduced- \(\chi ^2\) to 0.992 and visibly improved the fit, particularly in the vicinity of the 200 reflection and resulted in a texture index of 1.067. The lattice parameter was unchanged within error in the PO refinement, although the ADP did significantly reduce to \(U_{iso}(Ni)=0.0030(27)\) Å \(^2\) . This resulting fit is the one shown in Fig. 3 c. The good fit between a model of the well-known structure of Ni and our data demonstrates the reliability of our data-reduction procedures. Furthermore, this refinement quite clearly demonstrates that full, quantitative structural neutron refinements is achievable even at megabar pressures. In addition, the lack of sample-peak broadening, beyond the native instrumental resolution, suggests surprisingly low shear gradients, which are potentially being minimized by cupping of the (relatively large) diamond culets under our very high pressure conditions. Also notable is the low amount of PO, despite the absence of a pressure medium, which might be, at least partly, attributed to our ability to average over a large range of crystallite orientations for each Bragg peak we measure. Lastly, this angular range has the further advantage of delivering a good Q -range allowing for the detection of a good number of peaks even for high-symmetry systems such as Ni. Probing the high pressure behavior of graphite Clearly these experiments on Ni demonstrate the achievement of megabar neutron diffraction together with the fact that full crystallographic analysis remains possible at these pressures. Next it is interesting to demonstrate that these new neutron DACs can indeed address challenging questions in high pressure science that are not readily studied by X-ray diffraction. These studies often require gas loading, involve low-Z materials and will typically probe materials with lower symmetry than cubic Ni or D \(_2\) O. Here, we demonstrate this on the example of graphite loaded in a DAC with Ar as soft pressure-transmitting medium. This is chosen for several reasons: (i) Carbon is a low-Z element and hence past X-ray diffraction DAC studies on graphite always have had to contend with high backgrounds arising from Compton-scattering of the diamond anvils. (ii) A study involving gas-loading highlights that data collection is possible even if the gasket chamber is not entirely filled with sample material. Furthermore, while gas-loading involved an inert gas here, the same procedure can be extended to hydrogen, for example. (iii) Graphite exhibits a hexagonal structure. (iv) Clearly the phase diagram of carbon continues to intrigue and many open questions remain. Specifically here, the high pressure behavior of graphite remains of considerable interest as graphite does not transform directly to diamond upon compression. Instead intermediate phases form. A transparent phase forms between 14–18 GPa 36 , 37 , 38 , a formation that is critically dependent on the exact form of graphite used and on the compression conditions. These previous studies however, focused on optical methods, and thus the structure was not determined. While in situ synchrotron X-ray diffraction studies on graphite compressed within a DAC are challenging, some studies have been conducted. These indicate that this new structure is super-hard 39 , may possess the monoclinic structure of M-carbon but only nucleates very slowly over time 40 . The details of the nature of this structure, its dependence on the initial graphite used and the curious critical time-factor in nucleation still pose many open questions that could be addressed by high pressure neutron diffraction. Here, diffraction data from graphite from ambient pressure to \(\sim \) 50 GPa, are shown in Fig. 4 . The initial pressure of 0.7 GPa was sealed in during gas loading. Interesting differences to past high pressure X-ray diffraction data 39 , 40 are evident in this ambient pressure data set. Specifically, in our diffraction data, the 004 reflection is observed although it is typically absent in X-ray diffraction DAC data 39 , 40 . Closer inspection of our data resolved by angle rather than summarized over all detector banks shows that the 004 is not present in the same detector panels as the 002. Full crystallographic analysis is thus pending the development of multi-angle approaches to Rietveld refinement (see supplemental material ). Figure 4 High pressure neutron diffraction of graphite loaded with Ar pressure medium. ( a ) Powder diffraction patterns from ambient pressure close to 16 GPa indexed based on graphite’s hexagonal unit cell, \(P6_3/mmc\) . Finely binned data are shown by (grey) circles and the solid lines show more coarsely binned data for clarity. ( b ) 2D detector view extracted from MantidWorkbench (V6.1.0, ) 41 showing the diffuse single crystal features observed at 50 GPa in a d -range of 1.1–2.1 Å. Cell backgrounds such as single crystal diamond peaks are masked (greyed out areas). The areas integrated for each individual single crystal feature are indicated by red circles. This was performed in the exact same detector location for each pressure from 21–50 GPa. The pressure evolution of the three peaks is shown in ( c – e ). Full size image As shown in Fig. 4 a, upon compression to 10.6 GPa and 15.9 GPa, the typical large pressure-shift of the 002 and small pressure-shift of the majority of other reflections in observed. At the same time the 004 reflection was not readily observed any longer. Note that the pressures quoted are the starting pressure at data collection since pressure did drift up during collection, from an initial 8.7–10.6 GPa and from 14.3 to 15.9 GPa, respectively. Following the next pressure increase to \(\sim \) 21 GPa, structural disordering occurred and the hexagonal graphite structure dissolved over time (not shown here). This may be consistent with past X-ray diffraction studies that see significant peak broadening at similar pressure 39 or, following a wait time of 100+ hours, the appearance of new diffraction lines 40 . Subsequent continued compression to \(\sim \) 34 GPa resulted in the nucleation of a new crystal structure which persisted to the maximum pressure of \(\sim \) 50 GPa probed here. Interestingly, this high pressure crystalline phase did not nucleate as random powder but as highly oriented, diffuse single crystal-like features. This is clearly seen in the 2D view of the two detector banks at the maximum pressure of 50 GPa (see Fig. 4 b). Further pressure data sets and analysis details are available in the supplemental material. Integration of the three areas within the marked circles, i.e. of the diffuse single crystal features at various pressures, is shown in Fig. 4 c,d. The peaks at \(\sim \) 1.24 Å and \(\sim \) 1.70 Å clearly shift with pressure, which confirms they are sample peaks. It is not clear if the peak at \(\sim \) 1.91 Å shifts within the resolution-limit of SNAP, although very stiff behavior may be expected from a highly incompressible carbon phase. The number of peaks observed is insufficient for full structural analysis although the structure may yet be consistent with the previously observed M-carbon 40 . Further studies are needed to obtain a more complete set of diffraction peaks. It is noteworthy that detection of these diffuse single-crystal features was most certainly aided by our use of white-beam TOF neutron diffraction, i.e. an energy-discriminating diffraction technique, coupled with large area detectors and the large angular scattering aperture of our DACs. The fact that an extremely oriented structure forms, may indeed contribute to some of the past difficulties in detecting high-pressure polymorphs of graphite in X-ray DACs. Full analysis of our findings, including multi-angular approaches, and additional experiments are underway and will be presented once completed. Nonetheless, it is clear that our neutron DACs can contribute new understanding to phase diagrams of light elements such as carbon. Discussion Present developments in neutron DACs enable neutron diffraction above 1 Mbar for the first time. Remarkably, even at these high pressures data quality remains of sufficient quality for quantitative structure refinement, while retaining angular-resolved diffraction information on the sample. While previous breakout studies on ice have achieved pressures close to 1 Mbar 14 , 23 , the data quality in these studies was not of sufficient crystallographic quality to enable refinement and thus no scientific insight was gained. Yet, the possibility of refinement is a critical prerequisite for future works that require determination of detailed atomic positions and stoichiometry such as studies on metal hydrides, superhard nitrides, various ices, and many more. This capability for quantitative neutron diffraction at a megabar is developed within the context and framework of a neutron user facility. It is not only available to a few select researchers but is released and available to the wide community. This development thus opens many new possibilities and opportunities for a range of scientific areas and can aid in many pressing research questions that cannot readily be answered by synchrotron X-ray techniques. Furthermore, it is useful to consider these pressure records and the correspondingly reducing sample volumes in the context of past developments for in situ X-ray powder diffraction in a DAC. The first in situ X-ray diffraction was conducted in 1977 42 and 1 Mbar was achieved shortly after in 1978 1 . By the late 80s it was possible to perform multi-megabar experiments 43 , 44 . It took however, until the 2010s to perform experiments at maximum pressures of 6 Mbar 10 , 11 , 12 . In contrast, refinable neutron diffraction was only available up to \(\sim \) 3 GPa until the early 90s. The development of the Paris-Edinburgh (PE) cell for in situ neutron diffraction presented a major breakthrough that allowed for fully refinable data at 10 GPa in 1992 45 and \(\sim \) 25 GPa in 1995 13 . The current record for refined data measured in the PE cell stands at 40 GPa achieved in 2019 14 . Higher pressures yet maintaining refinable data have only been achieved recently in neutron DACs allowing for pressures of 52 GPa in 2013 28 and 62 GPa in 2019 29 . While other neutron DAC studies have also achieved pressures as high as 82 GPa 21 , data were not refined. It is illuminating to compare the sample volume versus maximum pressure for X-ray and neutron powder diffraction as shown in Fig. 5 . As is to be expected, the sample volumes required for high pressure neutron diffraction are considerably larger than those used for X-ray diffraction. For example, the first 25 GPa refinable data for neutron diffraction required close to 20 thousand times the sample volume of that used during the first synchrotron X-ray diffraction at similar pressures. The reduction in sample volume with increasing pressure has however, been significantly steeper for neutron diffraction. For synchrotron X-ray diffraction, the pressure increase from 25 GPa to 1 Mbar required a sample volume that was \(\sim \) 10 times smaller. In contrast, the same pressure increase for neutron diffraction required a sample volume decrease by a factor of 3500. As a result, the sample volume we required for neutron powder diffraction at 1 Mbar is only \(\sim \) 5–6 times the volume that had been used for the very first 1 Mbar synchrotron X-ray diffraction experiment. Figure 5 Sample volumes versus maximum pressure achieved during high pressure neutron and X-ray powder diffraction developments. For X-ray powder diffraction developments only DACs using gaskets are added here, for neutron powder diffraction developments, PE cells and neutron DACs (which are both using gaskets) are given. Lines indicating the volume reduction and increasing pressure records are added to guide the eye. Sample volumes during in situ high pressure powder diffraction development are taken from the literature as indicated by lower case letters: ( a ) first 10 GPa refined neutron powder data in a PE cell obtained in 1992 45 ; ( b ) first PE cell pressure data at 26 GPa obtained in 1995 13 ; ( c ) record 40 GPa pressure in the PE cell achieved in 2019 14 ; ( d , e ) refined data obtained in a neutron DAC at 52 GPa and 62 GPa achieved in 2013 and 2019, respectively 28 , 29 ; ( f ) the first in situ energy-dispersive X-ray powder diffraction performed up to 25 GPa in 1977 42 ; ( g ) the first in situ X-ray diffraction measurements at 1 Mbar in 1978 1 ; ( h , i ) very early 2 Mbar and 3 Mbar in situ X-ray diffraction performed in 1988 43 and 1990 44 ; and ( j , k ) representing the latest 6 Mbar and 6.15 Mbar records achieved in 2018 11 , 12 , respectively. Full size image Clearly, modern synchrotron and X-ray DAC techniques now enable significantly smaller sample volumes to be measured at 1 Mbar than these first experiments. Yet, neutron sources, neutron instrumentation and neutron DACs also continue to be developed. Advances are being made in overall neutron flux and brightness at new facilities such as the European Spallation Source or ORNL’s Second Target Station. Advances are made in instrumentation in terms of focusing guides and background reduction through collimation as well as the neutron DACs themselves. It is thus expected that sample volume requirements for neutron diffraction will decrease and will continue to approach those that are (or at least have previously been) used for synchrotron X-ray diffraction at multi-megabar pressures. The megabar neutron DACs presented here are a major step toward opening neutron diffraction to similar pressures as routinely probed by synchrotron X-ray diffraction. By developing and releasing this capability, it becomes now possible to address science questions that require the complimentary information that can only be provided by neutron diffraction to a wide community. Methods Diamond cell preparation Four different diamond cells were prepared as follows below. Cell 1 and cell 2 were loaded with Ni powder (<150 µm, 99.999% metal basis, Sigma-Aldrich), cell 3 was loaded with liquid D \(_2\) O (99.9 atom % D, Sigma-Aldrich) and cell 4 was loaded with graphite powder (graphite v598). Cell 1 was prepared with tungsten carbide seats and an upstream collimator made from hexagonal boron nitride with a final hole drilled to 400 µm diameter. The cell was equipped with CVD anvils from Washington Diamond (WD Lab Grown Diamonds, Laurel, MD, USA) polished to a 800 µm culet without any bevelling. A Re gasket was indented to 80 µm and drilled to 400 µm diameter. The gasket was balanced on thin Al rings to avoid cupping. No pressure transmitting medium was used. The cell blew out during pressure increase at an applied load of 5.13 tons. Cell 2 was prepared with tungsten carbide seats including an upstream collimator made natural amorphous boron powder mixed with epoxy with a final hole drilled to 250 µm diameter. The cell was equipped with CVD anvils from WD polished to a 600 µm culet with a double-bevel. A W gasket was indented to 70 µm and drilled to 300 µm diameter. No pressure transmitting medium was used. The cell blew out after holding an applied load of 5 tons for \(\sim \) 15 min. Cell 3 was prepared with tungsten carbide seats including an upstream collimator made from enriched boron powder ( \(^{10}\) B) mixed with epoxy with a 2 mm long hole drilled to 250 µm diameter. The cell was equipped with CVD anvils from Almax (Almax easyLab, Diksmuide, Belgium) polished to a 640 µm culet with a 700 µm bevel. A W gasket was indented to 68 µm and drilled to 295 µm diameter. No pressure transmitting medium was used. Cell 4 was prepared with tungsten carbide seats and an upstream collimator made from enriched boron powder ( \(^{10}\) B) mixed with epoxy with a 3 mm long hole drilled to 340 µm diameter. The cell was equipped with 5.5 \(\times \) 4.5 mm CVD from WD polished to a 800 µm culet with a 900 µm bevel. A W gasket was indented to 85 µm and drilled to 370 µm diameter. After loading sample and ruby balls for pressure measurement, the cell was gas loaded with Ar using the ORNL gas loader 26 . Upon an Ar gas pressure of 2.65 kbar, the cell was sealed inside the gas loader yielding a starting pressure of 0.7 GPa as per ruby fluorescence. Neutron diffraction High pressure neutron diffraction was conducted at SNS’s dedicated time-of-flight high pressure diffractometer, the SNAP beamline 25 . Measurements were performed using its neutron focusing guide and a center wavelength of 2.1 Å at 60 Hz chopper speed. This yields a bandwidth of \(\sim \) 0.5–3.65 Å. Detectors were placed at center positions of \(2\theta = 65^{\circ }\) and \(2\theta = 105^{\circ }\) for an optimized Q -coverage. The instrument was calibrated using a calibration standard to give correct lattice parameters, typically isotopic \(^{11}\) B-enriched NIST LaB \(_6\) or standard diamond powder. The diamond cells were placed into a hydraulic press for online compression 27 . Initial alignment of the cell on the beamline occurred via an alignment laser and via optics focused on the sample position for alignment along the beam. Alignment perpendicular to the beam was further refined using the neutron beam itself. Collection time at each pressure was as follows: 2.5 h for Ni in cell 1 (0.01 mm \(^3\) sample volume); 5 h and 0.5–1 h for long and short runs for Ni in cell 2 (0.005 mm \(^3\) sample volume), respectively; 0.33 h and 10–12 h for short and long runs for D \(_2\) O in cell 3 (0.0046 mm \(^3\) sample volume), respectively; and 8–13 h for graphite in cell 4 (0.001 mm \(^3\) starting gasket chamber volume). Data analysis The resulting diffraction data were reduced using the Mantid framework 41 employing the instrument calibration collected. As typical for all DAC data, the single-crystal diamond-anvil peaks were masked on the detector and removed from the powder data. The Ni data were calibrated, normalized and background corrected. A select data set was further subjected to corrections arising from diamond attenuation 34 . In this case, subsequent Rietveld refinement was performed using GSAS-II 35 . For full details on the intricacies of the reduction and analysis procedure as necessary for quantitative crystallographic analysis, refer to the supplemental material. The D \(_2\) O data were reduced using in Mantid the same methods as for our previous D \(_2\) O work 29 , namely calibration, vanadium normalization, absorption and background corrections, including those for diamond attenuation. The graphite data were reduced in Mantid using basic SNAP routines. Normalization and background correction was performed using vanadium pin and empty instrument measurements. Additional backgrounds were removed via a fit to the remaining experimental background in Mantid and subtraction of this fit. The mantid masking tool was used to extract the single-crystal peak intensities at pressures above \(\sim \) 30 GPa. Data availibility All relevant data are available from the corresponding author upon reasonable request.

For decades, scientists sought a way to apply the outstanding analytical capabilities of neutrons to materials under pressures approaching those surrounding the Earth's core. These extreme pressures can rearrange a material's atoms, potentially resulting in interesting new properties. A breakthrough resulted in 2022 when researchers at Oak Ridge National Laboratory's Spallation Neutron Source squeezed a tiny sample of material—sandwiched between two diamonds—to a record 1.2 million times the average air pressure at sea level, or approximately 1.2 megabar. But this was only the start—they still had to produce useful data from the experiments. Now those same scientists have implemented software that removes the signal interference affecting the neutrons as they pass through the diamonds before reaching the sample. The research is published in the journal Scientific Reports. "Researchers can now perform neutron experiments beyond 1.0 megabar and extract accurate data about extraordinary atomic structures of materials," said Malcolm Guthrie, computational instrument scientist at ORNL's SNAP beamline.

10.1038/s41598-023-31295-3

Nano

Scientists ID new catalyst for cleanup of nitrites

dx.doi.org/10.1039/C3NR04540D Journal information: Nanoscale

http://dx.doi.org/10.1039/C3NR04540D

https://phys.org/news/2013-11-scientists-id-catalyst-cleanup-nitrites.html

Abstract Nitrate (NO 3 − ) and nitrite (NO 2 − ) anions are often found in groundwater and surface water as contaminants globally, especially in agricultural areas due to nitrate-rich fertilizer use. One popular approach to studying the removal of nitrite/nitrate from water has been their degradation to dinitrogen via Pd-based reduction catalysis. However, little progress has been made towards understanding how the catalyst structure can improve activity. Focusing on the catalytic reduction of nitrite in this study, we report that Au NPs supporting Pd metal ("Pd-on-Au NPs") show catalytic activity that varies with volcano-shape dependence on Pd surface coverage. At room temperature, in CO 2 -buffered water, and under H 2 headspace, the NPs were maximally active at a Pd surface coverage of 80%, with a first-order rate constant ( k cat = 576 L g Pd −1 min −1 ) that was 15x and 7.5x higher than monometallic Pd NPs (∼4 nm; 40 L g Pd −1 min −1 ) and Pd/Al 2 O 3 (1 wt% Pd; 76 L g Pd −1 min −1 ), respectively. Accounting only for surface Pd atoms, these NPs (576 L g surface-Pd −1 min −1 ) were 3.6x and 1.6x higher than monometallic Pd NPs (160 L g surface-Pd −1 min −1 ) and Pd/Al 2 O 3 (361 L g surface-Pd −1 min −1 ). These NPs retained ∼98% of catalytic activity at a chloride concentration of 1 mM, whereas Pd/Al 2 O 3 lost ∼50%. The Pd-on-Au nanostructure is a promising approach to improve the catalytic reduction process for nitrite and, with further development, also for nitrate anions. 1. Introduction Nitrate (NO 3 − ) is found in the groundwater and surface water globally, especially in agricultural areas due to the use of nitrate rich fertilizers. 1–4 This anion, and its metabolic conversion to the nitrite anion (NO 2 − ), cause adverse human health effects, including methemoglobinemia or the ‘blue baby’ syndrome. 5 Nitrite can be also converted to carcinogenic N -nitroso compounds in the body, 6 which may result in cancer and hypertension. 7 The maximum contaminant levels (MCL) of NO 3 − and NO 2 − have been set at 10 mg N per L and 1.0 mg N per L (calculated by nitrogen weight, denoted as mg N per L), respectively, by the United States Environmental Protection Agency (US EPA). 8 For the European Union countries, the European Drinking Water Directive set concentration limits for NO 3 − and NO 2 − at 11.3 mg N per L and 0.15 mg N per L, respectively. 9 The World Health Organization guideline values of NO 3 − and NO 2 − are 11.3 mg N per L and 0.91 mg N per L. 10 Current methods to treat nitrate/nitrite contamination include ion exchange, reverse osmosis, biological denitrification, and catalytic reduction. 11–21 The EPA-approved treatment technologies for removing nitrates/nitrites in waters destined for drinking purposes are ion exchange and reverse osmosis, with the former being the most widely used. 22 The ion exchange approach removes nitrate/nitrite from water by replacing them with equivalent amount of anions such as chloride or bicarbonate ions. 11 The main disadvantage is that, after the ion exchange material is regenerated for further use by contacting with a brine solution, the resulting brine which contains the nitrate/nitrite anions must undergo further treatment prior to disposal. Reverse osmosis also results in a nitrate/nitrite-containing secondary stream that needs further treatment. The biological denitrification and catalytic reduction routes have been studied at the laboratory and pilot scales but they are not yet commercially practiced. Biological denitrification can selectively remove nitrate/nitrite from water, but it is affected by pH and the presence of other salts in water; it can generate undesired by-products; 15 and bacterial contamination of the treated water is a concern. 11 The catalytic reduction of nitrate/nitrite has received intense research interest ever since Pd-based catalysts ( e.g. , Pd–Cu, Pd–In and Pd–Sn) were shown capable of converting nitrate/nitrite to harmless nitrogen (N 2 ) using hydrogen. 23 This chemistry was first reported by Vorlop and Tacke in 1989 using supported Pd–Cu/Al 2 O 3 catalysts. 14 Vorlop and co-workers reported improved reduction activity using Pd–Sn/Al 2 O 3 and Pd–In/Al 2 O 3 catalysts. 24 The catalytic reduction of nitrate is generally accepted to occur in two stages: the reduction of NO 3 − to NO 2 − , and then reduction of NO 2 − to N 2 or NH 3 ( Scheme 1 ). 15,25–28 In greater detail, the second metal ( i.e. , Cu, In, Sn) reduces the chemisorbed NO 3 − species in the first step and the Pd then regenerates the oxidized metal via H 2 reduction (reaction (1) of Scheme 1 ). In the second stage, Pd by itself reduces chemisorbed NO 2 − species in several surface reaction steps, eventually to form N 2 (reaction (2)) or NH 3 (reaction (3)). Ammonia is an undesirable reaction product because it is a groundwater pollutant with a US EPA MCL of 0.66 mg N per L. 8 The Pd–Cu composition provides N 2 selectivity values as high as 80–95%, depending on the reaction testing conditions used. 19,29 Ammonia formation is favored at high pH values, and so N 2 selectivity is improved by operating the reaction at non-alkaline conditions, for example. 29 Scheme 1 General reaction pathway for the reduction of nitrate or nitrite anions using a Pd-based catalyst. M is a promoter metal such as Cu, In, and Sn. 27 While the compositional effects on catalytic nitrate/nitrite reduction have been amply studied, the effect of nanostructure is not well addressed yet. We have previously demonstrated that Pd catalytic activity for hydrodechlorination (a type of reduction reaction) can be varied with great control by depositing Pd metal on Au nanoparticles (to form Pd-on-Au NPs) at a pre-determined Pd surface coverage. 30–33 Motivating this work, we hypothesized that Pd-on-Au NPs catalyze nitrite reduction (the second stage of nitrate reduction) and that catalytic activity is correlated to Pd surface coverage. We synthesized ∼4.3 nm Pd-on-Au NPs of varying Pd coverages, and quantified their reaction rate constants for nitrite reduction at room temperature, under H 2 headspace, in water and at pH 5–7 (buffered using CO 2 gas). We quantified the corresponding N 2 selectivities, and studied the effect of chloride anions, a common groundwater constituent. 2. Materials and methods 2.1. Materials Tetrachloroauric( iii ) acid (HAuCl 4 ·3H 2 O, 99.99%), tannic acid (C 76 H 52 O 46 , >99.5%), potassium carbonate (K 2 CO 3 , >99.5%), palladium( ii ) chloride (PdCl 2 , 99.99%), 1 wt% Pd/Al 2 O 3 , modified Griess's reagent, and Nessler's reagent (K 2 HgI 4 ) were purchased from Sigma-Aldrich. Trisodium citrate (Na 3 C 6 H 5 O 7 , >99.5%, Fisher), sodium nitrite (NaNO 2 , 99.7%) and sodium chloride (NaCl, 99.99%) were obtained from Fisher. Hydrogen gas (99.99%) was purchased from Matheson. 0.9 wt% Au/Al 2 O 3 was obtained from Mintek Autek, South Africa. All experiments were conducted using Nanopure water (>18 MΩ cm, Barnstead NANOpure Diamond). 2.2. NP synthesis 4 nm Au NPs, Pd NPs and Pd-on-Au bimetallic NPs were synthesized as previously reported. 30 To synthesize Au NPs, 0.05 g tannic acid, 0.018 g K 2 CO 3 and 0.04 g trisodium citrate were dissolved in 20 mL of water. In a second flask, 200 μL of HAuCl 4 solution (0.127 mol L −1 ) was dissolved in 79.8 mL of water. Both solutions were heated to 60 °C, and the first solution was added to the second under vigorous stirring. The color of the resultant sol immediately changed from pale yellow to reddish-brown, indicative of the formation of Au NPs. The solution was heated to boiling, allowed to boil for 2 min and then cooled to room temperature ( Scheme 2 ). The Au NP concentration was 1.26 × 10 14 NP per mL. 30 Scheme 2 Two-stage synthesis of Pd-on-Au NPs with a specific Pd surface coverage. The Pd NPs were synthesized by replacing the HAuCl 4 solution with a H 2 PdCl 4 solution of the same molar concentration, 30 and instead of boiling for 2 min, the mixed solution was allowed to boil for 25 min. Pd-on-Au bimetallic NPs were prepared by adding specific amounts of H 2 PdCl 4 solution to the as-synthesized Au NPs. The various surface coverage percentages (sc%) (0, 10, 30, 60, 80, 100, 150, 300 sc%) of Pd-on-Au NPs were obtained by mixing 9, 18, 55, 73, 92, 150, 351 μL H 2 PdCl 4 solution (2.47 mM) and 2 mL of the as-synthesized Au NPs. 30 The mixed solution was stirred for 1 min and then bubbled with hydrogen gas for 2 min at a flow rate of ∼200 mL min −1 . For the calculation of surface coverage percentages (sc%), we modeled the 4 nm Pd-on-Au NPs as magic clusters, where the Au NP is a magic cluster of 7 shells, and the Pd forms the 8th shell or more ( Scheme 2 ). 30 2.3. Characterization UV-vis absorbance spectra of the NP solutions were measured on a Shimadzu UV-2401 PC spectrophotometer. Transmission electron microscopy (TEM) images were obtained using a JEOL 2010 transmission electron microscope operating at an accelerating voltage of 200 kV. The particle size distribution was calculated by counting around 200 particles. pH measurements were taken using a VWR sympHony SB20 meter with a standard pH electrode. Nitrite ions were analyzed using the Griess test. 34 A stock solution of the Griess reagent was prepared by dissolving 10 g of the powder (Sigma-Aldrich) in 250 mL Nanopure water, such that the final concentrations are 0.1 wt% N -(1-naphthyl)ethylenediamine dihydrochloride, 1 wt% sulfanilamide, and 5% H 3 PO 4 . 35 In a typical colorimetric assay, the Griess reagent solution (0.2 mL), a nitrite-containing solution (0.2 mL), and water (1.6 mL) are mixed together and kept in room temperature for 10 min. The sulfanilamide reacts with a nitrite anion, and the resulting compound further reacts with the amine, forming a colored solution. The absorbance at 540 nm is measured via UV-vis spectroscopy, and the NO 2 − concentration is determined in the 0 to 2.0 ppm range using a standard curve (Fig. S1 † ). Ammonia measurements were made using an ammonium (NH 4 + ) ion selective electrode (Cole-Parmer, detection limit 0.01 ppm, concentration range from 0.01 to 17 000 ppm) and using Nessler's test (concentration range from 0.02 to 2.5 ppm). 34,35 In a typical colorimetric assay, a Nessler's reagent solution (1 mL, 0.09 M of potassium tetraiodomercurate and 2.5 M potassium hydroxide, Sigma-Aldrich) and an ammonia-containing sample (1 mL) are mixed together and then kept at room temperature for 1 min. The ammonium reacts with the tetraiodomercurate at high pH, forming a colored solution. The absorbance at 420 nm is measured, and the NH 4 + concentration is determined in the 0 to 10.0 ppm range using a standard curve. (Fig. S2 † ). These concentrations were independently verified using an ammonium ion selective electrode (Cole-Parmer, detection limit 0.01 ppm, concentration range from 0.01 to 17 000 ppm) (Fig. S3 † ). 2.4. Catalytic experiments Batch nitrite reduction experiments were conducted in a three-neck 250 mL round bottomed flask. The amount of Pd-on-Au NPs added was chosen such that the total Pd amount per reaction was 0.0365 mg, the final catalyst charge concentration was 0.365 mg-Pd per L, and the final liquid volume was 99.6 mL (Table S1 † ). For example, 3.75 mL of a Pd-on-Au NP (80 sc%) sol was combined with 95.85 mL of water in the flask. The solution was then bubbled simultaneously with hydrogen gas (120 mL min −1 , to serve as reductant) and carbon dioxide gas (120 mL min −1 , to buffer the solution to a pH value of 5–7) for 5 min. For the Pd/Al 2 O 3 case, 36.5 mg was suspended in 10 mL of water, and then 1 mL of suspension was combined with 98.6 mL of water in the flask; the total Pd amount charged to the reactor was 0.0365 mg. For the Au/Al 2 O 3 case, 28 mg was combined with 99.6 mL of water in the flask; the total Au amount charged to the reactor was 0.25 mg. Additionally, a set of catalytic experiments were carried out in the absence of carbon dioxide gas. The catalytic reactions were conducted at room temperature under constant stirring (400 rpm) and continuous hydrogen gas (120 mL min −1 ) and carbon dioxide gas (120 mL min −1 ). The NaNO 2 solution (0.4 mL, 10 mg mL −1 of NO 2 − ) was injected to start the reaction, such the initial solution NO 2 − concentration was 40 mg NO 2 per L (or 12.2 mg N per L). The reaction was monitored periodically by withdrawing 1 mL aliquots from the reaction flask. By the end of the reaction test, ∼8 to 12 mL of reaction solution was removed from the flask, leaving behind 88–92% of the initial solution. For the Pd-on-Au NP tests, no separation of the particles from the reaction medium was performed. For the Pd/Al 2 O 3 test, the powder was separated from the reaction medium before nitrite and ammonium concentrations were determined. Because the H 2 is excess to nitrite, the observed reaction rate k meas (with units of min −1 ) was calculated by assuming first-order dependence on nitrite concentration: 36 (1) with the observed reaction rate constant k meas = k cat C cat , k cat is the Pd normalized reaction rate constant (with units of L g Pd −1 min −1 ), C NO2 is the concentration of nitrite (with units of mg L −1 ), t is reaction time (with unit of min) and C cat is the concentration of Pd (with units of g L −1 ). Accounting for only surface Pd atoms (calculated using the magic cluster model), k ′ cat comes from k meas = k ′ cat C ′ cat , where C ′ cat is the concentration of surface Pd. To check that the proper amount of catalyst was used, various aliquots (0, 1.25, 2.5, 5.0, 7.5, 10 mL) of 60 sc% Pd-on-Au NP sols were added to the reactor and the total reaction volume was set at 99.6 mL. The corresponding Pd amounts in reactor were 0, 0.00913, 0.0183, 0.0365, 0.0548, 0.073 mg, respectively. Reaction rates were determined as described above. To determine the effect of chloride, a series of experiments were performed by adding different amounts of 1 M NaCl solution was added to reactor with catalyst before bubbling H 2 and CO 2 gas, giving the final concentration of chloride from 0 to 0.05 M. The total reaction volume was kept at 99.6 mL. The actual volume of 1 M NaCl and nanopure water added to reactor were shown in Table S2. † The following catalyst was added to the reactor: 5 mL 60 sc% Pd-on-Au NPs solution or 7.2 mg Pd/Al 2 O 3 (1 wt%). The other conditions are the same with the typical catalytic experiment. 3. Results and discussion 3.1. Structure of Pd-on-Au NPs The synthesis and characterization of Pd-on-Au bimetallic NPs for hydrodechlorination reactions has been previously reported by our group. 30,32,37–39 Fig. 1 shows an example TEM image of as-prepared Pd-on-Au NPs. These NPs have a relative uniform size distribution, with mean diameter of 4.3 nm and a relative standard deviation of 12%. Our X-ray absorption spectroscopy (XAS) results reported in earlier publications indicated that all Pd atoms were located on the surface of the Au NPs. 32,40 At relatively low surface coverage (<30 sc%), Pd atoms were found mostly as scattered atoms on the Au surface. At higher surface coverages, some of Pd atoms formed non-oxidizable two-dimensional (2-D) ensembles, which we concluded to be especially the most active species for the hydrodechlorination (HDC) of trichloroethene (TCE). At surface coverage of 70% and higher, three-dimensional (3-D) ensembles appeared, in which oxidized Pd atoms were found on top of other Pd atoms, and the per-gram-Pd catalytic activity decreased. Fig. 1 (a) Transmission electron microscopy (TEM) image and (b) size distribution of Pd-on-Au bimetallic NPs (60 sc%). Each bar represents the total number of NPs of a particle diameter ±0.25 nm. 3.2. Effect of Pd surface coverage For all catalyst compositions, 100% nitrite conversion was reached within 30 min. Fig. 2 is representative of the evolution of nitrite, ammonia and nitrogen as a function of time. The decrease in nitrite concentration fit well to a pseudo first-order model, and observed reaction rate constant, k meas , was obtained from the slope of the natural log of nitrite concentration versus reaction time using linear least squares fitting. The selectivity to N 2 exceeded 99% for all catalysts ( Table 1 ). These near-100% selectivity values towards N 2 are similar to catalytic nitrite reduction studies from other research groups. 41 The Pd-on-Au NP catalyst remained active for nitrite reduction after spiking the reactor with additional nitrite solution two more times (Fig. S4 † ). Fig. 2 Concentration–time curves of NO 2 − , NH 4 + and N 2 . Reaction conditions : 60 sc% Pd-on-Au NPs with 0.365 mg L −1 Pd in reactor, 120 mL min −1 H 2 , 120 mL min −1 CO 2 , 400 rpm stirring rate, 1 atm pressure. The initial nitrite/Pd molar ratio was 252/1. Table 1 Rate constants and N 2 selectivity for Pd-on-Au NPs, Pd NPs, Pd/Al 2 O 3 , and Au/Al 2 O 3 Sample name k cat (L g Pd −1 min −1 ) k ′ cat (L g surface-Pd −1 min −1 ) N 2 Selectivity (at 90% conversion) 0 sc% Pd-on-Au NPs Not active a Not active a Not available 10 sc% Pd-on-Au NPs 32 32 99.3 30 sc% Pd-on-Au NPs 174 174 99.5 60 sc% Pd-on-Au NPs 479 479 99.8 80 sc% Pd-on-Au NPs 576 576 99.6 100 sc% Pd-on-Au NPs 466 466 99.5 150 sc% Pd-on-Au NPs 421 584 99.1 300 sc% Pd-on-Au NPs 188 460 99.3 Pd NPs 40 160 99.9 Pd/Al 2 O 3 (1 wt%) 76 361 99.5 Au/Al 2 O 3 (1.2 wt%) Not active a Not active a Not available a First-order rate constant k meas = 0 min −1 . The observed reaction rates determined for Pd-on-Au NPs (60 sc%) at various catalyst charges showed a linear dependence, showing that the reaction rate increased proportionately to the catalyst amount used ( Fig. 3 ). Here, the implication is that the reaction rate constant determined at the standard catalyst charge using eqn (1) (the middle point of the 5-point line) is equal to the rate constant determined from 5 catalyst amounts (the slope of this 5-point line, which is equivalent to k cat = k meas / C cat ), This linearity was seen for other catalyst compositions. Fig. 3 gave a slope of 416 L g Pd −1 min −1 , smaller in value than 479 L g Pd −1 min −1 ( Table 1 , data point circled in Fig. 3 ), which we attributed to experimental uncertainty. If the reaction rate does not increase proportionately to catalyst amount, then too much catalyst has been added and the mass transfer effect becomes noticeable. 42 Fig. 3 Observed reaction rate k meas determined at different catalyst charges of Pd-on-Au NPs (60 sc%). The circled data point (blue color) indicates the standard Pd amount used for all other catalysts in this study. The Pd-on-Au NP catalysts with various Pd surface coverage (sc%) were prepared by reducing appropriate amounts of Pd salt precursor in the presence of Au NPs using H 2 gas. The Pd surface coverages were calculated by modeling a magic cluster structure for the NPs, 30 such that all atoms could be accounted for. A ∼4 nm Au NP with surface coverages of 10, 30, 60, 80, 100, 150, and 300% had Pd weight loadings of 2.4, 5.8, 10.9, 15.5, 18.1, 19.7 wt%, respectively. The convenience and accuracy of the magic cluster model relies on the assumptions that all Pd precursor is fully reduced onto the Au surface and that the Au particles are perfectly monodisperse (which is not the case). Inductively coupled plasma-optical emission spectroscopy results of ∼4 nm Pd-on-Au NPs from an earlier study indicated the metals content was >95% of the expected values. 43 We note that the uncertainties in the amount of Pd salt added to the synthesis volume and in the amount of Au NPs used, lead to an uncertainty in the calculated Pd surface coverage of 0.01 × sc% (at one standard deviation). 33 Nitrite reduction activity of Pd-on-Au NPs varied significantly with Pd surface coverage ( Table 1 and Fig. 4 ). Monometallic Au NPs (0 sc%) and Au/Al 2 O 3 did not show activity for the reaction. Since Pd was the active metal, catalytic activity was quantified by normalizing k meas to total Pd content ( k cat ) or to total surface Pd content ( k ′ cat ). The rate constants k cat increased monotonically from 10 sc% to 80 sc% and decreased from monotonically from 80 sc% to 300 sc%, characteristic of a "volcano-shape" structure–activity plot. Fig. 4 Experimentally determined nitrite reaction rate constants for Pd-on-Au NPs as a function of Pd surface coverage and for monometallic Pd catalysts. The 80 sc% Pd-on-Au NPs had a rate constant 15x and 7.5x times greater than those of monometallic Pd NPs and Pd/Al 2 O 3 , respectively ( Table 1 ). Except for the 10 sc% case, all Pd-on-Au NPs were also more active, indicating a positive and direct interaction of the Au with the Pd metal. The volcano plots and higher catalytic activity of Pd-on-Au NPs observed here are consistent with observations for HDC of TCE and tetrachloroethene (PCE). 39 Even though nitrite is not structurally related to TCE (or PCE), the similarity in the volcano-shape plots suggests the reduction reactions occur on a common set of catalytically active sites. We attribute the promotional effect of Au on Pd catalysis of nitrite reduction to the high Pd dispersion on the Au NP surface and the presence of the non-oxidized 2-D Pd ensembles. At surface coverages of 150 and 300 sc%, the per-gram-Pd catalytic activity was lower due to a lower Pd dispersion and the topmost atoms of the 3-D Pd ensembles being oxidized. Accounting for surface Pd atoms, these Pd-on-Au NPs had higher rate constant k ′ cat values than those of monometallic Pd NPs and Pd/Al 2 O 3 (which also increased, Table 1 ). However, the trend in k ′ cat (which differ from k cat above 100 sc%) was not as clear. This is due to the over-simplication of the metal-on-metal NP structure when using the magic cluster model for calculations, the presence of oxidized Pd and the lack of experimental data quantifying the amount of surface Pd atoms ( e.g. , through titration methods). 32 That the surface-Pd-atom-normalized rate constants were larger than those for Pd NPs and Pd/Al 2 O 3 may be the result of the electronic effect, in which there is change in valence electron density of states in the Pd metal. 44–46 In the absence of the CO 2 buffer, nitrite reduction was extremely slow. For example, the k cat for 60 sc% Pd-on-Au NPs was 1.0 L g Pd −1 min −1 (compared to 479 L g Pd −1 min −1 ). In a sampling of 4 different compositions, the final conversion for all was ∼1 to 2%, and the final pH increased from an initial pH of 5–7 to >7. The low catalytic activity resulted from the alkaline condition, which was observed by other researchers. 28,29,47 The reduction of nitrite generates OH − species, which block the active sites of palladium based catalyst. Thus, the use of a buffer like CO 2 ( ref. 29 ) and formic acid, 15 or a water-solubilized conducting polymer 41 with buffering capacity is important to maintain high activities and accurate rate constant measurements. Recently, Werth and coworkers used isotopically labeled nitrate/nitrite to study the nitrate/nitrite reduction pathways using a Pd–In/Al 2 O 3 catalyst (5 wt% Pd, 0.5 wt% In). 27 Surface-bound NO and N 2 O were found as reaction intermediates, providing greater mechanistic insight into the general reaction pathway of nitrite/nitrate catalytic reduction ( Scheme 1 ). Derived from the further reduction of NO 2 − , the NO surface intermediate reduces in the presence of H 2 to form NH 4 + . 27 The NO intermediate also reacts with another NO to form the nitrous oxide surface intermediate, N 2 O, in a parallel reduction step. This N 2 O then reduces to form N 2 . The high N 2 selectivity values observed for all catalysts in this study suggest that the NO-to-N 2 O surface reaction step dominates the NO-to-NH 4 + step under the reaction conditions used, independent of the bimetal nanostructure. 27 3.3. Effect of chloride concentration Chloride is one of the most common ions in the drinking water, with a typical concentration of ∼50 mg L −1 (= ∼50 ppm = ∼0.001408 M). 48 Werth and coworkers reported that chloride drastically inhibited the activity of alumina supported Cu–Pd catalysts for nitrate/nitrite reduction. 48 We assessed the deactivation resistance of Pd-on-Au NPs to various chloride concentrations, and compared it to that of Pd/Al 2 O 3 . In the presence of 35.5 ppm chloride (0.001 M), the observed reaction rate of Pd/Al 2 O 3 decreased by approximately 50%, whereas the reaction rate of Pd-on-Au NPs decreased by ∼2% ( Fig. 5 ). Pd/Al 2 O 3 deactivated completely in the presence of 1775 ppm chloride (0.05 M). In contrast, Pd-on-Au NP activity decreased by ∼20%. Consistent with our previous study on chloride deactivation effects during TCE HDC, 38 Pd-on-Au NPs showed significantly enhanced resistance to chloride deactivation for nitrite remediation. Fig. 5 The reaction rate constants k cat of 60 sc% Pd-on-Au NPs and Pd/Al 2 O 3 at various chloride concentrations. The significant deactivation of Pd/Al 2 O 3 catalyst was attributed to the chemisorption of chloride to the Pd surface, blocking the active sites from reaction. 38 Chloride anions have been observed to oxidatively absorb to a Pd(111) surface in water under acid conditions. 49 Pd-on-Au NPs was much less affected by chloride, presumably due to a lesser extent of chloride chemisorption onto the Pd supported on the Au surface. 4. Conclusion This study demonstrates the application of Pd-on-Au NP catalysts on the reduction of nitrite, a common groundwater contaminant known to cause adverse health effects in humans. While monometallic Au NPs were inactive for the reaction, bimetallic Pd-on-Au NPs exhibited a volcano-shape dependence of activity on Pd surface coverage. 80 sc% Pd-on-Au NPs had maximum activity among all Pd-on-Au compositions, and was more active than monometallic Pd in both NP and Al 2 O 3 -supported forms. All catalyst compositions showed near-100% selectivity to N 2 over ammonia. The buffered reaction conditions were important to eliminate the pH rise arising from the formation of hydroxide anions during nitrite reduction. The Pd-on-Au catalyst was much more resistant to chloride deactivation compared to Pd/Al 2 O 3 , at concentrations typical of groundwater and at higher concentrations. These results suggest that bimetallic Pd-on-Au NPs have catalytic properties amenable for water pollution abatement and that the metal-on-gold nanostructure is a useful direction towards designing improved catalysts for the rapid and selective reduction of NO 3 − and other oxyanions like perchlorate (ClO 4 − ) and bromate (BrO 3 − ). Acknowledgements The authors gratefully acknowledge financial support from the J. Evans Attwell-Welch Postdoctoral Fellowship Program of the Smalley Institute of Rice University (to H.Q.), the National Science Foundation (CBET-1134535), the Welch Foundation (C-1676), and GSI Environmental, Inc. We thank Ms. S. Gullapalli for the assistance in TEM characterization.

Chemical engineers at Rice University have found a new catalyst that can rapidly break down nitrites, a common and harmful contaminant in drinking water that often results from overuse of agricultural fertilizers. Nitrites and their more abundant cousins, nitrates, are inorganic compounds that are often found in both groundwater and surface water. The compounds are a health hazard, and the Environmental Protection Agency places strict limits on the amount of nitrates and nitrites in drinking water. While it's possible to remove nitrates and nitrites from water with filters and resins, the process can be prohibitively expensive. "This is a big problem, particularly for agricultural communities, and there aren't really any good options for dealing with it," said Michael Wong, professor of chemical and biomolecular engineering at Rice and the lead researcher on the new study. "Our group has studied engineered gold and palladium nanocatalysts for several years. We've tested these against chlorinated solvents for almost a decade, and in looking for other potential uses for these we stumbled onto some studies about palladium catalysts being used to treat nitrates and nitrites; so we decided to do a comparison." Catalysts are the matchmakers of the molecular world: They cause other compounds to react with one another, often by bringing them into close proximity, but the catalysts are not consumed by the reaction. In a new paper in the journal Nanoscale, Wong's team showed that engineered nanoparticles of gold and palladium were several times more efficient at breaking down nitrites than any previously studied catalysts. The particles, which were invented at Wong's Catalysis and Nanomaterials Laboratory, consist of a solid gold core that's partially covered with palladium. Many areas of the United States are at risk of contamination of drinking water by nitrates and nitrites due to overuse of agricultural fertilizers. Credit: USGS Over the past decade, Wong's team has found these gold-palladium composites have faster reaction times for breaking down chlorinated pollutants than do any other known catalysts. He said the same proved true for nitrites, for reasons that are still unknown. "There's no chlorine in these compounds, so the chemistry is completely different," Wong said. "It's not yet clear how the gold and palladium work together to boost the reaction time in nitrites and why reaction efficiency spiked when the nanoparticles had about 80 percent palladium coverage. We have several hypotheses we are testing out now. " He said that gold-palladium nanocatalysts with the optimal formulation were about 15 times more efficient at breaking down nitrites than were pure palladium nanocatalysts, and about 7 1/2 times more efficient than catalysts made of palladium and aluminum oxide. Wong said he can envision using the gold-palladium catalysts in a small filtration unit that could be attached to a water tap, but only if the team finds a similarly efficient catalyst for breaking down nitrates, which are even more abundant pollutants than nitrites. "Nitrites form wherever you have nitrates, which are really the root of the problem," Wong said. "We're actively studying a number of candidates for degrading nitrates now, and we have some positive leads."

dx.doi.org/10.1039/C3NR04540D

Biology

Exotic signaling mechanism in pathogens

Sabine Bachmaier et al, Nucleoside analogue activators of cyclic AMP-independent protein kinase A of Trypanosoma, Nature Communications (2019). DOI: 10.1038/s41467-019-09338-z Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-019-09338-z

https://phys.org/news/2019-04-exotic-mechanism-pathogens.html

Abstract Protein kinase A (PKA), the main effector of cAMP in eukaryotes, is a paradigm for the mechanisms of ligand-dependent and allosteric regulation in signalling. Here we report the orthologous but cAMP-independent PKA of the protozoan Trypanosoma and identify 7-deaza-nucleosides as potent activators (EC 50 ≥ 6.5 nM) and high affinity ligands ( K D ≥ 8 nM). A co-crystal structure of trypanosome PKA with 7-cyano-7-deazainosine and molecular docking show how substitution of key amino acids in both CNB domains of the regulatory subunit and its unique C-terminal αD helix account for this ligand swap between trypanosome PKA and canonical cAMP-dependent PKAs. We propose nucleoside-related endogenous activators of Trypanosoma brucei PKA (TbPKA). The existence of eukaryotic CNB domains not associated with binding of cyclic nucleotides suggests that orphan CNB domains in other eukaryotes may bind undiscovered signalling molecules. Phosphoproteome analysis validates 7-cyano-7-deazainosine as powerful cell-permeable inducer to explore cAMP-independent PKA signalling in medically important neglected pathogens. Introduction The mechanisms of protein kinase regulation, specifically by small second messenger molecules, have been studied in great detail using protein kinase A (PKA) as a paradigm 1 . PKA, discovered 50 years ago, is present in most eukaryotes except plants, has a highly conserved structure, and is the main effector of the second messenger cAMP. Hence, the synonym cAMP-dependent protein kinase is commonly used. The inactive PKA holoenzyme is a complex of regulatory (PKAR) and catalytic (PKAC) subunits, either as R-C heterodimer or R 2 -2C heterotetramer. The R 2 homodimer is formed by an N-terminal dimerization/docking (DD) domain that also mediates sub-cellular localization via A kinase anchoring proteins (AKAPs). Two C-terminal cyclic nucleotide binding (CNB) domains cooperatively bind two molecules of cAMP, resulting in a conformational change of the R subunit that releases the active catalytic kinase subunit(s) from the inhibitory pseudo-substrate or substrate site of PKAR. The CNB domain is an ancient evolutionarily conserved domain family with >7500 members that confers ligand-dependent allosteric regulation to a diverse range of proteins 2 . In eukaryotes, CNB domains are bound and regulated by cyclic nucleotides. In bacterial transcription factors, some CNB domains can bind other ligands like heme in the case of the CO sensing transcription activator CooA 3 or chlorinated phenolic compounds in CprK, a member of the ubiquitous CRP-FNR family of transcription activators 4 . In metazoans, PKA has diverse functions ranging from metabolism and gene regulation to development, motility, and memory 5 . Many of these functions are tissue-specific and compartmentalized at the subcellular level 6 . In lower eukaryotes including fungi or apicomplexan protozoa like Plasmodium and Toxoplasma , PKA plays key roles in nutrient sensing, developmental switches, or infectivity processes 7 , 8 , 9 . Most species encode one or two PKAR isoforms and several PKAC isoforms 10 . The resulting holoenzyme isoforms differ in cell type-specificity, developmental expression, sub-cellular localization, and affinity to cAMP, thereby accounting for the pleiotropic functions of cAMP signalling. Trypanosoma brucei species are kinetoplastid parasites that infect a large variety of mammals, causing severe disease in domestic animals with important economic losses in endemic countries. The parasite is also causative of the deadly human African sleeping sickness, a neglected tropical disease 11 . Transmission is restricted to the habitat of the Tsetse fly in tropical Africa. Development of the parasite in the host and vector is a prerequisite for transmission. This developmental process can be induced by cAMP analogues 12 , 13 , 14 , although this is mediated by intracellular hydrolysis products of these analogues 15 operating via a complex network of effectors 16 . The parasite has been shown to release cAMP as a mechanism of evading the host’s innate immunity 17 . Essential roles of intracellular cAMP signalling have also been documented for cell division 12 , 18 , 19 , 20 and social motility 21 . It is therefore surprising that all attempts to detect cAMP-dependent kinase activity in African trypanosomes have failed 22 , 23 , 24 , 25 , 26 , 27 . Genes encoding three PKA catalytic subunit orthologues and one regulatory subunit orthologue have been identified in the T. brucei genome 22 , 26 , 28 , whereas alternative cAMP effectors like EPAC orthologues and cNMP-gated ion channels were not detected. By screening a genome-wide RNAi library for cAMP resistance in T. brucei , we identified a novel cAMP binding protein (CARP1) unique to kinetoplastids 29 , yet PKA was not among the hits of the screen. The catalytic subunits of T. brucei PKA are highly conserved with the presence of all 11 canonical kinase subdomains, the essential threonine in the kinase activation loop, and conserved residues implicated in mammalian PKAC’s binding to the regulatory PKAR subunits 30 . TbPKAR has a conserved C-terminal part with two CNB domains and the PKA substrate motif (RRT T V) that interacts with and inhibits PKAC. TbPKAR differs from its metazoan orthologues by an extended N-terminal domain with leucine-rich repeats (LRR) (Fig. 1a ). Some amino acid substitutions of consensus residues in the cAMP binding pockets have been noticed in sequence alignments 22 , 31 . The link between cAMP and PKA remains elusive in Trypanosoma in spite of the excellent overall conservation of the kinase. Fig. 1 PKA holoenzyme complexes in T. brucei . a Domain architecture of PKAR (top) and PKAC (bottom) orthologues from T. brucei (Tb) (TriTrypDB accessions: PKAR, Tb927.11.4610; PKAC1, Tb927.9.11100; PKAC2, Tb927.9.11030; PKAC3, Tb927.10.13010) compared to human (Hs) PKA (Uniprot accessions: PKARIα, P10644; PKACα, P17612). LRR leucine-rich repeat region, DD dimerization/docking domain, CNB cyclic nucleotide binding domain, kinase kinase domain. b Genotypes of cell lines with in situ tagged PKAC1 ( Ty1-C1 , cyan; Ty1 epitope tag in magenta) and PKAR ( R-PTP , blue; PTP-tag in black) compared to wild type (WT). The phleomycin resistance cassette ( BLE , grey) is indicated. c Two-colour fluorescent western blot analysis of the double tagged cell line ( ∆ c1/Ty1-C1 R-PTP ) using anti-PKAR and anti-PKAC1/2 (left panel, red signals), and anti-Ty1 and anti-PFR-A/C (loading control) (middle panel, green signals). The merge of both channels is shown on the right panel. Ty1 causes a mobility shift of PKAC1 and enables the detection of PKAC2 in cell lines devoid of wild type PKAC1. Note that PKAC1 and PKAR appear as doublet bands that we interpret as modification (PKAC1*) and allelic polymorphism, respectively, in the MiTat 1.2 cell line. d PTP affinity purification followed by western blot analysis of double tagged and control cell lines using antibodies as in c and anti-PKAC3. Equivalent amounts of soluble input material (IN), flow-through (FT), washes, and 13 equivalents of the eluate (Elu) were loaded. Source data to c and d are provided as a Source Data file Full size image There is a surprising deficit of knowledge in signalling mechanisms in these phylogenetically distant organisms and no complete pathway from receptor to effector has been elucidated to date. A possible explanation for this knowledge gap is suggested by the domain architectures found in the kinome of trypanosomes: few known signalling domains or protein–protein interaction domains are linked to the catalytic kinase domains of the 176 identified protein kinases. In addition, unusual domain combinations prevail 32 . Many conserved signalling effectors are likely to be differently connected and wired in various pathway architectures in these phylogenetically distant protozoa. This might also be the case for some second messenger dependencies. Here, we show that Trypanosoma PKA is not a cyclic nucleotide-dependent protein kinase. We use a chemical biology approach to identify highly specific activators of Trypanosoma PKA. The first crystal structure of a kinetoplastid PKAR explains the structural requirements for ligand selectivity. We suggest that this PKA has evolved to bind novel ligand(s), possibly taking the role of second messenger(s) in T. brucei . Our new activators are excellent tools to study this cAMP-independent PKA signalling in trypanosomes. Results PKA holoenzymes in Trypanosoma brucei We first established that the orthologous PKA subunit genes in T. brucei do encode proteins able to form the expected holoenzyme complexes of regulatory (R) and catalytic (C) subunits. One allele of PKAC1 was Ty1-epitope tagged in situ, while the second PKAC1 allele was deleted to generate T. brucei cell line ∆ c1/Ty1-C1 (Fig. 1b ). The absence of a wild type PKAC1 allele allowed simultaneous detection of the highly similar PKAC2 isoform by a PKAC1/2-specific antibody (Fig. 1c and Supplementary Fig. 1a ). PKAR was then C-terminally PTP-tagged in situ in cell line ∆ c1/Ty1-C1 to generate ∆ c1/Ty1-C1 R-PTP (Fig. 1b, c ). All three PKA catalytic subunit isoforms were pulled down by PKAR-PTP from lysates of cell line ∆ c1/Ty1-C1 R-PTP but not from the control cell line ∆ c1/Ty1-C1 (Fig. 1d ). Pull down from cell lines expressing Ty1- or HA-tagged PKAC or PKAR subunits independently confirmed the interactions between PKAR and each of PKAC1, 2, 3 in a heterodimeric complex (Supplementary Fig. 1b, c ). No co-precipitation of untagged PKAR or other PKAC isoforms was observed with tagged PKAR or PKAC1, 2, or 3 (Fig. 1d and Supplementary Fig. 1c ), indicating the absence of a tetrameric R 2 -2C complex that is found in mammalian PKA. Heterodimeric PKAR-C complexes are not unusual in lower eukaryotes 27 , 33 , 34 . The catalytic function of the kinase is essential for growth and viability, as RNAi-mediated repression of the catalytic subunits PKAC1 / 2 or PKAC3 is lethal or growth inhibitory, respectively 35 (Supplementary Fig. 2 ). Severe cell division defects are detected, as cytokinesis stages and multinucleated cells accumulate. This phenotype is frequently found when targeting essential trypanosome genes 36 . It does not necessarily indicate a specific role in cytokinesis. RNAi-mediated repression of PKAR has a very similar phenotype due to the rapid decrease of free PKAC1 (Supplementary Fig. 2 ). This indicates degradation of free PKAC released upon holoenzyme dissociation, as was observed for PKAC in mammalian cells 37 . Nevertheless, cell clones can be selected that maintain a basal PKAC1 level sufficient for survival after homozygous deletion of PKAR . The resulting ∆ pkar/ ∆ pkar cell lines 38 used later in this work show a mild growth phenotype with population doubling time (PDT) of 7.5 h versus 5.4 h for wild type cells. Trypanosoma PKA is not activated by cAMP The PKA holoenzymes were immunoprecipitated from trypanosomes expressing epitope-tagged PKA subunits to assay PKA activity. Contaminating kinase activities in the precipitate were excluded as (1) Ty1-affinity purification from wild type cells or cells expressing a catalytically inactive Ty1-PKAC1 N153A mutant 39 did not pull down PKA-specific activity and (2) phosphorylation of the PKA-specific substrate kemptide was inhibited by the PKA-specific pseudo-substrate peptide PKI(5–24) 40 (Supplementary Fig. 3a ). The basal activity of holoenzymes immunoprecipitated from cells expressing PKAR-Ty1 or Ty1-PKAC1 or HA-PKAC2 was not increased by cAMP (Supplementary Fig. 3b ), even when the cyclic nucleotide was added at unphysiologically high concentrations (1 mM). In the same experiments, cGMP activated at ≥100 µM (Supplementary Fig. 3b ). Some activation by cGMP had been noted before 22 , but several lines of evidence argue against the presence of cGMP or cGMP signalling in trypanosomatids 41 , minimizing the likelihood that cGMP is a physiological activator of the kinase. The unconventional cyclic nucleotides cXMP, cIMP, cCMP, and cUMP also did not show any significant effects (Supplementary Fig. 3c ). In order to exclude unsuitable assay conditions, an in vivo assay for PKA activity was established that is based on transgenic expression of the heterologous PKA substrate VASP (human platelet va sodilator s timulated p hosphoprotein) in T. brucei . Phosphorylation of VASP at the PKA site Ser-157 causes an electrophoretic mobility shift from 46 to 50 kDa 42 (Supplementary Fig. 3e ). Non-phosphorylated (46 kDa) and phosphorylated VASP (50 kDa) were quantified by western blot analysis. The ratiometric determination of VASP phosphorylation (non-phosphorylated/phosphorylated) is a reliable proxy for PKA activity. The myristoylated membrane-permeable peptide inhibitor myr-PKI(14–22) 40 reduced the measured activity to background (Supplementary Fig. 3f ). We then examined the effects of cAMP on PKA activity in live cells by three independent approaches. First, pharmacological elevation of intracellular cAMP was accomplished by inhibition of phosphodiesterases (PDEs) with a highly potent inhibitor of trypanosomal PDEs (CpdA, now renamed as NPD-001). The intracellular cAMP content increased up to 600-fold upon treatment with CpdA (Fig. 2a ) but did not elicit any change in in vivo PKA activity (Fig. 2b ). Second, reverse genetic elevation of intracellular cAMP was achieved by inducible RNAi-mediated depletion of the major cAMP-specific PDEs, PDEB1 and PDEB2 20 . The resulting 45-fold increase of cAMP content (Fig. 2c ) also did not stimulate PKA activity (Fig. 2d ). Third, VASP expressing cells were treated with membrane-permeable cAMP analogues 8-pCPT-cAMP, 8-pCPT-2′-O-Me-cAMP, or cAMP-AM (see Supplementary Table 1 ). The latter is a prodrug cleaved by esterases to deliver cAMP intracellularly 43 . No PKA activation was detected for any of the analogues up to 1 mM (Supplementary Fig. 3h ). Membrane-permeable cGMP derivatives also had no effect on PKA activity in vivo (Supplementary Fig. 3i ). Fig. 2 T. brucei PKA is not activated by cAMP. a Intracellular [cAMP] (mean ± SD of independent replicates; n = 5 (0 min, Mock); n = 6 (DMSO, CpdA)) and b in vivo PKA activity (with western blot inset) in cells treated or not (Mock) for 15 min with the PDE inhibitor CpdA 19 (10 µM; now renamed as NPD-001 74 ) or solvent (1% DMSO). Release of cAMP into the medium upon treatment was neglectable (Supplementary Fig. 3d ). c Intracellular [cAMP] (mean ± SD of n = 3 independent replicates) and d in vivo PKA activity (with western inset) upon inducible (1 µg ml −1 tetracycline) RNAi repression of PDEB1 and PDEB2 . Kinase activity cannot be determined at time 0, since the RNAi cell line (VASP i PDEB1/2 RNAi) and the control line (VASP i ) both harbour a tetracycline inducible VASP transgene. e Dose response of in vivo PKA activity upon 15 min treatment with dipyridamole (Dip). f Time course of in vivo PKA activity upon treatment with 100 µM dipyridamole (Dip) or dipyridamole + myr-PKI(14–22) or solvent (1% DMSO) or Mock. g Intracellular [cAMP] (mean ± SD of independent replicates; n = 5 (0 min, Mock 15 min); n = 3 (Mock 90 min); n = 6 (DMSO, Dip)) upon treatment with 100 μM dipyridamole (Dip) or 1% solvent (DMSO) or Mock for 15 or 90 min. h In vivo PKA activity upon treatment as in g for 15 min in wild type (WT), homozygous pkar knock out and PKAR add-back (in situ rescue) cells. For all in vivo kinase reporter assays, one representative western blot is shown as inset and data points are mean ± SD of n = 3 independent replicates. Source data are provided as a Source Data file Full size image The drug dipyridamole induces PKA activity in trypanosomes The most compelling evidence against cAMP-dependent PKA activity in trypanosomes is the lack of any change in kinase activity upon an up to 600-fold increase in intracellular cAMP content, caused by the PDE inhibitor CpdA. Prior to the availability of CpdA, the same experiment was done with dipyridamole, a potent inhibitor of mammalian PDEs 44 with modest activity against trypanosomal PDEs 15 , 45 . Initially, we were misled by a dose- and time-dependent induction of VASP phosphorylation by dipyridamole (Fig. 2e, f , Table 1 ) that correlated with a moderate increase in intracellular cAMP content (Fig. 2g ). As the data shown in Fig. 2a–d provide strong evidence that this effect by dipyridamole cannot be caused by the increase in cAMP, dipyridamole must induce PKA activity either directly or indirectly by a cAMP-independent mechanism. Yet, the effect is clearly mediated by PKAR, as no induction was detected in a VASP reporter cell line with homozygous deletion of PKAR (Fig. 2h ). Table 1 Summary of TbPKA activators Full size table Compound screening for activators of TbPKA Dipyridamole, a licensed drug inhibiting thrombocyte aggregation, is a PDE inhibitor but also interferes with adenosine transport and metabolism 46 . The possible link between PKA activation and purine metabolism motivated a targeted compound screen. We tested 13 different purine nucleoside or nucleotide analogues (Supplementary Table 1 ) for activity in the in vivo PKA reporter assay. Most of them are predicted to be membrane permeable due to lipophilic groups. Out of these, the four 7-deazaadenosine (tubercidin, Tu) analogues toyocamycin (Toyo), 5-iodo-tubercidin (5-I-Tu), 5-bromo-tubercidin (5-Br-Tu), and sangivamycin induced PKA activity with EC 50 values of 88 nM, 390 nM, 625 nM, and 39 µM, respectively (Fig. 3a , Table 1 ). The remaining compounds had either a slight inhibitory (8-pCPT-Ado, 8-pCPT-2-′O-Me-Ado) or no effect (Tu, 8-pCPT-Guo, 2-Cl-Ado, 8-pCPT-Ade, 6-Br-Tu, 8-Br-Ado, 6-Cl-PuR) in the PKA reporter assay (Fig. 3 , Supplementary Fig. 4a ). The most potent activator, toyocamycin, did not induce phosphorylation of VASP in the pkar knock out genetic background (Supplementary Fig. 4b ), demonstrating that PKA is the target kinase. Growth or viability of the parasites was affected by continuous treatment with all 7-deazaadenosine analogues (Table 1 ). However, this was apparently due to off-target effects, as the drugs had very similar effects on pkar knock out cells (Table 1 ). Since these analogues are known to target multiple proteins and processes in mammalian cells, including kinases and synthesis of DNA, RNA, and proteins (reviewed in ref. 47 ), growth or viability phenotypes were expected. Fig. 3 Activators of trypanosome PKA. a Hit compounds with their EC 50 of in vivo PKA activity (VASP reporter assay, mean ± SD, n = 3 independent replicates, representative western blots as inset) and their EC 50 of in vitro kinase activity (kemptide phosphorylation by T. brucei PKAR-PKAC1 holoenzyme expressed in L. tarentolae ). A representative dose response for Toyo ( n = 5), 5-I-Tu ( n = 4), 5-Br-Tu ( n = 2), tubercidin ( n = 2), and sangivamycin ( n = 1) is shown with SD of technical duplicates or triplicates. Binding parameters to T. brucei PKAR(199–499) expressed in E. coli were determined by isothermal titration calorimetry (ITC). The power differential (DP) between the reference and sample cells upon injection was measured as a function of time (inset). The main plot presents the total heat exchange per mole of injectant (integrated peak areas from inset) as a function of the molar ratio of ligand to protein. One out of three independent replicates is shown. b Data for 7-CN-7-C-Ino as in a ; for number of independent replicates see Table 1 . The EC 50 and K D values given in a and b are rounded values from Table 1 . c Thermodynamic signature (Δ G in blue, Δ H in green, and − T Δ S in red) compiled from ITC measurements (mean ± SD of n = 3 independent replicates) in a and b . Source data are provided as a Source Data file Full size image 7-Deazaadenosine analogues bind and activate TbPKAR The identified compounds may activate trypanosome PKA either directly or indirectly in the cell-based reporter assay. To address this, tagged PKAR and PKAC1 subunits were co-expressed in the heterologous Leishmania tarentolae expression system and the holoenzyme complex was isolated by tandem affinity purification (Supplementary Fig. 4e , inset). In vitro kinase assays with kemptide and [ 32 P]-ATP as substrates showed kinase activation upon addition of tubercidin and 5-substituted tubercidin analogues with the following order of potency: 5-I-Tu > Toyo > 5-Br-Tu > tubercidin > sangivamycin (Fig. 3a , Table 1 ). The EC 50 values measured with recombinant T. brucei holoenzyme purified from L. tarentolae were in the same order of magnitude as the EC 50 values of the in vivo PKA reporter assay. Up to 3.4-fold differences between in vitro kinase assay and cell-based assay are likely due to uptake, accumulation, or metabolism of the individual compounds in trypanosome cells. Tubercidin enters the trypanosome cell via nucleoside transporters 48 , while the 5-substituted analogues are more lipophilic and predicted to cross the cell membrane by passive diffusion. Tubercidin uptake might be too slow to cause PKA activation within the time frame of the in vivo reporter assay. The evidence for a direct mode of action provided by in vitro kinase assays was further corroborated by measuring binding parameters between the activating compounds and purified N-terminally truncated T. brucei PKAR(199–499) expressed in Escherichia coli (Supplementary Fig. 4h ). Isothermal titration calorimetry (ITC) determined a K D of 57 nM for toyocamycin with a ligand to protein molar ratio of 2.1 (±0.2), indicating availability of both CNB sites for binding (Fig. 3a , Table 1 ). A slightly lower K D of 32 nM was obtained for 5-I-Tu with a ligand to protein molar ratio of only 1.2 (±0.2). We suggest that 5-I-Tu has a preference for one binding site, whereas the K D value for Toyo averages over both binding sites. This is supported by the different thermodynamic signatures that indicate enthalpically-driven binding, which for 5-I-Tu is counteracted by a negative entropic effect (Fig. 3c ). It should be noted that all binding assays are performed with the ligand-free (apo) form of the PKAR subunit, whereas kinase assays probe the R-C holoenzyme. Nevertheless, the binding K D values for the analogues are close to the respective EC 50 for kinase activation in vitro. Most importantly, the values for Toyo and 5-I-Tu are similar to the K D for cAMP binding to mammalian PKARIα measured under identical conditions (28 ± 5 nM, Supplementary Fig. 4f and ref. 49 ). Dipyridamole, whose activation of T. brucei PKA had initially guided our compound screen, did not have any effect in the in vitro kinase assay (Supplementary Fig. 4e ), supporting an indirect mode of action in vivo. Development of a trypanosome-specific PKA activator In mammalian cells, some nucleoside analogues have been reported to be significantly less toxic when the adenosine moiety is replaced by inosine 50 , 51 . Therefore, as the first step in optimization, we changed the purine ring side groups of toyocamycin (7-cyano-7-deazaadenosine) to match inosine, resulting in the compound 7-cyano-7-deazainosine (7-CN-7-C-Ino). Compared to Toyo, 7-CN-7-C-Ino had 36-fold higher activation potency on the purified holoenzyme (EC 50 6.5 ± 4.6 nM) and 7-fold higher binding affinity to PKAR ( K D 8 ± 2 nM) in ITC measurements (Fig. 3b , Table 1 ). The increased affinity was mainly due to gain in the entropic component of binding (− T Δ S = −2.3 kcal mol −1 versus −0.9 kcal mol −1 for Toyo) (Fig. 3c ). This surprising increase in activation potency upon introduction of a single structural modification did not, however, translate into increased potency in the in vivo kinase reporter assay (EC 50 838 ± 139 nM), probably due to slower uptake or faster intracellular metabolism of the 7-deazainosine analogue. In a cell line with homozygous deletion of PKAC3 and simultaneous RNAi depletion of PKAC1/2 (Supplementary Fig. 4d ), phosphorylation of the PKA reporter substrate VASP was undetectable with or without treatment by 7-CN-7-C-Ino (Supplementary Fig. 4c ). This control fully corroborated the PKA specificity of the in vivo reporter assay based on VASP Ser-157 phosphorylation in trypanosomes. The cytotoxicity of 7-CN-7-C-Ino was drastically reduced compared to Toyo (>170-fold increase of EC 50 for cell viability, Table 1 ). 7-CN-7-C-Ino, Toyo, or 5-I-Tu did not bind to purified human PKARIα, whereas cAMP binding to the same PKARIα preparations ( K D 28 nM, Supplementary Fig. 4f ) was intermediate between the affinities of CNB-A and CNB-B that have been measured before separately 52 . Consistently, Toyo did not activate the mammalian PKARIα 2 –2PKACα holoenzyme, even at concentrations 270-fold above the EC 50 for the T. brucei PKA holoenzyme (Supplementary Fig. 4g ). As expected from the kinase assays (see Fig. 2b, d and Supplementary Fig. 3b ), recombinant T. brucei PKAR did not bind cAMP (Supplementary Fig. 4h ). In summary, a potent, highly specific, and nontoxic activator of the cAMP-independent trypanosomal PKA has been developed. Crystal structure of parasite PKAR with 7-CN-7-C-Ino To explore the molecular basis of binding and activation selectivity of Trypanosoma PKA for 7-deaza nucleoside analogues, we co-crystallized PKAR with 7-CN-7-C-Ino. The highest resolution X-ray diffraction data were obtained using truncated PKAR(200–503) of the related parasite Trypanosoma cruzi (Supplementary Fig. 5 , see Supplementary Table 2 for data collection and refinement statistics) that is highly homologous in primary sequence (79% identity) and structural alignment to the T. brucei PKAR fragment. The protein adopts a dumbbell-like structure (Fig. 4a ) consisting of two 8-stranded beta barrels (CNB-A and CNB-B) linked by an alpha helical element (R337 to N361, three helices). A 7-CN-7-C-Ino molecule can be clearly recognized on top of each beta barrel by the electron density in the Fo–Fc omit map (Fig. 4b, c ). In bovine PKAR (PDB 1RGS), the CNB-A and CNB-B domains are linked via allosteric communication 30 . When a cAMP molecule enters the CNB-B of PKAR, it is capped by a hydrophobic residue (Y371) via π-stacking interaction. This event triggers the rupture of a salt bridge (E261–R366) and consequently destabilizes the holoenzyme conformation. Tryptophan W260 thereby moves 30 Å closer to CNB-A where it caps the second cAMP molecule in that binding site by an analogous π-stacking interaction. The conformational change releases an active C-subunit. The amino acids taking part in these sequential ligand binding events linked to holoenzyme activation are readily identified in 7-CN-7-C-Ino-bound T . cruzi PKAR by alignment with cAMP-bound bovine PKARIα (Fig. 4a , Supplementary Movie 1 ). In TcPKAR(200–503) the capping residues in CNB-A and -B are Y371 and Y483, respectively, while a salt bridge might be formed by the pair R476–E372. Despite only 35% sequence identity, the C α alignment of both structures shows very high conservation (RMSD, root mean square deviation = 3.25 Å) including the fragmented αB/C helix (Supplementary Fig. 6 , Supplementary Movie 2 ) that is key of the conformation change mediating allosteric activation of bovine PKARIα 30 . Fig. 4 Co-crystal structure of trypanosome PKAR with 7-CN-7-C-Ino. a Structural alignment of T. cruzi PKAR(200–503) (chain representation in blue) and Bos taurus PKARIα(92–308) (PDB 1RGS, chain representation in grey). The two capping residues and the salt bridge pair are highlighted by their carbon atoms colour-coded as green and magenta in the B. taurus and T. cruzi PKAR structures, respectively. The overlay of ligand poses is shown in the blow-ups and the π-stacking interactions in both sites for both proteins are highlighted. b , c Fo–Fc (3 σ , green) and 2Fo–Fc (1 σ , blue) maps of TcPKAR CNB-A and CNB-B, respectively, showing 7-CN-7-C-Ino modelled to fit the electron densities. d , e Hydrogen bonding network (black dashed lines) of 7-CN-7-C-Ino bound to TcPKAR(200–503) CNB-A and CNB-B, respectively. The capping residues in CNB-A (Y371) and CNB-B (Y483) are labelled in magenta. 3D versions of a , d , and e are available as Supplementary Movie 1 , 3 , and 4 , respectively. f , g Sequence alignment of PKAR in CNB-A and CNB-B, respectively, of representative kinetoplastid parasites with a mammalian PKAR ( B. taurus PKARIα) as reference. Numbering refers to T . cruzi (top) and B . taurus (bottom), respectively Full size image Both CNB-A and -B pockets are occupied by 7-CN-7-C-Ino with the ribose moiety in a position very similar to the position of the ribose moiety of cAMP bound to bovine PKARIα (Fig. 4a , Supplementary Movie 1 ). The ribose moiety docks to the CNB-A site by donating two hydrogen bonds (hydroxyl groups O3 and O5) to E312 (Fig. 4d , Supplementary Movie 3 ). E310 contacts the O2 hydroxyl group and at the same time brings Y371 in close proximity to make a π-stacking interaction with the hypoxanthine moiety of the ligand. The free electron pair in the cyano moiety receives a hydrogen bond from the side chain of K294, while Y300 additionally shields the ligand from the solvent. In the CNB-B site (Fig. 4e , Supplementary Movie 4 ), E436 interacts with the 3′ and 5′ hydroxyl groups of the ribose ring. Y483 acts as a hydrophobic cap that seems to be positioned via a water-mediated (w07) hydrogen bonding network formed by the main chain carbonyl group of H431, the side chain of E434 and the O2 hydroxyl group from the ribose. The keto group at position 6 of the purine ring accepts a hydrogen bond from the main chain nitrogen atom of Y486, while the secondary amine in position 1 acts as a donor to the carbonyl of K482 (Fig. 4e , Supplementary Movie 4 ). In the CNB-B site, the cyano group does not engage in hydrophilic interactions but instead is sandwiched between the side chains of Y485 and Y486. A sequence alignment of kinetoplastid and mammalian PKAR (Fig. 4f , g) shows that alanine residues in both pockets (A202 and A326) of bovine PKAR are substituted by glutamates in the kinetoplastid PKAR orthologues. The positively charged arginine residues responsible for neutralizing the cyclic phosphate of cAMP (R209 and R333 in bovine PKAR) are substituted in kinetoplastid PKAs by neutral amino acids with shorter side chains (V, T, A, or N). Thus, few substitutions in a highly conserved signalling protein are hallmarks of a ligand selectivity switch from cAMP to 7-CN-7-C-Ino. Molecular docking of 7-deaza nucleosides For the series of 7-deaza nucleoside analogues, we found a good correlation between computational docking to the TcPKAR structure and the potency of activation of the purified kinase (Supplementary Fig. 7 and Table 1 ). Re-docking of the co-crystallized compound 7-CN-7-C-Ino to the CNB-A and -B sites of TcPKAR gave RMSD values of 0.266 and 0.252, respectively, for the best poses. Using the Glide E-model (GE) scoring system, the best docking poses of the compounds were ranked for the A site (7-CN-7-C-Ino > Toyo > 5-Br-Tu > 5-I-Tu > sangivamycin > Tu) and for the B site (7-CN-7-C-Ino > Tu > Toyo > 5-Br-Tu > 5-I-Tu > sangivamycin). Interestingly, Tu docks very well in the B site but is a weak activator (see Discussion). The cyano-, iodo-, or bromo- groups at position 5 of all other tubercidin analogues are accommodated in a hydrophobic pocket formed by the side chains of V444, V489, Y485, and Y486, the latter three being part of the αD helix (Supplementary Fig. 7c ). 7-CN-7-C-Ino forms two additional hydrogen bonds (K482/O and Y486/N) with the αD helix compared to Toyo, correlated with a 7-fold higher affinity and 36-fold higher activation potency (Table 1 ). PKA signalling and targets in trypanosomes PKA downstream signalling components and targets are so far completely unexplored in trypanosomes. To probe PKA target phosphorylation events, we first used a phospho-specific PKA substrate antibody detecting the phosphorylated consensus PKA sites RXXS*/T*. A 3–4-fold increase in global RXXS*/T* site phosphorylation was observed by western blotting after 10 min of Toyo or 7-CN-7-C-Ino treatment in wild type but not in ∆ pkar/ ∆ pkar cells (Fig. 5a ). Phosphoproteome analysis under these conditions (15 min ± 7-CN-7-C-Ino) showed 642 significantly (FDR ≤ 0.05, s 0 = 2) upregulated phosphosites, mostly containing PKA motifs (77%) (K/R-X-X-S/T, K/R-X-S/T) 5 , whereas these PKA motifs were underrepresented (19%) in the 84 downregulated phosphosites (Fig. 5b, c ; Supplementary Data 1 ). The frequency distribution of PKA motif subsets in the 7-CN-7-C-Ino-induced T. brucei phosphoproteome matches closely with that observed for human PKA motifs (Fig. 5c ). An unbiased motif discovery algorithm confirmed enrichment of the same PKA motifs among the upregulated phosphosites (Supplementary Fig. 8a, b ). Gene ontology (GO) enrichment analysis predicts functions of PKA in posttranscriptional regulation of gene expression, dynamics of cytoskeletal and organellar structures, signalling, and cell division and cytokinesis (Supplementary Fig. 8c ). Fig. 5 Target phosphorylation and expression changes. a Global display of proteins phosphorylated at RXXS/T sites. Wild type (WT) or ∆ pkar/ ∆ pkar ( pkar KO) cells were treated or not (−) with 2 µM Toyo or 7-CN-7-C-Ino (Ino) for 10 min and lysates were subjected to western blotting with anti-phospho-RXXS*/T* and anti-PFR-A/C as loading control. M: protein molecular weight marker. b Volcano plot representation of phosphopeptides quantified by label-free phosphoproteome analysis. Phosphopeptides are plotted according to p -value and fold change caused by treatment of T. brucei WT cells with 7-CN-7-C-Ino (8 µM, 15 min, n = 4 independent experiments) in comparison to untreated cells ( n = 4 independent experiments). Phosphopeptides that change significantly in abundance (FDR ≤ 0.05, s 0 = 2) and contain phosphosites matching PKA consensus motifs (R/K-X-X-S/T, R/K-X-S/T) 5 are shown as red dots; significantly changed phosphopeptides without PKA consensus motifs are shown in black. c Pie charts showing the fraction of PKA consensus motifs (red) within the downregulated ( n = 84, left) or upregulated ( n = 642, right) phosphosites. The frequency of individual subsets of PKA motifs (upregulated) was compared to the human PKA site motif frequency retrieved from the PhosphoSitePlus database ( ). d Abundance changes of metacaspase 4 (MCA4) and PKAC1/2 in WT or e pkar KO cells after treatment with 4 µM 7-CN-7-C-Ino (Ino) or Toyo in a 24-h time course. Western blot signals were normalized to the loading control PFR-A/C; untreated WT cells were set to 1. Source data to a and c – e are provided as a Source Data file Full size image We also explored changes in cellular protein abundance following PKA activation at later time points (6 and 12 h) by label-free quantitative proteomics. No significantly regulated proteins were detected 6 h post induction and only 14 proteins 12 h post induction in wild type cells (≥1.5-fold difference at p ≤ 0.05, Supplementary Fig. 9 ), including PKAC1/2 and metacaspase 4 (MCA4) 53 . The ∆ pkar /∆ pkar mutant trypanosomes served as negative control. As expected, PKAC1/2 levels decreased rapidly, resulting from ligand-induced dissociation of the holoenzyme complex and instability of the free C subunit, as reported for mammalian PKA 37 . MCA4, for which an antibody was available, informed on differential effects of 7-CN-7-C-Ino and Toyo: whereas both inducers decreased MCA4 abundance in wild type but not in ∆ pkar /∆ pkar cells 12 h post induction, only Toyo elevated MCA4 at later time points in a PKA-independent fashion (Fig. 5d ). This is compatible with observed PKA-independent effects of Toyo on growth (Table 1 ) due to multiple cellular targets. 7-CN-7-C-Ino, on the other hand, does not produce these PKA-independent effects on growth, viability, or visible phenotype. This compound is therefore proposed as novel and specific tool to study in vivo the essential processes regulated by PKA in trypanosomes. Discussion Functionally and biochemically uncharacterized PKAs are regularly annotated as cyclic AMP-dependent protein kinases due to high conservation in the eukaryotic kingdom. Here, we have identified novel potent activators of cAMP-independent trypanosome PKA and unambiguously show that ligand selectivity has evolved away from cAMP. This challenges the current view that all PKA orthologues are cAMP-dependent and explains why earlier attempts to detect cAMP-stimulated kinase activity 23 , 24 or binding of cAMP to recombinantly expressed domain fragments of trypanosomal PKAR orthologs 22 , 27 remained negative. The CNB domain seems to be a more versatile ligand-binding domain, not limited to cyclic nucleotides, as reported for some distant CNB family members in prokaryotes 3 , 4 . We expect that systematic surveys will identify novel eukaryotic CNB domain binding specificities, not only in phylogenetically distant trypanosomatids. Several conserved amino acids in the CNB domains of PKAR from T. brucei and related kinetoplastids depart from the consensus (see Fig. 4f, g ), as noted earlier 22 , 27 , 31 . Yet, the amino acid sequence alone did not allow the prediction of altered ligand specificity. For example, the consensus arginine interacting with the exocyclic phosphate of cAMP in almost all PKAs (R333 in bovine RIα) is one of the key replacements in kinetoplastid PKAR, but the mammalian RIα mutant of that residue retained cAMP activation at an only 5-fold decreased EC 50 54 . Our co-crystal structure shows that glutamate residues (E312/E436) present in both CNB pockets of TcPKAR may clash with the negatively charged phosphate of cAMP (Supplementary Movie 5 ) but strengthen the interaction with the ribose moiety of 7-CN-7-C-Ino. In addition, an important role of the αD helix, an extra helix only present at the C-terminus of trypanosome PKAR is suggested. These features are conserved in the Trypanosoma species T. brucei and T. cruzi that share the unusual ligand binding. The high affinity of the 7-deaza nucleoside analogues can be explained by interactions with side groups on the purine ring. For example, a hydrogen bond donor–acceptor pair (Y486/N–K482/O) from the backbone of the trypanosome-specific αD helix (see Fig. 4e ) favours the interaction with a ligand having an acceptor–donor pair at positions 1 and 6 of the purine ring. This is the case for the hypoxanthine-like purine ring in 7-CN-7-C-Ino (Fig. 4e , Supplementary Fig. 7b ) but not the adenine-like purine ring in toyocamycin. In silico docking shows that the bulky cyano-, iodo-, or bromo- side groups fill a small hydrophobic pocket formed at the interface between the beta barrel (V444) and the αD helix (V489, Y485, Y486) in CNB-B, and correctly predicts the relative affinities of these analogues (Supplementary Fig. 7c ). The low potency of sangivamycin can also be explained by the incompatibility of its bulky and hydrophilic side group with the small hydrophobic pocket. Interestingly, the second highest affinity to the B-site is predicted for tubercidin, while the kinase activation potency of this ligand is second lowest. Tubercidin has no side group at position 7; its ribose ring can sample more conformations and therefore possibly score higher in docking. The missing contact to the αD helix is however correlated with low potency. This suggests a role of this trypanosome-specific helix not only for binding but also for the conformational change that underlays the activation mechanism. Only recently, the critical residues for cyclic nucleotide selectivity (cAMP/cGMP) between PKA and PKG have been defined 52 , 55 . Mutations G316R/A336T in CNB-B convert hRIα from low to high cGMP affinity. Of note, the residue corresponding to A336 in hRIα is V444 in TcPKAR and is part of the hydrophobic pocket formed between beta barrel and αD helix. This pocket accommodates the substituent at the C7 position of our activators and may determine their selectivity. The structural alignment between TcPKAR and bovine PKARIα shows conservation of the general polypeptide chain folding and of the putative salt bridge and capping residues linked to the extensively studied allosteric activation mechanism in mammalian PKARIα 1 , 30 . The rapid decrease of PKAC1 levels upon treatment of cells with 7-CN-7-C-Ino (Fig. 5d , Supplementary Fig. 9a ) indicates that kinase activation involves holoenzyme dissociation in vivo. The paradigmatic features of the signalling protein PKA seem to be conserved and adapted to different activating ligands by minor substitutions in the PBC in concert with the trypanosome-specific αD helix extension in CNB-B. It follows that some of the “orphan” CNB domains identified in eukaryotic genomes 2 , 31 , 56 , 57 may be regulated by unexpected and novel ligands. Such discoveries will be as insightful as was the discovery of ligands for other classes of orphan receptors, e.g. from the G-protein coupled receptor 58 or nuclear hormone receptor 59 families. Our report on this unique cAMP-independent PKA may appear to question the role of cAMP as second messenger in trypanosomes. Homologs of other known effector proteins like Epac or cyclic nucleotide gated ion channels were not found in the T. brucei genome. Yet, cAMP and the enzymes for its production and degradation are present and play vital roles in cell division and host immunity subversion 12 , 17 , 18 , 19 , 20 . We have previously identified a number of potential cAMP response proteins (CARPs) by genome-wide RNAi library screening in T. brucei 29 . CARP1 has three CNB domains and the T. cruzi ortholog 60 binds cAMP. Interestingly, CARP1 is exclusively found in kinetoplastid genomes and possibly coevolved with the ligand swap of PKA away from cAMP. There are also additional candidate genes with predicted CNB domains, some of which may be cAMP effectors 41 , 60 . Divergent CNB domains as found in cAMP-binding Popeye-domain proteins 61 might be present but not recognized due to limited homology. Together, the cAMP-independence of PKA in trypanosomes is well compatible with cAMP signalling in these organisms. 7-CN-7-C-Ino is a potent activator and has low toxicity in vivo, in contrast to the 7-deazaadenosine analogues. The latter compounds are known to have pleiotropic effects in various cell types by interfering with cellular processes involving nucleosides or nucleotides. This comprises inhibition of kinases and inhibition of DNA, RNA, and protein synthesis (reviewed in ref. 47 ) and explains the broad antibiotic and antineoplastic activity profile. A possible reason for the reduced off-target effects of 7-CN-7-C-Ino might be lower in vivo stability of the drug. This should not affect the short term in vivo kinase reporter assay but reduce off-target effects on growth and viability. In trypanosomes, toxicity of the 7-deazaadenosine analogues is clearly not mediated by PKA, as we find no difference between WT and ∆ pkar/ ∆ pkar cells. A base level of catalytic PKA function is essential for growth and viability of T. brucei (Supplementary Fig. 2 ). Therefore, efficient inhibition of PKA activation by trapping the catalytic subunits in inactive holoenzyme complexes would reduce the free kinase to lethal level. The challenge in view of possible antiparasitic drug development will be to turn our novel activators to PKA inhibitors similar to what had been achieved for mammalian PKA 62 . As this work shows, TbPKA is essential, druggable, and has a ligand specificity differing from PKAR of the mammalian host. Due to conservation of the substitutions in the CNBs in different kinetoplastid PKAs, such inhibitors may also target the orthologues of related kinetoplastid pathogens like Leishmania . With 7-CN-7-C-Ino, we contribute a potent chemical tool that will spur further investigations on the cellular PKA targets and physiological functions in trypanosomes. Our 7-CN-7-C-Ino-induced phosphoproteome returned a surprising number of highly upregulated P-sites, mostly (77%) within PKA consensus motifs. The GO terms most enriched among the target proteins are related to posttranscriptional control of gene expression and signalling. PKA phosphorylates directly or indirectly 17 other kinases, indicating interconnection with other signalling pathways. Participation in posttranscriptional control is not surprising as many RNA binding proteins and regulators of RNA stability or translation are regulated by phosphorylation in many organisms 63 . In trypanosomes, gene expression is almost exclusively controlled at the posttranscriptional level 64 . Therefore, PKA that participates in transcriptional regulation in other eukaryotes, may be redirected to the targets that exert gene expression control in trypanosomes. The cytokinesis phenotypes observed upon genetic perturbation of TbPKA (Supplementary Fig. 2 and ref. 35 ) correspond well to enriched GO terms related to cell division and cytokinesis. The enrichment for cytoskeletal structures and the flagellum perfectly correspond to the subcellular localization of PKAR in the flagellum 65 , a motility phenotype 65 and cytokinesis, a process that depends on the flagellum 66 . Surprisingly, the quantitative proteome analysis upon induction with 7-CN-7-C-Ino returned only 14 proteins significantly regulated in abundance after 12 h. Among them, downregulation of MCA4, an unconventional metacaspase regarded as pseudopeptidase 53 is interesting, since T. brucei mutants with homozygous deletion of MCA4 have reduced virulence in animal infections 53 . The in vivo pathways and mechanisms upstream of PKA in trypanosomes are completely elusive to date. Is trypanosome PKA the effector of an unknown alternative second messenger? The direct activators of T. brucei PKA identified by this work include tubercidin, toyocamycin, and sangivamycin, natural antibiotics that are secondary metabolites of Streptomyces strains 67 . However, it seems unlikely that 7-deazapurines are the endogenous PKA ligands, since no homologs of the Streptomyces genes encoding enzymes required for synthesis of the precursor preQ0 68 were detected in the T. brucei genome. Of course, an alternative route of synthesis of a 7-deazapurine cannot be excluded. As we find indirect activation of T. brucei PKA by dipyridamole in live cells, this antimetabolite may perturb the parasite’s nucleoside metabolism, e.g. by blocking uptake of adenosine 69 . Nucleoside-related metabolites or by-products of purine salvage may have adopted the role of second messenger-like molecules targeting PKA in trypanosomes. The identity of the physiological PKA ligand and the respective pathway is an exciting line of current research. Methods Trypanosome culture conditions Bloodstream forms of the monomorphic Trypanosoma brucei brucei strain Lister 427, variant MiTat 1.2 70 , were cultivated at 37 °C and 5% CO 2 in modified HMI-9 medium 14 supplemented with 10% (v/v) heat-inactivated fetal bovine serum (FBS). Cell lines 13–90 71 or 1313–514 72 expressing T7 polymerase and Tet repressor were kept under continuous selection with 2.5 µg ml −1 G418 and 5 µg ml −1 hygromycin B or 0.2 µg ml −1 phleomycin and 2 µg ml −1 G418, respectively. Immunopurifications PTP purification was performed according to the protocol of Schimanski et al. 73 with a few modifications. Briefly, trypanosomes expressing PTP-tagged PKAR from the endogenous locus were lysed in PA-150 buffer (w/o DTT, supplemented with Complete Mini EDTA-free protease inhibitor cocktail (Roche) and 25 µg ml −1 pepstatin A) by three sonication cycles 30 s each with the Bioruptor device (Diagenode) at low power. After centrifugation (20 min, 20,000 × g , 4 °C), the cleared supernatant was incubated with IgG beads (pre-equilibrated with PA-150 buffer) for 4 h to overnight by overhead rotation at low speed. Two washes with PA-150 buffer and one with PBS were followed by elution of proteins bound to the IgG beads by incubation with 2× Laemmli sample buffer (125 mM Tris, pH 6.8, 4% (w/v) SDS, 20% (v/v) glycerol, 10% 2-mercaptoethanol, 0.02% (w/v) bromophenol blue) for 5 min at 95 °C. Immunoprecipitation of Ty1- or HA-tagged PKA subunits was performed by incubation of trypanosomes in lysis buffer (50 mM Tris, pH 7.2, 2 mM EGTA, 150 mM NaCl, 0.2% NP-40, 1 mM NaVO 4, 0.5% aprotinin, 2 µg ml −1 leupeptin, 1 mM PMSF) for 10 min on ice with subsequent clearing of the supernatant by centrifugation (20 min, 20,000 × g , 4 °C). The cleared lysate was incubated with the respective epitope-tag antibody coupled covalently (using the cross-linker DMP (Thermo Scientific) according to the manufacturer’s instructions) or non-covalently to protein G sepharose beads (Amersham Pharmacia) for 1 h to overnight. In vivo PKA reporter assay Trypanosomes were harvested (10 min, 1400 × g , 37 °C) and resuspended in HMI-9 medium (pre-heated to 37 °C) to a density of 5 × 10 7 cells ml −1 . After a 5–10 min recovery at 37 °C with mild shaking, test compounds or solvent were added to the cell suspension followed by careful mixing. After incubation at 37 °C for the specified time period, cells were lysed with 6× Laemmli sample buffer preheated to 95 °C and incubated for 5 min at 95 °C. Incubation of trypanosomes at this density for up to 30 min had no effect on VASP phosphorylation (Supplementary Fig. 3g ). GraphPad Prism 7.0 was used for EC 50 calculation by non-linear regression analysis using an equation for a sigmoidal dose–response curve with variable slope. The same software was used for visualization of all graphs and bar charts. Inhibitors and nucleoside analogues PKI(5–24) and myr-PKI(14–22) were obtained from Biomol. CpdA 19 (now renamed as NPD-001) 74 was synthesized by Geert-Jan Sterk, Mercachem. Dipyridamole was obtained from Sigma-Aldrich, toyocamycin and sangivamycin from Berry & Associates. All other nucleoside and cyclic nucleotide analogues were obtained from Biolog GmbH, Bremen. 7-CN-7-C-Ino was synthesized according to Hinshaw et al. 75 with 125 mg (429.1 µmol) Toyo as starting material. By variation to the original protocol, the raw product was purified by reversed phase medium pressure liquid chromatography (MPLC). Briefly, the raw product was dissolved in 20 ml water, filtrated and applied to a Merck LiChroprep ® RP-18 column (15–25 µM; 125 × 35 mm), previously equilibrated with water. The column was washed with water to remove excess of inorganic salts and hydrophilic impurities. Afterwards, 1% and 2% 2-propanol in water was used to elute the target compound. Product-containing fractions were concentrated by rotary evaporation under reduced pressure and subsequently freeze-dried to yield 89.34 mg (305.7 µmol) of 7-CN-7-C-Ino with a purity of 99.93% by analytical HPLC (ODS-A 120-11, RP-18 (YMC, Dinslaken, Germany); 250 × 4 mm; 9% acetonitrile, 20 mM triethylammonium buffer, pH 6.8; 1.0 ml min −1 ; UV-detection at 265 nm). Leishmania tarentolae expression system The L. tarentolae strain LEXSY T7-TR (Jena Biosciences) was cultivated at 26.5 °C in BHI medium supplemented with 10 µg ml –1 hemin, 100 U L −1 streptomycin and 100 mg L −1 penicillin according to the protocols provided by Jena Biosciences. For maintenance of T7 polymerase and Tet repressor, 10 µg ml −1 nourseothricin (NTC) and 10 µg ml −1 hygromycin B were added to the medium. For co-expression of T. brucei PKAR-10 × His and Strep-PKAC1, the full length ORFs were amplified from genomic DNA using primers introducing the respective epitope tag and cloned into pLEXSY_I-ble3 and pLEXSY_I-neo3 (Jena Biosciences), respectively. Details on primer sequences and cloning strategies are available upon request. Cells transfected with both constructs were cultivated in the presence of 100 µg ml −1 phleomycin and 50 µg ml −1 G418. L. tarentolae cells at mid log phase (2–3 × 10 7 cells ml −1 ) were induced with 10 µg ml −1 tetracycline for 24 h. Lysis of cells in 50 mM Tris, pH 7.4, 150 mM NaCl, 0.2% Triton X-100, 1 mM 2-mercaptoethanol was completed by a Dounce homogenizer. Tandem affinity purification of the holoenzyme complex was performed to guarantee subunit stoichiometry and highest purity: His-tag purification using Ni-NTA beads (Thermo Fisher Scientific) was followed by Strep-tag purification using gravity flow chromatography and StrepTactin sepharose beads (IBA), according to the manufacturers’ instructions. The eluted fractions were pooled and dialyzed against the kinase storage buffer (20 mM MOPS, pH 7.0, 150 mM NaCl, 1 mM 2-mercaptoethanol). In vitro kinase assay A radioactive PKA kinase assay was performed according to Hastie et al. 76 , using 100 µM Kemptide (LRRASLG) as kinase substrate and 100 µM ATP spiked with [γ- 32 P] ATP to give 200–400 cpm pmol −1 . GraphPad Prism 7.0 was used for EC 50 calculation by non-linear regression analysis using an equation for a sigmoidal dose–response curve with variable slope. The same software was used for visualization of all graphs and bar charts. Binding studies by isothermal titration calorimetry N-terminally truncated T. brucei PKAR (aa 199–499) was cloned into pETDuet-1 (Novagene) with an N-terminal His 6 -tag and expressed in E. coli Rosetta (DE3). The bacteria were grown in Luria–Bertani (LB) medium to OD 600 of ∼ 0.4 at 37 °C, followed by overnight induction with 0.4 mM IPTG at 20 °C. Cells were harvested by centrifugation and lysed in a French Press. Protein purification was done by Ni-NTA affinity chromatography, followed by elution in 50 mM HEPES, pH 7.5, 50 mM NaCl, and 250 mM imidazole. Eluted protein ( ∼ 7 mg ml −1 ) was dialyzed overnight in 50 mM HEPES, pH 7.5, 50 mM NaCl, 1 mM DTT, and stored at −80 °C. Ligand-free protein for binding studies was prepared by adding solid urea to the protein solution to a final concentration of 8 M. After 1 h at 25 °C, the solution was passed over a prepacked PD10 column (GE Healthcare) equilibrated with 8 M urea, 50 mM HEPES, pH 8.5, 50 mM NaCl. TbPKAR(199–499) was refolded overnight by dialysis against 50 mM Tris, pH 8.5, 240 mM NaCl, 10 mM KCl, 2 mM MgCl 2 , 2 mM CaCl 2 , 0.4 M sucrose, 1 mM DTT at 4 °C, followed by separation of monomers from aggregates by size-exclusion chromatography on a Superdex 200 Increase 10/300 GL column (GE Healthcare) equilibrated with 50 mM HEPES, pH 7.5, 50 mM NaCl, and 1% DMSO (buffer A). Eluted protein was diluted to 10 µM for ITC. The human PKARIα (full size) was expressed from the pETDuet-1 based plasmid 6H.tev/HsPKARα in E. coli Rosetta D3, prepared cAMP-free according to Buechler et al. 77 with the following modification: the refolding buffer was the buffer used for TbPKAR refolding. The purified protein was diluted to 5 µM. For binding assays, 13–19 injections of 2–3 µl were performed with a MicroCal PEAQ-ITC (Malvern) instrument. In each series, 100 µM of ligand (prepared using buffer A) was injected at 298 K into 5–10 µM protein freshly eluted from size exclusion chromatography. Binding constants and thermodynamic data were derived from best least square fit analysis, applying a model with two binding sites (performed with MicroCal PEAQ-ITC software). Quantitative phosphoproteomics For each sample, 6 × 10 8 T. brucei cells treated or not with 8 µM 7-cyano-7-deazainosine for 15 min were lysed in 300 µl 4% (w/v) sodium deoxycholate, 0.1 M Tris, pH 8.5 for 5 min at 95 °C (according to the protocol of Humphrey et al. 78 ). Samples were sonicated using a Bioruptor (Diagenode) (high power, two cycles of 10 min each, 30 s on/off). Protein concentration was determined by BCA protein assay and samples were adjusted to equal concentrations. Sample preparation and mass spectrometry were exactly carried out as described by Humphrey et al. 78 . MaxQuant 1.5.2.8 79 was used to identify proteins and quantify by LFQ with the following parameters: Database, TriTrypDB-39_TbruceiTREU927_AnnotatedProteins; MS tol, 10 ppm; MS/MS tol, 0.5 Da; Peptide FDR, 0.1; Protein FDR, 0.01 Min. peptide Length, 5; Variable modifications, Oxidation (M), Phosphorylation (STY); Fixed modifications, Carbamidomethyl (C); Peptides for protein quantitation, razor and unique; Min. peptides, 1; Min. ratio count, 2. For proteomic analysis, identified proteins were considered as statistically significant with FDR ≤ 0.05 and s 0 = 1 (two-sided Student’s T -test adjusted for multiple comparisons by Benjamini–Hochberg correction, Perseus 80 ). Phosphopeptide analysis was carried out in Perseus as suggested by Humphrey et al. 78 . The mass spectrometry phosphoproteomics data have been deposited to the ProteomeExchange Consortium via the PRIDE partner repository ( ) with the dataset identifier PXD012245 . GO enrichment analysis was performed in TriTrypDB with default settings and visualized in Revigo ( ). The motif discovery tool MoMo (using the motif-x algorithm) implemented in the MEME suite ( ) was used for unbiased motif discovery in the phosphoproteome dataset with the T. brucei TriTrypDB-40_TbruceiTREU927 protein database as background. Enriched sequence logos were visualized using CLC Main Workbench 7 ( ). Reporting summary Further information on experimental design is available in the Nature Research Reporting Summary linked to this article. Data availability The coordinates of the T. cruzi PKAR crystal structure bound to 7-CN-7-C-Ino have been deposited in the Protein Data Bank under the code PDB 6FTF . The phosphoproteome and proteome datasets are available in the PRIDE partner repository with the dataset identifiers PXD012245 and PXD009073 , respectively. Genome sequence and annotation information was obtained from TritrypDB ( ). Human PKA substrates and phosphorylation motifs were retrieved from the PhosphoSitePlus database ( ). Gene ontology (GO) enrichment analysis was visualized using Revigo ( ). The motif discovery tool MoMo implemented in the MEME suite ( ) was used for unbiased motif discovery in the phosphoproteome dataset. The source data underlying Figs. 1c, d , 2a–h , 3a–c , 5a, c–e , Table 1 , and Supplementary Figs. 1b , 2 , 3a–d, f–i , 4a–h , 8a–c , 9a–c are provided as Source Data file.

The unicellular parasite that causes sleeping sickness differs from other eukaryotes in the mode of regulation of an essential cellular signaling pathway. This provides a promising point of attack for drug development. Trypanosoma brucei, the insect-borne eukaryotic parasite that causes sleeping sickness in tropical Africa, is the best-known representative of a group of unicellular organisms known as kinetoplastids. Species belonging to this group are responsible for a number of potentially lethal infections in humans and other mammals, which are difficult to treat effectively. These include the trypanosome T. cruzi—the causative agent of Chagas disease, which is endemic to much of South America, while members of the related genus Leishmania attack the skin and mucous membranes. Leishmaniasis is a largely tropical disease, but is also found in Southern Europe. Research carried out by Ludwig-Maximilians-Universitaet (LMU) in Munich geneticist Professor Michael Boshart´s team with lead author Dr. Sabine Bachmaier and collaborators has now revealed that, in T. brucei, an essential intracellular signaling pathway is regulated differently from what has been established as a paradigm in eukaryotes. This discovery reveals a potential vulnerability, which could provide a promising target for new and specific therapies for infections caused by kinetoplastid parasites. The findings appear in the online journal Nature Communications. Protein kinase A (PKA) is an enzyme found in virtually all nucleated (eukaryotic) organisms apart from plants, and it plays a crucial role in cellular responses to external signals. In almost all eukaryotes, the enzymatic activity of PKA is dependent on binding of the molecule cAMP, which binds to the regulatory subunit of the protein. PKA is also essential for the survival of Kinetoplastida, as it has been shown to target cell division, cell motility and development in the group. However, all efforts to demonstrate a link between cAMP and PKA activity have failed. "We have known for several years that the PKA found in T. brucei cannot be activated by cAMP, but there have been several contradictory reports in the published literature," Boshart explains. "We have now comprehensively and conclusively verified this unexpected finding experimentally." To do so, the researchers developed an experimental test that allowed them to detect PKA activity in living trypanosomal cells. As in their earlier test-tube experiments, they found that there was no change in PKA activity when the intracellular level of cAMP was altered by chemical or genetic means. "In collaboration with a chemist who works for a small company, we tested an array of chemical compounds which we had identified as possible alternative regulators of the kinase enzyme, based on our experimental data," Boshart says. The search identified a number of candidate molecules that were able to activate the typanosomal PKA. "We then chose the most effective candidate from this set and optimized its performance by appropriately modifying its chemical structure," says Boshart. In this way, we obtained a compound that acted as a specific and highly potent activator of the enzyme." With the aid of 3-D structural analyses, they went on to show that the activator binds to the trypanosomal enzyme at the site in the regulatory subunit to which cAMP binds in PKAs from other eukaryotes. However, the detailed architecture of the binding pocket differs between the two proteins. The differences involve only two or three amino acid substitutions, but these changes suffice to ensure that cAMP no longer fits—while the synthetic alternatives fill the cavity neatly. With these alternative activators, Boshart and his colleagues now have useful tools at hand with which to probe the function(s) of PKA in the Kinetoplastida and identify the enzyme's target proteins. In addition, the results are of great interest from the therapeutic point of view. With appropriate modification, the parasite-specific activator could possibly be converted into an inhibitor of the trypanosome's PKA, without affecting the function of the mammalian host's enzyme.

10.1038/s41467-019-09338-z

Medicine

10-fold speed up for the reconstruction of neuronal networks

Kevin M Boergens et al. webKnossos: efficient online 3D data annotation for connectomics, Nature Methods (2017). DOI: 10.1038/NMETH.4331 Journal information: Nature Methods

http://dx.doi.org/10.1038/NMETH.4331

https://medicalxpress.com/news/2017-06-fold-reconstruction-neuronal-networks.html

Abstract We report webKnossos, an in-browser annotation tool for 3D electron microscopic data. webKnossos provides flight mode, a single-view egocentric reconstruction method enabling trained annotator crowds to reconstruct at a speed of 1.5 ± 0.6 mm/h for axons and 2.1 ± 0.9 mm/h for dendrites in 3D electron microscopic data from mammalian cortex. webKnossos accelerates neurite reconstruction for connectomics by 4- to 13-fold compared with current state-of-the-art tools, thus extending the range of connectomes that can realistically be mapped in the future. Main With the acceleration of 3D electron microscopic (EM) imaging of brain tissue 1 , 2 , 3 , image data sets sized tens of terabytes (TB) or even petabytes (PB) are becoming available. A cubic millimeter imaged at (15 nm) 3 voxel (vx) size corresponds to 0.3 PB of data ( Fig. 1a ); a mouse brain imaged at the same resolution corresponds to 110 PB of data. Single neurons typically extend over a large fraction of the data set ( Fig. 1a ), making it impracticable to distribute data on hard drives to large numbers of annotators who want to follow the processes of entire neurons. At the same time, data analysis in connectomics is limited by the amount of human annotation time that can be recruited for a given analysis project 4 . Thus, enabling efficient distributed 3D data annotation in PB-sized data sets, ideally in browser, is essential. Figure 1: In-browser 3D annotation of axons and dendrites for connectomics. ( a ) Sketch of mouse whole-brain 3D EM data set size (top) and 1 mm 3 of cerebral cortex (bottom) compared with the extent of a typical single pyramidal neuron (dendrites, magenta; axon, black). ( b ) Sketch of online data delivery modes using lateral prefetching in 2D (top, Google maps, CATMAID) and in 3D (bottom, webKnossos). ( c ) Comparison of data transmission when following a neurite using 2D image prefetching (black, in CATMAID) and 3D data prefetching in webKnossos (blue, ortho mode) under different bandwidth and latency conditions. Transcont, transcontinental access ( from Europe, custom CATMAID and webKnossos from South America, bottom to top). ( d ) webKnossos in-browser user interface with orthogonal viewports ( xy , yz , xz ), one 3D skeleton viewport, and the abstract tree viewer (right). ( e ) Sketch of viewing surface orientation (red) in orthogonal mode (top) and flight mode (bottom). ( f ) Flight-mode egocentric 3D image sampling on a hemisphere (top), yielding a single flight-mode data view (bottom). ( g ) Example of 3D prefetching in flight mode given flight direction and current position (gray, prefetched webKnossos cubes; red, flight-mode image surface). ( h ) Annotator speed training in ortho mode (black, n = 25 annotators) and flight mode (magenta, n = 26 annotators) on 40 neurites in cortex (randomly ordered per annotator). ( i ) Tracing speed test on 20 randomly selected cortical axons (including branches) performed by 26 annotators 8 weeks after training. Dashed line, first five neurites presented again from training. Solid line, 20 test axons. Box plots report tracing speed for these 20 test axons in ortho (black) and flight mode (magenta), reported over n = 26 annotators (left) and n = 20 axons (right); flight 1.51 ± 0.04 mm/h, ortho 0.96 ± 0.03 mm/h ( n = 520, mean ± s.e.m.). ( j ) Display of 20 test axons within data set boundary. ( k ) Illustration of tracing error measurement in one of the 20 axons—local errors (inset, less than 10-μm path length) and continuation errors (right). Black, ground truth; red, six-fold consolidated flight-mode tracing. ( l ) Tracing errors reported over tracing redundancy (using RESCOP 8 ) for 10 axons in ortho mode, flight mode and for 10 dendrites in flight mode (cyan), respectively. ( m ) same as l but only continuation errors (see k ). Dashed lines in l and m , path length corrected for each RESCOPed skeleton (see Online Methods ). ( n ) Relation between tracing speed and error rates for single-annotator reconstructions ( n = 30) of axons and dendrites (colors as in l and m ). Crosses indicate mean ± s.e.m. ( o ) same as n for continuation errors only. ( p ) Summary comparison of annotation time requirements for neurite reconstruction. Data from c , (black, webKnossos, ; red, CATMAID) and from i and n (crosses). Asterisk, annotation consumption documented in published work; pyK, consumption by experienced annotators 10 ; K, KNOSSOS annotation consumption from refs. 8 , 11 and 12 . Source data Full size image The existing in-browser annotation tool for connectomics, CATMAID 5 , 6 , uses efficient data storage and transmission in 2D image planes (comparable to Google Maps 7 ), which are sequentially browsed ( Fig. 1b ). While this approach makes data viewing and annotation seamless in the plane of imaging, 3D neurite tracing is slowed down by the time required to progress to the subsequent image plane. Under ideal high-bandwidth, low-latency connectivity conditions (like those within research institutions), this approach yields a reconstruction speed of about 470 μm/h when an expert follows an axon through 3D serial blockface EM (SBEM 1 ) data from mouse cortex ( Fig. 1c ). Under less optimal bandwidth and latency conditions (as are often experienced by student annotators at home and in mobile settings), however, reconstruction speed drops to 59–88 μm/h (measured for regular 3G connectivity and for transcontinental access, Fig. 1c ). webKnossos ( Fig. 1d , ) uses 3D data storage and transmission in small cubic packages of 32 3 vx ( Fig. 1b ). Cubic 3D image data storage using cubes of 128 3 vx is being employed in KNOSSOS 8 and pyKNOSSOS 9 , 10 , standalone data annotation applications for connectomics. We reduced 3D cube size to 32 3 vx for lag-free in-browser data transmission and for enabling the flight-mode data visualization introduced below. webKnossos enables data interaction in 3D EM images displayed in orthogonal planes ( Fig. 1d ) at a speed of about 2 mm/h ( Fig. 1c ), which drops to 0.7–1.2 mm/h for regular 3G connectivity and transcontinental annotation ( Fig. 1c ). Thus, 3D data visualization when following an axon is about 4-fold faster under ideal and up to 13-fold faster under nonoptimal connectivity settings than with existing in-browser tools ( Fig. 1c ). We next tested whether student annotators can be trained to interact with 3D brain image data at such speeds. Previously, annotators interacted with the image data using orthogonal image projections in the three cardinal planes for following the neuronal processes and for their annotation as 'skeletons' ( Fig. 1d,e , 'ortho mode', KNOSSOS 8 ). Effective tracing speed was 100–270 μm/h 4 , 8 , 10 , 11 , 12 for reconstructions in mouse retina, zebrafish olfactory bulb and mouse cortex. We asked whether annotators can be trained to annotate faster in ortho mode, and whether a more intuitive data presentation can further accelerate human annotation. For the latter, we developed 'flight mode', in which the 3D image data are sampled on a hemisphere centered at the annotator's current position ( Fig. 1f ). To enable such nonorthogonal data transmission and display in browser, we used (in addition to the small 3D cube size ( Fig. 1b )) a simple form of path prediction (the data being loaded in a stump in direction of flight; Fig. 1g ). Flight mode requires the EM image data to be sufficiently well aligned in 3D, as is routinely the case for neuronal tissue imaged using SBEM 1 . In flight mode, the annotator focuses on centering the target cursor onto the axon or dendrite being followed, steering the orientation with the mouse or keyboard while moving forward ( Supplementary Video 1 ). We suspected that this focusing on one intuitive egocentric visualization and interaction may accelerate annotation, since the user does not have to explicitly recenter the viewport and switch image plane orientation for processes running in off-axis directions ( Fig. 1e ). To investigate whether flight mode in fact accelerates human 3D image data annotation, we trained 51 student annotators on 40 neurites in a 3D EM data set from mouse cortex sized 93 × 60 × 93 μm 3 , imaged at 11.24 × 11.24 × 28 nm 3 using SBEM (data not shown; Fig. 1h ). The sequence of neurites to trace was shuffled for each student (total path length of all 40 neurites, 6.11 mm; number of branch points, 258; overall branch point rate, 42 per mm). The student annotators came from our pool of trained annotators and were thus experienced in connectomic data reconstruction ( Supplementary Fig. 1 ). They were randomly assigned to two groups of 25 and 26 students, respectively. One group was asked to trace in ortho mode, the other in flight mode. We supplied both groups with an 8.5-min tutorial movie (a separate movie for each group), which encouraged them to increase their movement speed whenever they felt they could go faster during reconstructions ( Supplementary Fig. 1b,d ). To enable constantly maximized tracing speed for each annotator, we automatically tracked the fraction of time during which the annotators proceeded at the preset movement speed (i.e., when holding the forward key pressed while navigating) and reminded them to increase their preset movement speed if they spent more than 75% of their tracing time constantly pressing the forward key. In ortho mode ( Fig. 1h ), annotators initially traced at 0.49 ± 0.04 mm/h (first ten processes, mean ± s.e.m.) and accelerated to 0.63 ± 0.05 mm/h (last ten processes traced, P < 10 −3 , Wilcoxon signed-rank test). In flight mode ( Fig. 1h ), annotators started at a speed of 0.84 ± 0.05 mm/h (faster than ortho mode, P < 10 −4 , Wilcoxon rank-sum test) and were able to increase their speed to an average of 1.11 ± 0.07 mm/h ( P < 10 −4 , Wilcoxon signed-rank test), 1.8 times faster than ortho mode tracing ( P < 10 −5 , Wilcoxon rank-sum test). These data indicate that a substantial reconstruction speed gain comes from per-user speed maximization, enabled by efficient 3D data handling, and an additional gain from the egocentric flight mode data interaction. To test whether this tracing speed can be routinely achieved for axons in mammalian cerebral cortex ( Fig. 1i,j ), we next randomly selected 20 axons from a (2.5 μm) 3 region in the same data set ( Fig. 1j ) and asked students trained in ortho mode to again use ortho mode and students trained in flight mode to again use flight mode. In each group, 13 of the trained students participated in this second experiment 8 weeks after the training (their initial training performance had been indistinguishable from the whole group, P > 0.24, Wilcoxon rank-sum test). We first presented five of the processes reconstructed during training to calibrate the persistence of the training effects, and then we presented the 20 new axon seeds in random order to all tracers ( Fig. 1i ; note that the five neurites from the training session were not included in the final speed measurement). Tracers resumed annotation at the speed attained during training and were able to further accelerate, yielding a reconstruction speed of 0.96 ± 0.03 mm/h in ortho and 1.51 ± 0.04 mm/h in flight mode ( Fig. 1i , mean ± s.e.m., n = 20 previously unseen randomly selected cortical axons; total path length, 2.53 mm; overall branch-point rate, 39 per mm). But were faster tracings more error prone? We next quantified the rate of errors for 10 randomly drawn axons out of the 20 test axons. For each axon, we manually counted the number of incorrect stops and incorrect continuations ( Fig. 1k , performed by two expert annotators blinded to tracing mode, see Online Methods ) and distinguished between errors yielding continuation mistakes (i.e., a premature stop or missed branch of a major part of the axon) or local errors (yielding less than 10 μm neurite loss or neurite addition). Figure 1l reports the rate of errors for flight-mode and ortho-mode tracings for single-annotator reconstructions and for consolidations of multiple reconstructions of the same axon (consolidated using RESCOP 8 ). We found that, first, the rate of tracing errors was not distinguishable between ortho and flight mode tracings ( P > 0.34 for all errors, P > 0.28 for continuation errors, Wilcoxon rank-sum test). Second, the average error rate for single-annotator reconstructions obtained at the achieved speed in webKnossos was not worse than the error rates reported previously in mouse retina 8 , 11 and cortex 13 . The rate of continuation errors ( Fig. 1m ) was 7.5 ± 3.4 per mm in ortho mode and 5.3 ± 3.0 per mm in flight mode for single-annotator tracings. We finally asked whether a speed–accuracy tradeoff could be observed in either of the tracing modes. For this we correlated the rate of errors with tracing speed in single-annotator tracings ( Fig. 1n ). No positive correlation could be found for ortho mode nor for flight mode (ortho r = −0.5, P = 0.007, flight r = −0.4, P > 0.05, Pearson's correlation). This also was true when only analyzing the continuation errors ( Fig. 1o , ortho r = −0.34, P > 0.05, flight r = −0.20, P > 0.28, Pearson's correlation). We thus conclude that annotators can be trained to trace cortical axons at 1.51 ± 0.04 mm/h in flight mode without a reduction in accuracy. webKnossos can support this speed online and in browser, and it provides a 6- to 15-fold improvement over published tracing speeds, depending on the reconstructed data sets (summarized in Fig. 1p ). To determine tracing speeds for dendrites, we reconstructed the shafts of ten randomly chosen dendrites (1.8-mm total path length, overall branch point rate 16 per mm) and measured tracing errors as described above. Since dendrites are about three times wider in diameter 14 , annotators could zoom out further and fly along dendrites faster than they do along axons. We found that tracing speed was 2.11 ± 0.16 mm/h including branch-point reconstruction, and single-annotator error rates were 2.7 ± 0.69 errors per mm dendrite ( Fig. 1l–o ). In mammalian brains, which constitute a main challenge of connectomics, about 90% of the neuronal processes are axons 14 , 15 . The speed gain for representative axon reconstruction in flight mode was therefore critical for the acceleration of connectomic reconstruction in mammalian cortex. However, the reconstruction of connectomes additionally requires the identification of synapses and the assignment of postsynaptic partners to the respective neuronal cell bodies and dendrites. Figure 2a,b illustrates a workflow for such connectome reconstruction. In this workflow, axons and dendritic shafts are reconstructed first (including branch points). Then a synapse movie mode is activated in webKnossos; in this mode, the user can fly along the pretraced axon and click into the postsynaptic process whenever a synapse is encountered. In the final step, the postsynaptic partner (in about 90% a spine head, Supplementary Fig. 2a ) is seeded for annotators to trace back to the main shaft of dendrites. In this workflow, the fraction of time spent on synapse annotation is small for sparse reconstructions (7–11% for typical network reconstructions, Fig. 2c , Supplementary Fig. 2b ) but is becoming more substantial for dense reconstructions, approaching about 50% of reconstruction time ( Supplementary Fig. 2b ). Since synapse detection requires only a local image classification (unlike neurite tracing), automated synapse detection is likely to soon replace manual synapse detection in dense connectome reconstructions (e.g., refs. 16 , 17 , 18 , 19 ). To exemplify a full connectomic reconstruction using webKnossos, we finally reconstructed 497 dendrites (total path length of 93.6 mm, tracing redundancy 3), and determined all synapses with 32 axons from the training set ( Fig. 1j , 4.55 mm path length of axons, tracing redundancy 6). We detected 104 synapses in this local connectome ( Fig. 2d,e ) (total annotation time was 27.3 h for axons, 133 h for dendrites, and 19.2 h for synapses). Figure 2: Connectome reconstruction using webKnossos. ( a ) Flow chart and illustration ( b ) of connectome reconstruction steps. Ves. cloud det., vesicle cloud detection; syn. partner, identification of postsynaptic partner. ( c ) Comparison of annotation times for axons, dendrites, spines and synapses including the speed gains for neurite reconstruction in webKnossos ( Fig. 1 ) for the local connectome shown in panels d and e and for a sparse example reconstruction (estimate, see inset, 100 layer-4 cortical axons innervating 300 layer-2/3 pyramidal cells within a cubic millimeter of cortex tissue, amounting to 2 m total path length). ( d ) Reconstruction of 497 dendrites and 32 axons in local SBEM data set from mouse cortex following the workflow in a . Colored spheres indicate excitatory (Exc., violet) and inhibitory (Inh., red) synaptic contacts (syn.). ( e ) Resulting connectome between 32 presynaptic axons (presyn.) and 497 postsynaptic (postsyn.) dendrites (only the 70 innervated dendrites are shown). Source data Full size image In summary, webKnossos accelerates human 3D data interaction for EM-based connectomics in browser by about 4- to 13-fold, which likely saturates human interaction speed with 3D EM data of nervous tissue using flight mode. While tested on well-aligned 3D SBEM data from mammalian cortex, these results are expected to be comparable for other neuropil with comparable neurite morphology (especially branch-point rates, e.g., in subcortical structures and ganglion cells in mammalian retina, see ). Reconstructions in highly anisotropic and potentially less well-aligned image data can still profit from the speedup because of faster display rates in ortho mode tracings (see Fig. 1i ). Thus webKnossos can serve as a versatile high-efficiency tool for 3D image data annotation in various 3D image analysis settings in connectomics and other fields. Methods 3D SBEM data, animal experiments. The 3D EM image data was acquired using serial blockface electron microscopy (SBEM 1 ) from primary somatosensory cortex layer 4 of a P28 C57BL/6 male mouse, same data set as in ref. 12 (93 × 60 × 93 μm 3 , 11.24 × 11.24 × 28 nm 3 resolution, data set 2012-09-28_ex145_07x2_new2). All animal experiments were carried out with approval of the local animal research authorities (Regierungspräsidium Oberbayern, Germany) and in accordance with the German Animal Welfare Act. Image data transmission. For efficient volume data transmission, data is requested in small cubic packages 32 3 voxel in size ('buckets') stored at original 8-bit depth. For bandwidth-limited settings, each voxel in a bucket is trimmed to the 4 most significant bits for transmission ('4-bit mode', user activatable). Buckets are requested along a priority ranking based on the current view point and direction of movement. In ortho mode, all buckets in the plane of the active viewport are loaded first, prioritized by the Manhattan distance to the viewport center. The buckets of the next two bucket layers in direction of movement are prioritized next, with two- and four-fold-reduced priority, respectively. All buckets are loaded at the user-specified magnification level. In flight mode the preview volume is a square frustum with basis sized 5 × 5 buckets, height of 2.5 buckets, oriented along the current movement direction, and top sized 4 × 4 buckets. All buckets fully or partially contained in this volume are requested at original magnification, prioritized by the Manhattan distance between the respective bucket center and the current viewpoint. The prioritized bucket request queue is updated on each user movement. Measurement of 3D data transmission speed: webKnossos. The speed of data transmission for 3D navigation in webKnossos ( Fig. 1d ) was measured as follows. webKnossos was run at a Hetzner (Gunzenhausen, Germany) data center on a server with the following specifications: Intel(R) Xeon(R) CPU E3-1245 V2 (4 × 3,4 GHz); 32 GB RAM; 15 × 3 TB HWRaid HDD. The EM data set was viewed in 4-bit mode in webKnossos run in Google Chrome (version 56) on a computer in the MPI for Brain Research. A neurite was picked and followed using the forward and arrow keys, keeping the forward key pressed where possible (which resulted in continuous image stream in webKnossos). To emulate reduced connectivity settings, the developer tools function of Google Chrome was used. The 'transcontinental' experiment ( Fig. 1c ) was performed on a computer connected to the network of the Instituto de Investigación en Biomedicina de Buenos Aires, Argentina, accessing the webKnossos instance running on the Hetzner server in Germany (see above). Measurement of 3D data transmission speed: sequential 2D. The speed of data transmission for sequential 2D image navigation ( Fig. 1c ) was measured as follows. We followed the instructions by the CATMAID authors to optimize server performance (published under ). A CATMAID instance (version 2016.12.16) was installed on a server in the compute center of the Max Planck Society (Garching, Germany) with the following specifications: Xeon E5-2630 12 cores, 128 GB RAM, 10 Gb network, JBOD of 4× Intel DC S3500 240 GB SSDs, Ubuntu 14.04. Postgres and data partitions resided on SSDs; XFS was used with noatime. The 3D image data set 2012-09-28_ex145_07x2_new2 (s. above) was converted to a series of 256 × 256 px jpg images, compressed by 75% with jpg headers removed to further reduce file size as suggested by the CATMAID authors. The data were resliced into three image series along the three cardinal directions. The 'transcontinental' experiment was performed by accessing CATMAID on from a desktop computer in the MPI for Brain Research and by accessing the custom CATMAID instance at the Max Planck datacenter in Germany from a computer connected to the network of the Instituto de Investigación en Biomedicina de Buenos Aires, Argentina. All tests were performed in Google Chrome (version 56). The viewports in CATMAID and webKnossos were set to similar size. Annotator training. For training annotators at high-speed annotation ( Fig. 1h ), 40 training neurites were selected from the cortex data set (see above). For this, a bounding box sized 4.5 × 4.5 × 4.2 μm 3 was chosen. Then, two annotators were asked to reconstruct all processes within this bounding box. Next, each of the reconstructed processes was classified as axon, dendrite or glia. Finally, 40 of the 68 processes classified as axons were randomly selected. For each process, an expert annotator defined a starting position and a starting direction (required for flight-mode annotation). 51 annotators were trained. These annotators were randomly assigned to two groups (flight ( n = 26) and ortho mode ( n = 25)). Annotators were asked to watch an introductory video, which instructed them to increase their maximum velocity setting (the speed at which the annotator progresses through the data when the space key is held down continuously). Each annotator was presented with the 40 training processes in random order. The annotators' preset maximum velocities were monitored during the annotation process. If the ratio of tracing speed and preset maximum velocity was higher than 0.75 for an entire tracing, the annotator was notified via e-mail and asked to increase the maximum velocity setting for the next annotation. Axon test reconstruction. For the test of axon reconstruction speed ( Fig. 1i ), the 51 previously trained annotators were asked to reconstruct 20 randomly selected axons 8 weeks after the initial training. 26 annotators signed up for this experiment (13 that had been trained on flight mode and 13 that had been trained on ortho mode). These annotators had not been faster in the final ten training iterations than the whole group of annotators ( P = 0.246 (Wilcoxon rank-sum test) und P = 0.250 ( t -test)) and had not been faster in the final training iteration ( P = 0.699 (Wilcoxon rank-sum test) and P = 0.649 t -test). To select a set of representative axons, a (2.5 μm) 3 bounding box (located randomly within the cuboid of 15 μm edge length centered to the data set center) was chosen that did not contain a soma, and all neuronal and glial processes within this bounding box were reconstructed by one expert annotator. Then all processes were classified as axonal, dendritic or glial. Three additional expert annotators proofread the annotation. Then, 20 of the 41 processes labeled as axonal were randomly selected, and for each axon a seed position and initial orientation were defined within the bounding box. The 26 annotators were first asked to again reconstruct five neurites from the training experiment (these five neurites were randomly chosen from the 40 training seeds and were the same for all annotators; the sequence in which these were presented was randomized for each annotator). Then each annotator was asked to reconstruct the 20 previously unseen test axons (in a sequence randomized per annotator) in flight mode or ortho mode. After all annotators had finished, all annotations were automatically scanned for open branch points (i.e., positions at which the annotator had set a branch point flag but had forgotten to jump back to for inspection) and seed nodes with a degree of 1 (i.e., starting points which had only been traced in one direction). 17 open branch points (12 at first node) and 22 unidirectional seeds were detected (5 in flight, 17 in ortho) of 2,170 fully annotated branch points, total. In these cases the annotators were asked to go back to the task and continue their annotations. The code for this automated annotation checking is provided in Supplementary Software 1 . Dendrite reconstruction. 497 dendrites were reconstructed in flight mode by 47 of the annotators previously trained in ortho or flight mode (see above). Those annotators that had worked in ortho mode before were asked to watch the instruction movie for flight mode before performing the dendrite reconstructions. Annotators were instructed to set the data set quality setting in webKnossos to medium (which means that image data is displayed at lower resolution) and not to reconstruct spines. The 497 dendrite seeds for the connectome reconstruction ( Fig. 2 ) were drawn randomly from a set of over 2,000 dendrites that had previously been reconstructed using webKnossos. For all dendrites, the z -axis pointing toward the data set center was used as initial flight orientation. Measurement of annotation speed. For measurement of annotation speed, the path length of a given neurite and the time it took to annotate that neurite were determined. To measure neurite path length from a skeleton annotation, the lengths of all edges within a skeleton were summed (as in refs. 8 and 11 ). However, this method has two caveats. First, noise in the placement of skeleton nodes will be biased to only increase apparent skeleton length, not decrease it, which could potentially lead to an overestimation of annotation speed. Second, this effect will depend on the density of placed skeleton nodes. Since in flight mode the skeleton nodes are placed automatically, the density of skeleton nodes is substantially higher in flight mode than in ortho mode tracings (flight, 6.3 ± 0.68 nodes/μm; ortho, 1.79 ± 0.67 nodes/μm, measured on the ten axons used for Fig. 1l–o ). To account for these potential biases, we first used nonuniform rational b-spline (NURBS 20 )-based skeleton smoothing to calibrate the effect of node placement noise on skeleton path length ( Supplementary Fig. 1f ); using the skeleton nodes as support knots, NURBS spline order (i.e., the degree of smoothing) NO = 4 and clamping the first and last node. We found that post-NURBS path length measurements of flight tracings are still on average 14.93 ± 1.08% (mean ± s.e.m.) longer than ortho tracings ( Supplementary Fig. 1g ). To correct for this and for the difference in node densities between tracing modes, we scaled NO in dependence of skeleton node density D s (in number of nodes per μm edge-based skeleton length), with N n (the total number of skeleton nodes per tracing) and parameters c 1 , c 2 and c 3 . The parameters were adjusted such that the average path lengths of neurites from the training set were similar for flight- and ortho-mode tracings (resulting in c 1 = 50, c 2 = 5, c 3 = 4; as a result, average ortho and flight path length for a given axon agreed within 1.79 ± 1.16%). See Supplementary Figure 1g and Supplementary Software 1 , 2 , 3 for comparisons of skeleton path length measurements based on edge length and NURBS smoothing with fixed NO and variable NO, respectively. The length measurements involving NURBS smoothing with fixed NO and variable NO reduced the path length obtained from the simple edge length addition method by less than 20% ( Supplementary Fig. 1g ). The variable NO path length measurement method was used for speed measurements in the axon test set and the dendrite tracings. To determine the annotation time of a given tracing, the administrative API of webKnossos was used ( Supplementary Software 4 ) to log autosave events. Autosave events are triggered when the annotator is actively tracing within the last 30 s and the last autosave was more than 30 s ago. Therefore, during annotation work, an autosave is submitted every 30 s (but not during pauses the annotator chooses to take). Annotation time was measured as the number of autosave events times 30 s. This is also the time used for determining annotator payment. Annotation redundancy: RESCOP. For determining the dependence of annotation error rates on annotation redundancy ( Fig. 1l–o ), multiple annotations of the same neurite from different annotators were consolidated using RESCOP 8 . Briefly, the priors and decision boundaries were fitted separately for axons traced in ortho and flight mode ( Supplementary Fig. 1h,i ). The priors were fitted using 20 randomly selected annotations for each neurite from the training annotations (i.e., 800 annotations for ortho and flight mode, respectively, total of 865,121 edges, 866,721 nodes for computing the vote histogram, Supplementary Fig. 1h ). The resulting decision boundaries are shown in Supplementary Figure 1i . Measurement of annotation error rates. For the measurement of annotation errors, 10 of the 20 test axons ( Fig. 1j ) were randomly selected. For these ten axons, a ground truth annotation was generated. To do this, the axon was first traced by one expert annotator. Then, all annotations of this axon from all tracers and tracing modes were superimposed; and all locations of discrepancy between the experts' annotation and all other annotations were inspected. Remaining errors in the expert annotation were corrected. Finally, two additional experts verified the ground truth annotation independently. Then, for each of the ten axons, three ortho-mode and three flight-mode annotations were randomly selected and their discrepancies to the ground truth annotation counted as in ref. 8 (Fig. 5c in this reference). Similarly, consensus skeletons at redundancies 2, 3, 4, 5, 6, 7, 10 and 13 were computed using RESCOP 8 (see above) for each of the ten axons and the two tracing modes, respectively. For each redundancy, 3 sets of tracings were randomly drawn from the available 13 tracings per axon and tracing mode. Thus, together, 540 reconstructions were error analyzed. Error analysis was done by one expert annotator and proofread by a second expert annotator. Both experts were blinded to the tracing mode in which the reconstructions were performed. For error analysis, the reconstructions were plotted in the three cardinal projections, overlaid with the ground truth reconstruction. Errors were classified into missing branches (false negatives) and wrongly added branches (false positives). Jumps from one process into another were counted twice, once as FP and once as FN. Errors were further classified according to the length of added or omitted neurite pieces (>10 μm, 5–10 μm, 3–5 μm, 1–3 μm; discrepancies smaller than 1 μm were not counted, as in ref. 8 ). Error segments larger than 10 μm were classified as continuation errors ( Fig. 1k ). For the measurement of errors in dendrite reconstructions, 10 of the 497 reconstructed dendrites were randomly selected and error-annotated as for the axon reconstructions. Since error rates were substantially lower for dendrites than for axons ( Fig. 1l–o ), only redundancies 1–6 were evaluated for dendrites. Path length of consolidated reconstructions. To make our results comparable to error rates reported in refs. 8 , 11 and 13 , we normalized the number of errors to the path length of the ground truth skeleton for each axon ( Fig. 1l,m ). However, since some annotations were shorter (due to missed neurite pieces) and others longer (due to added neurite pieces), we wanted to assure that our conclusions about error rates ( Fig. 1l–o ) were still correct when instead the neurite path length of the actual tracing or consolidation was used for error rate computation. To do so, we determined the path length for each RESCOP-consolidated reconstruction by generating a version of the ground truth reconstruction that matched the respective RESCOP-consolidated reconstruction (including its possible false negative errors), and we measured that skeleton's path length as described above. Synapse annotation and connectome reconstruction. To exemplify the full analysis workflow for reconstructing connectomes using webKnossos, we used all axons from the training reconstructions (32 axons, at RESCOP redundancy 6, step I in Fig. 2a–c ) and 497 dendrite reconstructions ( Fig. 2d , at redundancy 3, step II in Fig. 2a–c ). For synapse annotation (step III in Fig. 2a–c ), a synapse movie mode in webKnossos was used (this mode is automatically activated for webKnossos tasks of type 'synapseannotation'). This was built as an extension of flight mode in which the previously reconstructed skeleton was displayed. The annotator was asked to mark synapses by setting a single node into the postsynaptic process while navigating along the axon. For the synapse movie mode, the (consolidated) reconstruction was first cut into unbranched parts, and each of these parts was presented to the annotators (see Supplementary Software 1 for the corresponding MATLAB code). Ten annotators were trained for synapse annotation in an introductory 1-h seminar followed by two training axons for which they received immediate feedback. Then all annotators were asked to determine the output synapses of all 32 axons. To measure the precision and recall of synapse detection by student annotators, four of the axons were randomly selected, and synapse detection errors were determined by expert annotators. The student annotator with optimal precision and recall of synapse detection (precision 96%, recall 89%) was selected for the generation of the output connectome. In addition, annotators were instructed to mark axons as putative inhibitory axons if the majority of output synapses were made onto shafts. For the axons that the best annotator marked as inhibitory, a second annotator was asked to annotate the synapses of that axon. For the annotation of inhibitory synapses, the annotator was instructed not to focus on speed of synapse annotation. The resulting synapse annotations were reviewed by an expert annotator to establish error rates for inhibitory synapse annotation (no error in 20 reviewed synapse annotations). This procedure operated at 1.2 ± 0.5 h per mm candidate axon segment length ( n = 151, excitatory axons; 1.8 ± 1.0 h per mm for all axons, n = 178) To determine whether the postsynaptic targets of the reconstructed axons matched any of the 497 dendrites in the connectome, the annotation of the postsynaptic partner in synapse mode was used as a new seed for an annotation task (step IV in Fig. 2a–c ). The annotators for these tasks were asked to only reconstruct the postsynaptic structure (in about 90% of cases a spine) in ortho mode until it entered a dendritic shaft and to then place three additional skeleton nodes in the shaft center to simplify the matching to dendrite reconstructions. This annotation had a consumption of 31.1 ± 28.0 s annotation time per spine, mean ± s.d., n = 975; i.e. 2.3 ± 1.3 h per mm axon path length. Error rates of this postsynaptic process annotation were established by inspection of 30 randomly selected postsynaptic structures by an expert annotator (one wrong annotation). To match the postsynaptic partner reconstructions (step V in Fig. 2a ) with the 497 dendrite reconstructions, we finally measured the average distance d pd between all dendrites and the three shaft nodes of each postsynaptic partner reconstruction, and we detected the dendrite with the smallest average distance. To determine an attachment threshold—i.e. a maximum average distance d pd* up to which a postsynaptic partner reconstruction was considered to match a dendrite reconstruction—we used a randomly chosen set of 200 partner reconstructions. In these, the distribution of d pd ( Supplementary Fig. 2a ) indicated a threshold distance d pd* of 250 nm. To determine the error rate of postsynaptic partner matching, we evaluated the matching in an additional set of 200 randomly chosen spines and their closest dendrite (21 true positives, 1 false positive, 178 true negatives, no false negatives). All code for these procedures is available in Supplementary Software 1 . Connectome annotation time estimates. For the annotation time approximation of an example L2/3-L4 cortical connectome ( Supplementary Fig. 2c ), we used 1.5 mm/h reconstruction speed and six-fold redundancy for axons, and 2.1 mm/h reconstruction speed at three-fold redundancy for dendrites. For estimating the reconstruction time spent on synapse annotation ( Supplementary Fig. 2b ), two approaches for synapse annotation were considered. One, axon-based synapse annotation ( Fig. 2a,b ), proceeds along axons, marking synapses and identifying postsynaptic partners, which are then matched to dendrite reconstructions (see “Results”). The other, dendrite-based synapse annotation, proceeds along dendrites, reconstructing all spines along dendritic shafts. Spine annotation along dendrites proceeds at about 40 s per spine (time taken to reconstruct a spine and mark its presynaptic partner) at a spine density of about 1 per μm dendrite length. In both strategies, we assumed that only proximities of axons and dendrites at less than 5 μm distance need to be investigated for synapses. Therefore, depending on the density of axons and dendrites in a given reconstruction task, the length of axons and dendrites that need to be synapse searched varies, which yields an optimal strategy for any given volume density of axons and dendrites. Dendrite and axon reconstruction speed used for these estimates was 2.1 mm/h and 1.5 mm/h (at three- and six-fold redundancy, respectively). Statistical tests. The speed and the error comparison between ortho and flight tracers and the speed comparison between the subset of annotators for the second experiment and all annotators used Wilcoxon rank-sum test. The speed comparison for annotators between beginning and end of training used Wilcoxon signed-rank test. The correlation between tracing error rate and tracing speed was computed using Pearson's correlation. Software availability, code availability and licensing. webKnossos is available for testing at together with example data sets: the published retina data sets e2198 (ref. 21 ), k0563 (refs. 8 , 11 and 21 ) and e2006 (ref. 11 ), a 20 × 20 × 20 μm 3 sized subvolume of the data set 2012-09-28_ex145_07x2_new2 used for webKnossos testing (see above), and an example fluorescence data set (FD0149-2, data not shown). See Supplementary Video 2 for an introductory video. The webKnossos source code is provided as Supplementary Software 5 and is also available at . webKnossos is licensed under the AGPLv3 license (this applies to all source code files in Supplementary Software 1 , 2 , 3 , 4 , 5 and GitHub repository). webKnossos uses the following software packages and technologies: Scala, JDK 8, Play, mongoDB, WebGL, ThreeJS, Backbone, sbt. Data availability statement. webKnossos is openly accessible at , where data sets from retina and cortex can be browsed and annotated. The entire SBEM data set of the mouse cortex that support the findings in this study are available from the corresponding author upon reasonable request. webKnossos is open source, source code is available as Supplementary Software 5 and at . All reconstructions used in this study are available in Supplementary Software 1 , 2 and 3 . Source data for Figures 1 and 2 are available online. Additional information Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Scientists working in connectomics, a research field occupied with the reconstruction of neuronal networks in the brain, are aiming at completely mapping of the millions or billions of neurons found in mammalian brains. In spite of impressive advances in electron microscopy, the key bottleneck for connectomics is the amount of human labor required for the data analysis. Researchers at the Max Planck Institute for Brain Research in Frankfurt, Germany, have now developed reconstruction software that allows researchers to fly through the brain tissue at unprecedented speed. Together with the startup company scalable minds they created webKnossos, which turns researchers into brain pilots, gaining an about 10-fold speedup for data analysis in connectomics. Billions of nerve cells are working in parallel inside our brains in order to achieve behaviours as impressive as hypothesizing, predicting, detecting, thinking. These neurons form a highly complex network, in which each nerve cell communicates with about one thousand others. Signals are sent along ultrathin cables, called axons, which are sent from each neuron to its about one thousand "followers." Only thanks to recent developments in electron microscopy, researchers can aim at mapping these networks in detail. The analysis of such image data, however, is still the key bottleneck in connectomics. Most interestingly, human annotators are still outperforming even the best computer-based analysis methods today. Scientists have to combine human and machine analysis to make sense of these huge image datasets obtained from the electron microscopes. Virtual flight through the brain A research team led by Moritz Helmstaedter, director at the Max Planck Institute for Brain Research, has now found a novel highly efficient method of presenting these 3-dimenional images in-browser in such an intuitiv way that humans can fly at maximum speed along the cables in the brain. Achieving unprecedented 1,500 micrometers per hour, human annotators can still detect the branch points and tortuous paths of the axons (Boergens, Berning et al. Nature Methods, 2017). "Think of racing at 100 mph through a curvy, hilly village", compares Helmstaedter. Researchers think that this flight speed is the maximum humans can achieve in 3-D electron microscopic data of brain tissue – since the visualization is centered on the brain pilot, like in a plane, the steering is highly optimized for egocentric navigation. When combined with computer-based image analysis, the human part of data analysis in connectomics is now likely maximal, about 10-times faster than before. The recruiting software gives users the impression of sitting in the cockpit of an airplane. Credit: MPI f. Brain Research/ Julia Kuhl (Some Donkey) One key prerequisite for this success was the development of efficient data transmission and flight path prediction. The webKnossos software was developed in close collaboration with a computer science startup from Potsdam, Germany, called scalable minds. Over the five-year collaboration the unusual task of making gray-scale brain data usable in online data visualization was both challenging and rewarding, says Norman Rzepka, co-author of the study and one of the co-founders of the company. With human data analysis at its maximum, the researchers are now back to optimizing the computer part of the analysis – such that the precious human time is used most effectively for our scientific questions. Only when machines and humans collaborate optimally, connectomics can thrive in today's neuroscience. Flight through the brain. The "pilot" follows a dendrite and reaches 2,500 micrometers per hour almost effortlessly when following a dendrite. Sites of branching are indicated. Credit: MPI for Brain Research

10.1038/NMETH.4331

Medicine

Breastmilk antibody protects preterm infants from deadly intestinal disease

Maternal IgA protects against the development of necrotizing enterocolitis in preterm infants, DOI: 10.1038/s41591-019-0480-9 , www.nature.com/articles/s41591-019-0480-9 Journal information: Nature Medicine

http://dx.doi.org/10.1038/s41591-019-0480-9

https://medicalxpress.com/news/2019-06-breastmilk-antibody-preterm-infants-deadly.html

Abstract Neonates are protected from colonizing bacteria by antibodies secreted into maternal milk. Necrotizing enterocolitis (NEC) is a disease of neonatal preterm infants with high morbidity and mortality that is associated with intestinal inflammation driven by the microbiota 1 , 2 , 3 . The incidence of NEC is substantially lower in infants fed with maternal milk, although the mechanisms that underlie this benefit are not clear 4 , 5 , 6 . Here we show that maternal immunoglobulin A (IgA) is an important factor for protection against NEC. Analysis of IgA binding to fecal bacteria from preterm infants indicated that maternal milk was the predominant source of IgA in the first month of life and that a relative decrease in IgA-bound bacteria is associated with the development of NEC. Sequencing of IgA-bound and unbound bacteria revealed that before the onset of disease, NEC was associated with increasing domination by Enterobacteriaceae in the IgA-unbound fraction of the microbiota. Furthermore, we confirmed that IgA is critical for preventing NEC in a mouse model, in which pups that are reared by IgA-deficient mothers are susceptible to disease despite exposure to maternal milk. Our findings show that maternal IgA shapes the host–microbiota relationship of preterm neonates and that IgA in maternal milk is a critical and necessary factor for the prevention of NEC. Main NEC is associated with decreased diversity and increased numbers of Enterobacteriaceae in the intestinal microbiota; however, this is not sufficient to cause the disease 7 , 8 . Bioactive components of maternal milk, including IgA antibodies, shape the neonatal microbiota 9 , 10 , 11 , 12 . It is not known how the anti-bacterial IgA repertoire in maternal milk varies between women; however, IgA-producing B cells in the mammary glands originate in the intestine and thus may differ between mothers as a result of individualized microbiomes and infectious histories 13 , 14 , 15 . We hypothesized that differential binding of the preterm microbiota by maternal IgA is a central feature of NEC pathogenesis. To analyze immunoglobulin binding of gut bacteria in preterm infants, we stained fecal samples (Table 1 ) with anti-human IgA, IgM and IgG antibodies and measured the immunoglobulin-bound populations using flow cytometry 16 , 17 . This initial sample set contained 30 samples collected at the time of NEC diagnosis and 39 samples from age-matched controls. Surveyed across all samples, the percentage of IgA-bound bacteria was far greater than the percentages of IgM- and IgG-bound bacteria and samples from maternal-milk-fed infants contained a far greater abundance of IgA-bound (IgA + ) bacteria compared to formula-fed infants (Fig. 1a,b and Extended Data Fig. 1a,b ). Although a majority (11 out of 19) of formula-fed infants had less than 1% of their intestinal bacteria bound by IgA, some samples from formula-fed infants contained high amounts of IgA + bacteria (Fig. 1b ). Because B cells generally do not populate the intestine until about 4 weeks of age 18 , we hypothesized that fecal samples from formula-fed infants collected before this time point would not contain IgA-bound bacteria. Indeed, we found a significant temporal relationship between age and IgA binding in formula-fed infants that was not observed among maternal-milk-fed infants (Fig. 1c ). A dedicated analysis of samples from a single formula-fed preterm infant revealed no IgA + bacteria in the first 4 weeks of life, strongly indicating that maternal milk was the primary source of perinatal IgA (Extended Data Fig. 1c ). Limiting our analysis of this data set to 40 days after delivery, we found that samples from infants with NEC contained less IgA-bound bacteria than samples from age-matched controls (Fig. 1d ). However, infants with NEC in this cohort were more likely to be formula fed; additionally, their fecal samples were collected after NEC was diagnosed and treatment had been initiated with antibiotics and cessation of feeding. To eliminate the impact of these confounding variables, we selected and analyzed a prospectively collected longitudinal series of samples from 23 milk-fed preterm infants, of which 43.4% subsequently developed NEC (Table 1 ). Notably, we found that the fraction of IgA + bacteria decreased with time among infants that developed NEC, whereas IgA binding of fecal bacteria showed no relationship in controls (Fig. 1e and Extended Data Fig. 2a,b ). Thus, it appears that in infants who developed NEC, a change occurs in either the intestinal microbiota or the maternal IgA repertoire that leads to the ‘escape’ of intestinal bacteria from binding. Table 1 Study characteristics Full size table Fig. 1: IgA binding to the intestinal bacteria of preterm infants is positively correlated with maternal milk feeding and negatively correlated with the development of NEC. Flow cytometry analysis of bacterial IgA binding on fecal samples from preterm infants. a , Example of IgA staining of samples from preterm infants. b , Percentage of IgA-bound bacteria from infants fed with maternal milk fed ( n = 50) or formula ( n = 19). The box represents the number of samples with <1% IgA binding of intestinal bacteria. Two-sided Mann–Whitney U -test, data are mean ± s.e.m. c , Percentage of IgA binding was correlated by linear regression with time after delivery for infants fed with maternal milk ( n = 50) and formula ( n = 19); Pearson’s correlation coefficient. d , Percentage IgA-bound bacteria from controls ( n = 28) or infants diagnosed with NEC ( n = 23), from samples collected before DOL 40. Two-sided Mann–Whitney U -test, data are mean ± s.e.m. b – d , Samples from infants fed with maternal milk are indicated by black circles, samples from infants fed with formula are indicated by open red squares. e , Percentage IgA-bound intestinal bacteria from prospectively collected longitudinal samples of patients that will develop NEC ( n = 10 patients, n = 39 samples) and controls ( n = 13, n = 59) for each analyzed DOL. Patients are shown in different colors. Linear regression and Pearson’s correlation coefficient. Source Data Full size image We next sought to identify shifts in the microbiota that might explain this drop in IgA + bacteria. Analyzing all samples together, we observed a modest enrichment of Enterobacteriaceae and a reduction in Gram-positive anaerobes, such as Lachnospiraceae, in infants who developed NEC (Fig. 2a and Extended Data Fig. 3 ). This result aligns well with published reports that describe dysbiosis before the onset of NEC 7 . We therefore hypothesized that the loss of IgA binding to Enterobacteriaceae may be important to the decrease in IgA-bound bacteria in NEC. In support of this hypothesis, we found that the abundance of IgA-bound bacteria varied inversely with the abundance of Enterobacteriaceae among infants who developed NEC, but not controls (Fig. 2b ). Fig. 2: Reduced intestinal bacterial diversity driven by increased IgA-unbound Enterobacteriaceae precedes the development of NEC. Longitudinal fecal samples from preterm infants before the onset of NEC were selected from infants with NEC ( n = 10, 39 samples combined) and controls ( n = 13, 59 samples combined). a , Samples were analyzed for the relative abundance of different bacterial OTUs by targeted sequencing of 16S rRNA genes. The mean relative abundances of different taxa are shown for patients that will develop NEC and controls (pooled from all time points and infants). b , Relative abundance of Enterobacteriaceae in preterm fecal samples compared to the percentage of IgA-bound bacteria from controls and patients that will develop NEC. c – f , Fecal samples were separated into IgA + and IgA − pools before 16S rRNA sequencing and deconvolution based on post-sort analysis. c , Shannon diversity scores of IgA + and IgA − samples from patients that will develop NEC and controls for each analyzed DOL. d , Relative abundance of Enterobacteriaceae from IgA + and IgA − samples from patients that will develop NEC and controls for each analyzed DOL. e , Relative abundance of combined anaerobic bacterial OTUs from IgA + and IgA − samples from patients that will develop NEC and controls for each analyzed DOL. f , Ratio of reads (IgA − /IgA + ; log 2 -transformed values) from paired IgA + and IgA − samples for each analyzed DOL. The ratios of total number of bacterial reads, Enterobacteriaceae reads and combined anaerobe reads are shown. For b – f , each patient is shown in a different color and the R 2 value is based on a linear regression and Pearson’s correlation coefficient. Source Data Full size image To directly identify which bacterial taxa might be more or less bound by maternal antibodies in our longitudinal prospective cohort (Table 1 ), we physically separated the two fractions and measured microbial diversity by sequencing the V4 region of bacterial 16S rRNA genes (IgSeq) 19 . In accordance with published reports 16 , 20 , we did not achieve purity (>99%) in our IgA separations, complicating our ability to discriminate IgA + and IgA − bacteria (Extended Data Fig. 4a ). Post-sort flow cytometry analysis of each sample allowed us to deconvolve the contamination, significantly increasing the correlation between the percentage IgA + bacteria in the unsorted sample and the ratio of IgA + and IgA − bacteria reads from IgSeq (Extended Data Fig. 4b–d ). Deconvolution is particularly important in samples with low levels of IgA + bacteria, which both validates that our technique is correcting for contamination and improves analysis of the samples that are most critical to explaining our observation of reduced IgA + bacteria in NEC (Extended Data Fig. 4c,d ). Longitudinal IgSeq analysis of deconvolved samples revealed that control infants showed a significant increase in bacterial diversity among IgA − bacteria over time, whereas no significant changes were detected in the IgA + sample (Fig. 2c ). Conversely, among infants that progressed to NEC, the diversity of both IgA + and IgA − fecal bacteria significantly decreased over time 21 (Fig. 2c ). In accordance with our analysis of the total microbiome (Fig. 2b ), IgSeq also revealed that over time the IgA − intestinal microbiota of infants that will develop NEC became dominated by Enterobacteriaceae, whereas anaerobic bacteria (notably Clostridiales and Bifidobacteriales) were almost undetectable (Fig. 2d,e and Extended Data Fig. 5 ). Conversely, among controls, the relative abundance of IgA − anaerobes increased and the relative abundance of IgA − Enterobacteriaceae decreased, confirming that Enterobacteriaceae are bound at relatively higher frequencies over time in infants who did not develop NEC (Fig. 2b,d,e and Extended Data Fig. 5 ). Although the IgA + and IgA − fractions from controls differed in both diversity and the relative abundance of Enterobacteriaceae, in patients that progress to NEC these fractions were not discernibly different (Fig. 2c,d ). The lack of differences between IgA + and IgA − fractions found in infants that will develop NEC may be explained by very low diversity and domination by Enterobacteriaceae, such that there are few taxa available for IgA to bind (Extended Data Fig. 5 ). An advantage of IgSeq deconvolution is that it enables the comparison of taxon abundances between IgA + and IgA − samples, as their ratio corresponds to the abundance of IgA-bound bacteria in the unsorted sample (Extended Data Fig. 4 ). When we calculated the ratio of the paired IgA − and IgA + reads, we observe a unique increase over time in IgA − total and Enterobacteriaceae reads in infants who developed NEC that correlates well with IgA-binding data from flow cytometry (Figs. 1e , 2f ). By contrast, control infants showed no significant shifts in the relative abundances of IgA + and IgA − bacteria, Enterobacteriaceae or anaerobes and showed significant shifts in only less-abundant operational taxonomic units (OTUs) (Fig. 2f and Extended Data Fig. 6 ). As Enterobacteriaceae are the most abundant OTU in preterm infants and the only taxa uniquely increasing in the IgA − fraction of the microbiota before the onset of disease, the increase in IgA − Enterobactericeae is the most likely driver of the reduced IgA-bound bacterial fraction that is found to precede NEC (Figs. 1e , 2a,f and Extended Data Fig. 6 ). Thus, we have identified that infants who progress to NEC uniquely fail to diversify their intestinal microbiota with anaerobic bacteria and instead remain dominated by IgA-unbound Enterobacteriaceae. To test the possibility that the increase in Enterobacteriaceae was driven by a rapid bloom, we measured the number of each bacterial taxa per unit mass in fecal samples 22 , 23 . Although we saw increased numbers of Enterobacteriaceae among some infants who developed NEC, inter-individual variation was high and there were no statistically significant differences between infants with NEC and controls (Extended Data Fig. 7 ). However, our analysis may not adequately represent bacteria in the small intestine, so we cannot rule out the possibility of focal expansions of Enterobacteriaceae associated with NEC 24 . Nonetheless, we favor a model in which loss of IgA binding to the microbiota is induced either by mutation or by transcriptional modifications that allow sub-populations of Enterobacteriaceae to escape maternal IgA, but that do not lead to increases in the total number of Enterobacteriaceae. To further define the contribution of maternal IgA to disease pathogenesis, we turned to an experimental mouse model of NEC 25 , 26 . We bred mice so that heterozygote wild-type pups were fed by dams that either can (C57BL/6) or cannot produce IgA ( Rag1 −/− or Igha −/− ), and compared them to formula-fed positive controls (Fig. 3a ). We confirmed that mice, like humans, produce little IgA during the first two weeks of life and that dams are the primary source of neonatal IgA 18 , 27 , 28 (Fig. 3b ). We also determined that pups that received oral gavages of Enterobacter spp. to induce NEC and that were fed by C57BL/6 dams had enriched numbers of Enterobacter spp. in the IgA + fraction, indicating that mouse dams may produce protective IgA without being vaccinated (Extended Data Fig. 8 ). Notably, pups that underwent the NEC protocol that were fed by dams that lacked IgA ( Rag1 −/− or Igha −/− ) showed a phenotype that was indistinguishable from formula-fed controls. Specifically, pups fed by dams that lack IgA exhibited increased mortality compared to pups fed by C57BL/6 dams and severe intestinal damage characterized by shortened necrotic villi and mucosal sloughing (Fig. 3c–e ). Furthermore, pups fed by Igha −/− dams exhibited a significant reduction in weight gain compared to pups fed by wild-type mothers (Fig. 3f ). We have therefore shown, using an experimental model of NEC, that maternal milk only protects against NEC when it contains IgA. Fig. 3: IgA is a necessary component of breast milk for the prevention of the development of experimental NEC. a , Experimental NEC mouse model in which wild-type pups are fed by dams that either can (C57BL/6 mice ( n = 31)) or cannot produce IgA ( Rag1 −/− mice ( n = 22) or Igha −/− mice ( n = 14)). Formula-fed mice ( n = 48) are used as a positive control. b , Representative images of IgA staining of the fecal matter of 8-day-old pups from a . Images show the absence of IgA-bound bacteria in dam breast-fed and formula-fed pups. Representative images for mice fed by C57BL/6 ( n = 19), Rag1 −/− ( n = 9), Igha −/− ( n = 8) mothers or that were formula-fed ( n = 23) are shown. c , Representative images of hematoxylin and eosin staining of small intestines of pups that underwent the mouse protocol that induces NEC; pups were fed by C57BL/6 ( n = 23), Rag1 −/− ( n = 17), Igha −/− ( n = 17) mothers or were formula fed ( n = 40). d , Histology scores of the small intestines of pups from c . One-way ANOVA with multiple comparisons; data are mean ± s.e.m. e , Percentage of pups from a that survived the NEC protocol at different time points. Statistics determined by log-rank (Mantel–Cox) test. **** P = 0.00007. f , Weights of pups from a . Statistical difference calculated by one-way ANOVA with multiple comparisons on weights at experimental completion point; data are mean ± s.e.m. **** P = 0.0001. Data shown in b – f are grouped for three individual experiments with minimum n = 5 pups in each group for every experiment. Source Data Full size image Previous studies have shown associations between the abundance of Enterobacteriaceae and NEC 7 . We now show that IgA-unbound Enterobacteriaceae are more closely linked to NEC development than total Enterobacteriaceae abundance. Animal studies have indicated that both host and maternal IgAs are important in controlling Enterobacteriaceae and establishing a mature microbiota characterized by fastidious anaerobic bacteria 29 , 30 . Our results indicate that binding of bacteria by maternally derived IgA may promote diversity in the microbiome and the acquisition of anaerobic bacteria during the critical window when infants make little or no IgA of their own, perhaps by limiting inflammation driven by Enterobacteriaceae 31 , 32 , 33 . Future studies will be required to elucidate the mechanism by which IgA controls and modifies gut bacterial colonization in newborns. IgA has been shown to modify the expression of bacterial surface proteins and motility of bacteria, which may limit the ability of bacteria to gain access to the intestinal epithelium 17 , 34 . IgA may accomplish these tasks by ‘enchaining’ bacterial cells, allowing for easier expulsion and preventing gene transfer 35 . Notably, the current study did not discriminate between bacteria at the strain level. Thus, it remains to be determined whether the loss of IgA binding results from the appearance of new organisms that are not constrained by the existing IgA repertoire, or alternatively from changes to bacterial genomes and/or gene expression that allow early colonizers to escape IgA binding 36 , 37 . Temporal changes in bacterial binding could also result from shifts in the maternal IgA repertoire. Previous attempts to prevent NEC with intravenous immunoglobulins have largely failed to show efficacy 38 , 39 . However, the repertoire of intravenous antibodies may differ from that of secretory IgA and bacterial specificity was not accounted for in these studies. Future efforts might enable precision microbiome-informed strategies that enable augmentation of milk or formulas for preterm infants with rationally selected protective antibodies. Methods Mice C57BL/6 mice were purchased from Taconic. Rag1 −/− mice were obtained from Jackson Laboratories. Igha −/− mice were obtained from Y. Belkaid (NIH/NIAID). All mice were maintained at and all experiments were performed in an American Association for the Accreditation of Laboratory Animal Care-accredited animal facility at the University of Pittsburgh and housed in accordance with the procedures outlined in the Guide for the Care and Use of Laboratory Animals under an animal study proposal approved by the Institutional Animal Care and Use Committee of the University of Pittsburgh. Mice were housed in specific pathogen-free conditions. Human fecal samples The human study protocol was approved by the Institutional Review Board (protocol numbers PRO16030078 and PRO09110437) of the University of Pittsburgh. Fecal samples were collected fresh or from the diaper of preterm infants at the UPMC Magee-Womens Hospital and frozen immediately at −80 °C. The samples were later divided into age-matched controls and NEC depending on the incidence of NEC. Fecal IgA flow cytometry and magnetic sorting of IgA + and IgA − bacteria Either fecal pellets collected from mice after euthanasia or around 50 mg of frozen human fecal material was placed in 1.5-ml Eppendorf tubes and 1 ml phosphate-buffered saline (PBS) was added. The fecal material was disrupted by a combination of vortexing and pipetting and passed through a 40-µm filter to remove food and/or fibrous material. The fecal material was diluted with PBS to obtain an optical density of approximately 0.4 to maintain equality between samples and to prevent the magnetic columns from clogging. A volume of 200 µl of the suspended bacterial material was then frozen as an ‘unsorted’ control. An additional 200 µl of the suspended material was divided equally on a 96-well plate for IgA staining and isotype control for each sample to eliminate non-specific binding. The fractions were washed twice with staining buffer (1% bovine serum albumin (Sigma) in PBS; filtered through a 2.2-µm filter). The bacteria were stained with Syto BC (green fluorescent nuclear acid stain, Invitrogen; 1:400), APC anti-human IgA (Miltenyi Biotec, clone IS11-8E10; 1:10), APC anti-human IgA (Miltenyi Biotec, clone REA1014; 1:50), anti-human IgM BV421 (BD Biosciences, clone G20-127; 1:30), BV421 mouse anti-human IgG (BD Biosciences, clone G18-145; 1:10) or PE-conjugated anti-mouse IgA (eBioscience, clone mA-6E1; 1:500), rat anti-mouse IgM BV421 (BD Biosciences, clone R6-60.2; 1:30), rat anti-mouse IgG2a isotype (BD Biosciences, clone R35-95), anti-mouse IgG FITC (BioLegend clone, Poly4060; 1:30) and blocking buffer of 20% normal mouse serum for human or 20% normal rat serum for mouse samples (ThermoFisher). The isotype control was stained similarly using APC mouse IgG1 isotype control (Miltenyi Biotec, clone IS5-21F5; 1:10) or PE-conjugated rat anti-mouse IFNγ (eBioscience, clone XMG1.2). The stained samples were incubated in the dark for 1 h at 4 °C. Samples were then washed three times with 200 µl of staining buffer before flow cytometry analysis (LSRFortessa, BD Biosciences). For magnetic-activated cell sorting (MACS), we used 500 µl of the suspended fecal material to compensate for the loss of material during sorting and scaled our staining volume accordingly. Anti-IgA-stained fecal bacterial pellets were incubated in 1 ml per sample of staining buffer containing 45 µl of anti-APC or anti-PE MACS microbeads (Miltenyi Biotec) (20 min at 4 °C in the dark), washed twice with 1 ml staining buffer and centrifuged (8,000 r.p.m., 5 min, 4 °C), and then sorted using MS columns (Miltenyi Biotec). The flow-through was collected as the IgA-unbound (IgA − ) fraction. Columns were washed with 70% ethanol and sterile PBS between separations. The IgA-bound fraction was added in the column and the steps mentioned above were repeated four times for maximum enrichment. Then, 100 µl each of the IgA-bound and IgA-unbound fraction was used for post-sort flow cytometry analysis (along with unsorted sample). Absolute bacterial counts were determined by adding a known number of AccuCheck Counting beads (Life Technologies) to antibody-stained fecal samples of a given mass, which allows for the calculation of the total number of Syto + (DNA) events in any given sample. The number of total bacteria can then be multiplied by the measured abundance of any OTU to represent the number of bacteria of that taxon or mass in any sample. DNA extraction All microbial DNA was extracted using the MO BIO PowerSoil DNA Isolation kit (single-tube extractions). The unsorted, IgA-bound and IgA-unbound pellets were resuspended in solution TD1 by pipetting and vortexing, and approximately 200 µl of 0.1-mm diameter Zirconia/Silica beads (Biospec) were added and shaken horizontally on a lab mixer for 12–18 min at maximum speed using a MO BIO vortex adaptor. All remaining steps followed the manufacturer’s protocol. The extracted DNA was stored at −20 °C for further 16S amplicon PCR and sequencing. 16S amplicon PCR, sequencing and analysis PCR amplification of the small subunit ribosomal RNA gene (16S rRNA) was performed in triplicate 25-μl reactions. Reactions were held at 94 °C for 3 min to denature the DNA, with amplification performed for 30 cycles at 94 °C for 45 s, 50 °C for 60 s and 72 °C for 90 s; followed by a final extension of 10 min at 72 °C. Amplicons were produced utilizing primers adapted for the Illumina MiSeq. Amplicons target the V4 region and primers utilized the Illumina adaptor, primer pad and linker (forward primer) or Illumina adaptor, Golay barcode, primer pad and linker (reverse primer) followed by a sequence that targets a conserved region of the bacterial 16S rRNA gene as previously described 40 , 41 . The only deviation from the protocol was that PCR was run for 30 cycles. Amplicons were cleaned using the Qiagen UltraClean 96 PCR Cleanup Kit. Quantification of individual amplicons was performed with the Invitrogen Quant-iT dsDNA High Sensitivity Assay Kit. Amplicons were then pooled in equimolar ratio. Agarose gel purification was performed to further purify the amplicon pool and remove undesired PCR products before paired-end sequencing on the Illumina MiSeq. Read pairing, clustering and core diversity statistics were generated through PEAR, UPARSE and QIIME and R 42 , 43 . Linear discriminant analysis effect size was used to compare family level relative abundances between NEC and control groups 44 . Raw 16S rRNA data have been uploaded to NCBI BioSample/SRA and are available under accession number PRJNA526906 . Deconvolution and microbiome data analysis Flow cytometry was used to determine the percentage of IgA + and IgA − bacteria in each sample (unsorted, IgA + and IgA − ) after magnetic separation. We assumed that contamination affected each OTU equally and that the IgA + and IgA − samples are reciprocal (fractions of the same whole). The raw reads from sequencing of 16S rRNA genes were then deconvolved by summing the proportion of IgA bound or IgA unbound (as measured by flow cytometry) across the paired (infant and time point (DOL)) IgA + and IgA − samples for each OTU (Extended Data Fig. 3c ). Example to solve the fraction IgA + at a single time point and one OTU X : Total IgA + X = X correctly bound to IgA + + X contaminating IgA − Total IgA + X = (percentage of IgA + in bound fraction) × (number of X reads in bound fraction) + (percentage of IgA + in unbound fraction) × (number of X reads in unbound fraction) The deconvoluted data were processed through the QIIME2 workflow to create alpha diversity metrics with sampling depth chosen based on alpha rarefaction plotting. Abundance of individual or ‘pooled’ (anaerobes) OTUs was then calculated using the deconvolved values. The algorithm for deconvolution is available on GitHub ( ). If IgSeq is accurate, the ratio of the read numbers between IgA + and IgA − (IgA + /IgA + + IgA − ) samples should roughly correspond to the percentage of IgA + bacteria in the unsorted sample from which they were derived. This relationship does not hold for samples with low levels of IgA + bacteria (Extended Data Fig. 4c ) before deconvolution but is much improved after our method has been applied (Extended Data Fig. 4d ). We categorized all of the following OTUs as anaerobes: Bifidobacteriaceae, Prevotellaceae, Bacteroidiales S24-7, Clostridiaceae, Lachnospiraceae, Peptostreptococcaceae, Ruminococcaceae, Veillonellaceae and Tissierellaceae (Fig. 2e and Extended Data Fig. 6b ). qPCR for 16S rRNA PCR amplification of the 16S rRNA was performed in triplicate 10-μl reactions. Reactions were held at 95 °C for 3 min to denature the DNA, with amplification performed for 35 cycles (95 °C for 10 s and 60 °C for 30 s). The forward primer sequence of 16S was ACTCCTACGGGAGGCAGCAGT and the reverse primer sequence of 16S was ATTACCGCGGCTGCTGGC. qPCR for Enterobacter spp PCR amplification of the small subunit ribosomal RNA gene (23S rRNA) was performed in triplicate 10-μl reactions. Reactions were held at 95 °C for 3 min to denature the DNA, with amplification performed for 35 cycles (95 °C for 10 s and 60 °C for 30 s). The forward primer sequence of Enterobacter 23S was AGTGGAACGGTCTGGAAAGG and the reverse primer sequence of Enterobacter 23S was TCGGTCAGTCAGGAGTATTTAGC, as described previously 45 . Induction of NEC NEC is induced in 7- to 8-day-old mice by hand-feeding mice formula by oral gavage five times per day (22-gauge needle; 200 µl volume; Similac Advance infant formula (Ross Pediatrics):Esbilac canine milk replacer in a 2:1 ratio). The formula is supplemented with 10 7 CFUs of Enterobacter spp. (99%) and Enterococcus spp. (1%) and mice are rendered hypoxic (5% O 2 , 95% N 2 ) for 10 min in a hypoxic chamber (Billups-Rothenberg) twice daily for 4 days as described previously 46 , 47 . We used males and females in all experiments. Disease is monitored by weighing mice daily before the second feed. The severity of disease was determined on histological sections of the entire length of the small intestines stained with hematoxylin and eosin by trained personnel who were blinded to the study conditions according to previously published scoring system from 0 (normal) to 4 (severe) 48 . Statistics Statistical tests used are indicated in the figure legends. Lines in scatter plots represent the mean of that group. Group sizes were determined based on the results of preliminary experiments. Mouse studies were performed in a non-blinded manner. Statistical significance was determined with the two-tailed unpaired Student’s t -test or non-parametric Mann–Whitney U -test when comparing two groups and one-way ANOVA with multiple comparisons, when comparing multiple groups. All statistical analyses were calculated using Prism software (GraphPad). Differences were considered to be statistically significant when P < 0.05. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability Patient-related data not included in the paper were generated as part of clinical trials and may be subject to patient confidentiality. The human study protocol was approved by the Institutional Review Board (protocol numbers PRO16030078 and PRO09110437) of the University of Pittsburgh. All raw and analyzed sequencing data have been deposited in Sequence Read Archive (accession number PRJNA526906 ). Code availability The algorithm for deconvolution of IgSeq data is available on GitHub ( ).

A new study from the University of Pittsburgh and UPMC Children's Hospital of Pittsburgh finds that an antibody in breastmilk is necessary to prevent necrotizing enterocolitis (NEC)—an often deadly bacterial disease of the intestine—in preterm infants. Immunoglobulin A (IgA) antibodies bind to bacteria in the gut, and, according to the study, the more bacteria that's tied up with IgA, the less likely babies are to develop NEC. Since preterm infants get IgA only from mothers' milk in their fragile first weeks of life, the authors emphasize the importance of breastmilk for these babies. The study appears today in Nature Medicine. "It's been well known for a decade that babies who get NEC have particular bacteria—Enterobacteriaceae—in their guts, but what we found is that it's not how much Enterobacteriaceae there is, but whether it's bound to IgA that matters. And that's potentially actionable," said senior author Timothy Hand, Ph.D., assistant professor of pediatric infectious diseases at the R.K. Mellon Institute for Pediatric Research and Pitt's School of Medicine. The researchers looked at fecal samples from 30 preterm infants with NEC and 39 age-matched controls. Overall, breastmilk-fed babies had more IgA-bound gut bacteria than their formula-fed peers, and those who developed NEC were more likely to have been formula-fed. Tracking these infants' gut microbiomes over time, Hand's team found that for the healthy babies, Enterobacteriaceae was largely tied up by IgA, allowing diverse bacterial flora to flourish. But for the NEC infants in the days leading up to diagnosis, IgA-unbound Enterobacteriaceae was free to take over. To demonstrate causation between IgA and NEC, Hand and his team used a mouse model. "Mice, when they're born, are equivalent in their intestinal development to a human baby born at 24 weeks," said lead author Kathyayini Gopalakrishna, M.D., a Ph.D. student in the Pitt Graduate School of Public Health's Department of Human Genetics, "so they're a perfect model to study NEC in preterm infants." The researchers bred mice that couldn't produce IgA in their breastmilk. Pups reared on IgA-free milk were just as susceptible to NEC as their formula-fed littermates. So, breastfeeding in and of itself is not sufficient for NEC prevention. The milk must contain IgA to confer this specific benefit. But the solution for NEC may not be as simple as putting IgA into formula, Hand said. Because breastmilk has other benefits beyond IgA, donor milk is still the best option to fill the gap when breastfeeding or providing pumped maternal milk isn't an option. "What we showed is that IgA is necessary but may not be sufficient to prevent NEC," Hand said. "What we're arguing is that you might want to test the antibody content of donor milk and then target the most protective milk to the most at-risk infants."

10.1038/s41591-019-0480-9

Earth

Cities can cut greenhouse gas emissions far beyond their urban borders

Peter-Paul Pichler et al, Reducing Urban Greenhouse Gas Footprints, Scientific Reports (2017). DOI: 10.1038/s41598-017-15303-x Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-017-15303-x

https://phys.org/news/2017-11-cities-greenhouse-gas-emissions-urban.html

Abstract Cities are economically open systems that depend on goods and services imported from national and global markets to satisfy their material and energy requirements. Greenhouse Gas (GHG) footprints are thus a highly relevant metric for urban climate change mitigation since they not only include direct emissions from urban consumption activities, but also upstream emissions, i.e. emissions that occur along the global production chain of the goods and services purchased by local consumers. This complementary approach to territorially-focused emission accounting has added critical nuance to the debate on climate change mitigation by highlighting the responsibility of consumers in a globalized economy. Yet, city officials are largely either unaware of their upstream emissions or doubtful about their ability to count and control them. This study provides the first internationally comparable GHG footprints for four cities (Berlin, Delhi NCT, Mexico City, and New York metropolitan area) applying a consistent method that can be extended to other global cities using available data. We show that upstream emissions from urban household consumption are in the same order of magnitude as cities’ overall territorial emissions and that local policy leverage to reduce upstream emissions is larger than typically assumed. Introduction Cities worldwide strive to reduce their greenhouse gas emissions. A plethora of city networks such as the Global Covenant of Mayors for Climate & Energy, C40, International Council for Local Environmental Initiatives (ICLEI), and the Global Parliament of Mayors have emerged to foster and motivate cooperation in cutting urban GHG emissions. In 2017, 7,500 cities worldwide, representing 685 million people, were signatories of the Global Covenant of Mayors for Climate & Energy (the largest of those networks) and have declared emission reduction targets 1 . The 20 major cities in the Carbon Neutral Cities Network have pledged to become climate neutral 2 . Non-state Actor Zone for Climate Action (NAZCA), the UNFCCC online platform which tracks the climate commitments of non-state actors, lists more than 2,500 cities and many more have pledged to reduce their emissions by joining a growing number of dedicated city networks 3 . Although these initiatives increasingly recognize the inherent socio-metabolic openness of cities that inevitably leads to resource use and associated GHG emissions occurring outside the city boundaries 4 , 5 most cities still focus their reduction efforts entirely on the emissions released directly from their territory 6 . The vast majority of cities apply a perspective similar to the IPCC and OECD guidelines for national economies, where GHG emissions are attributed to the actors (households, firms, institutions) within the administrative territory on whose property or under whose legal control the emissions originate: e.g. the emissions from cement manufacturing are attributed to the cement producing company and the emissions from coal fired power plants to the electricity company. This traditional approach to allocate emissions is called territorial or production approach 4 , 7 . Figure 1a illustrates this for urban emissions. The territorial approach accounts for the direct emissions from all socio-economic actors within the city’s boundaries. There are the urban producers of goods and services and their associated transport (symbolized by a factory and a truck symbol in dark black in Fig. 1a ). There are the final consumers, typically broken down into household consumption, government consumption and fixed gross capital (investments in durable goods, e.g. public infrastructure). It is important to note that economically we distinguish between two types of actors, producers and final consumers, whereas in GHG accounting all economic actors are producers of direct emissions (symbolized by undulate lines in Fig. 1 ). Figure 1 Conceptual comparison between territorial GHG emission accounting ( a ) and the GHG footprint ( b ). Territorial emissions include the entirety of emissions that occur within the city boundary. These are direct emissions from production (goods & services, transport) and final consumption (households, government, gross fixed capital formation). Because they also include urban production for exports, territorial emissions are often indicative of the economic structure of a city (e.g. in the presence of heavy industry). The GHG footprint, instead, puts the focus on consumption within the city boundary. In this study it includes direct and upstream GHG emissions from household consumption. The former occur within the city boundary (e.g. heating and private transport), the latter may occur anywhere in the world (including within the city) and require analysing the entire supply chain of urban consumption. The GHG footprint is indicative of the consumption pattern of urban households. Full size image A consumption perspective takes a different view 7 . As the name suggests, here emissions are attributed not to the economic producer but to the economic (final) consumer of a good or service. For example, emissions are attributed to the person consuming electricity at home rather than to the power company, or to the person eating a steak, rather than to the farmer who produced it. In summary, the emissions attributable to the goods and services purchased by local final consumers but occurring along the entire production chains of the purchased goods and services are called upstream or embodied emissions. As global production chains continue to become longer and more complex, the difference between these two accounting perspectives has become considerable already at the national scale 8 . This observed separation of the geographic locations of production and consumption is even more pronounced for cities which are inherently open socio-metabolic systems. Today, urban upstream emissions often exceed those directly emitted on a city’s territory 9 , 10 . This fact limits the effectiveness of climate mitigation policies based on territorial emissions alone 11 , 12 . The reasons why current urban climate change mitigation initiatives overwhelmingly focus on territorial emissions are both pragmatic and political. Local decision makers are often unaware of the relevance of upstream emissions. The literature on upstream urban emissions is sparse and comparisons among cities are hampered by differences in methods, classifications and terminology 4 , 13 . The large number of city networks, each with different guidelines for urban emission inventories contribute to this problem (see Supplementary Information online for more details on urban GHG accounting guidelines). With few exceptions, accounts of upstream urban emissions for cities in emerging countries are not available. Most importantly though, there is a widely held view that local politicians don’t have much policy leverage to influence emissions outside their own territory 14 , 15 . To our knowledge this is the first international comparison of city specific GHG footprints from urban household consumption using a method that allows a near term, feasible and cross-city comparable inclusion of upstream emissions into urban GHG inventories. The GHG footprint is a composite indicator that combines direct emissions from local consumption sectors in the city with upstream emissions along global production chains attributable to local consumption. Different versions of GHG footprints are described in the literature 16 . In our study we calculate urban GHG footprints of household consumption, defined as the sum of direct and upstream GHG emissions associated with urban household consumption as defined for the national scale 8 and illustrated for cities in Fig. 1b . In principle, it would be desirable to include the other local final consumption sectors, government consumption and gross fixed capital formation, into the urban GHG footprint. However, as city level data of these consumption sectors are not available, those sectors could only have been included using national averages scaled to the city level. Such an approach does not add any urban specific information and those two categories are therefore not considered in our GHG footprint accounts. However, to facilitate comparison with other studies we do provide national per capita averages for GHG emissions from government consumption and gross fixed capital formation alongside our city results. We show that upstream emissions are relevant in cities in emerging and in developed countries and discuss ways in which local authorities could have substantial policy leverage to reduce both their territorial and their upstream emissions outside their geographic jurisdiction. Upstream emissions of urban household consumption and GHG footprints have been calculated for a number of cities 14 , 16 . Most published studies report results for single cities 6 , 17 , 18 or for multiple cities from a single country, e.g. UK 9 , Australia 10 , Finland 19 , USA 20 , 21 . With the exception of India 18 and China 22 , 23 studies for cities in developing or emerging economies are absent in the literature. The variation reported under the heading of urban GHG footprints is large ranging from 2.4 tCO 2 e/cap*yr (tons CO 2 equivalents per capita and year) in Delhi to 60 tCO 2 e/cap*yr in Luxembourg 17 , 18 . It is important to note however, that different definitions of urban GHG footprints prevail in the literature (see Supplementary Information online for more details) therefore comparability between published results across different studies is very limited. In addition to urban household footprints for individual cities, also aggregated national or regional urban household GHG footprints have been calculated 24 , 25 , 26 . This extremely important work serves a different purpose in providing statistically robust evidence of systemic patterns, such as persistent rural-urban differences or the dominant influence of income on GHG footprints of household consumption. This aggregated approach is, however, less useful for local policy that relies on site specific and comparable accounts, as the huge differences reported for individual cities clearly demonstrates. We compare the upstream GHG emissions from household consumption of four global cities to their territorial emissions from all sources, show GHG footprints of their household consumption, investigate the geographic reach of their global hinterlands, and discuss leverage points for urban policy to reduce their territorial and upstream emissions. Berlin, Delhi NCT (National Capital Territory), Mexico City and the New York MSA (metropolitan statistical area) - four cities from three continents - were selected to represent different size, history, urban form, income level and national culture 27 . To calculate upstream emissions we integrated data from household expenditure surveys for each of the four cities into Eora, a multi-regional input-output (MRIO) model with environmental extensions 28 . Our results are reported in tons of CO 2 equivalents and include all Kyoto gases (CO 2 , CH 4 , N 2 O, HFCs, PFCs, SF 6 , NF 3 ) for the year 2012 (2008 for Mexico City). Direct household emissions were taken from the respective local GHG emission inventories (see methods). Results Comparison between territorial and upstream emission The comparison between per capita total territorial emissions (TE) and upstream emissions from household consumption (UE) shows that they are of the same order of magnitude (Fig. 2 , Table 1 ). New York MSA (9.7 tCO 2 e/cap*yr TE and 10.6 tCO 2 e/cap*yr UE) and Berlin (5.6 tCO 2 e/cap*yr TE and 7.3 tCO 2 e/cap*yr UE), the two more affluent cities, have much larger per capita emissions from both accounting perspectives compared to Mexico City (2.8 tCO 2 e/cap*yr TE and 2.3 tCO 2 e/cap*yr UE) and Delhi NCT (1.6 tCO 2 e/cap*yr TE and 1.4 tCO 2 e/cap*yr UE). Upstream emissions from household consumption are substantial, ranging between 81% (Mexico City) and 130% (Berlin) of territorial emissions; and in the two more affluent cities (Berlin and New York) they surpass territorial emissions. Figure 2 Sectoral comparison of total territorial emissions (TE) and upstream emissions of household consumption (UE) among the four cities in units of CO 2 e per capita per year. Full size image Table 1 Upstream and direct household GHG emissions, GHG footprint and total territorial GHG emissions per sector [tCO 2 e/cap*yr]. Full size table Sectoral composition of territorial emissions The main sources of territorial emissions in the four cities are thermal services (space and water heating, cooking) in buildings and transport (Fig. 2 ). Building-related direct emissions are primarily determined by the living space per capita, thermal quality of the building stock, the heating technologies in use (e.g. on-site fuel combustion, district heating or electric heating) and the local climate 27 . Direct transport emissions are determined mainly by the emission intensity of the vehicle fleet and the share of private motorized trips in the modal split (i.e. the relative shares of different modes of transportation). The modal split, in turn, is influenced by urban form (including public transport infrastructure) and gasoline prices 29 . Territorial emissions summarized in the category “Other” include emissions from industry, commerce and public infrastructure services. Their share is highly variable and reflects the economic structure of a city (especially the presence of heavy industry) 4 , 9 . Sectoral composition of upstream emissions The average shares across the four cities are 28% for housing, 23% for transport, 26% for food, and 24% for all other sectors. Thus, housing and transport contribute over 50% to the upstream emissions from household consumption. Together with food those three sectors account for three quarters of total upstream emissions, while all other consumer goods and services contribute only one quarter. Upstream emissions in the housing category include those from electricity generation, remote heating, water supply, sewage and solid waste treatment, operational services to collect rent and provide accommodation and home maintenance and repair. Upstream emissions from transport include the production and maintenance of private cars as well as extraction, refining and transportation of gasoline, but exclude direct emissions from the operation of private vehicles. It also includes all emissions from private use of air travel, train, bus and other forms of public transport (including emissions for production, operation and maintenance of vehicles). Upstream emissions from food include production, processing and transportation of food items purchased by urban dwellers and emissions from visits to restaurants (see method section for a detailed breakdown of consumption categories). Urban household consumption GHG footprints The estimated GHG footprints are Delhi NCT 1.9, Mexico City 3.1, Berlin 8.9 and New York MSA 14.2 tCO 2 e/cap*yr (Fig. 3 ). The share of direct household emissions in the footprint is 25% in Delhi, Mexico City, and New York MSA and 18% in Berlin. Housing, transport and food are responsible for over three quarters of the GHG footprints of all cities. The upstream emissions attributable to all other consumer goods (e.g. electronics, clothes, etc.) and health services purchased by urban dwellers only make up between 9–24% of the GHG footprint. Figure 3 Per capita urban GHG footprints of the four cities and their sectoral composition. The shares of direct emissions in the footprint are indicated by criss-cross lines. Full size image Using the GHG footprint to account for urban household emissions offers two main benefits over a simple comparison between total territorial and upstream emissions as shown in Fig. 2 . Firstly, core infrastructure for electricity, heat, water and waste treatment is often situated at the urban periphery. Depending on whether such infrastructure is within or without the administrative territory of a city leads to notorious distortions in territorial emission accounting 30 . Secondly, territorial emissions and upstream emissions are not additive. Double counting would occur whenever parts of the supply chain of goods purchased in the city lie within the territorial boundary (e.g. the emissions of a district heating plant in a city are counted towards its territorial emissions as well as the upstream emissions of households). The GHG footprint used in this study resolves both of these issues. Geographical reach of upstream emissions Besides knowing the sectoral composition of the GHG footprint, urban policy makers can also benefit from knowing its geographic reach. Tracing the geographic locations of upstream emissions gives evidence on how “global” the supply chains of contemporary cities have become. Figure 4 shows that the shares of non-domestic upstream emissions range between 16% (Delhi NCT) and 52% (Berlin). Domestic refers to the nation state in which the city is located. The sample size of four certainly does not permit any generalizable conclusions on the global reach of urban supply chains but the results in Table 2 support the plausible hypothesis that the supply chains of wealthier cities (Berlin and New York MSA) are more international and those of cities in large nation states are more domestic. As shown in Table 2 , the income effect seems to be particularly strong for food and “other” goods and services (primarily manufactures) where the domestic emission shares in Berlin (47% and 37%) and New York (53% and 44%) are considerably lower than in Delhi (79% and 88%) and Mexico City (75% and 51%). Note, that the high shares of domestic emissions in Delhi and Mexico City still correspond to much lower absolute domestic emissions compared to Berlin and New York (see Table 2 ). The comparatively low share of domestic upstream emissions in all sectors in Berlin is partly attributable to the European single market and partly to the large energy imports (mainly from Russia) in Germany’s domestic energy supply. The high share of domestic upstream emissions in the housing and transport sector of New York might be attributable to the large size of the US economy. More details about the regional and sectoral distribution can be found in the Supplementary Information online. Figure 4 The global reach of urban GHG footprints. The four maps show the spatial distribution of the cities’ non-domestic upstream household GHG emissions. Maps are based on the Natural Earth public domain data set ( ) and were created in R 75 using the ggplot2 78 package. Full size image Table 2 Domestic shares in % of upstream emissions of urban households overall and in different consumption sectors (absolute values in tCO 2 e/cap*yr in brackets). Full size table Discussion This study for the first time presents urban household GHG footprints for four cities from three continents using a consistent and widely applicable method for urban consumption based GHG accounting. This is an important step toward creating the basis for comparable benchmarking, which in turn can enable more effective city collaboration and competition to reduce urban GHG footprints. Lack of methodological and terminological standardization, differences in data utilization, unequal inclusion of different GHGs, and differences in the definition of a GHG (or a carbon) footprint concept have so far prevented a meaningful comparison of urban GHG footprints between different studies (see also Supplementary Information online). Previous studies conducted for the same cities calculated GHG footprints at 2.4 tCO 2e /cap*yr for Delhi 18 (Delhi NCT 1.9 tCO 2 e/cap*yr, this study), between 15 and 16 tCO 2 e/cap*yr for Berlin 31 (8.9, this study), and ~16 tCO 2 e/cap*yr for New Yorks 32 (14.2 New York MSA, this study). Our results are consistently lower than those reported in the studies above. The main reasons for the discrepancies are fundamental differences in core definitions, methods, data sources, base years, and partly in geographic area of analysis (e.g. New York City vs New York MSA). The Berlin results were obtained from national consumption data downscaled to the regional level, applied to standard households, using a different input-output model and bottom-up calculations of emission intensities, the details of which are not sufficiently described in the publication. In addition the authors report that their values are approximately 27% higher compared to estimates by the German Ministry of the Environment 31 , 33 . The Delhi study 18 applied a different footprint concept, originally called transboundary infrastructure supply chain footprint 34 , that extends territorial accounting by selected upstream emissions for key infrastructure services and materials (such as electricity, water supply, air transport, cement etc.) provided to the city from outside its territory. Thus, a meaningful comparison to the household consumption footprint calculated here is not possible. Conceptually infrastructure supply chain footprints should be compared to territorial accounts, as the latter is a subset of the former. Because we took the territorial GHG accounts for Delhi directly from 18 this comparison simply repeats that Delhi’s territorial GHG emissions of 1.6 tCO 2 e/cap*yr make up 68% of the extended supply chain footprint of 2.3 tCO 2 e/cap*yr (see Supplementary Information online for more information about different urban GHG accounting approaches). Apart from the peculiarities of the different methods and definitions applied in the literature our results are certainly low end estimations, due to the exclusion of governmental expenditures and gross fixed capital formation in our GHG footprint calculations. Although not specific for the urban scale, we computed per capita upstream GHG emissions for governmental expenditure and gross fixed capital formation for the national level, to provide a first estimate of their scale. The national averages for these two categories are 0.2 and 0.7 for India, 1.2 and 2.8 for Germany, 2.7 and 3.3 for the US, and 0.2 and 1.0 for Mexico, respectively (all in tCO 2 e/cap*yr). Our method complies with the proposed British standard (PAS 2070) 35 and conceptually follows a well-established definition of the GHG footprint on the national level 8 . Eora is a freely available environmentally extended multi-regional input-output model used internationally 28 , 36 . This allows cities to calculate and monitor their household GHG footprint provided direct emissions inventories and urban consumer expenditure surveys are available. At present, the necessity to reconcile incompatible data sources (consumer expenditure data and national input-output tables) requires considerable time effort and the guesswork involved introduces uncertainty into estimated urban GHG footprints. This could be overcome by improved reporting on the urban level and by efforts to harmonize local and national accounting systems. Additional uncertainty comes from the simplifying assumptions in input-output modeling, the given sectoral and spatial resolution of Eora, and uncertainty induced by the balancing algorithm applied in Eora. These are discussed in more detail in the method section. Two further amendments to the presented method should be noted that would greatly improve the policy relevance of GHG footprints. Firstly, as discussed above, the present study considers only emissions from household consumption and disregards emissions for government consumption and investments (gross fixed capital formation). This gap could be closed by collecting city specific data for these two items and would allow the GHG footprint to capture the GHG emissions attributable to government and construction activities, the latter of which were shown to be substantial, particularly in the rapidly growing cities in emerging economies 37 . Secondly, the reliance on national input-output tables implies assuming uniform emission intensities in each sector across the national territory. Thus local efforts to supply low carbon goods and services are hidden in national averages. Ideally this problem could be addressed by providing local scale environmentally extended input-output tables 17 , 38 . However, it is unlikely that local input-output tables will be widely available and comprehensibly integrated into existing MRIOs any time soon. This creates a trade-off between widely adopting a method which can deliver comparable GHG footprint estimates for a large number of global cities and the necessary and continued efforts of the scientific community to increase the resolution and precision of state-of-the-art MRIO systems 38 . The urgency to substantially reduce GHG emissions in all parts of society, however, suggests that we begin with feasible extensions of urban GHG inventories immediately so that the perfect does not become the enemy of the good. We suggest that the method applied in this study is a reasonable balance between the conflicting goals of holistic urban GHG accounting and its near-term feasibility and the accuracy of urban specific GHG inventories and studies. Most importantly, though, our results suggest that urban leaders can, in fact, influence some of the main sources of extraterritorial upstream emissions. It is challenging to directly regulate household consumption choices but several aspects of city policy affect them indirectly. Housing and transport are the main sources of direct household emissions of urban citizens. The same two consumption categories are also responsible (together with food) for the majority of upstream emissions. This commonality suggests that local policies can be effective at reducing both. Transport policies aimed at direct urban emissions include land use and zoning policies to encourage higher-density settlements as well as a range of incentives to influence the modal split towards public transport, cycling or walking. Many of the specific measures, such as higher fuel prices or congestion taxes, fast and affordable public transport, or urban planning that encourages walking and cycling also incentivize fewer and smaller vehicles, thus reducing the upstream emissions of the car fleet. The upstream emissions of public transport infrastructure are directly amenable to local policy intervention. Urban public transport policies could, in addition to focussing on low operational energy and GHG emissions, include low carbon materials for public transport fleets and infrastructure into their climate mitigation goals. Cities’ options certainly depend on the supply of such alternatives, but with continuing volatile prices in global commodity markets 36 and growing concerns about the criticality of mineral raw materials supply 39 , material-efficient vehicles with low carbon emissions in the production phase could become economically attractive goals also for the vehicle manufacturing industry. The same logic applies to measures aiming to reduce emissions from housing. Building codes and construction standards that encourage energy efficiency in heating, cooling, and lighting also affect upstream emissions by influencing material choice. There are often trade-offs between the impacts of material choice on upstream and operational energy, but the point to note is that the policy lever for direct and upstream emissions associated with these aspects of housing is the same. Our findings suggest that it is important to revise policy goals based on lifecycle evaluation. “Zero waste” initiatives focused on reducing per capita environmental waste affect both direct and upstream emissions simultaneously. As in transport, city leaders have more direct control over the emissions from water and waste treatment services since these are often at least in part publicly financed. Finally, in parallel to market conditions motivating material efficiency and low carbon material choices, the rise and increasing ease of mobilisation of political protests against extraterritorial waste disposal may be a force against simply moving emissions outside the boundaries. Deep emissions reductions in both sectors will require revising regulatory regimes and focusing them on metrics that acknowledge both direct and upstream emissions. Technology-forcing regulations based on progressively improving performance standards rather than technology prescription can be a significant driver of innovation, as demonstrated by the success of California’s climate mitigation policy 40 . As these are rolled out, however, the performance parameters need to be carefully and comprehensively designed. Food consumption patterns seem less easily accessible for local policy. However, cities have some leverage via green procurement in public facilities (e.g. hospitals, schools, etc.). Considering that the livestock sector globally contributes 80% of all food related GHG emissions 41 , simply introducing or increasing vegan choices in communal catering could already have a significant impact. Further, a growing number of cities is starting to see the social, economic and ecological benefits of working more closely with their direct “hinterlands” (Sustainable Food Cities, Greenbelt Foundation), thereby reducing their food GHG footprint and strengthening local policy leverage. To properly reflect the effects of such promising policies in GHG inventories of cities still requires considerable efforts in data collection and harmonization. We think that only by incorporating GHG footprints into their routine planning will cities have the ability and incentive to help overcome some of the practical limitations of current urban GHG accounting. The Paris agreement of COP21 envisages cities as core elements in the UNFCCC process on climate change mitigation and many cities have already become committed and organized actors by joining city networks with a pledge to reduce GHG emissions 42 . The capacity and commitment of cities to act on climate change mitigation even in times of political uncertainty on the national and international level may prove essential to fulfilling the Paris agreement of keeping the increase of global mean temperature below the 2 °C guardrail. The method and results presented in this study provide an important first step towards internationally comparable benchmarking of the GHG footprints of cities and highlight why cities must both be encouraged and enabled to focus on their full emissions impact – upstream emissions as well as territorial emissions – as they continue to develop their climate mitigation plans. Methods The GHG footprints of urban household consumption reported in this study include direct emissions from urban consumption activities (space and water heating, cooking, fuel use from combustion engines) and upstream emissions, i.e. global supply chain emissions attributable to the goods and services purchased by local consumers. They exclude emissions attributable to government services and capital investments for which city specific data was not available. Direct emissions from private consumption are based on local emission inventories and upstream emissions were calculated using local consumer expenditure surveys (CES) and environmentally extended multi-regional input-output modeling (Eora) 28 . Each section of the technical method description is preceded by a short non-technical summary. Upstream Emissions CES data for Delhi NCT 43 , 44 , Berlin 45 , 46 , 47 , Mexico City 48 , 49 and New York MSA 50 were taken from local statistical sources. A number of pre-processing steps are necessary to make the raw CES data compatible to Eora. First, as the classification of consumption categories is different in each consumer expenditure survey and different between CES and the classifications used in Eora we constructed correspondence tables that map the corresponding categories of CES and Eora, simply called mapping in the remaining description. The classifications in Eora closely resemble the input-output tables provided by the national statistical offices of each country. Second, the local currency purchaser prices used in the CES are converted to USD base prices on a sector by sector basis. Finally, the import structure of urban final demand is mapped according to the national import structure. Data Pre-processing (1) The raw CES data need to be transformed into a final demand vector that conforms to the respective sectoral structure used in the national section of each city’s home country in Eora. The correspondence between CES categories and Eora sectors is many-to-many, meaning that one CES category can correspond to multiple Eora sectors and vice versa. This correspondence was manually established with the help of various statistical sources that contain descriptions of the respective CES categories and Eora sectors (Eora sectors are similar but not always equivalent to the national accounts statistics reported by most countries). The statistical sources used were: Delhi NCT 51 /India 52 , Berlin/Germany 53 , 54 , Mexico City 55 /Mexico 56 , and New York MSA 57 , 58 /USA 59 . For each city-country pair (e.g. Berlin and Germany) the result of this process is a binary correspondence matrix C nxm , where n is the number of city CES categories and m is the number of national Eora sectors. An element c ij is 1 if CES category i corresponds to Eora sector j and 0 otherwise. Given C , the system is still under-defined if we do not assume a uniform distribution between corresponding categories and sectors (e.g. expenditures in CES category “fruit” corresponds to Eora sectors “apples” and “pears” but not necessarily in equal shares). We assume the same ratio between corresponding categories/sectors (e.g. apples and pears) for urban and national final household demand. For this, we multiply each element in C column-wise with the Eora domestic final demand vector Y and normalize row-wise (dividing each element by its associated row sum) to arrive at a correspondence coefficient matrix C′. The city final demand vector Y ′ can then be calculated as $${Y}^{^{\prime} }=x\ast {C}^{^{\prime} }$$ (1) where x is the CES vector of the city. (2) The CES data must be converted from local currency purchaser prices into USD base prices for use with the emission coefficients provided with in the Eora satellite accounts. Currency conversion was performed via World Bank official exchange rates 60 which best reflect relative prices of tradeable goods 61 . Purchaser price (PP) to base price (BP) conversion was performed on a sector by sector basis using national BP/PP ratios calculated from Eora. Due to some inconsistencies in the data (i.e. negative purchaser prices or extremely high BP/PP ratios), base prices were capped at five times above or below purchaser prices (sensitivity analyses for factors 2 and 10 were performed). (3) In step 1 we have mapped all CES categories on domestic Eora sectors because CES data does not contain information on whether goods purchased in a city were produced in the domestic national economy or imported. We assume the urban import structure on a sector by sector basis to be equivalent to that on the national level. The practical problem presented by Eora is that it provides no correspondence tables between the heterogeneous national sector definitions across the 184 different countries (adding up to 14,838 sectors in total). While in principle the entire correspondence table could be constructed manually based on statistical sources, the large number of sectors in Eora makes this impractical. Instead, for each city’s home country we manually map those foreign sectors (sorted by decreasing size) that represent >90% of national final demand to 12 aggregate sectors (Agriculture, food, fossil fuel, manufacturing, furniture, electronics, paper, recreation, textiles, transport, health, housing). After also establishing correspondence tables between those 12 sectors and the domestic Eora sectors we obtain sectoral domestic/import ratios according to which we distribute city final demand Y ′ across domestic and foreign sectors to generate the internationalized city final demand vector Y c to be used in the input-output calculation. Computing upstream emissions Input-output tables are consistent, quantitative representations of the interlinkages (supply and use) among all production and final demand sectors within an economy, measured in monetary units. Eora, a multi-regional input-output model, represents the interlinked sectors of 184 countries (with a total of 14,838 sectors). Together with environmental extensions (i.e. of the total GHG emissions in physical units per sector), this allows to compute the upstream emissions along the entire supply chain for a given output (in USD) of any given sector. The Leontief Total Requirements matrix (LTRM), also called the Leontief inverse, L is computed as $$L={(I-A)}^{-1}$$ (2) where I is the identity matrix and A is the technical coefficient matrix. A is calculated as $$A=Z\ast {x}^{^{\prime} -1}$$ (3) where Z is the inter-industry matrix and x ′ is the total output vector x diagonalized into a matrix 62 . Eora provides satellite data 63 for each country and sector including annual emissions of Kyoto protocol greenhouse gases (GHGs) based on the Doha amendment 64 (CO 2 , CH 4 , N 2 O, HFCs, PFCs, SF 6 , NF 3 ). The satellite accounts for the F-gases provided by Eora had to be modified due to errors that are likely artefacts of Eora’s balancing and optimization algorithms. Instead of zeros, most cells uniformly contain 0.15 kt and some have negative emissions. Because overall F-gas emissions are very low but their conversion factors to CO 2 e are large, this leads to large distortions if uncorrected. The data was modified by subtracting 0.15 kt from each cell and replacing all negative values with zeros. This correction was successfully validated against national emissions reports 65 . Non-CO 2 gases are converted to CO 2 e using the common SAR GWP-100 66 (and 67 for NF 3 ). The coefficient vector k (for CO 2 e emissions per USD and sector) is defined by dividing the annual emissions by total sectoral output. The coefficients k , the LTRM L and the city final demand vector Y c , yield the upstream emissions e in international sectoral resolution: $$e=k\ast L\ast {Y}_{c}.$$ (4) Territorial emissions Territorial emissions were taken from municipal GHG inventories. Not all cities reported all Kyoto gases at the same level of detail: Berlin 46 (2012): CO 2, CH 4 , N 2 O; Mexico City 68 (2012): CO 2 , CH 4 , N 2 O; NCT Delhi 18 (2009): CO 2 , CH 4 , N 2 O). Direct emissions for New York MSA were gathered from four different greenhouse gas emission inventories (North Jersey 69 (2006): CO 2 , CH 4 , N 2 O, HFCs, PFCs, SF, Long Island 70 (2010): CO 2 , CH 4 , N 2 O, Mid-Hudson 71 (Putnam, Rochester and Westchester counties only) (2010): CO 2 , CH 4 , N 2 O, HFCs, PFCs, SF 6 , New York City 72 (2012): CO 2 , CH 4 , N 2 O, SF 6 ). The direct emission component in the GHG footprint includes only those territorial emissions emitted by the residents of the city. These include fuel use for space and water heating, cooking and private motorized transport. While those numbers could be taken directly from the GHG inventories of Delhi NCT and Mexico City, the statistical data for Berlin and New York MSA had to be disaggregated using additional sources from the literature. For Berlin, overall road transportation emissions had to be disaggregated to private motorized transport emissions. The nationwide private vehicle emissions share of total road transport emissions in 2010 (74%) was taken from the Federal Environmental Agency (UBA) 73 and multiplied by overall 2012 Berlin road transportation emissions. Similarly, private motorized vehicle emissions in the North Jersey and Mid-Hudson counties were calculated by multiplying the light duty vehicle emissions share of overall road transportation emissions in the US in 2005 (75.37%) and 2010 (73.45%) 74 with North Jersey and Mid-Hudson road transportation emissions, respectively. New York City’s taxi and for-hire car emissions (2006) share of total private motorized transport emissions (2005) (20.7%) 72 was subtracted from total private motorized transport emissions (2012). One overall caveat is that territorial transport emissions include local emissions by non-citizens (e.g. tourists) and exclude emissions of citizens outside the city. Result Aggregation and Visualization For reasons of space and clarity, the detailed sectoral results for upstream emissions were aggregated into four categories for the presentation of the GHG footprint results. The composition of the four categories is: Food : upstream: food, food away from home, non-alcoholic and alcoholic beverages, tobacco Housing : upstream: rent/shelter, energy, maintenance of buildings, utilities; direct: heating and cooking using fossil fuels or charcoal/wood Transport : upstream: purchase of cars, motor bikes, bikes, repair, maintenance, accessories, fuel, lubricants, public transportation, passenger transport, transport services, international transport, package holidays; direct: gasoline & diesel Other: upstream: services, health and all other manufactures; direct: industry, commerce, services, and government Map Visualizations All maps in Fig. 4 . are based on the Natural Earth public domain data set ( ). The figures were generated in R 75 using the rgeos 76 and rgdal 77 packages for geospatial calculations and ggplot2 78 for visualization. The clusters of countries and associated ranges used in the world maps of Fig. 4 were defined according to Fisher-Jenks using the classInt 79 package. Limitations and Uncertainties The main goal of this study was to obtain methodologically consistent and comparable estimates of the GHG footprints of international cities using a state of the art method that can be extended to additional cities with reasonable effort. At present, the chief way to accomplish this is using MRIO models. However, these models (like all models) incorporate a number of simplifying assumptions and presently have restricted sectoral and geographical resolution. From this follow some caveats which lead to uncertainties in the GHG footprint estimates. In input-output modeling, each sector is assumed to produce one uniform good (or a uniform basket of goods) at a uniform price and the same emission factor will apply to all goods within one sector of one country. For example, in the US input-output table, one dollar spent at an up-scale, locally sourced vegan restaurant in Manhattan will generate the same emissions as one dollar spent in an Alabama burger restaurant. Alleviating this problem requires not only input-output tables at higher sectoral and spatial resolution but also their integration into the MRIO model to accurately capture upstream emissions. Efforts to increase the sectoral and spatial resolution of MRIO models are currently underway 38 but are unlikely to cover many parts of the world, particularly in developing countries, in the near future. In addition to the model uncertainty in MRIO models, uncertainty is introduced into the GHG estimates through the mapping of CES to Eora as well as through the empirical uncertainty in the consumer expenditure data and all other statistical econometric and emissions data. The issue of mapping CES data to MRIO models can be addressed by local authorities by harmonizing their survey designs. The numerous sources of conceptual and empirical uncertainty call for a joint effort of the scientific community to develop a first quantification of the uncertainty range of urban GHG accounts. Data availability statement The datasets generated during and/or analysed during the current study which are not publicly available and referenced above are available from the corresponding author on reasonable request.

Greenhouse gas emissions caused by urban households' purchases of goods and services from beyond city limits are much bigger than previously thought. These upstream emissions may occur anywhere in the world and are roughly equal in size to the total emissions originating from a city's own territory, a new study shows. This is not bad news but in fact offers local policy-makers more leverage to tackle climate change, the authors argue in view of the UN climate summit COP23 that just started. They calculated the first internationally comparable greenhouse gas footprints for four cities from developed and developing countries: Berlin, New York, Mexico City, and Delhi. Contrary to common beliefs, not consumer goods like computers or sneakers that people buy are most relevant, but housing and transport - sectors that cities can substantially govern. "It turns out that the same activities that cause most local emissions of urban households - housing and transport - are also responsible for the majority of upstream emissions elsewhere along the supply chain," says lead-author Peter-Paul Pichler from the Potsdam Institute for Climate Impact Research (PIK). "People often think that mayors cannot do much about climate change since their power is restricted to city limits, but their actions can have far-reaching impacts. The planned emission reductions presented so far by national governments at the UN summit are clearly insufficient to limit global warming to well below 2 degrees Celsius, the target agreed by 190 countries, therefore additional efforts are needed." Housing and transport cause most city emissions, locally but also upstream Cement and steel used for buildings take a huge amount of energy - typically from fossil fuels - to be produced, for instance. If a city instead chooses to foster low carbon construction materials this can drastically reduce its indirect CO2 emissions. Even things that cities are already doing can affect far-away emissions. Raising insulation standards for buildings for example certainly slashes local emissions by reducing heating fuel demand. Yet it can also turn down the need for electric cooling in summer which reduces power generation and hence greenhouse gas emissions in some power plant beyond city borders. In transport, expanding public facilities can minimize local emissions from car traffic. This reduces the number of cars that need to be built somewhere else, using loads of energy. So this is a win-win. But, again, more can be done. Cities can decide from which sources they procure the power needed to run, for instance, their subway trains or electric buses. By choosing energy from solar or wind, city governments could in fact close down far-away coal-fired power plants. Comparison of New York, Berlin, Mexico City, Delhi - applicable to cities across the world Interestingly, while the greenhouse gas footprint in the four cities that the scientists scrutinized range from 1.9 (Delhi) to 10.6 tons (New York) of CO2 equivalent per person and year, the proportions of local to upstream household emissions as well as the relative climate relevance of housing and transport turn out to be roughly the same. The international reach of upstream emissions is vast but varies. In terms of emissions, Berlin's global hinterland is largest, with more than half of its upstream emissions occurring outside of Germany, mostly in Russia, China and across the European Union. But also around 20% of Mexico City's considerably smaller upstream emissions occur outside Mexico, mainly in the US and China. "Measuring indirect emissions of urban populations so far has often been considered to be unfeasible, at least in a way that makes it possible to compare different cities," says Helga Weisz, senior author of the study and a research domain co-chair at PIK. "We show that it is possible, but you have to invest the effort to actually do it." Her team analyzed huge amounts of existing data on economic input and output of different regions and successfully combined these with data on emission intensity of production in a lot of different sectors. The methodology that the scientists put together is in principle applicable in any place, enabling more effective collaboration between cities to reduce greenhouse gas emission footprints. "The power of cities, open interconnected systems of great density, to tackle climate change even in times of uncertainty on the national and international level has been underestimated by both many local decision-makers and most of the international community," says Weisz. "Cities must be encouraged and enabled to focus on their full emission spectrum - local and upstream - as they continue to develop their climate mitigation plans."

10.1038/s41598-017-15303-x

Medicine

Activation of two genes linked to development of atherosclerosis

Iwata H et al. "PARP9 and PARP14 cross-regulate macrophage activation via STAT1 ADP-ribosylation." Nature Communications DOI: 10.1038/NCOMMS12849 Journal information: Nature Communications

http://dx.doi.org/10.1038/NCOMMS12849

https://medicalxpress.com/news/2016-10-genes-linked-atherosclerosis.html

Abstract Despite the global impact of macrophage activation in vascular disease, the underlying mechanisms remain obscure. Here we show, with global proteomic analysis of macrophage cell lines treated with either IFNγ or IL-4, that PARP9 and PARP14 regulate macrophage activation. In primary macrophages, PARP9 and PARP14 have opposing roles in macrophage activation. PARP14 silencing induces pro-inflammatory genes and STAT1 phosphorylation in M(IFNγ) cells, whereas it suppresses anti-inflammatory gene expression and STAT6 phosphorylation in M(IL-4) cells. PARP9 silencing suppresses pro-inflammatory genes and STAT1 phosphorylation in M(IFNγ) cells. PARP14 induces ADP-ribosylation of STAT1, which is suppressed by PARP9. Mutations at these ADP-ribosylation sites lead to increased phosphorylation. Network analysis links PARP9–PARP14 with human coronary artery disease. PARP14 deficiency in haematopoietic cells accelerates the development and inflammatory burden of acute and chronic arterial lesions in mice. These findings suggest that PARP9 and PARP14 cross-regulate macrophage activation. Introduction Despite medical advances, the global burden of ischaemic heart disease is increasing 1 , 2 . Pro-inflammatory macrophage activation plays key roles in the pathogenesis of many disorders, including arterial disease 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 . Some pathways associated with macrophage activation may contribute to the shared mechanisms of inflammatory diseases, as demonstrated previously 11 , 12 . Despite potent therapies such as cholesterol-lowering by statins, substantial residual cardiovascular risk remains 7 , 13 , 14 , which drives the active search for novel solutions against pro-inflammatory macrophage activation. Dissecting complex and intertwined mechanisms for macrophage activation requires well-defined mechanistic models. The evidence suggests that distinct types of macrophage activation are functionally different in disease pathogenesis, a classification that has helped to assess the heterogeneity of macrophages 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 . For instance, pro-inflammatory and anti-inflammatory phenotypes can oppose one another, develop in response to distinct cytokines, differ in the activating stimuli and produce different cytokines. A recently proposed nomenclature suggests that each macrophage subpopulation can be called based on a specific stimulator, for example, M(IFNγ), M(LPS), M(IL-4), M(IL-10) 21 . This established paradigm demonstrates clear relationships between classical stimuli and their respective responses—interferon gamma (IFNγ) for pro-inflammatory activation in settings such as atherosclerotic vascular disease and interleukin (IL)-4 for activation that can counter that of M(IFNγ) or M(LPS) macrophages. Hence, we used this paradigm as a starting point to explore novel regulators through global proteomics. Proteomics screening and bioinformatics in mouse and human data sets found that poly ADP-ribose polymerase 14 (PARP14), also known as ADP-ribosyltransferase diphtheria toxin-like 8 (ARTD8), and PARP9/ARTD9 both increased in M(IFNγ) and decreased in M(IL-4) cells. The network analysis associated these PARP family members with human arterial disease. Sequence similarity to the PARP catalytic domain, which transfers ADP-ribose moieties from NAD to protein acceptors, characterizes the PARP family proteins 23 . The best-characterized member, PARP1/ARTD1, represents poly-ADP-ribosylation enzymes, which processively catalyse long and branching polymers of ADP-ribose additions starting from an initial post-translational modification, commonly of glutamate. Recent evidence also validates proteins that execute mono-ADP-ribosylation as having various functions 24 . PARP14/ARTD8 is an intracellular mono-ADP-ribosyltransferase. Previous reports indicated that PARP14 enhances IL-4-induced gene expression by interacting with the cytokine-induced signal transducer and activator of transcription 6 (STAT6) in B and T cells, thereby functioning as a transcriptional co-activator 25 , 26 that may mediate this effect. A recent study reported that PARP14 regulates the stability of tissue factor mRNA in M(LPS) in mouse 27 . Less information exists regarding the molecular function of PARP9/ARTD9. Although PARP9 appears to lack catalytic activity 28 , it increases IFNγ-STAT1 signalling in B-cell lymphoma 29 . This study employed a multidisciplinary approach, including proteomics, systems biology and cell and molecular biology to explore new mechanisms for modulating the functional profile elicited after macrophage activation. Mouse and human cell lines as well as primary macrophages were used for complementary analyses of PARP14-deficient mouse and human tissues. Ultimately, the analyses led to evidence that expression of PARP14 in haematopoietic cells restrains vascular inflammation in mouse models, which are not solely regulated by either IFNγ or IL-4. Our findings suggest a novel mechanism for regulating the balance of macrophage phenotypes in vascular disease, and potentially other disorders in which macrophage activation has an impact on outcomes. Results Proteomics screening for regulators of macrophage activation We used the tandem mass tagging (TMT) quantitative proteomics to identify regulators of pro-inflammatory and non/anti-inflammatory activation in mouse RAW264.7 and human THP-1 macrophage cell lines ( Supplementary Fig. 1a–c ). In this paradigm of macrophage heterogeneity, IFNγ and IL-4 promote distinctive subpopulations 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 . A pilot TMT proteomic study ( Supplementary Figs 1d and 2 ) analysed the changes in the proteomes at 0, 12 and 24 h, and observed the expected increase and decrease in STAT1 in M(IFNγ) and M(IL-4) cells, respectively, as determined by hierarchical cluster analysis ( Supplementary Fig. 2 ). Within this pilot study, we first noted that PARP14 co-clustered with STAT1 in the M(IFNγ) and M(IL-4) data ( Supplementary Fig. 2 ). To ascertain whether any changes in the M(IFNγ) and M(IL-4) proteomes were not because of cell culture conditions, we performed a second, more in-depth study that included an unstimulated macrophage control for both RAW264.7 and THP-1 experiments, and extended the stimulation period to up to 72 h, sampling six time points for a more detailed time-resolved proteomic study ( Supplementary Fig. 1d ). In this latter proteomic study, we quantified 5,137 and 5,635 proteins in RAW264.7 and THP-1 cells, respectively, across the three conditions: unstimulated control, IFNγ-stimulated macrophages and IL-4-stimulated macrophages—M(-), M(IFNγ) and M(IL-4), respectively ( Fig. 1a and Supplementary Table 1 ). An overview of the protein intensities across the three conditions revealed that the magnitude of protein abundance levels for each of the IFNγ and IL-4-stimulated RAW264.7 and THP-1 cells were generally higher than those of unstimulated cells ( Fig. 1b ), indicating that each stimulation promoted changes in protein abundance beyond those due to cell culture conditions alone. Figure 1: Bioinformatics to identify candidate regulators of macrophage activation. ( a ) Venn diagrams showing the distribution of quantified proteins from mouse RAW264.7 and human THP-1 cells in unstimulated control, IFNγ-stimulated and IL-4-stimulated macrophages: M(-), M(IFNγ) and M(IL-4), respectively. ( b ) Data set-filtering strategy. Upper panels: superimposition of the 0-h-normalized protein abundance profiles for M(-) (grey traces) versus M(IFNγ) (red traces) or M(IL-4) (blue traces) data sets in RAW264.7 and THP-1 cells. Lower panels: extracted protein profiles of interest generated by data filtering. Red traces only graphs: extracted profiles of proteins whose abundances exceed the M(IFNγ) threshold (+0.13, maximum protein abundance in unstimulated control at 8 h, dashed line). From these M(IFNγ)-filtered traces, those that decreased in IL-4 stimulation when compared with their unstimulated control are plotted to the right for RAW264.7 and THP-1 cells, respectively. ( c ) Hierarchical clustering of 38 proteins from that were identified in both RAW264.7 and THP-1 data sets. Each row corresponds to a protein gene ID. Full size image To pursue one class of potential upstream regulators based on protein abundances, we used the following criteria: proteins exhibited (1) an early increase in M(IFNγ) (within 24 h) followed by sustained levels until the later time points (up to 72 h), (2) a decrease in abundance in M(IL-4) and (3) no significant change in M(-). We employed two distinct informatics methods to explore proteins with such behaviours: data filtering (Method 1) and model-based clustering (Method 2; Supplementary Fig. 3 ). In Method 1, the M(-) data set permitted a subtraction of the background signal at all time points and thus facilitated data filtering ( Fig. 1b upper panels and Supplementary Fig. 3b ). Three proteins in the data set from mouse RAW264.7 cells and 12 proteins in human THP-1 cells ( Fig. 1b lower panels) met the three criteria above. PARP14 emerged from both data sets, and PARP9 appeared in the THP-1 data set ( Fig. 1b lower panels). In parallel, Method 2 ( Supplementary Fig. 3c ) produced 15 and 20 clusters for the RAW264.7 and THP-1 data sets, respectively ( Supplementary Fig. 4a ). Clusters mined for proteins whose abundances increased in M(IFNγ) but decreased in M(IL-4) with respect to M(-) revealed 490 proteins in the RAW264.7 data set and 414 proteins in THP-1 fulfilling these criteria ( Supplementary Fig. 4b and Supplementary Tables 2 and 3 ). The proteins were short-listed to those with similar time-resolved changes in both in human and mouse data sets, resulting in 38 candidate proteins ( Fig. 1c ). Both RAW264.7 and THP-1 data sets identified PARP9 and PARP14 ( Supplementary Fig. 4a ). Collectively, while Method 2 identified PARP9 and PARP14 in RAW264.7 and THP-1 cells, PARP14 was the only common protein that both Methods 1 and 2 identified among over 5,000 proteins in mouse and human data sets. The PARP9 and PARP14 network and coronary artery disease To understand the influence of PARP9 and PARP14 in a global interaction network (‘interactome’) and predict their potential clinical impact in disease mechanisms, we applied a network-based analysis ( Supplementary Fig. 5a ). Increasing evidence suggests that disease genes are not distributed randomly on the interactome but work together in similar biological modules or pathways 30 , 31 . Moreover, gene products (for example, proteins) linked to the same phenotype likely interact with each other and cluster in the same network neighbourhood 31 . We thus postulated that, if PARP9 or PARP14 influences the network neighbourhood of a disease, its immediate neighbours should be close to a disease module compared with random expectation 30 , 31 . Using the random-walk method, we defined a set of genes as a human disease module for each of the cardiovascular and metabolic diseases, and IFNγ-related diseases (see Methods). We then measured the average shortest distance of the immediate neighbours of the PARP9–PARP14 network to each disease module. An interactome mainly describes a set of physical or functional associations between proteins and does not provide cell or tissue specificity. In our study, we thus examined the genes of immediate neighbours reported in macrophages in the public databases. We included IFNγ-related autoimmune diseases as positive controls, as we used this cytokine to promote pro-inflammatory macrophage activation. The PARP9–PARP14 network is significantly close to systemic lupus erythematosus, dermatomyositis and polymyositis, as expected ( Fig. 2 ). The PARP9–PARP14 network had significantly greater proximity to the human coronary artery disease gene module compared with other cardiovascular and metabolic diseases ( Fig. 2 and Supplementary Fig. 5b ). The analysis also linked PARP9–PARP14 with osteoporosis. Moreover, we quantified the closeness of PARP9–PARP14 to autoimmune diseases, coronary artery disease and other diseases in the form of the distribution of shortest distances ( Supplementary Fig. 5c ) and demonstrated a clear separation between the distribution of diseases inside and outside the circle ( Fig. 2 ), indicating an enrichment of shorter distances ( d =1 and 2) for IFNγ-related autoimmune diseases (yellow) and coronary artery disease and osteoporosis (blue), compared with other cardiovascular and metabolic diseases (red; Supplementary Fig. 5c ). These results may indicate the potential impact of PARP9 and/or PARP14 on the pathogenesis of arterial disease or the onset of its clinical complications. Figure 2: Network analysis links PARP9–PARP14 with coronary artery disease. The PARP14 (blue)–PARP9 (purple) module consists of the first neighbours of each protein (light blue and orange nodes, respectively). The significance of closeness of the PARP9–PARP14 first neighbours in the interactome (PARP9–PARP14 module) and disease modules compared with random expectation is indicated by P values. The random expectation was same size-connected components of PARP9–PARP14 module and a disease module drawn randomly from the network. Closeness between PARP9–PARP14 modules and other diseases such as cardiovascular, metabolic and IFNγ-related diseases was evaluated in the network. The inner circle contains significantly close disease modules. Full size image In vitro validation in cultured macrophages qPCR and western blot analysis validated the proteomic screening data on PARP9 and PARP14. Consistent with the proteomics data ( Fig. 3a and Supplementary Fig. 4a ), mRNA and protein levels of PARP9 and PARP14 increased with IFNγ and decreased by IL-4 ( Fig. 3b and Supplementary Fig. 6 ). At protein levels, PARP14 increased before PARP9 in response to IFNγ ( Fig. 3c ). Figure 3: PARP9 and PARP14 expression in vitro and in vivo . ( a ) TMT-derived 0-h-normalized protein abundance profiles for PARP9 and PARP14 from mouse RAW264.7 and human THP-1 M(IFNγ) and M(IL-4) data sets. ( b ) PARP9 and PARP14 gene expression at 24 h after stimulation ( n =3). ( c ) PARP9 and PARP14 protein expression visualized by western blot. The time course in the relative protein abundances of PARP9 and PARP14 normalized to β-actin were quantified (graph, n =3). * P <0.05 and ** P <0.01, respectively, by Student’s t -test. Error bars indicate s.d. ( d ) Representative images of PARP9 and PARP14 expression in atherosclerotic plaques from the aorta of an Apoe −/− mouse ( n =3) fed a high-fat diet and from the carotid artery of a human ( n =5). Scale bars, 100 μm. Full size image PARP9 and PARP14 expression in plaque macrophages According to the network analysis that linked PARP9 and PARP14 with arterial disease, we performed immunohistochemistry in arterial lesions. Mouse ( Fig. 3d , left) and human ( Fig. 3d , right) atherosclerotic lesions exhibited PARP9 and PARP14 proteins, while they were less abundant in human carotid arteries with no apparent atherosclerotic changes ( Supplementary Fig. 7a ). Immunohistochemistry localized PARP9 and PARP14 expression in the majority of macrophages (CD68) of human atherosclerotic plaques, while few if any smooth muscle cells (SMα-actin) and endothelial cells (CD31) stained positively for these PARPs ( Supplementary Fig. 7b ). These results suggest that macrophages are a major source of PARP9 and PARP14 in human atherosclerotic lesions. Pro-inflammatory PARP9 and anti-inflammatory PARP14 A series of subsequent experiments examined whether PARP9 or PARP14 plays a causal role in macrophage activation. Expression levels of gene products typical of the M(IFNγ) phenotype (for example, IL-1β) and in the M(IL-4) phenotype (for example, MRC1) gauged the downstream effects of PARP silencing using small interfering RNA (siRNA) first in macrophage-like cell lines RAW264.7 and THP-1 since the original screening was performed in these cells, and then in human and mouse primary macrophages. In mouse RAW264.7 cells, PARP14 silencing enhanced the induction of tumour necrosis factor alpha (TNFα) and inducible nitric oxide synthase (iNOS) by IFNγ, and suppressed the response of MRC1 to IL-4 ( Supplementary Fig. 8a ). In human THP-1 cells, PARP14 silencing also increased TNFα and IL-1β mRNA induction in response to IFNγ and decreased MRC1 induction by IL-4 ( Supplementary Fig. 8b ). Increased levels of TNFα and IL-1β proteins in the supernatant of IFNγ-treated THP-1 cells with PARP14 silencing ( Supplementary Fig. 8c ) supported these data. Silencing of PARP9 suppressed the induction of TNFα, IL-1β and CCL2/MCP-1 mRNA under IFNγ stimulation in THP-1 cells, while MRC1 showed no significant change in IL-4 ( Supplementary Fig. 8d ). To obtain unambiguous evidence for the role of PARP9 and PARP14 in macrophage activation, we extended the in vitro validation studies to mouse and human primary macrophages. In human primary macrophages derived from CD14+ peripheral blood mononuclear cells (PBMCs), PARP9 silencing suppressed the expression of TNFα, IL-1β and CCL2/MCP-1 in IFNγ-treated macrophages, but exerted no significant effects on MRC1 induction by IL-4 ( Fig. 4a ). In contrast, PARP14 silencing accelerated the induction of TNFα, IL-1β and CCL2/MCP-1 by IFNγ, and suppressed MRC1 in IL-4-treated macrophages ( Fig. 4a ). In mouse bone marrow-derived macrophages (BMDMs), silencing PARP9 and PARP14 exerted the effects similar to those in human primary macrophages ( Fig. 4b ). Overall, siRNA experiments provided consistent results in macrophage-like cell lines and primary macrophages. Neither PARP14 nor PARP9 showed significant effects on viability, proliferation or apoptosis of mouse primary macrophages ( Supplementary Fig. 8e ). In addition, enforced expression of PARP14 in THP-1 cells suppressed the induction of TNFα, iNOS, TLR2 and TLR4 in M(IFNγ) ( Supplementary Fig. 8f,g ). Collectively, these findings indicate that PARP14 suppresses IFNγ-induced responses and augments IL-4-responses in macrophages. In contrast, PARP9 promotes responses to IFNγ. Figure 4: The molecular functions of PARP9 and PARP14 in macrophages in vitro . ( a ) The consequences of PARP9 and PARP14 silencing on IFNγ stimulated (TNFα, IL-1β and CCL2/MCP-1) and IL-4 stimulated (MRC1) gene expression in human primary macrophages ( n =8). ( b ) The consequences of PARP9 and PARP14 silencing on IFNγ stimulation (TNFα and iNOS) and IL-4 stimulation (Arg1 and MRC1) gene expression in mouse bone marrow-derived macrophages ( n =3). ( c ) The ratio of phosphorylated STAT1 and STAT6 protein levels to total STAT1 and STAT6 (pSTAT1/tSTAT1 ratio and pSTAT6/tSTAT6 ratio) in human primary macrophages ( n =6 and n =5, respectively) of the PARP9 and PARP14 silencing experiments. * P <0.05 and ** P <0.01, respectively, by Student’s t -test. Error bars indicate s.d. Full size image PARP9 and PARP14 regulate STAT1 and STAT6 activation Although expression patterns of PARP14 and PARP9 in M(IFNγ) and M(IL-4) were comparable, their respective siRNA experiments yielded opposing results, suggesting the involvement of distinct signalling mechanisms. IFNγ signalling involves activation (phosphorylation) of pro-inflammatory STAT1, while the IL-4 pathway uses anti-inflammatory STAT6 phosphorylation 32 , 33 . Immunofluorescence staining demonstrated enhanced intracellular colocalization of PARP14 and STAT1 in IFNγ-treated THP-1 cells compared with unstimulated cells ( Supplementary Fig. 9 ). Moreover, in human primary macrophages derived from CD14+ PBMCs, PARP14 silencing accelerated IFNγ-induced STAT1 phosphorylation and suppressed IL-4-promoted STAT6 phosphorylation ( Fig. 4c ). PARP14 may suppress pro-inflammatory macrophage activation by modulating the IFNγ–STAT1 axis, and promote the anti-inflammatory IL-4-STAT6 pathway. In contrast, PARP9 silencing decreased STAT1 phosphorylation in IFNγ-treated human macrophages ( Fig. 4c ), indicating that PARP9 may activate IFNγ–STAT1 signalling and induce pro-inflammatory activation. In addition, siRNA experiments in THP-1 cells that we used for proteomic screening produced similar results on the role of PARP9 and PARP14 on the STAT1 and STAT6 pathways ( Supplementary Fig 10a–c ). Evidence suggests the participation of STAT3 in immune responses in various contexts 34 . Neither PARP14 nor PARP9 silencing, however, exerted significant effects on STAT3 phosphorylation ( Supplementary Fig. 10d,e ). PARP9 and PARP14 interact with each other The results demonstrated above engendered the hypothesis of regulatory interplay between PARP9 and PARP14. Indeed, PARP14 silencing increased PARP9 mRNA expression in IFNγ-treated human primary macrophages ( Fig. 4a ) and THP-1 cells ( Fig. 5a ), while PARP9 silencing increased PARP14 mRNA in these cell types ( Figs 4a and 5a ). In contrast, enforced expression of PARP14 decreased PARP9 mRNA expression in IFNγ-treated THP-1 cells ( Fig. 5a ). Previous reports showed that, in B lymphocytes, IL-4 promotes catalytic activity of PARP14, leading to ADP-ribosylation of HDAC2, HDAC3 and p100 (a precursor of p52 that encodes the NF-κB2 protein), and activation of STAT6, thereby inducing its binding to IL-4-responsive gene promoters 25 , 26 , 35 , 36 . Co-immunoprecipitation (IP) revealed a PARP9/PARP14 complex ( Fig. 5b ) and immunofluorescence demonstrated the enhanced colocalization of PARP9 and PARP14 in THP-1 cells by IFNγ stimulation ( Fig. 5c ), suggesting that these two molecules interact in macrophages. Recombinant PARP14 protein induced ADP-ribosylation of PARP14 itself, PARP9, STAT1α and STAT6 ( Fig. 5d ). PARP9 did not promote ADP-ribosylation of either STAT1α or STAT6, which supports a previous report showing that PARP9 lacks catalytic activity 28 . Interestingly, PARP9 suppressed PARP14-induced ADP-ribosylation of STAT1α and STAT6 ( Fig. 5d ). While the majority of other PARP family members such as PARP1 are poly-ADP-ribosylation enzymes, PARP14 is a mono-ADP-ribosyltransferase 24 . Mass spectrometric analysis determined that Glu657 and Glu705 of STAT1α were mono-ADP-ribosylated by PARP14 ( Fig. 6a,b ). Glu657 and Glu705 neighbour Tyr701, a functionally critical phosphorylation site of STAT1α ( Fig. 6a ). Although we could not verify the precise ADP-ribosylation site in the STAT6 peptide, the conserved Glu makes it a plausible candidate ( Supplementary Fig. 11 ). Mass spectrometry further revealed that recombinant PARP9 inhibited PARP14-induced mono-ADP-ribosylation at Glu657 and Glu705 of STAT1α ( Fig. 6b , right panels). Figure 5: Potential interaction of PARP9 and PARP14. ( a ) PARP14 silencing and enforced expression significantly affected PARP9 gene expression in IFNγ-stimulated THP-1 cells ( n =3). PARP9 silencing increased PARP14 gene expression ( n =3). ( b ) Co-IP assay revealed a complex between PARP9 and PARP14. ( c ) Intracellular colocalization of PARP9 and PARP14 in the cytosol in M(-) and M(IFNγ). ( d ) PARP9 inhibits ADP-ribosylation of STAT1α and STAT6 by PARP14 (protein ribosylation assay). PARP14 auto-ribosylation is also indicated. * P <0.05 and ** P <0.01, respectively, by Student’s t -test. Error bars indicate s.d. Full size image Figure 6: Identification of PARP14-induced ribosylation sites in STAT1. ( a ) The amino-acid sequence of human STAT1α C terminus. Green amino acids indicate ribosylated peptides; confirmed ribosylation sites are underlined. STAT1 is phosphorylated at indicated tyrosine (red). ( b ; Left panels) MS/MS spectra for the mono-ADP-ribosylated peptides and corresponding unmodified forms. ADP-ribose fragments are annotated in green. *, ribosylation site; m, oxidized methionine. The grey circles indicate background or undetermined ions. (Right panels) MS1-based quantification of PARP9 inhibition of PARP14-mediated STAT1α ribosylation at E657 (upper panel) and E705 (lower panel), respectively. ( c ) Effects of mutated amino acids at E657 and E705 in STAT1 (ribosylation sites for PARP14) on its Tyr701 phosphorylation and pro-inflammatory gene expression in mouse bone marrow-derived macrophages ( n =4). * P <0.05 and ** P <0.01, respectively, by Student’s t -test. Error bars indicate s.d. Full size image The introduction of mutations at Glu657 and Glu705 to prevent ADP-ribosylation helped to investigate the functional relevance of these two sites within STAT1α. Overexpression experiments in mouse BMDMs revealed that mutations at Glu657 and Glu705 enhanced Tyr701 phosphorylation of STAT1α in IFNγ-treated cells as compared with wild-type STAT1α ( Fig. 6c ). Using the same mutant STAT1α, we have observed a similar response of STAT1α phosphorylation in HEK293 cells ( Supplementary Fig. 12a,b ). As a functional consequence, mutant STAT1α in mouse BMDMs increased expression of pro-inflammatory iNOS, IL-1β and CCL2/MCP-1 ( Fig. 6c ). Collectively, these data indicate that PARP14-mediated ADP-ribosylation of Glu657 and Glu705 may control Try701 phosphorylation of STAT1α, a potential mechanism for the interplay between ADP-ribosylation and phosphorylation in macrophage activation. PARP14 deletion enhances acute arterial lesion development To provide in vivo evidence that PARP14 participates in arterial lesion formation and macrophage activation, we further used PARP14 −/− mice. Peritoneal macrophages from PARP14 −/− or PARP14 +/+ mice enabled the examination of the macrophage phenotype. PARP14 −/− macrophages expressed higher mRNA and protein levels of iNOS and TNFα under IFNγ stimulation and lower levels of MRC1 and Arg1 mRNAs under IL-4 stimulation compared with PARP14 +/+ cells ( Fig. 7a,b ). PARP14 deficiency also enhanced phosphorylation of STAT1 induced by IFNγ and decreased STAT6 phosphorylation in IL-4-treated peritoneal macrophages ( Fig. 7c ). BMDMs of PARP14 +/+ and PARP14 −/− mice supported our results ( Supplementary Fig. 13a ). These results in PARP14 −/− macrophages are consistent with those of in vitro siRNA experiments. Figure 7: Role of haematopoietic PARP14 in acute arterial lesion formation in mice. ( a–c ) Cultured peritoneal macrophages derived from PARP14 −/− and PARP14 +/+ mice. ( a ) IFNγ and IL-4 pathway gene expression profiles ( n =3). ( b ) Secretion of inflammatory factors into culture media ( n =3). ( c ) Western blot and corresponding densitometry quantification of phosphorylated STAT1 and STAT6. Each data point is the average of triplicate samples per donor ( n =3). ( d ) Left: representative images of haematoxylin and eosin (H&E; top) and Mac3 (bottom) staining. Scale bars, 100 μm. Right: quantification of lesion formation in mechanically injured femoral arteries of PARP14 −/− and PARP +/+ mice. Mac3 staining represents macrophage accumulation ( n =4–5). ( e ) LCM of the neointima followed by gene expression analysis ( n =4). ( f ) Flow cytometry analysis of splenic CD11b+Ly6G− monocytes after induction of mechanically injured femoral arteries of PARP14 +/+ and PARP14 −/− mice ( n =3). ( g ) Representative H&E staining images and quantification of neointima formation in mechanically injured femoral arteries after bone marrow transplantation (BMT) PARP14 +/+→+/+ and PARP14 −/−→+/+ mice ( n =6). Scale bars, 100 μm. * P <0.05 and ** P <0.01, respectively, by Student’s t -test. Error bars indicate s.d. Full size image Our network analysis closely linked the PARP9–PARP14 module with coronary artery disease ( Fig. 2 ). To first examine whether PARP14 indeed plays a role in arterial diseases, we used two different models: (1) acute mechanical injury in femoral arteries of PARP14 −/− mice and (2) acute injury in mice that underwent bone marrow transplantation from PARP14 −/− mice. In a model of acute arterial responses due to wire-mediated mechanical injury, PARP14 deficiency enhanced neointima formation ( Fig. 7d , top panels) and increased macrophage accumulation ( Fig. 7d , bottom panels). Laser capture microdissection (LCM) of the intima of injured femoral arteries followed by real-time PCR demonstrated that PARP14 deficiency increased the levels of TNFα and iNOS mRNA and decreased Arg1 mRNA ( Fig. 7e ). The spleen is a reservoir of monocytes/macrophages that releases these cells in response to an event in remote organs 37 . Flow cytometry of splenic cells revealed that, in CD11b+Ly6G− monocytes/macrophages ( Supplementary Fig. 13b ), acute arterial injury increased the Ly6C high population in PARP14 +/+ mice, which was further expanded by PARP14 deficiency ( Fig. 7f and Supplementary Fig. 13c ). To examine the relative contribution of PARP14 in the haematopoietic lineage to the lesion development after acute mechanical injury, we performed bone marrow transplantation from PARP14 +/+ and PARP14 −/− mice. In lethally irradiated PARP14 +/+ mice whose bone marrow was reconstituted by PARP14 −/− bone marrow (BMT PARP14 −/−→+/+ mice), neointima formation after injury was accelerated, as compared with control PARP14 +/+ mice whose bone marrow was reconstituted by PARP14 +/+ cells (BMT PARP14 +/+→+/+ mice; Fig. 7g ). These results indicate that PARP14 derived from the haematopoietic cell lineage, the majority of which are macrophages in the neointima, plays an important role in the development of arterial disease. Haematopoietic PARP14 deficiency enhances atherogenesis To further examine the role of PARP14 in chronic arterial diseases, we used high-fat/high-cholesterol-fed low-density lipoprotein receptor-deficient ( LDLR −/− ) mice, an established model of atherosclerosis. Lethally irradiated LDLR −/− mice whose bone marrow was reconstituted with PARP14 −/− cells (BMT PARP14 −/−→ LDLR −/− ) or PARP14 +/+ cells (BMT PARP14 +/+→ LDLR −/− ) underwent high-fat/high-cholesterol feeding to develop chronic atherosclerotic plaques. Sixteen weeks after the initiation of an atherogenic diet, BMT PARP14 −/−→ LDLR −/− mice exhibited more plaque formation and macrophage accumulation in the aortic root compared with BMT PARP14 +/+→ LDLR −/− mice ( Fig. 8a ). The aorta of BMT PARP14 −/−→ LDLR −/− contained higher expression levels of the pro-inflammatory TNFα and CCL2/MCP-1, while MRC1 expression tended to be lower ( Fig. 8b ). BMT from PARP14 −/− mice also increased PARP9 expression in the aorta ( Fig. 8b ). These results are consistent with other in vitro and in vivo data that we have reported in the present study. Taken together, these findings indicate that PARP14 derived from the haematopoietic lineage, the majority of which are macrophages in arterial lesions, plays a protective role against the development of arterial diseases, which verifies our prediction by network analysis. Figure 8: Haematopoietic PARP14 in mouse atheromata and PARP9–PARP14 expression in human plaques. ( a ) Representative image and quantification of aortic root lesion formation and CD68+ macrophage accumulation (green, Alexa 488) in high-fat and high-cholesterol diet-fed LDLR −/− mice whose bone marrow was reconstituted by PARP14 −/− mice (BMT PARP14 −/−→ LDLR −/− mice, n =5), compared with LDLR −/− mice with PARP14 +/+ bone marrow (BMT PARP14 +/+→ LDLR −/− mice, n =6–7). Scale bars, 100 μm. ( b ) mRNA expression of the aorta from a . n =6–8. ( c ) Immunofluorescence staining of PARP14 and PARP9 proteins (green, Alexa 488) in human carotid plaques. CD68 (red, Alexa 594). Nuclei (blue, 4,6-diamidino-2-phenylindole, DAPI). Scale bars, 100 μm; insets, 10 μm ( n =5). Prevalence of PARP14+ or PARP9+ macrophages in macrophage-poor versus macrophage-rich plaques. * P <0.05 and ** P <0.01, respectively, by Student’s t -test. Error bars indicate s.d. Full size image Macrophage-rich human atheroma and PARP9 Seeking additional in vivo evidence for the potential role of PARP9 and PARP14 in arterial disease involved immunohistochemical analysis of human carotid atherosclerotic plaques surgically removed by endarterectomy. PARP9 and PARP14 signals were predominantly localized in plaque macrophages, as indicated by the overlapping CD68-positive signal ( Fig. 8c and Supplementary Fig. 14 ). Quantitative analysis demonstrated that more macrophages were immunoreactive for PARP9 in macrophage-rich plaques than in macrophage-poor plaques, while there was no significant difference in PARP14-positive macrophages ( Fig. 8c ). While some macrophages in the human plaques coexpressed PARP9 and PARP14, other cells stained positively for either PARP9 or PARP14 alone ( Supplementary Fig. 14 ). These lines of in vivo evidence indicate that macrophages in arterial lesions are heterogeneous, which may reflect diverse levels of pro-inflammatory activation in individual cells. Single-cell analysis of primary human macrophages The diversity in expression patterns of PARP9 and PARP14 in human plaques necessitated the additional assessment of macrophage subpopulations using single-cell gene expression profiling 38 of PBMC-derived human CD14-positive macrophages. Examined cell numbers were 86 in M(-) and 84 in M(IFNγ) of Donor 1, 93 in M(-) and 86 in M(IFNγ) of Donor 2 and 90 in M(-) and 81 in M(IFNγ) of Donor3, respectively. Our examination investigated whether responses of human primary macrophages towards IFNγ stimulation are heterogeneous. Comparing average levels of readouts (for example, inflammation-related factors) in the entire group of cells by qPCR cannot address this question. A workflow of single-cell gene profiling is shown in Supplementary Fig. 15a . The expression patterns of 91 target genes ( Supplementary Table 4 ) in M(-) ( n =268) and M(IFNγ) ( n =252; n =520 in total) derived from the three different donors. All cells were evaluated against each other based on their gene expression dissimilarity (Manhattan distance, ). We present the distance matrix as a distance-based graph. By combining the two conditions—M(-) and M(IFNγ)—across three donors (six conditions in total), M(-) cells (green) and M(IFNγ) (red) are clearly segregated, forming distinct clusters ( Fig. 9a ). However, there are ‘trails’ of M(IFNγ) that appear in the M(-) cluster, suggesting potential heterogeneity. Although this may be attributable to donor-to-donor variations, similar patterns were observed within individual donors as well ( Supplementary Fig. 15b ). As we are particularly interested in the expression profiles of PARP9 and PARP14, we examined variations/correlations in their expression levels in M(-) and M(IFNγ) human primary macrophages. In both phenotypes, PARP9 and PARP14 were correlated. However, M(-) cells showed lower levels of variation for PARP9 and PARP14 mRNA expression than did M(IFNγ) ( Supplementary Fig. 15c ). These findings suggest that the ‘IFNγ-polarized’ human primary macrophages are heterogeneous. Figure 9: Single-cell gene expression analysis of CD14+ macrophages. ( a ) Heterogeneity in IFNγ-stimulated compared with unstimulated cells in combined all donor cells ( n =520). ( b ) Similarity map of cells from all donors/conditions reveals three subpopulations; IFNγ-stimulated cells (Cluster 1), unstimulated cells (Cluster 2) and mixed populations (Cluster 3). Cluster 1 inset—there are two further subpopulations within IFNγ-stimulated cells, such as Groups 1 and 2. ( c ) The relative expression data for genes related to macrophage function are compared between Groups 1 and 2 ( n =112 and n =63, respectively). ( d ) Similarity maps of 91 genes examined. ( e ) Gene similarity map of PARP9/14, STAT, JAK and IRF genes based on analysis with all cells from all donors and conditions. * P <0.05 and ** P <0.01, respectively, by Student’s t -test. Error bars indicate s.d. Full size image It is possible to subdivide the cells into subpopulations based on expression profile similarity ( Fig. 9b ). Using Ward’s linkage, we observed the following three subpopulations: M(IFNγ) (Cluster 1), M(-) (Cluster 2) and Mixed (Cluster 3). It is important to note that these populations are not homogeneous and can be further divided into at least two subgroups (Groups 1 and 2, Fig. 9b ) in human primary M(IFNγ) (Cluster 1). Both PARP9 and PARP14 were significantly higher in Group 2 than in Group1. Considering the possible importance of PARP9 and PARP14 in the macrophage phenotype, we subsequently examined the genes related to macrophage functions in these two Groups 1 and 2. The potent matrix-degrading enzyme MMP-9 and the pattern recognition receptors CD36, TLR2 and TLR4 were higher in Group2, indicating that this subpopulation, which is associated with increased PARP9 and PARP14 expression, may possess a more pro-inflammatory phenotype ( Fig. 9c ). In our in vitro and in vivo experiments, RNA silencing or genetic deletion of PARP14 enhanced pro-inflammatory responses in IFNγ-stimulated macrophages. In the single-cell analysis, average levels of MMP-9 and TLR4 mRNA expression were also higher in IFNγ-stimulated macrophages treated with PARP14 siRNA ( Supplementary Fig. 15d ). However, their responses to PARP14 were heterogeneous. The gene expression correlation matrix of genes across all cells ( Fig. 9d ) demonstrates that both PARP9 and PARP14 are closely associated with genes known to participate in IFNγ signalling (for example, JAK1, JAK2, STAT1, STAT6, IRF1 and IRF5). This correlation is highly specific to these genes, and does not extend to other genes that we tested in this assay. Genes that are not correlated with PARP9 and PARP14, among all 91 genes tested, include JAK3, a mediator of IL-4 signalling ( Fig. 9e ). Of interest, these correlations are further enhanced in M(IFNγ) compared with M(-) ( Supplementary Fig. 15e ). Discussion The present study provides new evidence that PARP9 and PARP14 regulate macrophage activation. The specific novel findings demonstrated in this report include the following: (1) PARP9 promotes IFNγ-induced responses in mouse and human macrophages; (2) PARP14 suppresses IFNγ responses in mouse and human macrophages; (3) PARP14 induces IL-4-triggered responses in mouse and human macrophages; (4) PARP9 and PARP14 appear to have physical and functional interactions; (5) PARP9 and PARP14 closely interact with components of IFNγ signalling in macrophages; (6) PARP14-induced mono-ADP-ribosylation of STAT1 inhibits its phosphorylation; (7) the PARP9 and PARP14 interactomes have significant proximity to the coronary artery disease module; (8) PARP14 deficiency indeed accelerates macrophage activation and lesion development in mouse models of acute and chronic arterial diseases; (9) PARP14 in the haematopoietic lineage exerts protective effects against arterial diseases; and (10) single-cell analysis revealed that IFNγ-stimulated human primary macrophages derived from CD14+ PBMC contain subpopulations, in which PARP9 and PARP14 are closely associated with genes of IFNγ signalling. Our findings provide insight into new mechanisms for macrophage activation that play a critical role in the pathogenesis of inflammatory arterial diseases, a global health burden. We aimed to establish the unambiguous evidence for the role of PARP9 and PARP14 in macrophage activation using in vitro and in vivo studies. In this study, M(IFN) and M(IL-4) as two multidimensional models of macrophage heterogeneity were used in an unbiased proteomics approach. We then validated our key results on the functionality of PARP9 and PARP14 in mouse and human primary macrophages. Bioinformatic analysis of single-cell gene profiling in human primary macrophages revealed close links among PARP9, PARP14 and IFNγ pathway-related molecules (for example, JAK1, JAK2, STAT1 and STAT6), suggesting that these PARP family members contribute critically to the process of M(IFNγ) activation. We took advantage of the availability of a well-characterized paradigm of macrophage heterogeneity. The balance of diverse macrophage phenotypes (for example, pro-inflammatory versus anti-inflammatory subsets) may regulate normal homeostasis and disease mechanisms. Effective and safe anti-inflammatory therapies may require the fine-tuning of the imbalance of diverse macrophage phenotypes (for example, suppressing excessively activated pro-inflammatory macrophages without compromising protective functions or with enhancing anti-inflammatory activation). Our major goal was to identify molecules or pathways that may regulate the delicate balance of macrophage heterogeneity using global proteomics of M(IFNγ) and M(IL-4). We thus used stringent criteria to choose proteins that increased during M(IFNγ) activation and decreased in M(IL-4). Our in vitro experiments indicate that the interplay between PARP9 and PARP14 may indeed regulate the M(IFNγ)/M(IL-4) balance. We further validated the in vitro results in human arterial lesions and mouse models of arterial disorders, neither of which is solely regulated by IFNγ or IL-4, to provide clinically translatable evidence. Human and mouse atherosclerotic plaque macrophages express PARP9 and PARP14. We identified four macrophage subpopulations in human lesions: PARP9+/PARP14+; PARP9+/PARP14−; PARP9-/PARP14+; and PARP9−/PARP14−. Single-cell gene expression analysis further revealed that IFNγ-stimulated human primary macrophages—M(IFNγ)—are heterogeneous, which may support an emerging concept of the multidimensional model of macrophage activation 21 . In addition, subpopulations within these IFNγ-stimulated cells may have different functions, as shown by associations with MMP-9, CD36, TLR2 and TLR4. Future studies addressing the functional significance of the heterogeneity in human primary macrophages may lead to the development of personalized medical solutions. To explore the evidence for the anti-atherogenic role of PARP14 beyond in vitro assays, we used PARP14 −/− mice. Genetic deletion of PARP14 indeed promoted macrophage activation and accumulation in the intima of mechanically injured arteries, offering in vivo proof of concept. Moreover, significant acceleration of acute arterial lesion formation and chronic atherosclerosis was observed in the mice reconstituted with PARP14 −/− bone marrow. These findings suggest that haematopoietic PARP14 plays a central role in arterial disease. Our study further revealed the new biology for understudied members of the PARP family—PARP9 and PARP14. ADP-ribosylation assays demonstrated that PARP14 has the ability to mono-ADP-ribosylate STAT1 and mass spectrometric analysis identified two ADP-ribosylation sites proximal to its key phosphorylation site. Interestingly, PARP9 suppresses PARP14-dependent mono-ADP-ribosylation of STAT1. Furthermore, mutations of STAT1α ribosylation sites for PARP14 enhanced its phosphorylation and pro-inflammatory gene expression in macrophages. The interplay of PARP9 and PARP14 in STAT1α ADP-ribosylation and phosphorylation may in part explain why PARP9 and PARP14 exert opposing effects on IFNγ- and IL-4-induced macrophage activation. Supplementary Fig. 16 demonstrates an overview of regulatory mechanisms for the IFNγ-STAT1 and IL-4-STAT6 pathways by PARP9 and PARP14. In this study, unbiased global proteomics and bioinformatics of IFNγ- and IL-4-treated macrophages implicated PARP9 and PARP14 in novel mechanisms of macrophage activation. The subsequent network-based prediction of the close relationship between the macrophage PARP14–PARP9 module and human coronary artery disease genes supported the premise that our proteomics screen would effectively identify regulators of macrophage activation in the context of cardiovascular diseases. The approach was then followed by a series of in vitro and in vivo analyses on mouse and human cells/tissues that demonstrated the novel concept that PARP9 promotes and PARP14 suppresses IFNγ-induced activation of macrophages ( Supplementary Fig. 16 ). The present study also represents our strategy of target discovery research assisted by global proteomics screening and subsequent validation studies ( Supplementary Fig. 17 ). Collectively, our discoveries indicate that inhibition of PARP9 and/or activation of PARP14 may attenuate macrophage-mediated vascular diseases, and also provide new insight into the development of effective therapies for other inflammatory disorders. Methods Cell stimuli for cell culture and TMT sample preparation In this study, we used mouse and human IFNγ 10 ng ml −l and IL-4 10 ng ml −1 (R&D systems) as stimuli for macrophage activation, respectively. The murine monocyte/macrophage cell line RAW264.7 was obtained from American Type Culture Collection (ATCC, TIB-71, Rockville, MD) and maintained in 10% fetal bovine serum (FBS, Life Technologies) containing DMEM (Sigma) supplemented with penicillin and streptomycin (Corning) at 37 °C in humidified 5% CO 2 . Before stimuli (IFNγ 10 ng ml −1 and IL-4 10 ng ml −1 ), cells were starved for 24 h with 0.1% FBS-containing media. THP-1 was also purchased from ATCC (TIB-202) and maintained in RPMI 1640 medium in 10% FBS with penicillin and streptomycin at 37 °C in humidified 5% CO 2 . The macrophage-like state was obtained by treating the THP-1 monocytes for 48 h with PMA (200 ng ml −1 , Sigma). Mycoplasma contamination test was routinely performed (once a month). Each cell culture experiment, unstimulated, INFγ and IL-4, was prepared for isobaric labelling using the 6-plex TMT strategy (Pierce). For sample preparation the cells were lysed and proteolysed (Lys-C, Wako Chemicals) using the in-solution urea+ RapiGest (Waters) strategy detailed previously ( Supplementary Fig. 1a ) 39 . Tryptic peptides were labelled with TMT 6-plex reagent (Pierce), combined and desalted using Oasis Hlb 1cc (10 mg) columns (Waters). The peptides were then fractionated into 24 fractions based on their isoelectric focusing point (pH range of 3–10) using the OFF-gel system (Agilent; Supplementary Fig. 1b ). The fractions were dried using a tabletop speed vacuum (Fisher Scientific), cleaned with the Oasis columns and resuspended in 40 μl of 5% acetonitrile (Fisher Scientific) and 5% formic acid (Sigma-Aldrich) for subsequent analysis by liquid chromatography/mass spectrometry (LC/MS). Liquid chromatography tandem mass spectrometry The high-resolution/accuracy LTQ-Orbitrap Elite (Thermo Scientific) analysed TMT peptide samples, and the Q Exactive (Thermo Scientific) and Elite analysed in vitro ribosylated peptides. Both mass spectrometers are fronted with a Nanospray FLEX ion source, and coupled to an Easy-nLC1000 HPLC pump (Thermo Scientific). The peptides were subjected to a dual column set-up: an Acclaim PepMap RSLC C18 trap column, 75 μm × 20 cm (50 μm × 15 cm on the Q Exactive), and an Acclaim PepMap RSLC C18 analytical column 75 μm × 250 mm (Thermo Scientific). For TMT analysis the analytical gradient was run at 250 nl min −1 from 10 to 30% Solvent B (acetonitrile/0.1% formic acid) for 90 min, followed by 5 min of 95% Solvent B. Solvent A was 0.1% formic acid. For ribosylated peptides the gradient was run at 250 nl min −1 from 5 to 28% Solvent B for 10 or 30 min, followed by 5 min of 95% Solvent B. All reagents were HPLC-grade. For TMT analysis, LTQ-Orbitrap was set to 120 K resolution and the top 20 precursor ions (within a scan range of 380–2,000 m / z ) were subjected to higher-energy collision-induced dissociation (HCD, collision energy 40%, isolation width 3 m / z , dynamic exclusion-enabled, starting m / z fixed at 120 m / z and resolution set to 30 K) for peptide sequencing (MS/MS). The Q Exactive was set to 140 K resolution with a top 10 precursor selection method (scan range of 380–1,500 m / z ). HCD was set to a stepped normalized collision energy of 25±10%, isolation width of 1.6 m / z , dynamic exclusion-enabled and resolution set to 17.5 K for MS/MS. Ribosylated peptide candidates were screened in the MS/MS scan by the m6 peak of 348.1 (refs 40 , 41 ). Unmodified forms were calculated by subtracting the mass of the ADP-ribose (541.06 Da) from the observed precursor. Modified and unmodified m/z values and corresponding retention time windows were submitted to an inclusion list and analysed in using the data-independent acquisition module of the Q Exactive ( R =35 K). The MS/MS data were queried against the mouse or human UniProt database (downloaded on 27 March 2012) using the SEQUEST search algorithm via the Proteome Discoverer (PD) Package (version 1.3, Thermo Scientific) 42 , using a 10 p.p.m. tolerance window in the MS1 search space and a 0.02 Da fragment tolerance window for HCD. Methionine oxidation was set as a variable modification, and carbamidomethylation of cysteine residues and 6-plex TMT tags (Thermo Scientific) were set as fixed modifications. The peptide false discovery rate (FDR) was calculated using Percolator provided by PD: the FDR was determined based on the number of MS/MS spectral hits when searched against the reverse, decoy mouse or human database 43 , 44 . Peptides were filtered based on a 1% FDR. Peptides assigned to a given protein group, and not present in any other protein group, were considered as unique. Consequently, each protein group is represented by a single master protein (PD Grouping feature). Master proteins with two or more unique peptides were used for TMT reporter ratio quantification ( Supplementary Table 2 contains a summary of peptides and peptide-spectrum matches (PSMs) for PARP9 and PARP14). Ribosylation spectra were manually annotated 40 , 41 . Proteomics normalization and filtering strategy For each PSM the TMT ion channel intensities were normalized to the time-zero channel (reference normalization, for example, where i =1, 2, …, 6 and x 1 is the time-zero abundance). The protein’s abundance was then calculated by taking the median of its corresponding PSM ratios 45 . To extract the proteins that increase in IFNγ-stimulated (M(IFNγ)) but decrease in IL-4-stimulated condition (M(IL-4)), we applied a simple filtering logic that exploited the available unstimulated control data set. In this study, unstimulated control, M(-), serves as a control for the biological signal owing to the cell culture condition; that is, protein traces M(IFNγ) and M(IL-4) that exceed the edges of the baseline are more likely to be bona fide IFNγ- and IL-4-specific responses ( Supplementary Fig. 3b ). Protein profiles whose abundances surpassed the baseline, which was defined as maximum relative abundance at 8 h of M(-) (+0.13, log10 of relative abundance), after supplement of INFγ established the general threshold for the entire time M(IFNγ) data set. This cutoff value was the same that was used for both RAW264.7 and THP-1 M(IFNγ) data sets. Moreover, the proteins extracted from the IFNγ-stimulated filtering step were cross-referenced not only to M(-) but also to M(IL-4) data sets where proteins were expected to possess opposite profiles with respect to M(IFNγ) ( Supplementary Fig. 3b ). Therefore, the final list of candidate proteins ( Fig. 1b ) had profiles whose abundances increased in M(IFNγ) and decreased in M(IL-4), respectively, beyond their baseline controls. Proteomics data clustering Clustering was performed using the model-based algorithm 46 in R, which is based on finite mixture models; as such, it can successfully be applied to time series data analyses 44 such as those acquired for M(IFNγ) and M(IL-4) 47 ( Supplementary Fig. 3c ). In this approach, each time series (protein profile) y i , i =1, 2, …, N (where N is the number of TMT channels) is considered to be a single entity connected by a line. Clustering is achieved for a traditional finite mixture model by assigning each time series, y i , to a cluster. The non-supervised model-based clustering uses the Expectation Maximization algorithm to assign the profile to a specific cluster. MCLUST has the advantage of using the Bayesian Information Criteria to determine the number of clusters that best partition the data set by maximizing the intradata set variability. Finally, stronger covariance (and thus also dependencies) between sets of two time points is enabled by sum normalization 48 . After clustering was performed ( Supplementary Fig. 4a ) we focused on proteins shared in all three data sets (unstimulated control, IFNγ-stimulated and IL-4-stimulated) for both RAW264.7 and THP-1 cells, with the purpose of identifying proteins whose profiles increased in IFNγ-stimulated condition and decreased in IL-4-stimulated, but maintained at basal levels in unstimulated control condition. We inspected all clusters looking for an increase in the relative abundance of proteins with respect to time zero at any time point in IFNγ-stimulated, and subsequently cross-referenced the IL-4-stimulated and unstimulated control clusters for proteins whose profiles decreased and remained within the baseline, respectively. The proteins that fulfilled these three criteria are shown in the heat maps for RAW264.7 and THP-1 cell data sets ( Supplementary Fig. 4b ). Finally, to further narrow the list of proteins of interest, we prioritized those that were common to both species ( Fig. 1c ), as determined by identical Uniprot protein IDs. Heat maps were used only with the aim of ordering the protein expression levels according to the row obtained Z -score (that is, z =( x -mean)/s.d.). The proteins were clustered in the horizontal direction hierarchically using Euclidean distance and average linkage methods ( Fig. 1c and Supplementary Fig. 4b ). In these data matrices, each column represents the TMT time point analysis 0, 8, 12, 24, 48 and 72 h, and each row corresponds to a protein gene ID. PARP14–PARP9 network analysis As a parallel approach to further investigate the candidacy of PARP9 and PARP14 as cardiovascular and metabolic diseases, we turned to in silico or network-based prediction methods under the premise that a potential PARP14–PARP9 interactome would be close in vicinity to its pertinent vascular disease network/module. To evaluate the impact of PARP14–PARP9 neighbours, we used the HumanNet interaction database 49 . We then used the random-walk methodology 50 to construct the disease modules from the functional database using gene–disease associations extracted from genome-wide association studies and from the OMIM ( ) and the MalaCards ( ) databases for: Cardiomyopathy, Coronary Heart disease, Heart Failure, Hypercholestermia, Hypertension, Metabolic Traits, Osteoporosis, Cardiovascular Risk Factors, Sudden Cardiac Arrest, Systemic lupus erythematosus, Sjögren’s syndrome, polymyositis and dermatomyositis. The PARP9–PARP14 module was determined by first neighbours in the functional database. Further, we restricted the first neighbours to those that are expressed in macrophages based on gene expression. The motivation of this network analysis is based on two recent studies of ours that investigated the localization of disease in the network 51 and measure the separation of disease modules in the network 52 , both of which have proved instrumental in identifying disease modules. These network topology-based methods have been rigorously validated in both of these peer-reviewed publications by means of biological enrichment as well as compared with other network-based methodologies. On the basis of this selection, we considered 55 first neighbours for PARP9 and 149 for PARP14 that we annotated collectively as the PARP9–PARP14 module. Further, we measured the closeness of the first neighbours using the shortest-path topology measure with the disease modules above. For each disease module, we calculated the shortest-path distances to the PARP9–PARP14 module and compared these distances to the random distance distribution with the same module size (Wilcoxon test and Benjamini–Hochberg correction for multiple testing). Finally, we verified that the module size does not affect the significance of the P values for the Coronary Heart Disease and Osteoporosis modules by recalculating the random distances by reducing the module size to the top 35 genes from random walk. Mouse peritoneal macrophages Four days after intraperitoneal injection of 2.5–3 ml of 4% thioglycollate (Fisher Scientific), peritoneal macrophages were obtained by injecting 5 ml of ice-cold PBS into the peritoneal cavity using 25 G needle, followed by collection of the fluid using 23 G needle. In all, 5 × 10 5 cells were cultured on 24-well plates (Corning) with RPMI containing 10% FBS. Nonadherent cells were discarded after 16 h. After washing with PBS, cells were incubated with IFNγ 10 ng ml −1 and IL-4 10 ng ml −1 for 24 h until harvesting. Primary mouse macrophages derived from BMCs BMDMs were isolated and differentiated as described previously 53 . Briefly, whole-BMCs were harvested from femurs of 10–12-week-old male PARP14 −/− or PARP14 +/+ mice (C57BL/6 mice, Jackson Laboratory) under aseptic conditions and cultured in RPMI 1640 (Lonza) supplemented with 10% FBS, penicillin and streptomycin (Corning) in presence of macrophage colony-stimulating factor (M-CSF; 50 ng ml −1 ; Peprotech) at 37 °C in humidified 5% CO 2 . After 7-day culture, cells were incubated with IFNγ 10 ng ml −1 and IL-4 10 ng ml −1 for 24 h until harvesting. After 7 days of culture, cells were incubated with IFNγ 10 ng ml −1 and IL-4 10 ng ml −1 for 24 h. Whole-BMCs were cultured in M-CSF (50 ng ml −1 ) for 7 days and analysed. FACS analysis showed that most of cells (>80%) were positive for CD11b, CD115 and weakly positive for F4/80 and negative for CD11c and Ly6G, indicating the differentiation of BMCs into macrophages. FACS was performed on the FACSAria2 (BD Bioscience), and data were analysed using the Flowjo software (Tree Star). Isolation of CD14+ human PBMCs PBMCs were isolated from buffy coat using lymphocyte separation medium (MP Biomedicals). After isolation of PBMCs, CD14 + cells were isolated by Dynabeads in combination with anti-CD14 antibody (Dynabeads FlowComp human CD14, Invitrogen), according to the manufacturer’s instruction. Briefly, CD14+ dynabead-bound cells (CD14 + cells) were isolated using a magnet. After releasing Dynabeads from cells, bead-free CD14 + PBMCs (5 × 10 5 cells) were cultured in 24-well dish plates with 0.5 ml of RPMI supplemented with 10% FBS at 37 °C in 5% CO 2 . Immunohistochemistry and immunofluorescence Samples were cut into 7 μm thin slices, and cryo-sections were fixed in acetone. After blocking in 4% appropriate serum, sections were incubated with primary antibodies (Mac3 (1:200, M3/84, BD Pharmingen), PARP14 (1:50, HPA01206, Sigma-Aldrich), PARP9 (1:100, ab53796, Abcam) and human CD68 (1:200, M0876, Dako), followed by biotin-labelled secondary antibody (1:250, Vector Laboratories, Burlingame, CA, USA) and streptavidin-coupled Alexa Fluor 488 antibody (Life Technologies). For immunofluorescence double labelling, after avidin/biotin blocking (Vector Laboratories), the second primary antibody was applied overnight at 4 °C, followed by biotin-labelled secondary antibody and streptavidin-coupled Alexa Fluor 594 antibody (1:250, Life Technologies). Sections were washed in PBS and embedded in mounting medium containing 4,6-diamidino-2-phenylindole (Vector Laboratories). For bright-field immunohistochemistry on tissue sections, following the first biotin-labelled secondary antibody incubation, sections were incubated with streptavidin-labelled horseradish peroxidase (HRP) solution (Dako), followed by 3-amino-9-ethylcarbazole (AEC) solution. Slides were examined using the Eclipse 80i microscope (Nikon, Melville, NY, USA) or the confocal microscope A1 (Nikon). All images were processed with the Elements 3.20 software (Nikon). Human tissue and counting PARP14–PARP9-positive macrophages Atherosclerotic carotid arteries ( n =10) were collected from patients undergoing endarterectomy procedures at Brigham and Women’s Hospital according to IRB protocol # 1999P001348 (PL). Informed consent was not necessary because all samples are considered as ‘discarded material.’ Samples have two groups, such as macrophage-rich ( n =5) and no macrophage-rich ( n =5) plaques, respectively. Samples were embedded in optimal cutting temperature (OCT) compound and stored at −80 °C until use. In three different fields of each sample ( n =5), 200 cells (nuclei) with CD68-positive cells were evaluated whether they express PARP14 and/or PARP9 (600 cells per sample). The quantification of immunofluorescence was performed by examiners who were blinded to group allocation (macrophage-rich versus no plaque). Animal ethics All animal experiments used in this study were approved by and performed in compliance with Beth Israel Deaconess Medical Center’s Institutional Animal Care and Use Committee (Protocol 024-2014). PARP14-deficient mice PARP14 −/− mice were backcrossed into the C57BL/6 genetic background over 10 generations 25 , 54 . Male PARP14 −/− mice and age-matched PARP14 +/+ mice were used for a vascular injury model or as donors for BMT. Wire-induced acute vascular injury Transluminal arterial injury was induced for 10-week-old mice by inserting spring wire (0.38 mm in diameter, C-SF-15-15, Cook) into the femoral artery under microscopic observation (leica M80) 55 . Right common femoral artery, superficial femoral artery and deep femoral artery (DFA) were first exposed. After looping SFA with 9-0 nylon suture, a small hole in the DFA was made and a wire was inserted through the hole and advanced to the iliac artery (10 mm). DFA was ligated with 9-0 nylon suture when the wire was removed at the closer point to DFA and common femoral artery bifurcation than the hole. Bone marrow reconstitution Bone marrow was reconstituted as described previously 56 with minor modifications. Briefly, recipient mice were lethally irradiated with a total dose of 1,000 rads. The next day, unfractionated bone marrow cells (BMCs, 1 × 10 6 ) that had been harvested from donor mice and suspended in 0.3 ml of phosphate-buffered saline were administered to each recipient mouse via the tail vein. We confirmed bone marrow reconstitution by examining gene expression of PARP14 in the bone marrow of BMT PARP14 +/+→+/+ and BMT PARP14 −/−→+/+ mice (93.3% reduction of the PARP14 gene in BMT PARP14 −/−→+/+ compared with BMT PARP14 +/+→+/+ mice). Four weeks after bone marrow reconstitution, wire-mediated femoral artery injury was performed. Two weeks after injury, to collect arterial samples, mice were killed by intraperitoneal administration of an overdose of Pentobarbital and then perfused with 0.9% NaCl solution at a constant pressure via the left ventricle 57 . High-fat diet-induced chronic atherosclerosis First, we examined PARP9 and PARP14 expression in Apolipoprotein E-deficient ( ApoE −/−) mice, which were backcrossed with C57BL6 mice over 10 generations and then fed by high-fat and high-cholesterol diet for 20 weeks. As a chronic atherosclerosis model, we used low-density lipoprotein receptor (LDLR)-deficient mice ( LDLR −/−) fed a high-fat and high-cholesterol diet for 16 weeks. Bone marrow of PARP14 −/− or sex and age-matched PARP14 +/+ mice was transplanted to lethally irradiated LDLR−/− mice, as described above (BMT PARP14 +/+→ LDLR −/− and BMT PARP14 −/−→ LDLR −/− mice). These mice were fed a high-fat and high-cholesterol diet for 16 weeks after bone marrow reconstitution, and tissues were harvested for histological and molecular analyses. Evaluation of arterial lesions The collected tissue was embedded in OCT compound (Tissue-Tek) and then frozen in liquid nitrogen and stored at −80 °C until further use. Intima and media area and their ratio in injured femoral arteries were measured by manual tracing in at least three sections of three different levels with 100 μm interval in each animal ( n =5), using the Elements 3.20 software (Nikon). The quantification of histology and immunohistochemistry was performed by examiners who were blinded to group allocation ( PARP14 +/+ versus PARP 14−/− and BMT PARP14 +/+→+/+ versus BMT PARP14 −/−→+/+ and BMT PARP14 +/+→ LDLR −/− versus BMT PARP14 −/−→ LDLR −/− mice). No randomization method was used. RNA interference RNA silencing was performed as described previously 58 . Briefly, 20 nmol l −1 siRNA against PARP14 (L-023583 for human cells and L-160447 for mouse cells) and PARP9 (L-014734 for human cells and L-05024901 for mouse cells; all ONTARGETplus SMART-pool, Thermo Scientific) or non-targeting siRNA (scramble control siRNA, ON-TARGET Non-Targeting Pool, Thermo Scientific) was transferred into macrophages using SilenceMag (BOCA Scientific, Boca Raton, FL), according to the manufacturer’s instruction. Target sequences of siRNA pool were follows: Human PARP14: 5′-UAAUCAAAGGUCUCUUAUG, UAAUGCUUAAGGUCCUCAU, UCAUUAUACUGCCAUUCUA-3′ and 5′-GAACUCUUGACAUCAUUUC-3′, Human PARP9: 5′-AAUUACAUCUGCCGUCUGC-3′, 5′-UUUGUGGCAAGAAAUUCCG-3′, 5′-UUAAUCAACAGGGCUGCCA-3′ and 5′-UACAGCCAAACUUAUUCUG-3′ Mouse PARP14: 5′-CUUGAAAGCUUUACGUAUA-3′, 5′-CAGCAAUAGGAACGGGAAA-3′, 5′-CCAAAGAACUUGAUCAACA-3′ and 5′-CGUAGUAGCAAAAGCGAUA-3′, Mouse PARP9: 5′-ACACAAUGUCUUCGAAAUU-3′, 5′-CCAGACAGCUAUCGAAUUA-3′, 5′-CCAAAUAUGAUCUACGCAU-3′ and 5′-CGUACACAUUUCAACGAUA-3′. Control scramble: 5′-UGGUUUACAUGUCGACUAA-3′, 5′-UGGUUUACAUGUUGUGUGA-3′, 5′-UGGUUUACAUGUUUUCUGA-3′ and 5′-UGGUUUACAUGUUUUCCUA-3′. Cell growth and cell viability Cell proliferation, viability and apoptosis were assessed using the CellTiter 96 AQueous Nonradioactive Cell Proliferation Assay Kit (MTS), Cell Titer Blue assay and Apo-ONE Homogeneous Caspase-3/7 Assay Kit, respectively (Promega), according to the manufacturer’s instructions. RNA preparation and real-time PCR Total RNA from the cell culture was isolated using TriZol (Life Technologies), and reverse transcription was performed using the QuantiTect Reverse Transcription Kit (Qiagen, Hilden, Germany). The mRNA expression was determined by TaqMan-based real-time PCR reactions (Life Technologies). The following TaqMan probes were used: Hs99999902_m1 (human RPLP0), Mm00725448_s1 (mouse RPLP0), Hs00981511_m1 (human PARP14), Mm00520984_m1 (mouse PARP14), Hs00967084_m1 (human PARP9), Mm00518778_m1 (mouse PARP9), Hs00174128_m1 (human TNF), Mm00443258_m1 (mouse TNF), Hs00174097_m1 (human IL-1β), Mm01336189_m1 (mouse 1L1β), Mm00475988_m1 (mouse ARG1), Hs00267207_m1 (human MRC1), Mm00485148_m1 (mouse MRC1), Mm00440502_m1 (mouse NOS2) and Hs00234140_m1 (human CCL2). The expression levels were normalized to RPLP0. Results were calculated using the Delta-Delta Ct method, and presented as arbitrary unit. LCM and RNA amplification LCM was performed on the Leica LMD6500 Microdissection System. Neointima were cut using the following LCM parameters: power, 50 mW; pulse duration, 2 ms; and spot size, 20 Hm. RNA was isolated using the PicoPure RNA Isoaltion Kit, followed by RNA amplification using the RiboAmp HS Plus RNA Amplification Kit (both Arcturus, Mountain View, CA, USA), according to the manufacturer’s protocol. PCR array was performed using the Fluidigm PCR system. Western blot analysis Cells were lysed with RIPA buffer containing protease inhibitor (Roche). Protein concentration was measured using the bicinchoninic acid method (Thermo Scientific). Total protein was separated by 8–10% SDS–PAGE and transferred using the iBlot western blotting system (Life Technologies). Primary antibodies against human and mouse PARP14 (1:250, HPA01206 Sigma-Aldrich), human and mouse PARP9 (1:250, ab53796, Abcam), human and mouse STAT1 (1:1,000, #9172, Cell Signaling), phosphorylated STAT1 (1:1,000, #9167, Cell Signaling), human and mouse STAT6 (1:2,000, #9362, Cell Signaling), mouse (1:1,000, ab54461, Abcam) and human (1:2,000, #9361, Cell Signaling) phosphorylated STAT6 and human and mouse β-actin (1:5,000; Novus) were used. For secondary antibodies, we used anti-mouse (1:1,000–5,000, A4416, Sigma) and rabbit (1:1,000–5,000, NA934-1ML, GE Healthcare Life Sciences) IgG antibodies. Protein expression was detected using Pierce ECL Western Blotting substrate reagent (Thermo Scientific) and ImageQuant LAS 4000 (GE Healthcare). Uncropped images of western blots are demonstrated in Supplementary Figs 18–20 . ELISA The amounts of TNFα and IL-1β released into the culture media after stimulation were measured using an ELISA kit following the manufacturer’s instructions (Duoset Kit, R&D). The culture medium of unstimulated macrophage was used as the negative control. Standard, control or sample solution was added to the ELISA-well plate, which had been pre-coated with specific monoclonal capture antibody. After being shaken gently for 3 h at room temperature, the polyclonal anti-TNFα antibody, conjugated with horseradish peroxidase, was added to the solution and incubated for 1 h at room temperature. A substrate solution containing hydrogen peroxidase and chromogen was added and allowed to react for 20 min. The levels of cytokines were assessed by a plate reader at 450 nm and normalized with the abundance of standard solution. Nitrate quantification To quantify nitric oxide in cell culture media of macrophages, mimicking iNOS concentration, the Griess Reagent Kit (G-7921, Life Technologies) was used. Co-IP Cells were lysed in IP lysis buffer (Thermo Scientific). A volume of 100 μg of protein was incubated with PARP14 antibody (5 μg, Invitrogen) and Dynabeads streptavidin (Life Technologies) by rotation overnight at 4 °C, followed by washing three times with PBS/Tween 20 (0.02%), using a magnet to collect the beads after each wash. Five per cent of the precipitated protein sample was subjected to SDS–PAGE. Protein expression was detected using Pierce ECL western blotting substrate Reagent and ImageQuant LAS 4000. ADP-ribosylation assays Recombinant human PARP14 and PARP9 (BPS Bioscience Inc.) proteins and BSA (Sigma-Aldrich) were incubated with recombinant human STAT1α (OriGene Technologies Inc.) or STAT6 protein (Sino Biological Inc.) at a final concentration of 5 ng μl −1 in the presence of 100 μM β-nicotinamide adenine dinucleotide hydrate (NAD; Sigma-Aldrich) or 6-biotin-17-NAD (Trevigen Inc.) in 50 mM Tris-HCl buffer (pH 7.4) for 1 h at room temperature. Ribosylation of STAT1α and STAT6 was detected by liquid chromatography tandem mass spectrometry (LC-MS/MS) after trypsin digestion (using the biotin-free NAD reaction) or by western blotting using Streptavidin-HRP (Abcam) after SDS–PAGE. Quantification of the relative abundances of ribosylated STAT1α peptides was completed by calculating the area under the curve (AUC) of the extracted ion chromatograms of the monoisotopic peaks of the modified versus unmodified peptides. The ratios were reported as AUC mod./(AUC mod. + AUC unmod). Construction and enforced expression of mutant STAT1 Human pcDNA-GFP-STAT1 was purchased from Addgene (Cambridge, MA, USA). Step-wise mutations (glutamic acid, E, to glutamine, Q) was introduced at the two ribosylation sites flanking the phosphorylation at Tyr701—E657 and E705—by recombinant PCR mutagenesis. Mutated constructs were verified by DNA sequencing. Mouse bone marrow macrophages were differentiated from bone marrow stromal cells using 10 ng ml −1 M-CSF for 12 days. pcDNA-GFP-STAT1 (WT) and the mutant pcDNA-GFP-STAT1 E657Q, E705Q, were transferred by Magnetofection (OzBioscience, San Diego, CA, USA). HEK293 cells were transfected using Lipofectamine LTX (Invitrogen, USA). Twenty-four hours after transfection, cells were serum-starved (0.1% FBS) for 2 h and stimulated with IFNγ for 1 h (phospho-STAT1) or 24 h (mRNA expression of pro-inflammatory factors). The overexpressed STAT1 was immunoprecipitated using anti-GFP antibody, clone 9F9.F9 (1:1,000, ab1218, Abcam). STAT1 phosphorylation at Tyr 701 was detected by anti-phospho-STAT1 (Tyr701; 1:1,000, mAb #7649, Cell Signaling). Antibodies against STAT1 (ab3987, Abcam) and GFP (ab290, Abcam) served as loading controls. Transfection into THP-1 cells was performed using the magnetofection method described above. Flow cytometry The spleen was removed and homogenized to isolate splenocytes. Splenocytes were incubated in red-blood-cell lysis buffer (ACK lysing buffer, Gibco) to remove erythrocytes. After incubation with anti-mouse CD16/CD32 mAb (BioLegend, 0.5 μg per million cells) to block the Fc receptor, cells were then stained with antigen-presenting cell-conjugated CD11b, fluorescein isothiocyanate-conjugated Ly6c and phycoerythrin-conjugated Ly6G (BioLegend) in autoMACS running buffer containing bovine serum albumin, EDTA, PBS and 0.09% azide (Miltenyl Biotec) for 30 min. After washing cells with autoMACS running buffer, stained cells were analysed by FACSAria2 (BD Bioscience) and Flowjo software (Tree Star). Ly6c expression was evaluated in CD11b+Ly6G− splenic monocytes (apparent b). Single-cell gene expression analysis For single-cell analysis of CD14+ PBMCs derived from three donors, cell capture and target pre-amplication steps were performed using the C1 system (Fluidigm) according to the manufacturer’s instructions. Quantitative real-time PCR was performed using the BioMark 96.96 Dynamic Array platform (Fluidgm) 38 . After isolation of CD14+ PBMCs from buffy coat, cells were cultured for 10 days. Cells in unstimulated control condition were harvested on day10 and IFNγ-stimulated cells were harvested on day 11 after 24 h incubation with IFNγ 10 ng ml −1 . The cell capture rate by the C1 chip was 89.6% (86/96) in unstimulated control and 83.3% (84/96) in IFNγ-stimulated of donor 1, 96.9% (93/96) in unstimulated control and 89.6% (86/96) in IFNγ-stimulated of donor 2, and 92.7% (89/96) in unstimulated control and 85.4% (82/96) in IFNγ-stimulated of donor3. The raw real-time PCR reads for each array were transformed into n × m matrices using Python’s Pandas libraries ( ; where n =cell index and m=gene). Each data matrix was then processed and analysed using an in-house developed platform. We first run a check for each column (genes) to see whether there are undetected values (CT=999) interspersed among positive reads (CT<25). If less than 10% of the reads are positive, we substitute those values with 999, and consider the entire gene undetectable for this array. On the other hand, if more than 10% of the genes are positive, their corresponding reads with undetected values are substituted with the average value of the positive reads. Granted, more sophisticated missing-value imputation (MVI) techniques exist 59 ; however, we may not have enough features (96 features) to fully benefit from MVI nor is it clear whether this necessarily leads to improvement in signal. Moreover, genes requiring missing-value estimation tend to fall near the limit of detection, and are unlikely to benefit fully from MVI. For the purpose of having a fully populated matrix with no missing values, the averaged value therefore should suffice. The final missing-value adjusted reads are converted into log2exp via the following equation (equation 1), where LOD stands for the limit of detection and set at recommended default value of 24. Although we do not normalize using housekeeping genes (see below for normalization method), they can be good indicators of the overall read quality for a given cell. Cells without the housekeeping gene expression (that is, GAPDH ) were removed from analysis. Next, we calculate the mean expression value of the GAPDH gene (average log2exp of all cells for GAPDH ). Cells with outlier GAPDH expression (more than 3 s.d.’s from the mean) were also excluded from the analysis. Individual cells may exhibit extreme reads because of transcriptional burst, or because of high-instrument sensitivity [5]. Such extreme values need to be contained. Trimming methods (removal of extreme outliers based on the 5% quantile on either side of the read distribution) are unsuitable, as they would alter the dimensions of the matrix as well as lead to the loss of data. For example, given 100 genes with 100 observations each. If there is a 5% chance for every gene that at least one observation would contain an extreme outlier, then via the process of trimming, only 95 of the observations could be used for analysis since the outliers, and hence the entire column (containing useful information as well) would be discarded. To circumvent this, and maintain the overall integrity of the matrix, we performed Winsorization 60 , where we set boundaries of the extreme values to the values of the 5 and 95%th quartiles, respectively. To compare the chips, individual genes are converted into z -scores by subtracting the mean from the log2exp, and division by the s.d. Arrays are compared on a per-donor basis (Donor_i unstimulated control, IFNγ-stimulated, where i =1,2,3). For each pair, the arrays are merged into a combined matrix. We calculate the Manhattan distance 61 between all cells based on their gene parameters, and represent the clustering using the minimum spanning tree 62 . Statistical analysis Data are given as mean ±s.d. Moreover, ‘ n ’ indicates the number of independent experiments or number of animals/samples. Tests with a P value less than 0.05 were considered statistically significant. Pairwise group comparisons were performed using a Student’s t- test (GraphPad prism 5, Prism Software Inc. (La Jolla, CA)). If F test showed the variance was significantly different, unpaired t -test with Welch’s correction was performed. Exclusion criteria were set by Grubbs’ test. The experiments were not randomized. No statistical method was used to predetermine sample size. The experiments were not randomized. Data availability The data that support the findings of this study are available within the article and its Supplementary Information Files . Additional information How to cite this article : Iwata, H. et al . PARP9 and PARP14 cross-regulate macrophage activation via STAT1 ADP-ribosylation. Nat. Commun. 7, 12849 doi: 10.1038/ncomms12849 (2016).

Researchers at Brigham and Women's Hospital have found two new potential drug targets for treating arterial diseases such as atherosclerosis. By using proteomics to screen a vast number of molecules, the researchers identified PARP9 and PARP14 - two members of the PARP family of proteins - as regulators of macrophage activation, which has been linked to arterial disease by systems biology. Though the mechanisms that activate macrophages, a type of digestive white blood cell that targets foreign cells, remain incompletely understood, previous research shows that macrophages play an important role in the development of atherosclerosis and its thrombotic complications. Masanori Aikawa, MD, PhD, director of the Center for Interdisciplinary Cardiovascular Sciences (CICS) at the Brigham, his research fellow Hiroshi Iwata, MD, PhD, and colleagues studied atherosclerosis on the protein-level to determine which molecules were most involved in the regulation of macrophages. Once Aikawa and his colleagues narrowed down their search to these two proteins, they silenced each gene in cultured macrophages and found that tamping down PARP14 increased macrophage activation while tamping down PARP9 had the opposite effect. Aikawa founded CICS and hopes that this hypothesis-generating method can be used to streamline the lengthy process of drug development. Aikawa and CICS are using a more systematic approach which hinges on network analysis; this analysis predicts which pathways are most likely to control their studied effect so that they can prioritize these pathways. Ideally, this process would take a fraction of the time in comparison to searching through each individual pathway unaware of their likelihood of affecting their studied effect. Aikawa and his colleagues plan to augment these findings to develop targeted therapeutics for atherosclerosis and other diseases. "Macrophage activation plays a role in not only vascular disorders but also various inflammatory and autoimmune diseases," said Aikawa. "These results could provide important information about the mechanisms of these diseases and help to develop much needed new therapeutics."

10.1038/NCOMMS12849

Chemistry

Zinc oxide: Key component for the methanol synthesis reaction over copper catalysts

Núria J. Divins et al, Operando high-pressure investigation of size-controlled CuZn catalysts for the methanol synthesis reaction, Nature Communications (2021). DOI: 10.1038/s41467-021-21604-7 Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-021-21604-7

https://phys.org/news/2021-03-zinc-oxide-key-component-methanol.html

Abstract Although Cu/ZnO-based catalysts have been long used for the hydrogenation of CO 2 to methanol, open questions still remain regarding the role and the dynamic nature of the active sites formed at the metal-oxide interface. Here, we apply high-pressure operando spectroscopy methods to well-defined Cu and Cu 0.7 Zn 0.3 nanoparticles supported on ZnO/Al 2 O 3 , γ-Al 2 O 3 and SiO 2 to correlate their structure, composition and catalytic performance. We obtain similar activity and methanol selectivity for Cu/ZnO/Al 2 O 3 and CuZn/SiO 2 , but the methanol yield decreases with time on stream for the latter sample. Operando X-ray absorption spectroscopy data reveal the formation of reduced Zn species coexisting with ZnO on CuZn/SiO 2 . Near-ambient pressure X-ray photoelectron spectroscopy shows Zn surface segregation and the formation of a ZnO-rich shell on CuZn/SiO 2 . In this work we demonstrate the beneficial effect of Zn, even in diluted form, and highlight the influence of the oxide support and the Cu-Zn interface in the reactivity. Introduction The utilization of fossil fuels as the main energy source gives rise to serious environmental issues, including global warming caused by the continuously increasing level of atmospheric CO 2 . The hydrogenation of CO 2 is an important process to valorize CO 2 by converting it into useful chemicals and fuels, such as methanol 1 , 2 , 3 , 4 . Methanol is presently industrially synthesized from CO 2 , CO, and H 2 at pressures between 50 and 100 bar and temperatures between 200 and 300 °C using Cu/ZnO/Al 2 O 3 catalysts 5 , 6 , 7 . Although Cu/ZnO-based catalysts have been long studied, the role and the dynamic nature of the active sites that originate at the metal-oxide interface under the strongly reducing reaction conditions are still under debate. It has been shown that Cu-based catalysts containing ZnO are more active than their ZnO-free counterparts 8 and that ZnO both enhances the dispersion of the Cu nanoparticles (NPs), and acts as an electronic promoter 8 , 9 through a strong metal-support interaction 8 , 9 , 10 . This promoting effect of Zn has been attributed to the Zn being present in a number of different configurations: the formation of graphitic ZnO x layers on Cu NPs 11 , metallic Zn atoms within the Cu NP surface 12 , CuZn alloy formation 13 , Zn-decoration of stepped Cu surfaces 1 , or the formation of ZnO at the interface with Cu in single crystals 3 , 14 . A strong correlation between the Zn coverage on Cu NPs and the methanol synthesis activity was also reported 2 . Furthermore, the structure-sensitivity and size-dependent activity for this reaction were shown for Cu and CuZn-based catalysts 4 . Heterogeneous catalysts are dynamic entities and adapt their morphology and electronic structure to the chemical potential of the surrounding gaseous atmosphere 15 , 16 . Thus, for methanol synthesis catalysts, operando studies in pressure regimes that are industrially relevant (20–100 bar), as the ones discussed in the present work, are key to overcome the “pressure gap” 15 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , as they probe Cu and Zn under their actual working state, unraveling their active chemical and structural state. Investigations on Cu/ZnO systems revealed the presence of poorly crystalline ZnO interacting with Cu 15 , reversible changes of the Cu NP morphology depending on the reaction conditions 17 , 23 , CuZn bulk alloy formation under severe reduction conditions (600 °C in CO/H 2 ) 17 , 26 , and ZnO reduction and a stronger CuZn interaction for catalysts supported on SiO 2 19 , 20 , 21 . Interestingly, even at sub-atmospheric pressures (1–2 mbar) 27 , the dynamic behavior of these catalysts has also been corroborated with the observation of wetting/non-wetting of Cu NPs on ZnO 28 . Despite the tremendous effort dedicated to the investigation of this reaction, questions still remain on whether metallic Zn/brass is present under industrially relevant reaction conditions, its role in the reaction selectivity, and whether ZnO is needed as a part of the support, or small ZnO ensembles in direct contact with Cu within a NP can also serve to activate CO 2 . To address these questions, here we use morphologically and chemically well-defined Cu and CuZn NP catalysts that help close the materials gap between the heterogeneous industrial catalysts and the model single-crystal systems previously studied, while also addressing the pressure gap through operando spectroscopic characterization at industrially relevant high-pressure conditions. The small NP size chosen in our study allows us to investigate the changes occurring at the surface and near-surface region of the NPs by the bulk-sensitive technique X-ray absorption spectroscopy (XAS), since the fraction of atoms at the surface of Cu NPs of ca. 3 nm is ~60% 29 . In this way, we are able to shed light on the dynamic evolution of the metal NP/oxide support interface during CO 2 + CO hydrogenation. Results Model catalysts synthesis and characterization Five catalysts, designed to be analogous to the commercial Cu/ZnO/Al 2 O 3 catalyst, composed of size-controlled NPs were synthesized to investigate the different components and interactions of the industrial catalyst. Monometallic Cu NPs were deposited on (i) ZnO/Al 2 O 3 , (ii) SiO 2 , (iii) γ-Al 2 O 3 , and bimetallic Cu 0.7 Zn 0.3 NPs on (iv) SiO 2 , and (v) γ-Al 2 O 3 . The NPs were prepared following an inverse micelle encapsulation method. Figure 1a shows an atomic force microscopy (AFM) image of the Cu NPs with a mean height of 3.1 ± 0.8 nm. Similar data were obtained for the Cu 0.7 Zn 0.3 NPs (Supplementary Figure 1 ). A Cu/Zn ratio of 70/30 was chosen for the bimetallic NPs to assure sufficient promotion of Cu by Zn 4 . The powder catalysts were synthesized by incipient wetness impregnation of the pre-formed NP solutions on the oxide supports (Supplementary Figure 2 ) and calcined at temperatures ≥400 °C for 6 h in 20% O 2 /(Ar or N 2 ) to completely remove the polymer. Figure 1b shows a scanning transmission electron microscopy (STEM) image of the CuZn/SiO 2 catalyst. The mean particle size determined by TEM after calcination is 2.7 ± 0.7 nm. Energy-dispersive X-ray spectroscopy (EDX) maps in Fig. 1 c and d confirm that the Cu and Zn signals, respectively, originate from the same NP. Figure 1e displays a STEM image of the CuZn/Al 2 O 3 sample showing isolated NPs. Figure 1f shows an EDX map overlayed with a high-angle annular dark-field image. The mean NP diameter determined by TEM is 3.3 ± 0.7 nm. In Fig. 1g an EDX map for the Cu/ZnO/Al 2 O 3 catalyst is displayed, showing Cu distributed over the ZnO/Al 2 O 3 support. Additional TEM images are shown in Supplementary Figures 3 – 5 . Fig. 1: Microscopy characterization of the catalysts. a AFM image of the Cu NPs supported on SiO 2 /Si(111). b STEM image of the CuZn/SiO 2 catalyst; c , d show the corresponding Cu and Zn EDX maps. e STEM image and f EDX map of the CuZn/Al 2 O 3 catalyst. g EDX map of the Cu/ZnO/Al 2 O 3 catalyst. In EDX maps f , g red represents Cu, green Zn and blue Al. The scale bars correspond to a 400 nm, b – g 10 nm. Full size image X-ray diffraction (XRD) patterns of the Cu/ZnO/Al 2 O 3 , CuZn/SiO 2 , and CuZn/Al 2 O 3 catalysts were recorded in their calcined state and after reaction. A detailed description of the findings is included in the Supplementary Discussion, and Supplementary Figure 6 . Note here that XRD is not sensitive to small atomically disordered NPs, and therefore, the results obtained are biased to the minority fraction of agglomerated NPs in these samples. Catalyst structural evolution under a high-pressure environment To gain insight into the interaction between Cu and Zn and into the evolution of their chemical state and structure under methanol synthesis conditions, the Cu/ZnO/Al 2 O 3 , CuZn/Al 2 O 3 , and CuZn/SiO 2 catalysts were investigated by means of operando XAS at the Cu and Zn K-edges. The study of nano-sized pre-formed CuZn NPs allowed us to clearly track the evolution of the Zn species. This would be challenging to implement for the industrial catalyst as it contains Zn as the bulk support. XAS data of the fresh calcined samples were acquired under He at room temperature, followed by a reduction (activation) under 10% H 2 /He for 2 h at 245 °C for the Cu/ZnO/Al 2 O 3 and the CuZn/Al 2 O 3 catalysts and at 325 °C for the CuZn/SiO 2 catalyst 19 . A higher reduction temperature was used for the NPs deposited on SiO 2 owing to the higher stability of the CuO x species on this support. Further details on the effect of the reduction temperature are given in the Supplementary Discussion. Next, the samples were studied under an industrially relevant methanol synthesis mixture containing 4% CO 2 , 10% CO, 72% H 2 , balanced in He, at: (i) 20 bar at 220 °C, (ii) 20 bar at 280 °C, (iii) 40 bar at 280 °C, and (iv) 40 bar at 320 °C. The last step at 320 °C was performed in order to mimic an ageing treatment. Figure 2 shows Cu K-edge and Zn K-edge X-ray absorption near-edge structure (XANES) data and Fig. 3 shows the magnitude of the Fourier transform (FT) of the extended X-ray absorption fine-structure (EXAFS) data of the three catalysts measured in their initial state and under the methanol synthesis mixture at 20 bar and 220 °C (XAS data for the other reaction regimes are shown in Supplementary Figures 9 – 16 ). At the Cu K-edge, we observe that Cu species are initially oxidized and exhibit a local structure similar to CuO. Upon reduction in hydrogen and under reaction conditions, these species are reduced and remain reduced (Cu 0 ) (see Figs. 2 a and 3a , Supplementary Figures 10 and 11 and Supplementary Table 8 ). EXAFS data (see Supplementary Discussion) indicate that, although some NP sintering was revealed by XRD and TEM (Supplementary Figures 3 – 6 ), the majority of Cu is available within well-dispersed NPs that remain stable during the reaction. This follows from the observation that the 1 st shell coordination numbers (CNs) for reduced Cu are ca. 8–10, in agreement with NP sizes of ca. 3 nm 30 . Importantly, no changes in the CNs were observed under the different reaction conditions, suggesting that the majority of particles retained their size, shape, and structure. Fig. 2: Operando XANES spectra. a Cu K-edge and b Zn K-edge XANES spectra for Cu/ZnO/Al 2 O 3 , CuZn/SiO 2 and CuZn/Al 2 O 3 catalysts in their initial state (dashed lines), and under operando conditions in CO 2 +CO+H 2 , p = 20 bar, T = 220 °C (solid lines). The reference spectra for CuO and ZnO (dashed lines), Cu and Zn foils (solid lines) are also shown. The vertical arrow in b marks the signature of the zinc spinel structure. Full size image Fig. 3: Operando Cu K-edge and Zn K-edge EXAFS spectra. The FT magnitude of the a Cu K-edge and b Zn K-edge EXAFS spectra for Cu/ZnO/Al 2 O 3 , CuZn/SiO 2 , and CuZn/Al 2 O 3 catalysts in their initial state (dashed lines), and under operando conditions in CO 2 +CO+H 2 , p = 20 bar, T = 220°C (solid lines). The magnitude of the FT of the EXAFS spectra for CuZn/SiO 2 during c a reduction in 10% H 2 /He for 20 h at 245°C and d at temperatures between 245 and 600°C. The reference spectra for a Zn foil, ZnO, CuO, and Cu foil are also shown. The highlighted areas in b – d indicate the region where the Zn–M bond develops under reaction conditions. The latter is also shown in the inset in b , with spectra α , β , γ , δ, and ε corresponding to consequent measurements at 220°C/20 bar, 280°C/20 bar, 280°C/40 bar, and 320°C/40 bar. To enhance the display of the referred features, different k-ranges were used for the Fourier transform in a , b 1–10.5 Å −1 , c 3–12 Å −1 , and d 1–10 Å −1 . The observed changes are schematically presented in e for CuZn/Al 2 O 3 and f for CuZn/SiO 2 . Full size image Notably, the Zn K-edge XAS data are substantially different for CuZn/SiO 2 and CuZn/Al 2 O 3 , revealing the distinct interaction between the CuZn NPs and the silica and alumina supports. The Zn K-edge XANES spectra of CuZn/SiO 2 and CuZn/Al 2 O 3 catalysts in their initial state can be aligned well with the spectrum of wurtzite ZnO (where Zn is in 2+ state and tetrahedrally coordinated), two main features appearing at 9664 eV and 9669 eV in the CuZn/SiO 2 catalyst. Nevertheless, the relative intensities of these features are different for CuZn/SiO 2 and bulk ZnO, and in conjunction with the weak contribution of distant coordination shells in the EXAFS data, indicate that Zn is present in a disordered oxide phase 31 . For the CuZn/Al 2 O 3 sample, the small peak in the Fourier-transformed Zn K-edge EXAFS spectra between 2.5 and 3.0 Å is more pronounced and the Zn K-edge XANES spectra have an additional feature at ca. 9674 eV (see arrow in Fig. 2b ). Similar XANES fingerprints were previously observed for various spinel structures such as ZnAl 2 O 4 32 , indicating the migration of Zn atoms into the alumina support. Under operando conditions, only minor changes can be observed for the Zn K-edge XAS of CuZn/Al 2 O 3 , suggesting the lack of significant reduction of the Zn species, which already incorporated in the Al 2 O 3 matrix during the calcination. More pronounced changes are observed for the CuZn/SiO 2 : a decrease of the intensity of the main Zn K-edge XANES features and the appearance of a shoulder at the edge onset, characteristic for metallic Zn (Fig. 2b and Supplementary Figure 14 ). These changes can be related to the partial reduction of Zn species in this sample, and can be more easily observed in the Zn K-edge EXAFS spectra (Fig. 3b and Supplementary Figure 16 ), where an additional feature appears at ca. 2.2 Å in the FT and systematically increases under reaction conditions. This feature can be linked to the formation of Zn–Zn or Zn–Cu bonds and its intensity is low after a 2 h reduction at 325 °C (or during the first measurements under operando conditions), but becomes more pronounced with increasing reaction time. To confirm that the development of this feature was related to the prolonged exposure to reducing conditions, we investigated the effect of reducing atmospheres at different temperatures on the reduction of Zn. We collected the Zn K-edge XAS data for CuZn/SiO 2 after 20 h under 10% H 2 /He at 245 °C and 1 bar (Fig. 3c and Supplementary Figure 17 ), and observed similar changes: the peak at ca. 2.2 Å related to Zn–M bonds clearly develops. This long-term study thus shows an increasing metallic Zn contribution for the CuZn/SiO 2 NPs after the 20 h, as compared with that after 1–2 h of the reduction (Fig. 3d ), indicating slowly progressing reduction of Zn species. To further corroborate this finding, we performed an additional experiment where the CuZn/SiO 2 sample was investigated under 10% H 2 /He at 1 bar and temperatures from 245 °C up to 600 °C (Fig. 3d and Supplementary Figure 17 ). EXAFS data fitting (Supplementary Figures 12 , 13 , and 18 , and Supplementary Table 9 ) confirmed the gradual transformation of initially tetrahedrally coordinated oxidized Zn species into a material with coexisting ZnO and reduced Zn species. Surface reorganization and oxidation state evolution To complement the bulk chemical and structural information obtained by XAS and gain further insight on the outermost layers of the NPs, synchrotron near-ambient pressure X-ray photoelectron spectroscopy (NAP-XPS) measurements were carried out. With this method, further knowledge was attained on the evolution of the oxidation state and surface composition of 6 nm Cu 0.7 Zn 0.3 NPs deposited on SiO 2 /Si(100) (Fig. 4 and Supplementary Figure 19 ). In particular, depth profile information of the NP composition was obtained by choosing two different incident photon energies: 1250 eV, corresponding to inelastic mean free paths (IMFP) of ~0.8 nm for Cu 2p and 0.7 nm for Zn 2p photoelectrons, and 1580 eV corresponding to ~1.2 nm IMFP for Cu 2p and Zn 2p 33 . The series of in situ NAP-XPS data consisted of: (i) oxidation at 400 °C in 0.13 mbar O 2 to remove adventitious carbon, (ii) reduction at 350 °C in 1.3 mbar H 2 , and (iii) reaction at 250 °C in 1.3 mbar of the reaction gas mixture. The effects of different reaction mixtures (CO 2 +CO+H 2 versus CO 2 +H 2 ) and the presence of water 34 on the surface reorganization and chemical state of different species present during the reaction were studied (see concentrations in Supplementary Table 7 ). Fig. 4: Near-ambient pressure X-ray photoelectron spectra of CuZn NPs on SiO 2 /Si(100). a Cu 2p and b Zn 2p regions recorded using a photon energy of 1250 eV. c Copper and zinc atomic fractions obtained under different atmospheres as indicated on the plot. The two-photon energies used for probing the surface (1250 eV) and subsurface (1580 eV) regions of the CuZn NPs are displayed. Error bars correspond to the standard deviation. The inset shows an AFM image of the sample. Full size image The series of Cu 2p 3/2 and Zn 2p 3/2 NAP-XPS spectra, as well as the X-ray generated Cu and Zn LMM Auger regions are shown in Figs. 4 a, b and Supplementary Figures 20 – 23 . In the initial state under ultra-high vacuum (UHV) conditions, the Cu and Zn atomic ratios are 75:25, close to the nominal 70:30 synthesis ratio. Cu 2p shows the typical satellites of Cu 2+ 35 , but can only be adequately fitted with an additional component of Cu + or Cu 0 (ca. 25%); Zn 2p indicates the presence of ZnO. During the initial treatment in O 2 at 400 °C, copper gets almost fully oxidized and the Cu/Zn ratio at the surface increases slightly, indicating a slight Cu surface segregation (Fig. 4c ). The deconvolution of the Zn Auger region shows that zinc is present as ZnO (Supplementary Figure 22 ). Upon exposure to H 2 at 350 °C, strong segregation of Zn toward the surface and of Cu toward the core of the NPs is observed, indicating the formation of a Zn-rich overlayer on the NPs. The Zn atomic content at the surface reached ~72%. This is in agreement with previous reports 1 , 36 and responds to the higher stability of ZnO as compared with CuO in a reductive atmosphere. A shift toward higher binding energies (BEs) of the Zn 2p and Cu 2p and the Auger regions was observed, which can be due to charging effects of the SiO 2 substrate. Under H 2 atmosphere, copper gets fully reduced, which can be confirmed by the Cu LMM Auger spectrum (Supplementary Figure 23 ) 35 . At the same time, the shift of the Zn 2p region towards higher BEs indicates that Zn mainly remains in Zn 2+ state, either in the form of ZnO or Zn(OH) 2 . Analysis of the Zn LMM region, which is more sensitive to the Zn oxidation state, suggests a maximum of 5% metallic Zn coexisting with ZnO. It has been previously reported that the ZnO overlayer formed corresponds to distorted ZnO 11 , 37 , which is also in good agreement with our in situ XAS results. Under reaction conditions at 250 °C with the H 2 +CO+CO 2 mixture (as used before), no further reorganization of the metals is observed. Cu remains reduced and the Zn LMM lines seem to indicate a slight increase (up to 9%) in the concentration of metallic Zn. This indicates that under methanol synthesis conditions, the surface of the NPs remains Zn-rich, with a distorted ZnO x layer located on top of the Cu NPs. Interestingly, this applies to the water-containing mixtures as well. Only under H 2 +CO 2 , a slight Cu surface segregation is observed, reaching 33 at.% Cu at the surface. To bridge the pressure gap and display the relevance of these surface-sensitive experiments for the high-pressure regime, quasi in situ XPS measurements in combination with a high-pressure reactor cell (HPC) were carried out (Supplementary Figure 25 ). For this purpose, the same experimental conditions as for the synchrotron-based operando NAP-XPS measurements were applied in the HPC, but at more relevant pressures (1 bar for the reduction and 20 bar in the reaction mixture), before measuring the sample with XPS in UHV. A detailed description of these experiments can be found in the Supplementary Discussion. It is shown that there is no pressure effect for the segregation trends of these samples and therefore, our NAP-XPS synchrotron results stay valid even at higher pressures. Catalytic performance The activity of the catalysts in powder form was tested in a fixed-bed reactor at 20, 40, and 60 bar, at temperatures from 220 °C to 280 °C (later described as “reaction cycle”) in a reactant mixture of CO 2 /CO/H 2 /He=4/10/74/12, with a weight-to-flow ratio W/F = 0.18 g s/mL. Figure 5a and Supplementary Figure 7 display the steady-state reactivity data of our catalysts at different temperatures and pressures after an activation pre-treatment. Fig. 5: Catalytic performance and evolution of the reaction products during the activation. a Comparison of the methanol production and overall selectivity of all catalysts at 250°C and 40 bar. The C 2+ products are mainly C 2 H 6 . Error bars represent standard deviation. b Temporal evolution of the main reaction products of the CuZn/SiO 2 and CuZn/Al 2 O 3 catalysts at 40 bar and 280°C during an activation treatment until the steady state was achieved. The dashed lines serve as a guide for the eye. Full size image As shown in Fig. 5a , under steady-state operation and regardless of the support, the bimetallic CuZn NPs produce always more methanol than their monometallic Cu counterparts. The highest methanol yield was obtained for CuZn/Al 2 O 3 , followed by CuZn/SiO 2 , Cu/ZnO/Al 2 O 3 , and Cu/Al 2 O 3 , and a lower rate was observed for Cu/SiO 2 . Interestingly, CuZn/SiO 2 is significantly less active than CuZn/Al 2 O 3 , but comparable in activity to Cu/ZnO/Al 2 O 3 . In addition, the stabilization of the methanol production for the CuZn/SiO 2 catalyst took longer, and a significant deactivation after an initial increase of the methanol production was observed. The temporal evolution of the methanol production of the CuZn/SiO 2 and CuZn/Al 2 O 3 samples (at 40 bar and 280 °C) for 250 h is displayed in Fig. 5b , featuring the deactivation of CuZn/SiO 2 until the steady-state was achieved. Considering the selectivity trends, we observe that our most-active catalyst (CuZn/Al 2 O 3 ) is the least selective for methanol (selectivity towards methanol is ca. 58%, at 250 °C, 40 bar) with significant production of dimethylether (DME) at the highest reaction temperatures (250 °C, 280 °C) and to a lesser extent of CH 4 , as shown in Fig. 5a and Supplementary Figure 7 . This is due to the presence of the acidic sites of γ-Al 2 O 3 , which are able to further catalyze the dehydration of the produced methanol to DME 38 , 39 . Its monometallic counterpart, the Cu/Al 2 O 3 catalyst, shows higher DME and CH 4 selectivity, at the expense of methanol, and the minor (1%) production of longer hydrocarbons, such as C 2 H 6 . On the other hand, the silica- and ZnO/Al 2 O 3 -supported catalysts are highly selective toward methanol. Owing to the low amount of Cu and Zn with respect to the amount of the support material in our samples (<5 wt.% total concentration), the contribution from the support is more significant for our catalysts as compared with the industrial bulk-like samples (see the methods section for details). Discussion Our XAS data acquired at relevant industrial pressures allow the pressure effect on the local structure of the CuZn/SiO 2 catalyst to be determined. By comparing the results obtained during the activation under hydrogen (1 bar) and reaction steps at 20 and 40 bar, both from the XANES and EXAFS data analyses (Supplementary Figures 9 and 11 and Supplementary Table 8 ), it can be seen that the Cu structural parameters for the activation step (1 bar, 325 °C) are similar to those obtained under the methanol synthesis mixture at the highest temperature studied (40 bar, 320 °C, step that corresponds to the ageing treatment performed). Therefore, we conclude that no significant pressure or chemical potential effect can be detected on the Cu local structure for the CuZn/SiO 2 catalyst. Combining these XAS results with the XRD data of the catalysts after reaction (these techniques are sensitive to short and long-range order, respectively), the majority of copper remains stable (XAS) but a small fraction of the NPs sinter (XRD and TEM). It is worth mentioning here that the after reaction TEM and XRD data were recorded after more than 150 h under operation, including steps at 280 °C, which could favor the aggregation of Cu. In addition, the XRD data also confirmed the formation of the CuZn phase on SiO 2 after reaction (see Supplementary Figure 6 and Supplementary Discussion). The temporal evolution of the catalytic performance displayed in Fig. 5b , where the CuZn/SiO 2 gradually deactivates until it reaches the steady state after ca. 200 h under operation, can be linked to the gradual reduction of ZnO under operando conditions, as inferred from XAS and corroborated by NAP-XPS. As stated earlier, from the spectroscopic results under methanol synthesis conditions we observe the partial reduction of ZnO, leading to the coexistence of ZnO with reduced Zn species. Furthermore, from the NAP-XPS results, we can conclude that ZnO and reduced Zn species are mostly located at the surface of our CuZn NPs. By XAS, we observe a similar evolution of Zn after 20 h under 10% H 2 /He at 245 °C and 1 bar pressure (Fig. 3c ). Thus, we can conclude that the partial reduction of ZnO and formation of either metallic Zn or Zn–Cu bonds are not dependent on the pressure or the specific composition of the reducing reaction (H 2 versus CO 2 +CO+H 2 ) mixture. Zn reduction is, however, significantly enhanced at higher temperatures (>250 °C). The temperature effect is also demonstrated in the aging experiment performed under 10% H 2 /He at 1 bar pressure and 245 °C, 320 °C, 450 °C, and 600 °C (Fig. 3d and Supplementary Figure 17 ), where the contribution of the Zn–M feature is low at 320 °C, but is increasingly more pronounced at 450 °C and 600 °C, demonstrating and emphasizing the temperature effect on the reducibility of the Zn species. The unambiguous attribution of the latter species to metallic Zn or CuZn alloy formation is not possible owing to their relatively low contribution to our XAS data, together with the similar scattering cross-sections of Cu and Zn, making them indistinguishable by XAS. However, the EXAFS Zn–M bond length is similar to the Cu–Cu bond length in metallic Cu, thus, the incorporation of Zn into a Cu-rich alloy is a viable hypothesis. Note that the formation of CuZn alloy (brass) in CuZn catalysts was previously reported under CO/H 2 atmosphere at 600 °C 17 , and other references also suggested that some alloying of Cu and Zn can take place at temperatures as low as 250 °C 26 . Our results confirm that this process also takes place at lower temperatures, namely 245 °C, although then the rate of ZnO reduction is slower. Regarding the catalytic activity, higher intrinsic methanol formation rates have been reported for Cu/Al 2 O 3 than for Cu/SiO 2 39 , which is in agreement with our data and has been attributed to the Lewis acid nature of the Al 2 O 3 support. Nevertheless, the presence of the stronger acid sites of γ-Al 2 O 3 leads to the dehydration of methanol to DME. From our data, we observe that the addition of Zn promotes the methanol formation for both, the CuZn/Al 2 O 3 and the CuZn/SiO 2 catalysts and its selectivity in the case of CuZn/Al 2 O 3 versus Cu/Al 2 O 3 . This is remarkable and highlights the importance of the intimate CuZn interaction for the CO 2 +CO hydrogenation reaction, since in our catalysts zinc is not present as a bulk support, as done in previous studies, but as a highly diluted phase forming the CuZn NPs. For CuZn/SiO 2 , the addition of Zn causes a methanol production rate increase ca. three times with respect to its monometallic counterpart, and a similar increase is also obtained for CuZn/Al 2 O 3 . It is interesting to note here that the Zn K-edge XAS data revealed for the CuZn/Al 2 O 3 catalyst the migration of Zn atoms to the alumina support and the lack of significant Zn reduction. Nevertheless, our data clearly show that Zn is still able to promote the methanol formation rate and improve its selectivity from 35% for Cu/Al 2 O 3 to 58% for CuZn/Al 2 O 3 . By comparing the methanol yields of our NPs with those of the bulk-like industrial Cu/ZnO/Al 2 O 3 catalyst reference measured under identical conditions (Supplementary Figure 8 ), a clear size-effect is observed, with a lower methanol production for the 2–3 nm NPs employed here, as previously reported elsewhere 4 . This result highlights that operating with the highest possible surface area materials (smallest NPs <5 nm) is not always beneficial for the catalytic process. In addition, the lower Cu and Zn concentrations used in our study (<5 wt.% total metal loading) could also lead to lower methanol productions than those obtained with the industrial catalyst 40 . We finally conclude that morphologically, structurally, and chemically well-defined catalysts composed of ca. 3 nm Cu and CuZn NPs supported on ZnO/Al 2 O 3 , γ-Al 2 O 3 , and SiO 2 were synthesized in order to investigate CuZn and NP/support interactions. By using operando XAS data acquired at high pressure (up to 40 bar), we were able to follow the evolution of the NPs’ structure and composition under reaction conditions. For both, CuZn/SiO 2 and CuZn/Al 2 O 3 , the majority of the zinc species are present in the form of a disordered oxide, whose structure depends strongly on the support. In particular, the incorporation of Zn in the Al 2 O 3 support was observed for CuZn/Al 2 O 3 , forming a spinel-like structure that hinders the reduction of the cationic Zn species under reaction conditions. On the other hand, for CuZn/SiO 2 , we observed the gradual reduction of the oxidic zinc species under reaction conditions, and the formation of Zn–Cu or Zn–Zn bonds. Although this reduction is accelerated at higher temperatures, it was already observed at temperatures commonly used for methanol production (i.e., 250 °C). The slow reduction of ZnO species in this sample observed by XAS is paralleled by a decrease in methanol yield as a function of the reaction time. After an initial deactivation, the steady-state methanol production of CuZn/SiO 2 is still similar to that of Cu/ZnO/Al 2 O 3 , but CuZn/SiO 2 is far more active than Cu/SiO 2 . This indicates the beneficial effect of the CuZn interaction even if the ZnO species get partially reduced or if they are present highly diluted as part of the NPs. In addition, NAP-XPS data revealed Zn surface segregation for CuZn/SiO 2 , and the formation of a distorted ZnO x overlayer, virtually independent from the chosen reactant mixture. Our data indicate that a strong interaction of Cu and ZnO is necessary for a highly selective and active catalyst, which is affected by the choice of the support and the nature of the CuZn interface. Methods Synthesis of powder catalysts The size-controlled NPs were synthesized by an inverse micelle encapsulation method. Monodisperse micellar NP solutions were prepared by dissolving [poly(styrene)-block-poly-(2-vinylpyridine), PS-P2VP, Polymer Source Inc.] polymers in toluene and stirred for 2 days. In parallel, the metal salts CuCl 2 ·2H 2 O (Sigma Aldrich), ZnCl 2 (Alfa Aesar), or Zn(OAc) 2 (Sigma Aldrich) were dispersed in tetrahydrofuran and stirred for 2 days. Nominal Cu:Zn=70:30 atomic ratios were used. Subsequently, the solutions containing the metal salts were incorporated into the micellar solutions and stirred for 2 additional days. The PS-P2VP molecular weights and the polymer-head (P2VP)-to-metal salt ratios are reported in Supplementary Table 1 . The NP height was determined via AFM (Bruker MultiMode 8 microscope) on dip-coated SiO 2 /Si(111) wafers after polymer removal by oxygen plasma (0.3 mbar) (Fig. 1a and Supplementary Figure 1 ). The Cu and Cu 0.7 Zn 0.3 NPs were supported on nanocrystalline powders via incipient wetness impregnation of the NP solutions. The supports employed are commercial SiO 2 (STREM chemicals), γ-Al 2 O 3 (Inframat Advanced Materials), and ZnO/Al 2 O 3 (containing 10% mol Al 2 O 3 ) synthesized from the precipitation of Zn and Al following a variation of the method reported elsewhere 41 . For this, 12.34 g Zn(CH 3 CO 2 ) 2 and 2.52 g Al(NO 3 ) 2 ·9H 2 O were dissolved in 60 ml deionized water. Then 1 M Na 2 CO 3 was added dropwise until a pH value of 9 was reached. After 1 h, the solution was washed, filtered, and dried to obtain a white powder. In the final step, the powder was calcined at 600 °C. XRD analysis showed only reflections corresponding to zincite (ZnO). Temperatures for calcination and cleanliness (removal of polymeric carbon) after the calcination of the NPs supported on the oxide powders were determined by thermogravimetric analysis (TGA) (see Supplementary Figure 2 ). The catalysts were calcined for 6 h in a rotating tubular oven under a flow of 20% O 2 in Ar at the temperatures obtained from the TGA analysis. The metal content of the synthesized powders (after calcination) was determined by inductively coupled plasma-mass spectrometry (ICP-MS). To prepare the samples for ICP, a precise amount of the calcined powder catalyst was dissolved in 10 ml of a 1:1:3 mixture of H 2 SO 4 , HNO 3 , and HCl. This solution was digested in a microwave (Anton Paar GmbH, Multiwave GO) at 180 °C for 30 min. Then, the solution was further diluted with water. The results of the ICP measurements are given in Supplementary Table 2 . TEM imaging TEM and STEM images of the samples before and after reaction were acquired using the microscopes at the Fritz Haber Institute (Thermo Fisher Talos F200X, JEOL ARM200F, operated at 200 kV) in Berlin, at the Ernst Ruska-Centrum (FEI Titan 80-200, operated at 200 kV) in Jülich, and at the Ruhr-University (JEOL JEM-2800) in Bochum. The samples after reaction were imaged by transferring them under inert atmosphere to a glove box and loading them onto Au grids with a holey carbon film. The samples were transferred to the Talos TEM using a vacuum transfer holder. This procedure ensured no exposure to air/O 2 after the reaction. For STEM analysis with a Cs-probe corrected FEI Titan 80-200 microscope, a probe semi-angle of 25 mrad and an inner collection semi-angle of the detector of 88 mrad were used to achieve high-angle-annular dark-field conditions. Compositional maps were obtained with EDX using four large-solid-angle symmetrical Si drift detectors. Additional STEM EDX maps are shown in Supplementary Figures 3 – 5 . XRD characterization The XRD patterns were recorded using a Bruker-AXS D8 Advance diffractometer equipped with a Cu K α source and a position-sensitive energy-dispersive LynxEye XE-T detector. XRD patterns were recorded in continuous scanning mode in a 2θ range of 20–90 °, applying an increment 0.02 ° and a variable divergence slit configuration ensuring constant sample illumination. Rietveld refinement was performed using the software package TOPAS ® (Bruker-AXS) to analyze the diffraction patterns taking into account instrumental broadening, zero error, and sample displacement. Owing to the structural complexity of the Al 2 O 3 , no Rietveld refinement was performed on the CuZn/Al 2 O 3 diffraction pattern. Furthermore, the diffraction signals of the SiO 2 support were considered as convolution of individual peaks, which made a Rietveld quantification impossible. The results of the XRD experiments are shown in Supplementary Figure 6 and Supplementary Tables 3 – 6 . Catalytic testing The catalytic activity was measured in a high-pressure fixed-bed flow reactor. About 50 mg of the catalyst were diluted with ~300 mg SiC (ratio 6:1) and then placed in a glass-lined steel tube. Before testing, all catalysts were reduced by flowing 10% H 2 in He for 2 h at 245 °C. The activity was measured at pressures of 20, 40, and 60 bar and temperatures of 220 °C, 250 °C, and 280 °C. The reaction gas mixture consisted of a 10% CO, 4% CO 2 , 72% H 2 , and 14% He, which was used as an internal standard. The total flow was 17 ml/min. The reaction products were measured online by gas chromatography (GC) with an Agilent Technologies 7890B gas chromatograph equipped with a flame ionization detector and two thermal conductivity detectors. All reported values are the average of at least three consecutive injections. A carbonyl trap was used to ensure that the catalyst remains Ni-free. Our carbonyl trap consisted of a stainless steel tube filled with SiC, which was heated up to 300 °C. The trap was placed on the CO line directly before the mixing with the other gases took place. The absence of Ni after reaction was confirmed by TEM-EDX. The values for the methanol production are normalized using the Cu content (grams of Cu) in the powder catalyst determined by ICP-MS. Supplementary Figure 7 shows additional results obtained during the catalytic measurements. The activity for each catalyst was measured in multiple consecutive reaction steps. After the reduction, the catalyst was cooled down to 220 °C and the reactant mixture was introduced, then the reactor was pressurized to 20 bar. The reactivity was then measured at 220 °C, 250 °C, and 280 °C, each temperature step lasting for 8 h. After this the catalyst was cooled down to 220 °C, before going to the next pressure. This procedure was repeated for the 40 bar and 60 bar data points, for a complete cycle of all reaction steps. During the initial reaction steps, the catalytic activity was not stable, and a long activation period (50–140 h) was needed until the steady-state operation was achieved. The steady-state data for all catalysts are included in the main text (Fig. 5a ) and in Supplementary Figure 7 . The activity of the Cu/SiO 2 and Cu/ZnO/Al 2 O 3 catalysts was found to become stable relatively fast, since already during the step at 40 bar no major changes were observed. For the catalysts supported on γ-Al 2 O 3 , the stabilization took longer time, and the catalyst only showed stable methanol production values during the 60 bar reaction step. During the activation of the Cu/SiO 2 , Cu/ZnO/Al 2 O 3 , Cu/Al 2 O 3 , and CuZn/Al 2 O 3 catalysts, the production and selectivity toward methanol increases and they reach a steady state after the first cycle of all reaction conditions. The data displayed in Fig. 5 shows the results of a second cycle for all reaction conditions once the steady state was reached. However, the activation period is longer for the CuZn/SiO 2 catalyst than for the others, as it does not become stable during the first run through all reaction temperatures and pressures. For this reason, we stayed at one reaction condition (280 °C and 40 bar) until a steady state was reached, which in this case took 150 h (see Fig. 5b ). The same was done for the CuZn/Al 2 O 3 catalyst for comparison (see Fig. 5b ). Using the knowledge gained from the operando XAS experiments, we can conclude that this slow change, which is accelerated at higher temperatures is related to the formation of a metallic Zn phase. Another unique feature of the activation of the CuZn/SiO 2 catalyst is that the methanol production first increases with time, as was the case for all the other catalysts, but then only for this catalyst the activity decreases after ~50 h. The subsequent reduction of the catalytic activity of the CuZn/SiO 2 sample can then be associated with the reduction of Zn, which was only found to take place for this catalyst. The values given in Fig. 5a and Supplementary Figure 7 are collected after the activation period, during the steady-state operation, when no changes in activity over time were observed. The empty reactor was also tested before each experiment to ensure the absence of any background contribution or residuals from previous measurements. In this test, the empty reactor was heated up and pressurized while flowing the reaction gas mixture. To further ensure the reliability of the measurements carried out here on low-metal loading NPs, a commercial copper-based methanol synthesis catalyst (Alfa Aesar, 45776) with a sieve fraction 100–200 μm, was tested under the same conditions in our reactor, showing the high selectivity and yield expected for this high-copper loading sample (Supplementary Figure 8 ). Operando XAS The operando XAS measurements were performed at beamline 2–2 at SSRL (operando high-pressure experiments) and Quick X-ray Absorption and Scattering (QAS) beamline at NSLS-II synchrotron (ambient pressure/high temperature control experiments). Beryllium tube reactors (for measurements at pressures up to 40 bar) and quartz capillary reactors (for measurements at pressures up to 20 bar) were used to emulate the packed bed reactors used for the catalytic tests. The reactors were loaded with the sample, which was fixed in position with quartz wool plugs. The reactor was connected to a gas manifold, consisting of multiple mass flow controllers for precise gas dosing. A backpressure regulator adjusted the pressure in the reactor. The beryllium reactor was heated with a tubular oven with a window for the beam. This configuration only allows for measurements in transmission. The quartz reactors were heated with heating coils and enables, dependent on the configuration, measurements in fluorescence. The samples were diluted with boron nitride to optimize the absorption of the sample for transmission measurements. Control experiments at the QAS beamline were carried out in fluorescence mode using a PIPS detector. A 30% detuned Si(220) double crystal monochromator was used for energy selection at SSRL, whereas a Si(111) monochromator was used at NSLS-II. The ionization chambers used as X-ray detectors for transmission measurements were filled with N 2 . Cu and Zn reference foils were also measured together with the spectra of the samples for energy calibration. During the operando studies at SSRL, all samples were initially measured under He atmosphere and at room temperature to record their initial state. Subsequently, the 10%H 2 /He mixture was dosed and the temperature was raised until the reduction temperature was achieved, which was kept for 2 h. The reduction temperature was 245 °C for the Cu/ZnO/Al 2 O 3 and the CuZn/Al 2 O 3 catalysts and 325 °C for the CuZn/SiO 2 catalyst. A higher activation temperature was used for the sample deposited on SiO 2 in order to reduce the CuO x species, which were found to be more stable on this support. In a separate follow-up experiment (see below), we used 245 °C activation temperature also for the CuZn/SiO 2 catalyst and obtained similar results as for the sample activated at 325 °C, thus the difference in activation temperatures is not crucial here. Next, the temperature was lowered to 220 °C and the reaction mixture (10% CO, 4%CO 2 , 72% H 2 , 14% He) was introduced and the reactor was pressurized to the required working pressures, i.e., 20 and 40 bar. Heating rates of 5 °C/min were used for all temperature changes. The reaction series consisted of: (i) 20 bar at 220 °C, (ii) 20 bar at 280 °C, (iii) 40 bar at 280 °C, and (iv) 40 bar at 320 °C. The last step at high temperature (320 °C) was performed in order to mimic an aging treatment. A mass spectrometer recorded the gaseous effluent of the reactor. For the CuZn/SiO 2 sample, we performed additional experiments where the sample was activated at 245 °C in a H 2 /He mixture at 1 bar for 2 h. The sample was then cooled down to room temperature, and XAS spectra were recorded at room temperature to obtain high-quality data that are not obstructed by high thermal disorder. Subsequently, the sample was exposed to reaction conditions (220 °C and 20 bar) for 2 h, then depressurized and cooled down to room temperature, where XAS spectra were also recorded. Moreover, in a separate experiment we followed the reduction of the CuZn/SiO 2 sample during a long-term treatment, where the sample was kept at 245 °C in a H 2 /He mixture (at atmospheric pressure) for 20 h. Afterwards, the sample cell was cooled down to room temperature, where XAS spectra were collected. Finally, at the QAS beamline we also performed control experiments for the CuZn/SiO 2 sample, where the sample was kept in a H 2 /He mixture at 1 bar, and the temperature was gradually increased from 245 °C to 320 °C, then to 450 °C and 600 °C, to enhance the reduction of the Zn species. In situ NAP-XPS The NPs for the NAP-XPS study were synthesized using the same method described above, using a PS-P2VP (48,500:70,000) polymer to obtain the micelles. Deposition of the particles on the SiO 2 /Si(100) support was done by dip coating (1 cm/min). The polymer ligands were removed by an O 2 plasma treatment (SPI Plasma Prep III Plasma Etcher, 20 min, 20 W, 350 mTorr). The procedure of dip-coating and plasma treatment was repeated three times to increase the density of particles on the support. The particle size and distribution on the SiO 2 /Si(100) was measured by AFM. The NP height values were obtained with the open-source software Gwyddion (Supplementary Figure 19 ). The NAP-XPS measurements with synchrotron radiation were performed at the HIPPIE beamline of the MAX IV synchrotron in Lund (Sweden) and the IOS beamline at NSLS-II in Brookhaven (NY, USA). Additional lab-based NAP-XPS measurements were carried out at the FHI Berlin using Al K α radiation and a hemispherical analyzer (Phoibos 150, SPECS GmbH). The synchrotron measurements were performed at X-ray energies of 1250 eV and 1580 eV to obtain a depth profile, as different kinetic energies result in different escape depths of the photoelectrons. 1250 eV corresponds to an IMFP of ~0.8 nm for Cu 2p and 0.7 nm for Zn 2p photoelectrons, probing the outermost layers of the NPs, and 1580 eV corresponds to ~1.2 nm for Cu 2p and Zn 2p , thus probing deeper regions 33 . The energy values were chosen to avoid overlap with peaks resulting from the Auger electrons. The LMM X-ray generated Auger regions for Cu and Zn were also recorded. All peaks were aligned to the elemental Si 2p 3/2 peak at a BE of 99.4 eV. The Si peak itself was fitted using a doublet, with an energy splitting of 0.6 eV. For the quantification of the elemental composition at given photon energy, the peaks of the Cu 2p 3/2 and Zn 2p 3/2 regions were fitted. The obtained values were corrected with the relative sensitivity factors of the different elements. For quantitative analysis of the peak area, the IMFP and ionization cross-section for the corresponding photon energy were taken into account. The series of NAP-XPS measurements were as follows: the first step was oxidation in pure O 2 (0.13 mbar at 400 °C) to remove any adventitious carbon on the surface of the sample. The C 1 s peak was monitored during the oxidation and the experiment only continued when no traces of carbon were detected. This was followed by a reduction in H 2 (1.3 mbar, 350 °C) for the activation of the catalyst by reduction of the Cu and afterwards, the reaction gas mixture (1.3 mbar, 250 °C) was introduced. The measurements were carried out under different reaction mixtures, whose concentrations are shown in Supplementary Table 7 . Between each of these steps, the sample was cooled down in the gas atmosphere to near room temperature, then the chamber was evacuated, and filled with the new gas before heating up. For each reaction mixture, a new measurement series was performed introducing an identically prepared but fresh sample. An important aspect to be considered when synchrotron NAP-XPS investigations are conducted is the strong influence of the X-ray beam on the sample (Supplementary Figure 20 ). During the measurements in a gas atmosphere, besides the reversible surface segregation discussed in the main text, irreversible radiation-induced segregation trends were also observed. In fact, we detected the loss of the metal segregating to the surface (e.g., Cu in O 2 atmosphere, and Zn in H 2 atmosphere and under the reaction gas mixture) when the X-ray beam stayed on the same spot of the sample surface for several minutes. To avoid any radiation-induced effects on our results, the spot on the sample was changed to a new, fresh location after each set of scans and the Cu 2p and Zn 2p regions were measured in alternating order in consecutive rounds of measurements. The reproducibility and reliability of the results presented here was confirmed in a large number of independent measurements at the synchrotron, as well as analogous investigations with a lab-based NAP-XPS system. Data availability All data are available from the authors on reasonable request.

The current commercial production of methanol through the hydrogenation of the green-house gas CO2 relies on a catalyst consisting of copper, zinc oxide and aluminum oxide. Even though this catalyst has been used for many decades in the chemical industry, unknowns still remain. A team of researchers from the Interface Science Department of the Fritz-Haber-Institute of the Max Planck Society, the Ruhr-University Bochum, Stanford Linear Accelerator Center (SLAC), FZ Juelich and Brookhaven National Laboratory have now elucidated the origin of intriguing catalytic activity and selectivity trends of complex nanocatalysts while at work. In particular, they shed light on the role of the oxide support and unveiled how the methanol production can be influenced by minute amounts of zinc oxide in close contact with copper. Methanol can serve as an energy source or as a raw material for the production of other chemicals, with over 60 million metric tons produced yearly. The traditional copper, zinc oxide and aluminum oxide catalyst converts synthesis gas, which is composed of H2, CO and CO2, into methanol. Though reliable, this specific catalyst's efficiency changes over time, thus affecting its longevity, as is the case with many catalysts. "We therefore studied copper and mixed copper-zinc nanoparticles on various oxide supports to understand how they interact and evolve and unravel the role of each catalyst constituent. This knowledge will serve to improve future catalysts." says Núria Jiménez Divins, one of the lead authors of the study. The team investigated the catalytic process under realistic reaction conditions reproducing those applied in the industrial process, meaning high pressures (20-60 bar) and mild temperatures. This required synchrotron-generated X-ray radiation. Simon R. Bare from the Stanford Synchrotron Radiation Lightsource, who contributed to the experiments, explains: "Reactions at such temperature and high pressures need to take place in a closed container which should also be transparent for the X-rays, which makes the measurements challenging. The special reactor design in combination with synchrotron radiation allowed us to undertake so-called operando-measurements, where we watched live what happens to the catalytic components at the industrially relevant reaction conditions." This allowed the researchers to follow not just the birth and death of the catalyst, but also its development and transformations leading to changes in its activity and selectivity. By combining results from microscopy, spectroscopy and catalytic measurements, the team found that some supports had a more positive influence on the performance of the catalyst than others because of how they interacted with zinc oxide, which was available in a highly dilute manner as part of Cu-Zn nanoparticles. On silicon oxide supports, zinc oxide was partially reduced to metallic zinc or gave rise to a brass alloy during the catalytic process, which over time proved to be detrimental for the methanol production. When using aluminum oxide as a support, zinc interacts strongly with the support and gets incorporated in its lattice, giving rise to a change in reaction selectivity towards dimethyl ether. "This is an interesting finding", says David Kordus, the other lead author of the study and Ph.D. student at the Interface Science Department at FHI. "We know now that the choice of support material has an influence on how the active components of the catalyst behaves and dynamically adapts to the reaction conditions. Especially the oxidation state of zinc is critically influenced by this, which should be considered for future catalyst design." This work published in Nature Communications demonstrates that zinc oxide does not need to be available as part of the support, but that it still has a beneficial function even when available in highly dilute form as part of the nanoparticle catalyst itself. This will help elucidate the methanol synthesis catalysts better and potentially lead to an improvement of the catalyst for this important industrial process.

10.1038/s41467-021-21604-7

Earth

Global warming, El Nino could cause wetter winters, drier conditions in other months

Robert J. Allen et al, 21st century California drought risk linked to model fidelity of the El Niño teleconnection, npj Climate and Atmospheric Science (2018). DOI: 10.1038/s41612-018-0032-x

http://dx.doi.org/10.1038/s41612-018-0032-x

https://phys.org/news/2018-09-global-el-nino-wetter-winters.html

Abstract Greenhouse gas induced climate change is expected to lead to negative hydrological impacts for southwestern North America, including California (CA). This includes a decrease in the amount and frequency of precipitation, reductions in Sierra snow pack, and an increase in evapotranspiration, all of which imply a decline in surface water availability, and an increase in drought and stress on water resources. However, a recent study showed the importance of tropical Pacific sea surface temperature (SST) warming and an El Niño Southern Oscillation (ENSO)-like teleconnection in driving an increase in CA precipitation through the 21st century, particularly during winter (DJF). Here, we extend this prior work and show wetter (drier) CA conditions, based on several drought metrics, are associated with an El Niño (La Niña)-like SST pattern. Models that better simulate the observed ENSO-CA precipitation teleconnection also better simulate the ENSO-CA drought relationships, and yield negligible change in the risk of 21st century CA drought, primarily due to wetting during winter. Seasonally, however, CA drought risk is projected to increase during the non-winter months, particularly in the models that poorly simulate the observed teleconnection. Thus, future projections of CA drought are dependent on model fidelity of the El Niño teleconnection. As opposed to focusing on adapting to less water, models that better simulate the teleconnection imply adaptation measures focused on smoothing seasonal differences for affected agricultural, terrestrial, and aquatic systems, as well as effectively capturing enhanced winter runoff. Introduction In response to anthropogenic climate change, climate models from the Coupled Model Intercomparison Project (CMIP) version 3 and 5 indicate a likely transition to a more arid climate over many land areas, 1 resulting in increased frequency and intensity of drought. 2 , 3 Severe and widespread drought during this century are of particular concern for southwestern North America, including California (CA). 4 , 5 , 6 , 7 , 8 This drying is a consequence of reduced precipitation in the subtropics, and a poleward expansion of the subtropical dry zones. 9 , 10 , 11 , 12 , 13 Moreover, in addition to reduced precipitation, warmer temperatures will lead to an increase in evapotranspiration 3 and a decrease in mountain snow mass. 6 All of this translates into reduced surface water availability, soil moisture and runoff, 14 implying significant stress on water resources. Future drought risk in southwestern North America may even exceed that during the driest centuries of the Medieval Climate Anomaly. 15 , 16 However, despite the large body of literature suggesting an increase in drought under greenhouse gas forcing, uncertainty still remains. Overall, there is medium confidence that warming will increase the duration and intensity of droughts in some regions, including the Mediterranean, Central America and Mexico, northeast Brazil, southern Africa and central North America. 13 , 17 The only regions with a consistent increase in drought are those where precipitation decreases. 18 Furthermore, many of the above studies quantify drought using the Palmer Drought Severity Index (PDSI), which is calculated from a simple water balance model, 19 and may overestimate the increase in global drought 20 due to an oversensitivity to warmer temperatures. 21 Consistent with this uncertainty, a recent study suggests increased radiative forcing may lead to a decrease in the likelihood of CA agricultural drought. 22 This study, however, was based on a single model and did not focus on 21st century climate change. An important component of drought is precipitation, and relatively large uncertainty exists for future CA precipitation projections. 23 , 24 Relative to CMIP3, however, CMIP5 CA precipitation projections tend to yield a more consistent increase. 25 , 26 , 27 This is related to a coherent extratropical response, involving a southeastward extension of the the upper level winds in the east Pacific, 25 an increase in storm track activity, 28 and an increase in CA moisture convergence, 29 all of which promote an increase in CA precipitation. Moreover, a robust dynamical response in the tropics also exists, 27 including an increase in central/eastern tropical Pacific divergence and a poleward propagating Rossby wave, both of which are reminiscent of an El Niño-like teleconnection. Combined with warming of the tropical Pacific sea surface temperatures (SSTs), CMIP5 models that better simulate the observed El Niño-CA precipitation teleconnection yield larger, and more consistent increases in CA precipitation through the twenty-first century. 27 Using a multitude of models from the CMIP5 30 archive, we build off of prior work 27 and evaluate future CA drought risk using multiple metrics under business as usual warming. We primarily focus on annual drought to be able to compare our results to previous studies of future water availability in California and the Southwestern US from GCM simulations. 4 , 31 More importantly, the question of potential changes in seasonal versus annual (and inter-annual) drought greatly affects the feasible adaptation measures to climate change. If long-term water availability decreases, this will likely result in greater competition between agricultural, environmental, and urban users for California water and greater perennial stress on human and ecological systems. In this scenario, adaptation measures will need to focus on adapting to less water (including managing ecosystems as they transition to a drier state). If annual drought risk doesn’t change, adaptation measures can focus more on smoothing seasonal differences for affected agricultural, terrestrial, and aquatic systems as well as effectively capturing enhanced winter runoff. These two scenarios have greatly different potential losses due to climate change as well as differing adaptation costs. Results Tropical pacific sea surface temperatures and California hydrology Multiple studies 2 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 have showed that SSTs are the dominant driver of changes in precipitation and drought in many world regions over the 20th century. For CA, the dominant source of SST variations that contribute to precipitation variations is El Niño/La Niña. Observations from 1948/49 to 2014/15 show the correlation between Niño 3.4 SSTs (5S-5N; 190-240E) and CA winter (DJF) precipitation is ~0.3–0.4, which is significant at the 99% confidence level. Although not all El Niño (La Niña) winters are wetter (drier) than normal since other factors also contribute to CA hydrological variations (e.g., meteorological noise and atmospheric rivers), and some evidence suggests the El Niño-CA precipitation relationship may not be stationary through time, 40 , 41 , 42 this teleconnection represents a reasonably strong pathway by which CA precipitation variations occur. We also note that El Niño influence on California precipitation is strongest during late winter, and is stronger in the south than the north. 43 As previously showed, models that better simulate the El Niño-CA precipitation teleconnection yield a very different projection of CA precipitation through the 21st century ( Supplementary Discussion ; Table 1 ). 27 CMIP5 models with a 21st century detrended CA DJF precipitation versus Niño 3.4 SST correlation of at least 0.30 (14 models and 29 realizations in total), referred to as “CMIP5 HIGH–r”, yield larger and more consistent increases in CA precipitation (Supplementary 1 lists the models and correlations). The ensemble mean annual mean (ANN) CA precipitation trend is positive at 0.16 mm day −1 century −1 , significant at the 99% confidence level (Methods). Furthermore, 73% of the model-realizations yield an increase in CA ANN precipitation (Table 1 ). In contrast, models that yield a corresponding correlation less than 0.20 (12 models and 17 realizations), referred to as “CMIP5 LOW−r”, yield more consistent decreases in 21st century CA precipitation. The ensemble mean ANN precipitation trend is −0.15 mm day −1 century −1 , significant at the 99% confidence level, and 76% of the model-realizations yield a decrease in CA ANN precipitation. Table 1 21st century ensemble mean annual and seasonal CMIP5 hydrological trend statistics for California Full size table Most of the difference in CA ANN precipitation projections between CMIP5 HIGH–r and LOW–r occurs during DJF. CMIP5 HIGH–r yields a significant increase of 0.84 mm day −1 century −1 (80% realization agreement) whereas CMIP5 LOW–r yields a much weaker, and nonsignificant, increase of 0.09 mm day −1 century −1 , with low realization agreement of 59% (Table 1 ) Non-winter seasons generally feature precipitation reductions in both model subsets, with drying trends during both March–April–May (MAM) and September–October–November (SON), particularly for CMIP5 LOW–r models (Table 1 ). Consistent with “robust spring drying” of the southwestern U.S. due to strengthening and expansion of the subtropical high pressure in the Pacific and Atlantic, 44 the precipitation decrease is particularly robust during MAM. We expand upon these results by first showing the importance of central/eastern tropical Pacific SSTs to other hydrological variables over CA. Figure 1 shows the 1979–2015 correlation between detrended annual mean soil moisture and SSTs in CMIP5 HIGH–r, CMIP5 LOW–r, and observations, including European Space Agency Climate Change Initiative (ESACCI) soil moisture observations 45 from passive microwave sensors and Hadley Centre Sea Ice and Sea Surface Temperature data. 46 Model correlations are based on 1950–1999, but similar results are obtained over different 20th century time periods (e.g., 1901–1999), as well as the 21st century (not shown). Simulated soil moisture is based on shallow/intermediate depths (0.2–1.0 m), which corresponds to the depth over which the ESACCI passive soil moisture observations represent. A comprehensive evaluation of CMIP5 soil moisture simulations 47 show that CMIP5 models are able to simulate the seasonal variability in soil moisture over the United States. However, models tend to overestimate near-surface (0–10 cm) and soil column soil moisture (0–1 m) in the western US. Fig. 1 Late-20th century correlation maps between California soil moisture and sea surface temperatures. Detrended correlations between (left panels) annual mean sea surface temperatures (SST) and annual mean California (CA) soil moisture; and (right panels) December–January–February (DJF) SSTs and the subsequent June–July–August (JJA) CA soil moisture for (top panels) observations; (middle panels) the model subset that yield a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (CMIP5 HIGH–r); and (bottom panels) the model subset that yields a corresponding correlation less than 0.20 (CMIP5 LOW–r). Soil moisture observations are from European Space Agency Climate Change Initiative (ESACCI) and HadISST is used for observed SSTs. All correlations are based on the 1950–1999 time period, except the ESACCI observed correlations, which span 1979–2015 Full size image ESACCI soil moisture observations show an El Niño-like correlation pattern, implying anomalously high (low) CA soil moisture corresponds with anomalously warm (cold) central/eastern tropical Pacific SSTs. This correlation patterns also exists in the models. CMIP5 LOW–r underestimates the observed correlation, whereas CMIP5 HIGH–r better reproduces it. Figure 1 also shows that similar, but somewhat stronger correlations are obtained between DJF SSTs and the subsequent June–July–August (JJA) CA soil moisture. Thus, the effects of El Niño/La Niña, which peak during Northern Hemisphere winter, are realized throughout the year, including during CA’s dry season, when municipal, agricultural, and natural systems are most water stressed. A similar El-Niño-like relationship exits for other observed drought indices over CA ( Supplementary Discussion ). This includes the Standardized Precipitation Index (SPI), the World Meteorological Organization’s recommended index for monitoring abnormal dryness and wetness, 48 as well as the observed Standardized Precipitation Evapotranspiration Index (SPEI) 49 and the self-calibrated Palmer Drought Severity Index (scPDSI) 50 (Supplementary Fig. 1 ). A similar relationship also exists between central/eastern tropical Pacific SSTs and both CA precipitation minus evapotranspiration (P–E) and total runoff using the North American Land Data Assimilation System (NLDAS) version 2 models. 51 , 52 Moreover, CMIP5 HIGH–r models better simulate these relationships, relative to CMIP5 LOW–r models (Supplementary Figs. 2 – 7 ). To further isolate the relationship between SSTs and extreme CA dryness (drought) and wetness, Fig. 2 shows the normalized SST anomaly pattern associated with the five largest and five smallest CA annual mean soil moisture anomalies based on ESACCI observations and CMIP5 models. Again, a distinct ENSO-like relationship exists in the observations, with CMIP5 HIGH-r models better reproducing this observed pattern, relative to CMIP5 LOW-r models. Similar results exist based on the SPI and SPEI (Supplementary Figs. 8 and 9 ). Furthermore, Fig. 2 shows that similar conclusions apply between DJF SSTs and the subsequent JJA CA soil moisture. Fig. 2 Late-20th century sea surface temperature anomaly composite maps for the five largest and smallest California soil moisture anomalies. (Top panels) Annual mean sea surface temperature (SST) anomaly pattern associated with the five ( a – c ) largest and ( d – f ) smallest annual mean CA soil moisture anomalies for (left panels) observations; (middle panels) the model subset that yield a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (CMIP5 HIGH–r); and (right panels) the model subset that yields a corresponding correlation less than 0.20 (CMIP5 LOW–r). (Bottom panels) December–January–February (DJF) SST anomaly pattern associated with the subsequent June-July-August (JJA) five ( g – i ) largest and ( j – l ) smallest CA soil moisture anomalies. SSTs are normalized by the absolute value of CA soil moisture. Soil moisture observations are from European Space Agency Climate Change Initiative (ESACCI) and HadISST is used for observed SSTs. All analyses are based on the 1950–1999 time period, except ESACCI, which span 1979–2015 Full size image Thus, a significant relationship exists between anomalously warm (cool) central/eastern tropical Pacific SSTs and anomalously wet (dry) conditions in CA. These relationships exist on longer time scales (annual and longer) for all hydrological indicators, and also seasonally (DJF SSTs versus JJA hydrology) for soil moisture and runoff metrics, including the scPDSI. Moreover, CMIP5 HIGH–r models better simulate these relationships, relative to CMIP5 LOW–r. Surface water availability In the context of drought, not only is precipitation important, but so is evapotranspiration. Under business-as-usual (Representative Concentration Pathway 8.5, RCP8.5), CA is projected to warm significantly, ranging from 2.3 to 7.4 K century −1 , with a multi-model mean of 5.0 K century −1 and 100% model realization agreement (Supplementary Table 2 ). This warming results in enhanced evaporative demand of the atmosphere, which acts to reduce surface water availability (i.e., precipitation minus evapotranspiration, P–E). In terms of P–E, the CMIP5 RCP8.5 annual mean CA trend ranges from −0.77 to 0.53 mm day −1 century −1 , with a non-significant multi-model mean of −0.02 mm day −1 century −1 . A decrease in CA ANN P–E occurs in 53% of the model realizations (Table 1 ). Thus, although significant uncertainty exists, the CMIP5 multi-model mean suggests enhanced evapotranspiration exceeds the increase in precipitation, resulting in a reduction in surface water availability. Extending this analysis to surface water availability, CMIP5 HIGH–r yields an ensemble mean ANN P–E increase of 0.09 mm day −1 century −1 , significant at the 99% confidence level, with 70% of the model realizations yielding an increase (Table 1 ; Fig. 3 ). As with precipitation, CA P–E increases more from south to north, with P–E decreasing by −0.06 mm day −1 century −1 in southern (32.0–34.9 °N; 239.4–245.6 °E) California, but increasing to 0.11 and 0.22 mm day −1 century −1 in central (34.9–38.6 °N; 236.9–243.1 °E) and northern (38.8–42.4 °N; 235.6–240.6 °E) California, respectively, all significant at the 99% confidence level. In contrast, CMIP5 LOW–r shows the opposite response–an ensemble mean ANN P–E decrease of −0.22 mm day −1 century −1 , significant at the 99% confidence level, with 81% model realization agreement. All three CA regions experience a significant P–E decrease (Supplementary Table 2 ). Thus, unlike the entire CMIP5 database–and in particular CMIP5 LOW–r–the increase in CA precipitation in CMIP5 HIGH–r exceeds the increase in CA evapotranspiration. Fig. 3 Coupled Model Intercomparison Project version 5 RCP8.5 2006–2100 precipitation minus evapotranspiration trends Ensemble mean precipitation minus evapotranspiration (P–E) a , b Annual; c , d December–January–February; e , f March–April–May; g , h June–July–August; i , j September–October–November mean trends [mm day −1 century −1 ] for two CMIP5 model subsets. Left panels show the model subset that yield a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (CMIP5 HIGH–r); Right panels show the model subset that yield a corresponding correlation less than 0.20 (CMIP5 LOW–r). Symbols represent trend significance at the 90% (diamond), 95% (X) or 99% (+) confidence level, accounting for autocorrelation. Blue/green (brown) colors represent an increase (decrease) in surface water availability. Also included are the three regions comprising California, denoted with thick black lines Full size image As with precipitation, however, the increase in surface water availability in CMIP5 HIGH–r models primarily occurs during DJF (Table 1 ; Fig. 3 ). CMIP5 HIGH–r yields a significant increase in DJF CA P–E of 0.70 mm day −1 century −1 , with 79% realization agreement. In contrast, CMIP5 LOW–r yields a nonsignificant decrease of −0.14 mm day −1 century −1 (56% realization agreement). Non-winter seasons, particularly MAM, generally feature significant P–E reductions in both model subsets, which tend to be larger and more robust in CMIP5 LOW–r. The tendency for a decrease in surface water availability in CMIP5 models, particularly CMIP5 LOW–r, implies an enhanced risk of CA drought. In contrast, the tendency for an increase in surface water availability in CMIP5 HIGH–r, particularly for central and northern CA, implies a muted risk of CA drought under warming. However, even in CMIP5 HIGH–r, a decrease in surface water availability is projected during the non-winter months, particularly MAM and SON. Precipitation type and snowpack Another important consideration for future changes in drought risk is how the type of precipitation (i.e., solid versus liquid) is projected to change. Consistent with the large future warming under RCP8.5, significant and robust decreases in CA solid precipitation (i.e., snowfall) are projected. The entire CMIP5 database yields a significant decrease of ANN solid precipitation at −0.12 mm day −1 century −1 , with 100% model realization agreement. CMIP5 HIGH–r and LOW–r also yield significant decreases of −0.10 and −0.15 mm day −1 century −1 , respectively, with 100% model realization agreement (Table 1 ). For the entire CMIP5 database, the ensemble mean ANN increase in total (solid + liquid) precipitation is 0.06 mm day −1 century −1 , and the decrease in solid precipitation is −0.12 mm day −1 century −1 . This yields a significant increase in liquid precipitation (i.e., rain) of 0.18 mm day −1 century −1 (76% realization agreement), three times larger than the increase in total precipitation. In the case of CMIP5 HIGH–r, the ensemble mean increase in (solid + liquid) precipitation is 0.16 mm day −1 century −1 and the decrease in solid precipitation is −0.10 mm day −1 century −1 . This yields a significant increase in liquid precipitation of 0.26 mm day −1 century −1 (83% realization agreement), which is ~60% larger than the increase in total precipitation. With CMIP5 LOW–r, where total and solid precipitation both decrease by −0.15 mm day −1 century −1 , the change in liquid precipitation is negligible at −0.006 mm day −1 century −1 (47% realization agreement). Thus, models project a robust decrease in snowfall, and in turn, snowpack. Furthermore, in most models, especially CMIP5 HIGH–r, the proportion of precipitation falling as rain is projected to increase. This has implications for water storage systems that rely on the gradual melt of winter snow throughout the spring and early summer. Change in Sierra snowpack is another important aspect of California climate change, and it has major impacts on hydrology. Consistent with prior studies, 23 , 53 , 54 , 55 we find large and robust decreases in CA surface snow amount. For both CMIP5 HIGH-r and LOW-r, 100% of the realizations yield a negative 21st century CA trend in surface snow amount based on the annual mean (Table 1 ), and all seasons, including MAM. In terms of the percent change in CA snow amount, defined as 2050–2099 minus the 1950–1999 climatology, divided by the 1950–1999 climatology, both model subsets yield similar CA percent changes, including −89 and −90% based on annual means for CMIP5 HIGH–r and LOW–r, respectively. Similarly for MAM, the percent change in CA snow amount is −94% for both model subsets. Thus, models project robust and large decreases in CA snowpack, and this is independent of the model subset. This again suggests continued caution is warranted regarding future changes in CA drought. Land surface hydrology In addition to quantifying meteorological drought (i.e., changes in precipitation), hydrological and agricultural drought also exist. Hydrological drought is often quantified in terms of streamflow, 56 whereas agricultural drought is quantified in terms of soil moisture. 57 The entire CMIP5 database (32 models and 65 realizations archived total runoff) yields a significant increase in CA ANN runoff of 0.05 mm day −1 century −1 , with 60% model realization agreement (Table 1 ). Figure 4 shows that the increase in CA total runoff is more robust and larger in CMIP5 HIGH–r, with an ensemble mean ANN increase of 0.13 mm day −1 century −1 and 73% model realization agreement. In contrast, CMIP5 LOW–r yields the opposite response–CA ANN runoff decreases by −0.18 mm day −1 century −1 , with 78% model realization agreement. These changes in runoff are consistent with the corresponding changes in precipitation. Fig. 4 Coupled Model Intercomparison Project version 5 RCP8.5 2006–2100 total runoff trends. Ensemble mean total runoff a , b Annual; c , d December–January–February; e , f March–April–May; g , h June–July–August; i , j September–October–November mean trends [mm day −1 century −1 ] for two CMIP5 model subsets. Left panels show the model subset that yield a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (CMIP5 HIGH–r); Right panels show the model subset that yield a corresponding correlation less than 0.20 (CMIP5 LOW–r). Symbols represent trend significance at the 90% (diamond), 95% (X) or 99% (+) confidence level, accounting for autocorrelation. Blue/green (brown) colors represent an increase (decrease) in total runoff. Also included are the three regions comprising California, denoted with thick black lines Full size image Similar results are obtained for regional CA runoff changes (Supplementary Table 2 ). CMIP5 LOW–r models yield significant reductions in each region, with larger reductions from south to north. CMIP5 HIGH–r yields nonsignificant runoff increases in southern CA, with significant increases of 0.13 mm day −1 century −1 in central CA and 0.26 mm day −1 century −1 in northern CA. Thus, CMIP5 LOW–r models imply a reduction in streamflow and an enhanced risk of hydrological drought. CMIP5 HIGH–r models, however, imply the opposite response. As with the other hydrological metrics, however, the increase in runoff in CMIP5 HIGH–r models primarily occurs during DJF (Table 1 ; Fig. 4 ). CMIP5 HIGH–r yields a significant increase in DJF CA P–E of 0.57 mm day −1 century −1 , with 90% realization agreement. In contrast, CMIP5 LOW–r yields a nonsignificant decrease of −0.27 mm day −1 century −1 (67% realization agreement). Non-winter seasons generally feature runoff reductions in both model subsets, which tend to be larger and more robust in CMIP5 LOW–r. For example, during MAM, CMIP5 HIGH–r (CMIP5 LOW–r) yields a nonsignificant (significant) decrease of −0.03 (−0.36) mm day −1 century −1 with 57% (100%) realization agreement. Agricultural drought is quantified using both shallow (~0–0.25 m) and deep (~1.5–3 m) soil moisture (Supplementary Table 3 ). Of the two, deep soil moisture is most relevant for agricultural impacts, as the deeper layer controls moisture supply for most deeper rooted, perennial CA plants and crops ( Supplementary Discussion ). Deep soil moisture (SMB) shows a significant ensemble annual mean decrease using all CMIP5 models (27 models and 51 realizations), at −4.6 mm century −1 , with 61% realization agreement (Table 1 ). Figure 5 shows that CMIP5 HIGH–r yields a weaker and nonsignificant decrease at −2.4 mm century −1 with low realization agreement at 43% (i.e., the majority of the trends are actually positive). CMIP5 LOW–r yields a larger and more robust decrease, at −9.7 mm century −1 with 92% realization agreement. An SMB decrease exists for each subregion, with CMIP5 LOW–r yielding trends of −4.1, −12.5, and −11.7 mm century −1 for southern, central and northern CA (all significant at the 99% confidence level), with realization agreements of 92% for each subregion. CMIP5 HIGH–r yields corresponding trends of 0.2, −4.8, and −1.9 mm century −1 for southern, central and northern CA, with central CA yielding the lone significant trend (Supplementary Table 2 ). Corresponding model realization agreements continue to remain low at 47, 43, and 43%, respectively. Thus, CMIP5 LOW–r yields robust decreases in CA deep soil moisture, implying enhanced risk of agricultural drought. CMIP5 HIGH–r yields weaker decreases in SMB, with minimal realization agreement. Fig. 5 Coupled Model Intercomparison Project version 5 RCP8.5 2006–2100 deep soil moisture trends. Ensemble mean deep (1.5–3 m) soil moisture a , b Annual; c , d December–January–February; e , f March–April–May; g , h June–July–August; i , j September–October–November mean trends [mm century −1 ] for two CMIP5 model subsets. Left panels show the model subset that yield a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (CMIP5 HIGH–r); Right panels show the model subset that yield a corresponding correlation less than 0.20 (CMIP5 LOW–r). Symbols represent trend significance at the 90% (diamond), 95% (X) or 99% (+) confidence level, accounting for autocorrelation. Blue/green (brown) colors represent an increase (decrease) in deep soil moisture. Also included are the three regions comprising California, denoted with thick black lines Full size image Seasonally, the CMIP5 HIGH–r models yield a nonsignificant increase in CA SMB during DJF (Table 1 ; Fig. 5 ) of 1.3 mm century −1 , with 62% realization agreement. In contrast, CMIP5 LOW–r yields significant a decrease of −11.0 mm century −1 (92% realization agreement). Non-winter seasons generally feature SMB reductions in both model subsets, which tend to be larger and more robust in CMIP5 LOW–r. For example, during JJA, CMIP5 HIGH–r (CMIP5 LOW–r) yields a significant decrease of −4.5 (−9.7) mm century −1 with 48% (92%) realization agreement. Figure 6 shows that shallow soil moisture (SMT) is also projected to decrease, and these changes are more robust relative to changes in deep soil moisture. CMIP5 models yield a significant SMT decrease of −0.92 mm century −1 with 80% realization agreement. All three regions yield a significant decrease in SMT at −0.74, −1.0, and −0.94 mm century −1 for southern, central and northern CA, with 82, 80, and 75% realization agreement, respectively. CMIP5 LOW–r yields a stronger and more robust decrease at −1.6 mm century −1 with 100% model realization agreement. CMIP5 HIGH–r yields a decrease of −0.6 mm century −1 with 62% model realization agreement. Both model subsets yield significant SMT decreases for each CA region, with larger realization agreement in CMIP5 LOW–r (Supplementary Table 2 ). Fig. 6 Coupled Model Intercomparison Project version 5 RCP8.5 2006–2100 shallow soil moisture trends. Ensemble mean shallow (0–0.25 m) soil moisture a , b Annual; c , d December–January–February; e , f March–April–May; g , h June–July–August; i , j September–October–November mean trends [mm century −1 ] for two CMIP5 model subsets. Left panels show the model subset that yield a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (CMIP5 HIGH–r); Right panels show the model subset that yield a corresponding correlation less than 0.20 (CMIP5 LOW–r). Symbols represent trend significance at the 90% (diamond), 95% (X) or 99% (+) confidence level, accounting for autocorrelation. Blue/green (brown) colors represent an increase (decrease) in shallow soil moisture. Also included are the three regions comprising California, denoted with thick black lines Full size image Seasonal SMT results show that CMIP5 HIGH–r (LOW−r) models yield non-significant (significant) negative trends during DJF (Table 1 ; Fig. 6 ), with low (high) realization agreement. Non-winter seasons generally feature significant SMT reductions in both model subsets, which tend to be larger and more robust in CMIP5 LOW–r. For example, during JJA, CMIP5 HIGH–r (CMIP5 LOW–r) yields a significant decrease of −0.52 (−1.37) mm century −1 with 52% (83%) realization agreement. The decrease in soil moisture is consistent with the decrease in surface water availability, particularly in CMIP5 LOW–r. In CMIP5 HIGH–r, despite an increase in surface water availability, soil moisture–particularly near the surface–still decreases. This is likely due to the change in precipitation, with more rain instead of snow–the increase in surface water is not able to be absorbed into the soil, and instead runs off. Furthermore, the more robust decrease in near-surface soil moisture, as opposed to deep soil moisture, is consistent with the importance of enhanced evaporation due to warming in dictating changes in near-surface soil moisture. 22 In contrast, deeper soil moisture responds most sensitively to changes in precipitation, which also helps to explain the robust SMB decrease in CMIP5 LOW–r (where precipitation decreases), but weaker SMB change in CMIP5 HIGH–r (where precipitation increases). California annual and multi-annual drought risk Figure 7 shows the enhanced risk of CA annual drought (Methods; a −0.5 σ departure from the 1950–1999 baseline) based on several metrics, including surface water availability (P–E), precipitation (P), runoff, shallow (SMT) and deep soil moisture (SMB). CMIP5 models show enhanced annual drought risk for all metrics, ranging from modest (but still generally significant) increases based on P–E, P and runoff (3–8%), to more substantial increases based on SMB (12.8%) and SMT (26.1%). The enhanced drought risk based on SMB represents an increased probability of drought from 34.2% at the end of the 20th century, to 47.0% at the end of the 21st century (Supplementary Table 4 ). Much of the enhanced drought risk comes from CMIP5 LOW–r models, which yield significant increases in drought risk for all metrics. Based on P–E, P, and runoff, CMIP5 LOW–r yields enhanced CA drought risks of 22.4, 12.8, and 28.2%, respectively. Based on soil moisture metrics, the enhanced drought risk is larger, at 29.4% for SMB and 42.6% for SMT. The enhanced drought risk based on SMB represents nearly a doubling of the probability of drought, from 34.3% at the end of the 20th century, to 63.7% at the end of the 21st century (Supplementary Table 4 ). Fig. 7 Coupled Model Intercomparison Project version 5 RCP8.5 enhanced California drought risk. Ensemble mean ( a ) annual and ( b ) multi-annual (3-year) California drought risk [%] based on precipitation minus evapotranspiration (P–E), precipitation (P), runoff, shallow (0–0.25 m) soil moisture (SMT) and deep (1.5–3 m) soil moisture (SMB). Colored bars show the ensemble mean enhanced drought risk based on all CMIP5 models (blue; ALL), the model subset that yields a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (green; HIGH–r), and the model subset that yields a corresponding correlation less than 0.20 (red; LOW–r). Bars with a thick black outline indicates significance at the 95% confidence level, based on a standard t -test Full size image In contrast, CMIP5 HIGH–r models yield a muted, and generally non-significant, change in CA annual drought risk. Based on P–E, P and runoff, the change in annual drought risk is 0.3, −2.1, and 1.3%, respectively. For soil moisture metrics, the values are larger–particularly for surface soil moisture–but considerably smaller than those based on CMIP5 and in particular, CMIP5 LOW–r, at 3.4% for SMB and 16.7% for SMT. Similar results are obtained based on multi-annual (3 year) drought (Fig. 7b ; Supplementary Table 5 ) and more severe (Methods; a = −1.0 σ departure from the 1950–1999 baseline) drought (Supplementary Fig. 10 ; Supplementary Table 6 ). Thus, similar conclusions apply for the duration and intensity of drought, in addition to the frequency. We have also verified that the change in drought risk is similar when based on the SPI ( Supplementary Discussion ; Supplementary Fig. 11 ). These results are consistent with the prior discussion that showed CMIP5 LOW–r tends to yield decreases in precipitation, surface water availability, runoff and soil moisture. In contrast, CMIP5 HIGH–r generally yields the opposite responses, particularly for P, P–E, and runoff. The main exception is soil moisture, particularly SMT, with CMIP5 HIGH–r models yielding relatively large and significant increases in SMT drought at 16.7%. This is consistent with the strong sensitivity of drying of the upper soil to enhanced evaporation due to warming, as opposed to changes in precipitation. However, in the context of agricultural drought, the deep soil moisture is most relevant for agricultural impacts (Supplement), since the deeper layer controls the moisture availability for most California plants and non-vegetable crops. 22 CMIP5 HIGH–r yields a non-significant increase in SMB drought at 3.4%, suggested minimal change in the risk of agricultural drought. The results are robust to the criteria used to define CMIP5 HIGH–r and LOW–r models. For example, re-defining the HIGH–r (LOW–r) models as those with a late 20th century DJF CA precipitation versus Niño 3.4 SST regression coefficient within one sigma (less than 1 sigma) of observations 27 leads to similar conclusions. CMIP5 HIGH–r continues to yield a negligible increase in CA drought risk (Supplementary Fig. 12 ). Finally, an alternative “HIGH–r” model subset is defined that satisfies multiple criteria: 1. late-20th and 21st century correlations between DJF CA precipitation and Niño 3.4 SSTs are significant at the 90% confidence level; 2. late-20th century DJF CA precipitation versus Niño 3.4 SST regression coefficient falls within 1-sigma of the observed range; and 3. late-20th century DJF CA precipitation climatologies fall within 1-sigma of the observed range. These criteria result in three models, with 12 realizations. As with the other methods used to define HIGH–r models, minimal increase in CA drought risk exists in this model subset (Supplementary Fig. 12 ). Thus, models that better simulate CA precipitation statistics yield negligible change in CA drought risk. Regionally, a larger increase in annual meteorological drought risk exists in southern CA, based on P and P–E (Supplementary Fig. 13 ). This is consistent with a decrease in both P and P–E in southern CA for all model subsets. Moreover, CMIP5 LOW–r yields the largest enhanced meteorological drought risk in southern CA, consistent with the maximum decrease in southern CA P and P–E in this model subset. In central and northern CA, based on CMIP5 HIGH–r, the annual meteorological drought risk generally decreases, particularly for northern CA, where P–E and particularly P indicate decreased drought risk. This, too, is consistent with the increase in P and P–E in central, and particularly, northern CA in CMIP5 HIGH–r. Based on soil moisture metrics, each region is projected to experience an increased risk of agricultural drought, independent of model subset, with the smallest increase in CMIP5 HIGH–r. Moreover, CMIP5 HIGH–r continues to yield non-significant increases in drought risk based on deep soil moisture. The increase in SMB drought risk is relatively similar for all three regions, although there is a weak south to north decrease in the enhanced risk. For example, CMIP5 yields an increase in SMB drought risk of 18.5, 14.1, and 10.2% from south to north, which is consistent with the corresponding change in precipitation, and the importance of P in driving changes in deep soil moisture. In other words, regions that receive a larger increase (decrease) in precipitation–including central and northern CA (southern CA)–yield a weaker (stronger) increase in drought risk based on deep soil moisture. Compared to SMB, the drought risk based on shallow soil moisture shows the opposite gradient–an increase from south to north. For example, CMIP5 yields an increase in SMT drought risk of 20.8, 20.3, and 28.7% from south to north. This is consistent with the importance of evapotranspiration in dominating the SMT response, as E increases from 0.07 mm day −1 century −1 in southern and central CA to 0.10 mm day −1 century −1 in northern CA. Similar results are obtained using alternative RCPs, including RCP4.5 (Supplementary Fig. 14 ) and RCP6.0 (Supplementary Fig. 15 ). Generally, a smaller increase in drought risk occurs under scenarios with less end of the century radiative forcing (e.g., RCP4.5 versus RCP8.5), but the difference between CMIP5 HIGH–r and LOW–r remains. CMIP5 LOW–r models yield the largest increase in CA drought risk, particularly based on soil moisture metrics (both deep and shallow). CMIP5 HIGH–r yields much weaker increases, especially based on meteorological and hydrological drought. The largest increase of CA drought risk in CMIP5 HIGH–r models continues to be with the SMT metric. However, deep soil moisture, which is most important for agricultural drought, does not significantly increase in any RCP under CMIP5 HIGH–r. California seasonal drought risk As described above, the bulk of the wettening trend in CMIP5 HIGH–r occurs during DJF. Non-winter seasons generally feature drying trends in both model subsets, which tend to be larger and more robust in CMIP5 LOW–r. This shift in seasonal precipitation implies a shortening of the wet season, which could stress water resources and ecosystems, irrespective of changes in the annual mean. Figure 8 shows seasonal changes in CA drought risk for surface water availability (P–E), precipitation (P), runoff, shallow (SMT) and deep soil moisture (SMB). Consistent with the relatively large increase in DJF CA precipitation in CMIP5 HIGH–r models, DJF CA drought risk is generally reduced. CMIP5 LOW–r models, however, still yield a significant increase in DJF CA drought for most metrics. During SON, models yield a significant increase in CA drought risk, including CMIP5 HIGH–r models (except based on SMB). CMIP5 LOW–r models yield the largest increase in SON CA drought risk, particularly based on hydrologic (runoff) and agricultural (SMT and SMB) metrics. During MAM, models also generally yield a significant increase in drought. All CMIP5 models show that this season experiences the largest increase in meteorological drought (P and P–E), particularly in CMIP5 LOW–r models. CMIP5 HIGH–r also yields significant increases in MAM drought based on P, P–E and SMT, but less than that in other models. MAM drought in CMIP5 HIGH–r models based on both runoff and SMB, however, yield non-significant increases. During summer (JJA), meteorological drought is generally muted, except in CMIP5 LOW–r models based on precipitation. However, models show a significant increase in hydrological and agricultural drought, which is again largest in CMIP5 LOW–r models. Consistent with the other seasons, CMIP5 HIGH–r models yield non-significant increases in SMB. Fig. 8 Coupled Model Intercomparison Project version 5 RCP8.5 enhanced seasonal California drought risk. Ensemble mean a December–January–February (DJF); b March–April–May (MAM); c June–July–August (JJA); and d September–October–November (SON) California drought risk [%] based on precipitation minus evapotranspiration (P–E), precipitation (P), runoff, shallow (0–0.25 m) soil moisture (SMT) and deep (1.5–3 m) soil moisture (SMB). Colored bars show the ensemble mean enhanced drought risk based on all CMIP5 models (blue; All), the model subset that yields a detrended DJF Niño 3.4 sea surface temperature versus California precipitation correlation of at least 0.30 (green; HIGH–r), and the model subset that yields a corresponding correlation less than 0.20 (red; LOW–r). Bars with a thick black outline indicates significance at the 95% confidence level, based on a standard t -test Full size image Thus, although the increase in annual and multi-annual CA drought is muted in CMIP5 HIGH–r models, significant increases still occur in the dry season, particularly MAM and SON for most metrics. These results imply continued concern is warranted regarding future CA drought risk. The wintertime increase in precipitation and other hydrological variables in CMIP5 HIGH–r models implies adaptation measures focused on smoothing seasonal differences for affected agricultural, terrestrial, and aquatic systems, as well as effectively capturing enhanced winter runoff, to alleviate possible drought increases during the dry season. Discussion Building off previous work, 27 CMIP5 models that better simulate the observed correlation between CA precipitation and El Niño interannual variability yield larger and more consistent increases in California precipitation, surface water availability and runoff under global warming, with relatively small changes in soil moisture (particularly SMB). This, in turn, translates into negligible change in the risk of annual and multi-annual CA drought under warming. In contrast, models that poorly simulate the observed El Niño–CA precipitation teleconnection yield significant enhancement of drought risk, for all five drought metrics, for CA and each of the three subregions. This is consistent with projected reductions in surface water availability, runoff, and soil moisture. Similar conclusions are obtained when additional criteria are used to subsample the models, including the ability to simulate CA precipitation climatologies, and the observed sensitivity of CA precipitation to El Niño variations. With CMIP5 HIGH–r, the lone exception is the shallow soil moisture estimate of drought, which is projected to significantly increase. This is consistent with the strong sensitivity of drying of the upper soil to enhanced evaporation due to warming, as opposed to changes in precipitation. However, in the context of agricultural drought, the deep soil moisture is most important, since the deeper layer controls the moisture availability for most California plants and non-vegetable crops. Although this model subset suggests a muted risk of enhanced annual and multi-annual CA drought, it does suggest an increased risk of flooding. This is due to several factors, including an increase in precipitation, as well as an increase in the proportion of precipitation falling as rain, as opposed to snow. Consistently, runoff is projected to increase. These changes are generally largest in central and northern CA, and muted in southern CA. On a seasonal basis, CMIP5 HIGH–r yields significant increases in CA drought risk for several metrics during the dry season. This is related to shortening of the wet season, which could stress water resources and ecosystems, irrespective of changes in the annual mean. Furthermore, both model subsets project large and robust decreases in Sierra snowpack. Thus, continued concern is warranted regarding future CA drought risk. However, the increase in precipitation and other hydrological variables in CMIP5 HIGH–r models during the winter implies adaptation measures focused on smoothing seasonal differences for affected agricultural, terrestrial, and aquatic systems, as well as effectively capturing enhanced winter runoff. Models possess uncertainties, including shortcomings in their land surface modeling 47 , 58 and a lack of soil moisture simulation assessments, 13 which could impact the conclusions based on soil moisture metrics. Additional model uncertainties include possible overestimation of tropical convection, 59 which could impact the El Niño-like tropical/extratropical dynamical response that is important for producing the increase in CA precipitation, and muted drought risk in CMIP5 HIGH–r models. Furthermore, CMIP5 models may be deficient in their ability to simulate the tropical Pacific SST response to warming. 60 , 61 , 62 Although the lack of significant warming in the central/eastern tropical Pacific over the last few decades may support such a deficiency, it is also possible these recent tropical Pacific SST anomalies are driven by coupled interactions in the tropical Pacific. 63 Moreover, there are fundamental arguments that support the CMIP5 tropical Pacific SST projections. The tropical overturning circulation is expected to weaken due to thermodynamical constraints–tropical precipitation increases at a slower rate than water vapor, so the tropical overturning circulation, including the Walker circulation and the equatorial easterly trade winds, slow down. The Bjerknes feedback, a positive feedback between trade wind intensity and the zonal SST gradient, implies that the above changes would lead to a reduced zonal SST gradient (i.e., relative warming of the central/eastern tropical Pacific). Nonetheless, if central/eastern tropical Pacific SSTs do not warm in response to increasing greenhouse gases as projected by the CMIP5 models, CA may experience larger increases in drought risk than reported here. Methods Trend and correlation significance Ensemble mean trend significance is based on a standard t -test, accounting for the influence of serial correlation by using the effective sample size, n (1 − r 1 )(1 + r 1 ) −1 , where n is the number of years and r 1 is the lag-1 autocorrelation coefficient. Significance of enhanced drought risk is based on t -test for the difference of means. Significance of correlations is also based on a t -test with N – 2 degrees of freedom, with the t -statistic equal to r / [(1 − r 2 )/( N − 2)] 0.5 . N is the sample size (e.g., number of years) and r is the correlation. Detrended correlations are estimated by first detrending the corresponding time series, and then calculating the correlation. Drought definition A drought year is defined as one in which a metric (e.g., P–E, SMB) falls below a 0.5 standard deviation ( σ ) departure from the 1950–1999 climatological mean; the change in drought risk is defined as the difference in the probability of drought from 2050–2099 relative to 1950–1999. “Extreme” drought requires a −1 σ departure from the 1950–1999 mean. Multi-annual drought is defined similarly, but requires three consecutive −0.5 σ departures (−1 σ for multi-annual extreme drought) from the 1950–1999 climatological mean. Probabilities are calculated for each model, and then averaged to form the ensemble mean. Significance of the change in drought risk is based on a standard t -test. Data availability The data and code that support the findings of this study are available from the corresponding author upon reasonable request.

So here's the good news: Despite fears to the contrary, California isn't facing a year-round drought in our warming new world. However, UC Riverside Earth Sciences Professor Robert Allen's research indicates that what precipitation the state does get will be pretty much limited to the winter months—think deluge-type rainfall rather than snow—and non-winter months will be even dryer than usual, with little or no rain at all. "It is good news," Allen said. "But only relative to the alternative of no rain at all." Allen's latest findings build on his 2017 research that concluded global warming will bring increased winter precipitation to California through the end of this century. The findings are outlined in a paper by Allen and his co-author Ray Anderson, research soil scientist at the USDA-ARS US Salinity Lab, titled "21st century California drought risk linked to model fidelity of the El Niño teleconnection." It was published September 3 in Climate and Atmospheric Science. The paper focuses on how "greenhouse-gas-induced climate change" will affect drought conditions in the state. The findings are based on 40 climate models that were compared to actual precipitation, soil moisture, and streamflow in the state between 1950 and 2000. Historically, about 90 percent of California's rain and snow have come during the winter months of December, January, and February, Allen said, with sporadic rain scattered over the rest of the year. But now, warming surface temperatures in the tropical Pacific Ocean are expected to amplify the rainy season by sending stormy El Niño conditions over the state in the winter. Bottom line, Allen said, the flooding and mudslides that accompanied the heavy winter rains of 2017 shouldn't be considered an aberration, but potentially California's new weather norm. The trick will be finding a way to capture excess water for dry periods, he said. "It's all about smoothing the seasonable differences. If we can take advantage of the enhanced winter rainfall, we can hopefully get through the drying trends the rest of the year." Trapping that winter precipitation will be a challenge, however, especially since it's likely to come more in the form of rain than snow due to the warming climate. Historically, snow in the mountains feeds reservoirs and provides water to agriculture when it is needed in the summer, but rain will just run off unless it is captured. Allen's findings also bode ill for California's fire season. The state's new norm could mimic—or surpass—the fire season of 2017, the worst in California's history, as wet winters encourage lush spring growth that will quickly parch during the hot and dry season, becoming wildfire fuel. In fact, Allen said, these "new norm" projections aren't for a distant future. "I think it's here now, so we need to start acting as quickly as possible," he said. "Adaptation is incredibly important in response to climate change, and in this case it means enhancing our water storage capabilities, our reservoirs and dam structures, because things are going to become drier in the nonwinter months." And for ordinary citizens? This might be a great time to start investing in rain barrels. "In Southern California, it could mean having native plants in your yard because a grass yard has to be irrigated, and that's probably not the wisest use of water," he said. "It's all about living sustainably."

10.1038/s41612-018-0032-x

Biology

Animals and plants jointly coexist

Jörg Albrecht et al. Plant and animal functional diversity drive mutualistic network assembly across an elevational gradient, Nature Communications (2018). DOI: 10.1038/s41467-018-05610-w Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-018-05610-w

https://phys.org/news/2018-08-animals-jointly-coexist.html

Abstract Species’ functional traits set the blueprint for pair-wise interactions in ecological networks. Yet, it is unknown to what extent the functional diversity of plant and animal communities controls network assembly along environmental gradients in real-world ecosystems. Here we address this question with a unique dataset of mutualistic bird–fruit, bird–flower and insect–flower interaction networks and associated functional traits of 200 plant and 282 animal species sampled along broad climate and land-use gradients on Mt. Kilimanjaro. We show that plant functional diversity is mainly limited by precipitation, while animal functional diversity is primarily limited by temperature. Furthermore, shifts in plant and animal functional diversity along the elevational gradient control the niche breadth and partitioning of the respective other trophic level. These findings reveal that climatic constraints on the functional diversity of either plants or animals determine the relative importance of bottom-up and top-down control in plant–animal interaction networks. Introduction All species are involved in mutualistic and antagonistic interactions with other species 1 . Collectively these interactions between pairs of species form complex networks that structure ecological communities 2 and maintain essential ecosystem functions, such as pollination, seed dispersal or biological control 3 . Within these networks the matching of species’ traits determines whether pairs of species are able to interact and how effective their interactions are 4 . Therefore, phenotypic traits related to species’ interactions are thought to determine the realised set of interactions within species-rich interaction networks 4 . In a broader sense, the functional traits that regulate species interactions can be viewed as coexistence traits that govern niche breadth (the diversity of a species’ interaction partners) and niche partitioning (the complementary specialisation of several species on exclusive interaction partners) in complex ecological networks and determine the ecosystem functions derived from ecological communities 4 , 5 , 6 . In theory, we would expect that species’ niche breadth and the degree of niche partitioning among species, increases with functional diversity 7 , 8 . If trait matching structures plant–animal interaction networks, the functional trait spaces of plants and animals should reciprocally control the niche breadth and partitioning of species in the respective other trophic level (Fig. 1a–c ). Hence, a reduction of functional diversity in the lower trophic level is expected to cause a reduction in niche breadth and partitioning (niche contraction and convergence) in the higher trophic level and vice versa (bottom-up and top-down control, respectively; Fig. 1d ). The convergence of interaction niches also implies that a reduction of functional diversity in one trophic level may cause increased competition for mutualistic partners in the other trophic level. The functional diversity in one trophic level may thus not only constrain the interaction niches but also the functional diversity in the other trophic level through biotic filtering and competitive exclusion 9 . Fig. 1 Potential bottom-up and top-down effects of plant and animal functional diversity on the assembly of mutualistic networks. a Example of a mutualistic bird–fruit network, in which plants and animals are ordered along two corresponding size-matching trait axes according to bill and fruit size, respectively. The grey lines represent bird–fruit interactions, which are constrained by trait matching so that the small-billed bird can only consume small fruits, whereas the large-billed bird can also consume large fruits. Trait matching determines the realised interaction niches of plants and animals (represented on the trait axis of the other trophic level). b Removal of plant species P 3 causes a loss of plant functional diversity and a contraction and convergence of the birds’ interaction niches corresponding to a reduction in niche breadth (diversity of a species’ interaction partners) and niche partitioning (complementary specialisation of several species on exclusive interaction partners; red histograms A 1 and A 2 at the top). c Likewise, removal of bird species A 2 causes a loss of animal functional diversity and a contraction and convergence of plants’ interaction niches (blue histograms P 1 and P 2 at the bottom). d Consequently, a loss of functional diversity in one trophic level should cause a reciprocal reduction in niche breadth and partitioning in the other trophic level. Note that the convergence of interaction niches also implies that a reduction of functional diversity in one trophic level may cause increased competition for mutualistic partners in the other trophic level (e.g., between A 1 and A 2 in b or between P 1 and P 2 in c ) Full size image Our framework implies that bottom-up and top-down effects may simultaneously control the assembly of plant–animal interaction networks and that abiotic constraints on either plant or animal functional diversity may determine the relative importance of bottom-up or top-down control 10 . This prediction goes beyond those from biodiversity experiments, as the latter have mainly focussed on either bottom-up or top-down effects, but have not yet assessed the relative importance and context dependence of both mechanisms simultaneously 10 , 11 , 12 . Despite a general consensus on the relevance of functional diversity for species interactions and associated ecosystem functions in real-world ecosystems 13 , 14 , 15 , it is unknown to what extent shifts in plant and animal functional diversity along environmental gradients alter the structure of species interaction networks and whether network assembly is primarily bottom-up-controlled or top-down-controlled. Here, we ask whether trait matching is a general phenomenon across mutualistic networks and whether bottom-up and top-down forces simultaneously control the assembly of these networks in real-world ecosystems. To address these two questions, we recorded a unique dataset of mutualistic bird–fruit, bird–flower and insect–flower interaction networks and associated functional traits along a 3.5 km elevational gradient (872–4396 m above sea level [a.s.l.]) of near-natural and anthropogenic habitat on the southern slopes of Mt. Kilimanjaro. The dataset comprises a total of 14,728 interactions between 200 plant and 282 animal species (99 bird and 183 insect species) sampled across the three types of mutualisms. First, we tested whether trait matching generally structured the assembly of these networks by analysing multivariate associations between plant and animal functional traits. Second, we tested with a Bayesian hierarchical structural equation model to what extent changes in plant and animal functional diversity along the elevational gradient control network assembly. We find consistent evidence that trait matching determines pair-wise interactions across mutualisms. Plant functional diversity is primarily limited by precipitation, while animal functional diversity is mainly constrained by temperature. We further discover that shifts in plant functional diversity along the elevational gradient control the niche breadth and partitioning of animals in the interaction networks, while animal functional diversity determines interaction niches of plants. Therefore, our findings reveal that environmental constraints on either plant or animal functional diversity determine the relative importance of bottom-up and top-down control in plant–animal interaction networks. Results Trait-associations in plant–animal mutualistic networks To characterise the functional trait spaces of plants and animals in each of the three types of mutualistic networks, we selected traits related to size matching (matching traits), to energy provisioning and requirements (energy traits), as well as to foraging stratum and mobility (foraging traits; Fig. 2 and Methods section) 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 . We projected plants and animals in each mutualism into their multidimensional trait spaces and assessed functional relationships between plant and animal trait spaces using a combination of RLQ and fourth-corner analyses, taking into account which combinations of species pairs were observed to interact 24 , 25 , 26 (see Fig. 2 and Methods section). Fig. 2 Associations of plant and animal functional traits in different types of plant–animal mutualistic networks. Results of a combination of RLQ and fourth-corner analyses for a , b bird–fruit, c , d bird–flower, and e , f insect–flower mutualisms. We use RLQ analysis ( a , c , e ) to map the multivariate trait space of animals on the multivariate trait space of plants based on plant–animal interaction networks. The eigenvalues of the RLQ analysis in a , c , e indicate the proportion of the cross-covariance between plant and animal trait spaces explained by each RLQ axis 25 . Vectors in a , c , e depict the coefficients of plant (blue) and animal (red) traits on the first two axes of the plant and animal trait spaces from RLQ analysis. If two vectors are long and point into the same or opposite directions the absolute correlation coefficient between the corresponding traits is large. Different line types in a , c , e indicate different types of functional traits, related to size matching (continuous line), energy provisioning of plants and energy requirements of animals (long-dashed line), as well as foraging stratum and mobility (dash-dotted line). b , d , f Representation of the mutualistic networks in the multivariate trait spaces of plants and animals. Dots in b , d , f represent species scores of plants (blue) and animals (red) in their multivariate trait spaces and grey lines depict interactions between plants and animals. The size of the dots is proportional to the number of links that a species has (i.e., species degree) Full size image A null model, in which we randomised species’ identities 27 , indicated that the global associations between plant and animal trait spaces were larger than expected by chance in the bird–fruit (sum of RLQ-eigenvalues: ∑ λ i = 0.18, P < 0.01), and in the insect–flower mutualism (∑ λ i = 0.083, P < 0.05), and tended to be larger than expected in the bird–flower mutualism (∑ λ i = 0.16, P = 0.074). The marginal trend in the bird–flower mutualism is likely due to a lack of statistical power owing to the relatively small number of species (26 plant and 24 bird species, respectively) 27 . The first ordination axis explained most of the cross-covariance between plant and animal trait spaces (range across the three mutualisms: 90–99%), whereas the second axis explained only a minor proportion (range: 0.75–9.3%). In separate analyses of the two ordination axes, associations between the first axes of plant and animal trait spaces were stronger than expected by chance in all three mutualisms (bird–fruit: Pearson’s r = 0.26, P < 0.001; bird–flower: r = 0.33, P < 0.01; insect–flower: r = 0.23, P < 0.01). Associations between the second axes were generally weaker and significant only in the bird–fruit mutualism (bird–fruit: r = 0.13, P < 0.05; bird–flower: r = 0.040, P = 0.67; insect–flower: r = 0.063, P = 0.40). Across the three mutualisms, the first axes of plant and animal trait spaces were most strongly correlated with matching traits (absolute Pearson’s | r | = 0.21 ± 0.016 [mean ± s.e.m.], Moran’s test 28 : P < 0.001, see Methods) and energy traits (| r | = 0.19 ± 0.024, P < 0.001), whereas correlations with foraging traits were weaker and more variable in magnitude (| r | = 0.11 ± 0.034, P = 0.031; Supplementary Table 1 ; Supplementary Fig. 1 ). The second axes of plant and animal trait spaces were not correlated with any of the trait types (matching traits: | r | = 0.035 ± 0.0074, Moran’s test: P = 0.74; energy traits: | r | = 0.031 ± 0.013, P = 0.74; foraging traits: | r | = 0.055 ± 0.014, P = 0.23; Supplementary Table 1 ; Supplementary Fig. 1 ). Bottom-up and top-down effects of functional diversity We assessed whether shifts in plant and animal functional diversity along gradients of climate and land use drive network assembly in the three mutualisms using Bayesian hierarchical structural equation models with a stochastic variable selection procedure (Fig. 3 ; Supplementary Fig. 2 ; see Methods). The structural equation models tested for consistent direct and indirect ‘functional diversity’-mediated effects of mean annual temperature, mean annual precipitation and land use on niche breadth and partitioning of plants and animals across the three mutualisms. We quantified niche breadth as the mean effective number of partners based on the exponent of the Shannon diversity of links ( e H ), and niche partitioning as the mean standardised Kullback–Leibler distance ( d′ ) across the plants and animals in each network, respectively 29 . Fig. 3 Reciprocal bottom-up and top-down effects of functional diversity on the assembly of plant–animal mutualistic networks. a Variation in mean annual temperature (MAT [°C], red) and mean annual precipitation (MAP [mm yr −1 ], blue), as well as land use (LU; filled circles, near-natural habitats; open circles, anthropogenic habitats) along the elevational gradient of Mount Kilimanjaro, Tanzania. The Bayesian hierarchical structural equation models in b , c tested for direct and indirect ‘functional diversity’-mediated effects of mean annual temperature, mean annual precipitation and land use on ( b ) niche breadth (partner diversity, e H ) and ( c ) niche partitioning (complementary specialisation, d′ ) of plants and animals via functional diversity of plant and animal communities (functional dispersion, FD; subscripts p and a for plants and animals, respectively). The lines in a represent loess smooth functions (degree = 2, span = 0.5) fitted to the temperature and precipitation data across the elevational gradient. In b , c only paths that were supported by the Bayesian variable selection (2log e (Bayes factor) > 2) are shown (see Supplementary Table 2 ). Path colours depict bottom-up-mediated effects (blue), top-down-mediated effects (red) and direct abiotic effects (grey) on network structure. Grey double-headed arrows depict covariance terms that account for correlated errors due to common unmeasured sources of variance and due to reciprocal effects of functional diversity on the other trophic level. Path widths are proportional to standardised effect sizes. The values near the endogenous variables depict the marginal ( r 2 ) variance explained by fixed factors only, as well as the conditional ( r c 2 ) variance explained by fixed and random factors combined (see Methods for details) 70 . Sample sizes are n obs = 126 observations, n site = 53 study sites and n mutualism = 3 mutualisms Full size image Plant functional diversity was positively related to mean annual precipitation, whereas animal functional diversity increased with mean annual temperature. In line with our prediction (Fig. 1 ), we found that plant functional diversity was positively associated with niche breadth and partitioning of animals, while animal functional diversity was positively related to niche breadth and partitioning of plants (Fig. 3b, c ; Supplementary Table 2 ; Supplementary Fig. 3 ). Mean annual temperature was also directly positively associated with the niche breadth of animals, as well as with the niche partitioning of plants and animals (Fig. 3b, c ). We found no consistent direct or indirect effects of land use on functional diversity, niche breadth or partitioning of plants and animals across the three mutualisms (Fig. 3b, c ). Structural equation models including univariate functional diversity metrics based on matching, energy or foraging traits showed that the increase in plant functional diversity with mean annual precipitation was primarily driven by an increase in the variability of foraging strata (i.e., plant height), while the increase in animal functional diversity with mean annual temperature was primarily driven by an increase in the variability of matching traits (i.e., bill width, bill length and proboscis length, respectively; Supplementary Fig. 4 ). Niche breadth and partitioning of plants, as well as niche breadth of animals were related to the functional diversity of all trait types (matching, energy and foraging traits), whereas niche partitioning of animals was mainly related to functional diversity of foraging strata in plant communities. In addition to the multivariate analysis, univariate models indicated an increase in the variability of size-related energy traits in response to land use (i.e., body mass in the bird mutualisms and head width in the insect mutualism; Supplementary Fig. 4 ). Discussion Our study provides a general assessment of the importance of trait matching and functional diversity for the assembly of mutualistic networks. We show that matching of species’ functional traits is a general mechanism regulating interactions in mutualistic networks. Importantly, we discover that plant and animal functional diversity are related to distinct climatic factors and constrain the realised niche breadth and partitioning of the respective other trophic level. Hence, our study reveals that environmental constraints on either plant or animal functional diversity drive the relative importance of bottom-up and top-down effects on mutualistic network assembly. We found that functional traits related to size matching were strongly associated with network structure across mutualisms, because size matching imposes critical barriers that either directly prevent interactions between plants and animals or strongly constrain their effectiveness 17 , 30 . Traits related to energy provisioning and requirements were also closely associated with network structure, which can be explained by optimal foraging theory 31 , because larger animals with higher energetic requirements should prefer spatially clustered resources (i.e., plants with a high resource density) to reduce the energetic costs of foraging 32 . Yet, the direction of the relationships between traits related to energy requirements of animals and resource provisioning by plants differed between bird and insect mutualisms (Fig. 2a, c, e ): while larger birds tended to forage on plants with higher resource density, larger insects foraged on plants with lower resource density. In the insect–flower mutualism matching and energy traits were positively correlated in animals, but negatively correlated in plants (Fig. 2e ). These contrasting trait associations in plants and animals suggest that trade-offs with matching traits might alter associations between energy traits in plant–animal mutualisms. In addition, associations between traits related to energy requirements and provisioning could be more strongly affected by the environmental context, if high competition at low resource availability forces animals to forage less selectively 33 , 34 . The more variable associations of traits related to foraging stratum and animal mobility with network structure, may be due to the fact that animals frequently cross habitat boundaries to obtain resources 35 , even though certain bird 16 , 24 and insect 18 , 19 , 20 , 21 , 22 , 23 species preferentially forage in specific habitats or vegetation layers for flowers and fruits. These results suggest that traits related to size matching between plants and animals, as well as traits related to resource provisioning and energy requirements, have the strongest effect on the assembly of pair-wise interactions and that a few trait dimensions may be sufficient to characterise the functional structure of mutualistic networks 36 . In line with our prediction, changes in the functional diversity of plant and animal communities were the principal drivers of changes in network structure along the studied elevational gradient. More specifically, we discovered that biotic interactions are simultaneously bottom-up and top-down controlled, as the functional diversity of plants and animals limited the niche breadth and partitioning of the respective other trophic level. The fact that the niche breadth and partitioning of plants, as well as niche breadth of animals were related to the functional diversity of matching, energy and foraging traits, suggests that multiple assembly processes determine these niche properties 37 . In contrast, the niche partitioning of animals was mainly related to the functional diversity of foraging strata in plant communities (i.e., variability in plant height). This finding supports the idea that vertical stratification in tropical forest ecosystems fosters niche partitioning among different functional guilds in mutualistic networks 16 . Overall, these results question the general prevalence of bottom-up effects of producer diversity on consumers, as suggested by biodiversity experiments 12 . Our results rather suggest that the prevalence of bottom-up and top-down control depends on the abiotic context 10 . Thus, our study demonstrates the value of trait-based network approaches for gaining mechanistic insights into how abiotic constraints on community assembly affect biotic interactions in real-world ecosystems 15 . We found that the functional diversity of plant communities increased with precipitation, while the functional diversity of animal communities increased with temperature. Both the dependence of plant diversity on water availability and the dependence of animal diversity on ambient temperature are well documented patterns in the literature 38 , 39 , 40 , 41 , 42 , 43 , 44 . In our study, changes in plant functional diversity were mainly related to an increase in the variability of plant height with precipitation, which may indicate that abiotic filtering in arid ecosystems constrains the range of plant growth forms, whereas in wet ecosystems increased competition (e.g., for light) leads to vertical stratification in plant communities 37 , 45 . On the other hand, changes in animal functional diversity were primarily related to an increase in the variability of matching traits with temperature, which may indicate that energetic constraints in cold ecosystems favour less selective foraging strategies and less specialised morphologies of traits that are related to resource use 46 . In general, the relationships between plant and animal functional diversity and climatic factors suggest that the niche space of these communities expands and the abundance of functionally distinct species increases under favourable climatic conditions that allow for a wider range of functional strategies 37 , 45 , 46 , 47 , which is consistent with the ‘physiological tolerance hypothesis’ in plants 39 and the ‘energy constraints hypothesis’ in animals 46 . The fact that plant and animal functional diversity were related to distinct environmental factors indicates that the responses of both trophic levels to changes in abiotic conditions could be decoupled. Despite these potentially decoupled responses of plants and animals to abiotic changes, the functional diversity of each trophic level reciprocally constrained interactions with the other trophic level via bottom-up and top-down control. According to our findings, the extent to which biotic interactions are bottom-up- or top-down-controlled depends on whether abiotic factors primarily limit plant or animal functional diversity. In particular, our results suggest that top-down effects are more limiting in cold and wet ecosystems (at mid and high elevations), whereas bottom-up effects limit network assembly in hot and arid ecosystems (at the mountain base). Interestingly, this pattern resembles the latitudinal shift in the relative importance of precipitation and temperature constraints on plant and animal diversity 38 : While water availability is the principal factor limiting plant and animal diversity in warm tropical and subtropical ecosystems, temperature is more limiting in cold temperate and boreal ecosystems. This suggests that the latitudinal shift in the importance of these abiotic factors for plant and animal diversity might be partly mediated by the relative importance of bottom-up and top-down control in different environments. This context-dependence of the prevailing mechanism is also likely to be an important driver of the observed variation in the effects of global change on biotic interactions and associated ecosystem functions 48 . In addition to the indirect effect of temperature through animal functional diversity, temperature was also directly negatively associated with niche breadth and overlap in the networks. This direct temperature effect may be explained by changes in resource availability and/or foraging behaviour along the temperature gradient 33 , 34 , 40 , 49 . Lower resource availability and increased competition, as well as higher energetic costs of foraging in cold environments, may force animals to forage less selectively on plants and might cause an increase in niche breadth 33 , 49 and overlap 34 at low temperatures. We found no consistent effects of land use on multivariate functional diversity of plants or animals across the three mutualisms. However, we detected an increase in functional diversity of traits related to energy requirements of animals in anthropogenic compared with near-natural habitats. Previous studies reported a higher availability of plant resources (flowers and fruits) in anthropogenic compared with near-natural habitats on Mount Kilimanjaro 45 , 50 . Increased resource availability in anthropogenic habitats may release animal communities from energetic constraints and might allow for a wider range of metabolic niches 51 and an increased variability in animal body sizes 46 , 50 . Plant functional diversity may be less related to effects of land use, because many of the anthropogenic habitats on Mount Kilimanjaro host a high plant diversity (e.g. the traditional homegarden agroforestry systems in the lower montane forest belt) 52 , 53 . Here, we integrated multiple types of mutualistic interaction networks with functional traits to assess the effects of trait matching and functional diversity on the assembly of species-rich plant–animal networks. Our study demonstrates that trait matching is a key determinant of network assembly and that the relative importance of bottom-up and top-down control in mutualistic networks is determined by whether environmental conditions limit the functional diversity of resources or consumers. This has important implications for the response of interaction networks and associated ecosystem functions to environmental change. As species have to adapt to their environment, but also depend on interactions with other species, trait matching and the functional diversity of interaction partners constrain species’ responses to environmental change. Trade-offs between functional adaptations to environmental conditions and biotic interactions may cause non-equilibrium dynamics, in which species are far from their optimum in terms of environmental conditions or interaction partners 54 . This trade-off between abiotic and biotic constraints might limit the adaptive capacity of species to environmental change. Therefore, projections of biodiversity and ecosystem functions in response to abiotic changes are inaccurate if they do not account for the manifold interactions among species in ecological communities. Methods Study area The study was conducted on the southern and south-eastern slopes of Mt. Kilimanjaro (Tanzania, East Africa; 2°45′–3°25′S, 37°00′–37°43′E). Mt. Kilimanjaro rises from the savannah plains at an elevation of 700 m a.s.l. to a snow-capped summit at an elevation of 5895 m a.s.l. Precipitation is bimodal with the main rainy season occurring from March through May and more variable short rains around November. The mean annual temperature decreases almost linearly with elevation, whereas mean annual precipitation peaks at an elevation of ~2200 m a.s.l. (Fig. 3a ) 52 , 53 . Due to a long history of human settlement, natural habitats in the lowlands have been subject to various forms of human disturbance including fire, logging, and agroforestry practices 52 , 53 . Habitats above 2700 m a.s.l. are protected as a national park since 1973 (Mt. Kilimanjaro National Park); since 2006 also the forests above 1800 m a.s.l. are included in the National Park. Study design We collected data on a total of 53 study sites along five transects on the southern slopes of the mountain (minimum pair-wise distance of 300 m). The study sites cover six near-natural and six anthropogenic habitat types: savanna ( n = 5) and maize fields ( n = 4; 870–1150 m a.s.l.); lower montane forest ( n = 5), traditional homegarden agroforestry systems ( n = 5), coffee plantations ( n = 5) and grassland ( n = 7; 1150–2050 m a.s.l.); natural ( n = 5) and disturbed Ocotea forest ( n = 4; 2150–2750 m a.s.l.); natural ( n = 5) and disturbed Podocarpus forest ( n = 3; 2750 – 3000 m a.s.l.); Erica forest ( n = 2; 3950–4000 m a.s.l.), as well as alpine Helichrysum vegetation ( n = 3; 3850–4400 m a.s.l.). A detailed description of vegetation and land-use types on Mount Kilimanjaro is given by Hemp 52 , 53 . No statistical methods were used to predetermine sample size. Temperature and precipitation data All study sites were equipped with temperature sensors that were installed ~2 m above the ground. Temperature sensors measured temperatures in 5 min intervals for a time period of ~2 years. We calculated the mean annual temperature (MAT, °C) as the average of all measurements per study site. Mean annual precipitation (MAP, mm yr −1 ) was interpolated across the study area using a co-kriging approach based on a 15-year data set from a network of about 70 rain gauges on Mt. Kilimanjaro 53 . As we did not have data of MAT and MAP for one study site, we predicted these data using a linear model with the observed MAT and MAP data as the response variables and elevation (third-order polynomial) and habitat type as additive explanatory variables (MAT: R 2 = 0.99, n = 52, P < 0.0001; MAP: R 2 = 0.98, n = 52, P < 0.0001; Fig. 3a ). Plant–animal interactions We studied bird–fruit and bird–flower interactions on 52 of the 53 study sites between November 2013 and October 2015. To do so, we established one plot of 30 × 100 m 2 size on each site, covering a representative amount of the flowering and fruiting plant community typical for each habitat type. Each site was sampled once, but replicate sites in each habitat type were sampled both in the cold and in the warm dry season to account for seasonal variability 55 . We observed birds using binoculars to record interactions with fruiting and flowering plants. We identified birds using Zimmerman et al. 56 . On each site, birds were observed for 25 h in total, distributed over 4 consecutive days 55 . Observations were conducted for 7 h (1–5 h after sunrise, 2 h before sunset) on the first 3 days and for 4 h on the last day (1–4 h after sunrise). We recorded the number of visits of each bird species on each fruiting or flowering plant species, respectively, and recorded their behaviour. In the analysis we considered only visits that were classified as legitimate seed dispersal or pollination events, i.e., swallowing or carrying away of fruits from mother plants, as well as pollen uptake. In total, we conducted 1300 h of bird observations during the study period, during which we recorded 9194 bird–fruit interactions between 68 plant and 86 bird species, as well as 3124 bird–flower interactions between 30 plant and 28 bird species, respectively. We had to restrict our analysis to a subset of 8085 interactions between 63 plant and 85 bird species on 39 study sites in the bird–fruit mutualism and to a subset of 2583 interactions between 26 plant and 24 bird species on 20 study sites in the bird–flower mutualism for which we were able to obtain trait data. We studied insect–flower interactions on 19 of the 53 study sites between January 2011 and November 2012. To do so, we established one plot of 100 × 100 m 2 size on each site, covering a representative amount of the flowering plant community typical for each habitat type. If possible, we sampled each site several times to account for seasonal variability between the cold and warm dry seasons. In the analysis, we accounted for the repeated sampling on the study sites by including study site as a random factor (see section Bayesian hierarchical structural equation model). During each sampling round on a given study site, we conducted a 4-h transect walk. In case of rain, strong wind or dense fog, transect walks were interrupted and continued later on that day or on the next day. During each transect walk, we moved slowly through the vegetation of each study site and recorded each interaction in which an insect touched the reproductive part of a plant species (herbaceous plants and bushes up to 9 m height). Thus, we assumed that all flower-visiting insects contribute to pollination. We collected specimens of each plant species for identification in the lab. Whenever possible, flower-visiting insects were caught with sweep nets. Thereby, we restricted our sampling to major groups of pollinators (including Apiformes, the paraphyletic group of non-bee aculeates, Symphyta and Syrphidae). We excluded other Diptera from the analyses, for which identification below family level was not feasible. We identified the caught insect specimens to the lowest taxonomic level possible. We identified 79% of all flower-visiting insects to species level and 97% to genus level (for simplicity, we refer to all morphospecies as species). In total, we conducted 320 h of insect observations during the study period, during which we recorded 4236 insect–flower interactions between 149 plant and 188 insect species. We had to restrict our analysis to a subset of 4060 interactions between 131 plant and 183 insect species on 19 study sites in the insect–flower mutualism for which we were able to obtain trait data. Plant and animal functional traits We quantified functional traits of plant and animal species that are known to structure the mutualistic interactions between these species groups via trait matching 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 . Thereby, we distinguished between three types of traits: traits related to size matching and the effectiveness of interactions (matching traits), traits related to resource provisioning by plants and energy requirements of animals (energy traits), and traits related to foraging stratum and animal mobility (foraging traits). Traits related to size matching were fruit diameter and bill width in the bird–fruit mutualism, corolla depth and bill length in the bird–flower mutualism, and corolla depth and proboscis length in the insect–flower mutualism 17 , 24 . Traits related to resource provisioning and energy requirements were crop mass and body mass in the bird–fruit mutualism, the number of flowers per plant and body mass in the bird–flower mutualism, and the number of flowers per inflorescence and head width in the insect–flower mutualism. Traits related to foraging stratum and mobility were plant height and the Kipp’s index 57 (i.e., the ratio between Kipp’s distance and wing length as a measure of the pointedness of the wing) in the bird–fruit and bird–flower mutualisms, as well as plant height and the forewing index (i.e., the ratio of forewing length to body length) in the insect–flower mutualism 21 , 22 , 23 . We selected head width instead of body length as a proxy for energy requirements in the insect–flower mutualism in order to avoid using the same morphological variable two times in the analysis. As body length and head width were highly correlated (Pearson’s r between log e -transformed variables: r = 0.84, n = 91 species), the decision of whether to include body length or head width did not affect our conclusions. For fruiting plants, we measured the maximal diameter of 15 fruits for each plant species (five fruits each from three different individuals) using sliding callipers (precision of 0.01 mm) and weighed the fruits using a digital scale (precision 0.01 g). Moreover, we estimated the total number of ripe fruits per plant individual for each species on the study sites. On plants with very large crop sizes, we counted the number of fruits for representative branches and used these to estimate the crop size of the whole plant. We calculated crop mass by multiplying the mean fruit mass by the mean number of fruits per individual for each plant species. We measured plant height for each species on the study sites using a laser range finder (precision 1 m). For flowering plants, we measured corolla depth using sliding callipers (precision 0.01 mm) on herbarium specimens. For the bird–flower mutualism, we estimated the total number of flowers per plant individual for each species on the study sites and measured plant height using a laser range finder (precision 1 m). For the insect–flower mutualism, we counted the number of flowers per inflorescence and measured plant height on herbarium specimens or compiled data about plant height from the literature if only parts of plant specimen were available in herbarium samples 58 . When no trait information was available for a plant species, we used the trait values from closely related species in the same genus (bird–flower mutualism: n = 4 out of 31 plant species; insect–flower mutualism: n = 7 out of 136 plant species). For fruit-eating and flower-visiting birds, we measured bill width, bill length, Kipp’s distance and wing length using sliding callipers (precision 0.01 mm) on museum specimens following Eck et al. 57 and extracted information about body mass from the literature 59 , 60 . Measurements were taken on an average of four specimens per species (range = 1–16). We measured bill length as the distance from the commissural point of the upper and lower bill to the tip of the closed bill, and bill width as the external distance between the two commissural points. We calculated Kipp’s index as the ratio of Kipp’s distance (distance between tip of the first secondary and tip of the longest primary of the folded wing) and wing length. For flower-visiting insects, we measured proboscis length (Hymenoptera: length of glossa; Diptera: length of labellum) 61 , forewing length, body length, head width and intertegular distance (for aculeate bees) of collected specimens using a binocular microscope with a calibrated ocular micrometre (precision 0.01 mm). When no trait information was available for an insect species, we used the mean trait values from related species in the same genus ( n = 11 cases), from species in the same family ( n = 12 cases), or from species in the same order ( n = 2 cases). Excluding the insect species for which we only had trait information at the genus, family or order level from the analyses led to identical conclusions. For 75 bee species, we estimated body length based on the relationship with intertegular distance (log e (body length) = 1.89 + 0.518 × log e (intertegular distance), r 2 = 0.48, n = 49, P < 0.0001). For 10 hymenoptera species (families: Pompilidae, n = 3; Tiphiidae, n = 2; and Vespidae, n = 5), we estimated proboscis length based on the relationship with head width (log e (proboscis length) = −1.32 + 1.54 × log e (head width), r 2 = 0.47, n = 112, P < 0.0001). We calculated the forewing index as the ratio of forewing length and body length. We square-root-transformed all plant and animal traits for the bird–fruit, the bird–flower and the insect–flower mutualisms before the analysis. Trait associations in mutualistic networks To test for functional relationships between the multidimensional trait spaces of plants and animals in the mutualistic networks, we adopted a combination of the RLQ and the fourth-corner analyses 24 , 25 , 26 . The RLQ analysis builds on the simultaneous ordination of three tables: a table R ( m × p ) describing p traits for m plant species, a table Q ( n × a ) describing a traits for n animal species, and a third table L ( m × n ) containing qualitative or quantitative information about the occurrence or frequency of pair-wise interactions between the m plant and n animal species. Here we defined L as a binary matrix based on whether an interaction between a pair of plant and animal species had been observed at least once across all study sites in each mutualism. Therefore, table L is analogous to a metaweb that describes the potential for pair-wise interactions between plants and animals based on all available information from the observations. We first applied correspondence analysis (CA) to table L and principal components analyses (PCA) to tables R and Q 25 . In the PCAs of R and Q each plant and animal species was weighted by its number of links in L (i.e., species degree) 62 . Then, we combined the three separate ordinations of R , L , and Q using the RLQ approach to identify the main relationships between plant and animal trait spaces as mediated by their interactions 25 . The RLQ analysis is based on a p × a matrix Ω containing measures of the intensity of the links between plant and animal traits 63 . The further eigendecomposition of Ω T Ω allows identifying the main associations between plant and animal traits 26 , 64 . For the first dimension, this analysis finds a vector u 1 containing coefficients for the plant traits and a vector v 1 of coefficients for the animal traits. These loadings measure the contributions of individual traits and are used to compute scores for plant ( x 1 = RD p u 1 ) and animal species ( y 1 = QD a v 1 ) where D p and D a are diagonal matrices of variable weights (here species degree). RLQ chooses the coefficient vectors u 1 and v 1 so that the derived species scores have maximum squared cross-covariance cov P ( x 1 , y 1 ) 2 = ( x 1 T Py 1 ) 2 = λ 1 , where λ 1 is the first RLQ eigenvalue. In other words, RLQ finds linear combinations of plant and animal traits (i.e., trait syndromes) so that their squared cross-covariance is maximum. The same quantity is maximised for the k dimensions with the additional constraints of orthogonality ( u i T D p u j = v i T D a v j = 0 for i ≠ j ). Results are stored in matrices U = [ u 1 |…| u k ], V = [ v 1 |…| v k ], X = RD p U = [ x 1 |…| x k ] and Y = QD a V = [ y 1 |…| y k ]. We used the eigenvalues of the RLQ analysis to select those ordination axes that explained most of the cross-covariance between plant and animal trait spaces for our analyses 25 . We selected the first two RLQ axes, because in all three mutualisms these axes together explained more than 99% of the cross-covariance between plant and animal trait spaces (bird–fruit: 90% and 9.3%; bird–flower: 99% and 0.75%; insect–flower: 94.8% and 5.1%, Fig. 2a, c, e ). Then, we applied the fourth-corner permutation test (model 6 with 9999 permutations) 25 to evaluate the statistical significance of the associations between plant and animal trait spaces using three different approaches. First, we used the sum of eigenvalues of the RLQ analysis as a multivariate statistic to test for global associations among plant and animal trait spaces 25 , 26 . Second, we tested for correlations between the first and second dimensions of plant and animal trait spaces by using the RLQ scores on the first two axes of the plant and animal trait spaces as variables in the fourth-corner analysis 25 . Third, we assessed which of the different trait types (matching, energy and foraging traits) were correlated with each of the first two RLQ axes 25 . Because we aimed at generalising our results across mutualisms and because the direction of effects in multivariate space is arbitrary, we compared the absolute magnitude of correlations between matching, energy and foraging traits and the RLQ axes across the three mutualisms. Moreover, instead of assessing the significance of individual correlations, we assessed the overall support for the hypotheses that matching, energy and foraging traits are related to the first and second RLQ axes across the three mutualisms. To do so, we used the equation given by Moran 28 , based on a Bernoulli process, to calculate the probability, P , of obtaining a given number of significant tests from a given number of trials just by chance. This probability is given by the equation P = [ N !/( N − K )! K !] × α K (1 − α ) N − K , where N is the number of tests conducted and K is the number of tests below the significance level α (here α = 0.05). The rationale behind this equation is that the evidence against the null hypothesis from a given number of statistical tests increases with the number of significant tests 28 . For instance, four out of six matching traits were significantly correlated with the first RLQ axes at α = 0.05 (Supplementary Table 1 ). According to Moran’s equation the probability of obtaining four significant results out of six tests at α = 0.05 by chance is P = 8.5 × 10 −5 , providing strong support for the hypothesis that matching traits structure pair-wise interactions in mutualistic networks. Previous work has shown that the sequential permutation approach for statistical testing of the ‘fourth-corner problem’ (model 6) has good power (0.88) for 100 species, reasonable power (0.60) for 50 species and some power (0.40) for 30 species 27 . As the number of plant and animal species in two of our datasets is relatively small, the statistical power of the permutation test is low. For instance, although the magnitude of the absolute correlations between matching traits and RLQ axis 1 in the bird–flower mutualism is similar to those in the bird–fruit and insect–flower mutualisms, the permutation test is not significant at α = 0.05 (Supplementary Table 1 ; Supplementary Fig. 1 ). Therefore, the statistical tests we conducted can be considered highly conservative with respect to Type I errors. Functional diversity effects on network assembly We used functional dispersion to measure the functional diversity of plants and animals within the mutualistic networks (FD p and FD a hereafter) 65 . Compared to other measures of functional diversity functional dispersion has several advantages: Functional dispersion is abundance-weighted and therefore less influenced by extreme values and it is by definition unaffected by species richness 65 . We calculated multivariate and univariate functional dispersion of plant and animal communities in each mutualism. To do so, we first calculated distance matrices based on the Gower distance between species based on the combination of all three functional traits (multivariate FD) or based on each trait type separately (univariate FD based on matching, energy or foraging traits). The Gower distance equals the mean character difference across traits after standardisation of the trait values by their ranges and has been recommended for calculation of functional diversity metrics based on multiple traits, because it is less sensitive to extreme trait values than the Euclidean distance 66 . Moreover, the standardisation of the trait values by their ranges yields an empirical maximum value of the distance function that equals one 66 , which allows for a meaningful comparison among multiple species groups with different sets of functional traits. Then we projected species into a multidimensional functional trait space using a principal coordinates analysis (PCoA), and calculated functional dispersion as the mean abundance-weighted distance of each species in a given community to the abundance-weighted centroid of all species in this community 65 . To estimate the abundance of plant and animal species in the networks, we used their marginal interaction totals in each network. Functional dispersion was only weakly correlated with species richness of plants and animals in our data (absolute Pearson’s | r | < 0.41 for plants and | r | < 0.33 for animals across multivariate and univariate FD metrics; n = 126 in all cases). To quantify the interaction niches of plants and animals in the networks, we used two different measures of niche breadth and niche complementarity. We quantified niche breadth of plants and animals as the mean effective number of partners based on the exponent of the Shannon diversity of links ( \({\it {e}}^{H_{\it {\mathrm p}}}\) and \({\it {e}}^{H_{\it {\mathrm a}}}\) ), and niche partitioning as the mean standardised Kullback–Leibler distance ( d′ p and d′ a ) across the plants and animals in each network, respectively 29 . Bayesian hierarchical structural equation model To test for bottom-up and top-down effects of plant and animal functional diversity on niche breadth and partitioning, respectively, we used Bayesian hierarchical structural equation models. The general structure of the models that we used here is reviewed in the literature 67 and described in more detail in the supplementary material (Supplementary Note 1 ). We fitted two separate structural equation models, one including \({\it {e}}^{H_{\it {\mathrm p}}}\) and \({\it {e}}^{H_{\it {\mathrm a}}}\) and the other including d′ p and d′ a as measures of the interaction niches of plant and animal communities on each study site (Supplementary Fig. 2 ). In these structural equation models, we treated mean annual temperature (MAT, °C), mean annual precipitation (MAP, mm yr −1 ) and land use (LU, binary variable) as exogenous predictor variables. We treated FD p and FD a as well as measures of plant and animal interaction niches ( \({\it {e}}^{H_{\it {\mathrm p}}}\) and \({\it {e}}^{H_{\it {\mathrm a}}}\) or d′ p and d′ a , respectively) as endogenous variables. The models included all potential direct effects of MAT, MAP and LU on FD p and FD a , as well as on \({\it {e}}^{H_{\it {\mathrm p}}}\) and \({\it {e}}^{H_{\it {\mathrm a}}}\) or d′ p and d′ a , respectively. Moreover, the models included the effects of FD p and FD a on \({\it {e}}^{H_{\it {\rm p}}}\) and \({\it {e}}^{H_{\it {\rm a}}}\) or d′ p and d′ a , respectively. We also included covariance terms between FD p and FD a , as well as between \({\it {e}}^{H_{\it {\rm p}}}\) and \({\it {e}}^{H_{\it {\rm a}}}\) or d′ p and d′ a to account for correlated errors due to common unmeasured sources of variance and due to reciprocal effects of functional diversity on the other trophic level. The total number of samples included in the analysis was n = 126. To account for the hierarchical structure of the data we included study site ( n = 53) and mutualism type ( n = 3) as random factors into the structural equations. The measures \({\it {e}}^{H_{\it {\rm p}}}\) and \({\it {e}}^{H_{\it {\rm a}}}\) were transformed to their natural logarithm and all variables were scaled to zero mean and unit variance before analysis. To separate informative from non-informative paths, we used a Bayesian indicator variable selection with global adaptation 68 (Supplementary Note 1 ). We used 2log e (Bayes factor) as a measure of evidence for a given effect (BF hereafter) 69 . Values of BF < 2 indicate no support; values between 2 and 6 indicate positive support; values between 6 and 10 indicate strong support; and values >10 indicate decisive support. We report the marginal variance, r m 2 , that is explained by the fixed factors, as well as the conditional variance, r c 2 , that is explained by the fixed and random factors combined as measures of model fit 70 . The models were implemented in JAGS 71 , and run in R 72 through the rjags package 73 . We ran eight parallel chains for the models. We used uninformative priors for all parameters and the initial values for the chains were drawn randomly from uniform distributions. Each chain was run for 51,000 iterations with an adaptive burn-in phase of 1000 iterations and a thinning interval of 100 iterations, resulting in 500 samples per chain, corresponding to 4000 samples from the posterior distribution. The chains were checked for convergence, temporal autocorrelation, and effective sample size using the coda package 74 . Residuals were checked for normality and variance homogeneity. Code availability The computer code of the analyses is available in figshare with the identifier . The JAGS code for the Bayesian hierarchical structural equation model is also given as part of the Supplementary Information (Supplementary Note 1 ). Data availability The data that support the findings of this study are available in figshare with the identifier . Due to the Data and Publication Policy of the Research Unit FOR1246, the figshare data are embargoed for public release until 1 January 2020. Until the embargo date for public release the data are available from the corresponding author upon reasonable request. Change history 05 September 2018 This Article was originally published without the accompanying Peer Review File. This file is now available in the HTML version of the Article; the PDF was correct from the time of publication.

The tropical rainforest, with its permanently wet climate, brims over with an abundance of plant species. Flowers of all sizes with shallow to deep tubes offer a wide variety of food sources to pollinators. Sunbirds, for example, developed long, down-curved beaks enabling them to reach the nectar at the bottom of long tubes. But who is the leading actor in this interplay of plants and animals? "Ecosystems with high precipitation typically contain a wide variety of plants and show a high level of functional diversity. The higher this functional diversity, the more their animal partners are able to specialize on particular plant species. The plants therefore serve as the deciding factor for the animals' specialization," explains Dr. Jörg Albrecht of the Senckenberg Biodiversity and Climate Research Center, lead author of a new study regarding this topic. Albrecht and his colleagues from Germany, Switzerland and Tanzania observed and analyzed almost 15,000 interactions between plants, their pollinators and their seed dispersers over a period of five years. The study area comprised survey plots from the foothills of the Kilimanjaro up to the mountain's summit. Since 2010, an international team of scientists in the joint project "Kilimanjaro ecosystems under global change," sponsored by the German Research Foundation, has been working here to find out how climate and land use affect the biological diversity on Africa's tallest mountain. Credit: Eike Lena Neuschulz However, on Kilimanjaro, it is not always the plants that make the rules. The newly published study shows that there also exist climatic conditions in which the animals set the agenda. Albrecht explains, "Mean temperature is more important than precipitation for animal diversity. In cold climates, animal species tend to resemble each other more closely and show a lower level of functional diversity. Therefore, in cold regions, plants are not able to specialize on specific animal partners and frequently interact with the same species." Since the amount of precipitation and the mean temperature in many areas is expected to change due to climate change, researchers suspect that changes will also occur in the networks between plants and animals. "Depending on how the climate will change in a particular region, either plants or animals could drive the structure of these networks in the future. A change in the flora could thus also have an impact on the animal partners of the plants," explains the study's senior author, Dr. Matthias Schleuning of the Senckenberg Biodiversity and Climate Research Center. According to the authors, the ability of species to adapt to new environmental conditions is not only determined by the species themselves, but also depends on the responses of their plant or animal partners. Precise predictions regarding future changes in biodiversity are essential to mitigate future environmental changes caused by climate change. "In order to better predict how species and ecological communities will respond to global change, it is important to consider the environmental conditions together with the mutual interdependencies between species," says Matthias Schleuning.

10.1038/s41467-018-05610-w

Earth

Dark fiber lays groundwork for long-distance earthquake detection and groundwater mapping

Jonathan B. Ajo-Franklin et al, Distributed Acoustic Sensing Using Dark Fiber for Near-Surface Characterization and Broadband Seismic Event Detection, Scientific Reports (2019). DOI: 10.1038/s41598-018-36675-8 Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-018-36675-8

https://phys.org/news/2019-02-dark-fiber-groundwork-long-distance-earthquake.html

Abstract We present one of the first case studies demonstrating the use of distributed acoustic sensing deployed on regional unlit fiber-optic telecommunication infrastructure (dark fiber) for broadband seismic monitoring of both near-surface soil properties and earthquake seismology. We recorded 7 months of passive seismic data on a 27 km section of dark fiber stretching from West Sacramento, CA to Woodland, CA, densely sampled at 2 m spacing. This dataset was processed to extract surface wave velocity information using ambient noise interferometry techniques; the resulting V S profiles were used to map both shallow structural profiles and groundwater depth, thus demonstrating that basin-scale variations in hydrological state could be resolved using this technique. The same array was utilized for detection of regional and teleseismic earthquakes and evaluated for long period response using records from the M8.1 Chiapas, Mexico 2017, Sep 8th event. The combination of these two sets of observations conclusively demonstrates that regionally extensive fiber-optic networks can effectively be utilized for a host of geoscience observation tasks at a combination of scale and resolution previously inaccessible. Introduction Hydrogeologic and seismological data collection are two domains for which the absence of high spatio-temporal resolution data is particularly acute, with significant impacts on our ability to characterize near-surface soil properties, groundwater systems, and seismic events. Even relatively basic subsurface hydrological parameters such as water table depths in surficial aquifers suffer from severe undersampling in both space and time. While heavily monitored basins often have a multitude of wells providing subsurface access, they are neither uniformly distributed nor frequently monitored resulting in heterogeneous datasets requiring manual quality control, curation, and analysis. The few basin-wide hydrogeological data sources, typically based on satellite remote sensing technologies, provide only surficial property estimates like soil moisture 1 , 2 , integrated strain response (e.g. InSAR 3 ), or low-resolution volumetric datasets (e.g. GRACE 4 , 5 ) which require assimilation with point measurements to provide finely resolved operational parameters 6 . Remote sensing measurements often also suffer from temporal undersampling due to satellite pass frequency. More recently, seismic ambient noise interferometry 7 has been leveraged to provide broader information on ground water storage; unfortunately limited permanent seismic networks present a challenge for these approaches. Likewise, seismological data collected using existing permanent networks often have spatial regions which suffer from significant spatial undersampling, particularly in areas distant from major plate boundaries, resulting in challenges when attempting to detect and locate small natural and induced events. The case of small magnitude induced events is particularly problematic since the basins where oil and gas production, wastewater injection, and carbon dioxide sequestration occur are often distant from historically seismogenic faults and the associated permanent seismic networks. Network sparsity increases the minimum event size for detection, results in statistical biases in the catalog, and greatly increases depth uncertainty for local events. Recent studies focusing on seismic catalog completeness in California have determined that even M2 events cannot be detected in the majority of the Sacramento and San Joaquin Basins using the existing network stations 8 . These spatio-temporal undersampling problems, for both hydrological and seismological measurements, can be remedied by re-purposing ubiquitous sensing platforms already deployed at scale. A recent example of such an approach is the utilization of smartphone accelerometers to measure strong ground motion as part of earthquake early warning applications as shown in previous studies 9 , 10 ; other examples include using social media proxies as sensors 11 or MEMS accelerometers in pervasive stationary devices such as personal computers 12 . Broader efforts to leverage networking and sensor technologies related to the Internet-of-Things (IoT) for seismology are developing but still in their infancy 13 . An alternative approach is to exploit components of the built environment to serve as distributed sensor networks. In this case we couple the use of unlit subsurface fiber-optic cables, commonly referred to as “dark fiber” since they are not utilized for data transmission, and distributed acoustic sensing (DAS) to provide such a spatially extensive sensing platform. The vast majority of fiber-optic cables in the earth’s near-surface were installed exclusively for the purpose of telecommunications. Due to high cost of fiber-optic installation, typical commercial practice is to deploy significantly more capacity, as measured by fiber count, than required; this practice, combined with advances in bandwidth available per fiber, have yielded a surplus of available fibers that remain unused. The US footprint of such unused fiber networks is massive with tens of thousands of linear kilometers of long distance fiber-optic cables available for lease or purchase in the current environment. One notable aspect of such dark fiber network components is that they tend to utilize existing “right-of-way” corridors along roads and rail connections 14 , environments rich in ambient seismic noise. Given the ubiquitous nature of installed telecom fibers, few studies have explored use of this resource for sensing applications. An early experiment explored the use of Brillouin Optical Time Domain Analysis (BOTDA) to monitor temperature over previously installed telecom fiber 15 ; however, these studies were conducted primarily to provide network integrity information rather than for environmental sensing. In a seismological context, several recent studies 16 , 17 , 18 have demonstrated the benefits of leveraging urban telecom infrastructure at a small scale. Most recently, a study in southwest Iceland 19 provided an excellent example of utilizing telecom fiber for detecting local earthquakes and measuring co-seismic strain measurement over a short transect. Distributed Acoustic Sensing (DAS) is a recently developed technique which utilizes coherent optical time-domain reflectometry to accurately measure the phase and amplitude of vibrations along an optical fiber 20 , 21 , 22 . The technique exploits changes in Rayleigh scattering induced by extensional strain; these measurements have now been quantitatively compared to point seismic recordings at both intermediate 23 and low frequencies 24 and utilized for a host of tasks including vertical seismic profiling 20 , 21 , 25 , near-surface soil property estimation 26 , 27 , 28 , 29 , 30 , surface refraction tomography 31 and earthquake seismology 16 , 17 , 18 , 19 , 32 . DAS has created a recent paradigm shift in applied geophysics by enabling seismic measurements at a combination of high frequency (kHz range), large distances (tens of km), and fine spatial sampling (as small as 1 m), a combination previously unavailable with conventional sensors at moderate costs. We should note that DAS is distinct from long-range optical interferometry approaches which provide even greater measurement distances but sacrifice spatial localization; this class of techniques was recently demonstrated 33 as an approach for seismic detection utilizing trans-oceanic cables. While prior studies have convincingly demonstrated the value of dense networks for seismic imaging as well as a range of other purposes 34 , the high costs associated with massive nodal deployments over long time periods has precluded their use in many contexts. In this study, we demonstrate the application of DAS utilizing dark fiber for measurement of seismic wavefields at the sub-basin scale with an extremely fine spatial sampling (2 m) over long time periods; the resulting ultra-dense dataset is utilized for both hydrogeological/near-surface characterization, using ambient noise interferometry, and the detection of seismic events, both regional and global. This combination is perhaps a new frontier which leverages investment in built infrastructure to greatly extend the reach and sampling of existing permanent monitoring networks. Seismic Monitoring with Dark Fiber Networks Our study utilized dark fiber components of ESnet’s Dark Fiber Testbed. ESnet, a US Department of Energy (DOE) user facility, provides high-performance unclassified network infrastructure to connect DOE research sites including high performance computing (HPC) facilities and data-intensive instrumentation e.g. x-ray, neutron, and nanoscience facilities. The Dark Fiber Testbed is a 20,920 km (13,000 mile) network of short and long haul telecommunication fiber designed for testing novel network communication equipment and protocols. The network consists of single mode telecommunication fibers of varying age and installation technologies and hence is an excellent proxy for existing commercial network components. This study is one of the first experiments that utilizes this massive network for sensing purposes. Figure 1A , depicts the long haul regional sections of the Dark Fiber Testbed in California, as well as the segment exploited for our test (1B), which runs from West Sacramento, CA to Woodland, CA. Figure 1 Map of a section of the ESNet Dark Fiber Testbed ( ). ( A ) The regional network within CA and western NV; zone of panel (B) shown in black dashed box. ( B ) The subsection of the network used in this study. The red segment in ( B ) is the area of focus for ambient noise analysis; W1 and W2 are reference wells for water table and soil horizons, respectively. The study fiber (blue) is aproximately co-linear with an active rail line. Dashed green line labeled I-5 is Interstate 5, a major source of ambient noise beyond the rail corridor. Full size image Field Deployment The site of our study was a transect located in the Sacramento River flood plain, north and west of Sacramento, CA. The geology of the site consists largely of Quaternary sediments including a sequence of silts and clays underlain by fine sands. Prior regional studies 35 have mapped the surface sediments as a mixture of poorly sorted Holocene alluvium near the Sacramento and finer-grained Holocene basin deposits deeper in the flood plain. Partially lithified sediments from the Tehama formation have been mapped from approximately 50 m to greater depths 36 . The segment of dark fiber we utilized for this study, shown in Fig. 1 panel B in blue, runs from West Sacramento CA to the small town of Woodland CA. As can be seen from the fiber network map, the recording profile extends from an urban environment into a section of farmland near the Sacramento River, crossing Interstate 5 before bending westward towards Woodland. For the length of the fiber route shown in Fig. 1B , installed cables utilize the right-of-way associated with a rail line and are roughly co-linear with the train tracks. The agricultural areas sampled by this profile are partially irrigated through a variety of methods and groundwater is actively extracted from both the shallow surficial aquifer as well as deeper sources. The seismic dataset presented in this study was recorded between July 28th, 2017 and Jan. 18th, 2018. The DAS interrogation unit (IU; Silixa iDAS, Elstree, UK) was installed in a telecommunication Point-of-Presence (PoP) facility in West Sacramento. Hardware details on the installation are described in Methods Section 0.2. Ambient seismic noise was recorded using the DAS IU at 500 Hz sampling with a spatial sampling of 2 m; the gauge length was 10 m and fixed in hardware. Data was streamed continuously to large capacity (8 TB) USB-3 external hard drives that were exchanged on a weekly basis. While the surface geometry of the dark fiber network was known before deployment, the mapping to linear fiber location was established by sequential impact tests at surface locations surveyed with high accuracy differential GPS. This is necessary due to the common practice of including spools of slack cable during telecom installation, an approach that makes the mapping of surface geometry to linear fiber location more complicated. Impact locations were observed on individual DAS gathers in terms of fiber distance, coordinated by GPS time; by establishing true DGPS coordinates for these locations, we were able to compensate for slack effects when mapping back to the previously surveyed deployment geometry. The resulting geometry likely has an uncertainty on the order of 5 m due to the interpolation process along the transect. After the network geometry was established, DAS signal strength, amplitude, and periodicity were examined to evaluate noise characteristics along the array. Dominant noise features include several regional highways, diffuse urban noise, and energy from local railroad activity. Qualitatively, the highest quality data was observed on a straight section of fiber starting beyond West Sacramento and extending to the noise field of Interstate 5, shown as the highlighted red profile in Fig. 1B . Zones to the Southeast of this section suffered from non-optimal installation conditions (e.g. the fiber was attached to surface structures including a bridge) as well as incoherent noise in the urban transition zone around West Sacramento. Zones to the North and West suffered from both optical fading, insufficient return photons which decreased measured S/N, and broadside noise interference from Interstate 5. This illustrates the potential heterogeneity of signal quality across the existing telecom network. High Resolution Kilometer-Scale Near-Surface Imaging Using Ambient Noise Prior studies have demonstrated the potential of utilizing ambient noise interferometry and classical sensors to detect fluctuation in groundwater state 7 . Ambient noise interferometry is an established group of seismological techniques that utilize environmental vibrations from either near-surface or subsurface sources to retrieve coherent seismic information, referred to as empirical Green’s functions 37 , 38 , 39 , 40 , 41 . More recently, several authors 26 , 28 have demonstrated that such techniques could be utilized to transform infrastructure noise in the 2–30 Hz band (e.g. surface waves generated by cars, trucks, and trains) into accurate and stable 1-D estimates of shear wave velocity from 0–30 m depth using DAS. While prior studies have examined data acquired on a fit-for-purpose array, the processing strategy proposed can be easily adapted to dark fiber deployments. The present study images near-surface shear wave structure using infrastructure noise generated by freight trains operating along a 6600 m subsection of the dark fiber array (bold red line in Fig. 1B ). Similar to earlier work which exploited road noise, we use ambient noise interferometry to transform the raw noise records into virtual common-shot gathers; we then apply multichannel analysis of surface waves (MASW) analysis 42 , 43 , 44 , 45 to infer shear-wave velocity ( V S ) profiles with a multimodal inversion strategy. Figure 2 is an illustration of this workflow (see Fig. S1 for more details). All examples use only 40 minutes of ambient noise data with the caveat that only noise from trains are utilized; a more detailed discussion of data selection is detailed in the Methodology section. After generation of dispersion images from virtual common-shot gathers, we auto-pick dispersion curves for high energy modes. These experimental curves are inverted using a previously developed Monte Carlo (MC) inversion strategy 46 which utilizes a novel multimodal objective function 47 which does not require mode numbering. Supplementary Fig. S2 shows a detailed example of solutions across the array and associated fits to experimental curves. Of the resulting solutions, the best-fitting model is selected for interpretation; however, the family of accepted models can be used to evaluate solution uncertainty as discussed in the Supplementary section and shown in Fig. S3 . Figure 2 Illustration of data processing workflow for ambient noise interferometry. ( a ) Example of train noise shown via an 8 second time domain slice. The red box in ( a ) highlights subsection of the array used for ( b ) noise correlation gather, ( c ) dispersion analysis, and ( d ) inversion of the shear-wave velocity ( V S ) profiles. Black and white markers in ( c ) denote observed and model-predicted multimodal dispersion curves respectively. In ( d ), the yellow dashed lines denote upper and lower bounds of the parameter space used in Monte Carlo sampling; the bold red line marks the best-fit V S profile; the yellow/blue lines denote the top 0.1% best-fitting V S profiles (color coded by their corresponding inversion misfits); Misfit † denotes normalized misfit values (min-max normalized by misfits of the top 0.1% V S profiles). Full size image The results of our processing flow are a series of V S profiles, each representing 1-D approximations of the subsurface underneath each of the 120-meter-long fiber subsections (Fig. 3 ). This length was selected to provide sufficient array size for dispersion analysis yet some degree of lateral spatial resolution 42 . As is to be expected, the installation conditions of the dark fiber are not always ideal for V S imaging, hence gaps are left in the pseudo 2-D section shown in Fig. 3b . Reasons for the gaps include a small section with localized strongly directional coherent noise, likely a pump in one case, as well as poorly-structured dispersion curves which could not be fit within reasonable tolerance. However, with 57% of the 6600-meter-long fiber transect providing useful data (inverted sections in Fig. 3b ), our lateral coverage of 3760 meters warrants its place as one of the longer high-resolution MASW profiles obtained using ambient noise alone. For all offsets shown, the forward modeling based on the inversion results effectively predicts the picked dispersion curves as can be seen in Fig. 2 as well as Fig. S2 for the fundamental mode and 1st overtone which can be reliably observed. We should note that the S-wave velocity of the lowest interface (underlying half space) is poorly constrained due to the limited offset, minimal recovered energy at low frequencies, and biases in the autopicker under these conditions; autopicker behavior within our present implementation tends to bias this velocity high without manual intervention. Zones above 40 m are considerably better constrained as discussed in the Supplementary section 2 . Figure 3 Shear-wave velocity ( V s ) inversion results and ground truth comparison. ( a ) Depths of groundwater levels (GWL) (upper) and V s 30 estimates (lower) extracted from the surface wave inversion results. ( b ) Pseudocolor display of V s profiles and comparisons against well data. In ( b ), W1 and W2 mark surface locations of the two reference wells shown in Fig. 1 ; the blue dashed line denotes the depth of the ground water level provided by water well W1; the black dashed lines correspond to lithological horizons obtained from geotechnical well W2. Full size image Comparisons to our limited ground truth database reveal that the interfaces of the V S profiles match reasonably with the groundwater level and deeper lithological horizons, which confirms the validity of the inversion results. The lithological horizons consist of transitions from near-surface silty clay units to deeper sand/gravel units and underlying clay-rich horizons, as inferred from drilling logs from an unmonitored local well (W2) as shown in Fig. 3b . The depth to the surficial aquifer is estimated from the thickness of the top low velocity layer recovered from the surface wave inversion; in Fig. 3a , these estimates are compared to the water table depths measured in well W1, located slightly to the southeast of the measurement section. Details of well W1 are included in the Supplementary Information. More detailed analysis of a sub-array in the central section of the profile (see Fig. S3 ) show that the depth uncertainty for the top interface at that location is approximately 4.62 ± 0.8 m , hence the measured water table matches within error. Also shown in Fig. 3a is an estimate of V S 30 , the travel-time average of V s over the top 30 meters, calculated directly from the surface wave inversion results. V S 30 is a widely used indicator of seismic site conditions and is readily obtainable from surface wave inversion 48 ; both the values and the lateral variations of V S 30 are useful data products that the dark fiber array can provide and can be compared to local estimates made as part of geotechnical studies. To evaluate the utility of the DAS ambient noise inversion results for hydrogeological monitoring, we compare the transition in V s identified as the water table to direct point measurements in the one well with sufficient temporal sampling information located on our transect. The well (W1) is located at the southeastern end of the high S/N section of dark fiber near the intersection of County Road 126 and the Old River Road, shown as a red dot on Fig. 1B . Water level in this well varies by as much as 7.6 m (25 ft) over annual hydrologic cycles and is impacted by Sacramento River levels, agricultural irrigation and surficial aquifer pumping. As can be seen in Fig. 3a , the depth of the first V s increase (3.8–4.6 m below the surface) correlates to the measured surficial water table depth. We should note that an increase in V s at the water table is not predicted by classical rock physics models which assume that shear modulus is not sensitive to saturation; despite this, past field studies 26 , 49 have provided detailed observations of such sensitivity, presumably due to the impact of moisture on soil cohesion for dry surficial materials. Time-Lapse Monitoring of Groundwater Level As mentioned previously, the Sacramento Basin is a dynamic hydrologic environment with multiple productive aquifers, active groundwater production, irrigation, and river interactions. Besides static profiles quantifying V s , the dark fiber array also allows for time-lapse monitoring of the subsurface, enabling measurement of hydrologic transients in the near-surface. Traditionally, because sensors are rarely dense enough for imaging, time-lapse monitoring results are typically presented as an apparent-velocity perturbation along the path connecting two sensors, without identifying where on this path the changes are occurring. This level of detail is insufficient for proactive management of groundwater resources, given that both vertical and lateral changes of the groundwater level could affect the overall sustainability of the aquifer. With the dark fiber profile enabling time-lapse imaging of the near surface, changes in groundwater levels could be resolved with spatial-temporal resolutions pertinent to groundwater management. Figure 4 demonstrates the repeatability of time-lapse imaging chiefly by comparing the groundwater levels retrieved from a monitoring well against what were obtained from the closest dark-fiber section. Note that no major rainfall events had occurred during the three-month monitoring period processed for this study, and as a result, the maximum changes in the measured water table did not exceed 0.9 m, on the order of our current error estimates for determination of surficial layer thickness (see Supplementary section, Table 1 ). The high repeatability of interferometric gathers, both in the offset-time domain (Fig. 4a ) and frequency-velocity domain (Fig. 4b ) suggests that such levels of variation will not be resolvable using this approach. The high time domain repeatability for single offsets are also shown in the Supplementary Materials (Fig. S4 ). An inspection in the model space (Fig. 4c ) confirms the inversion’s insensitivity to these sub-meter changes. However, such repeatability provides assurance for reliably resolving larger changes in groundwater levels, because the well data, when examined over a year-long period, exhibit groundwater level changes up to 8 m. Such marked changes should be resolvable with the dark fiber array with a more extended monitoring period. Figure 4 Time-lapse repeatability demonstration of ambient noise analysis in ( a ) space-time domain, ( b ) frequency-velocity domain, and ( c ) in terms of groundwater levels obtained from the model domain. Color sequence of red, green, and blue denote chronological orders of the monitoring period. In ( c ), the median, min-max range, and percentiles are calculated based upon all the topmost best-fitting models associated with the monitoring period. Full size image Earthquake Seismology with a Dark Fiber DAS Array Seismic network detection thresholds are highly heterogeneous, even across regions known for dense seismic monitoring like the western United States and Japan 50 , 51 , in part because broadband seismic stations are sited in hard-rock locations where background noise is low 52 , 53 . Areas of less-competent geology, like sedimentary basins, therefore correlate with poor catalog completeness; the magnitude of completeness is Mc 2–3 in the Sacramento and Southern San Joaquin Basins compared to Mc 0.5–2.4 in the San Francisco Bay Area or Mc 0.5–1.8 in Southern California 8 , 54 , 55 . Thus, despite the greater Sacramento area hosting significant gas production, underground gas storage, and high-volume waste water disposal, all of which can impact seismicity, the Sacramento Dark Fiber DAS array is located 30 km away from the nearest networked short-period seismometer (NDH) and 62 km away from the nearest broadband seismometer (AFD). A relevant question when examining seismic events on telecommunications networks in contrast to fit-for-purpose installations is the impact of installation conditions. As recently demonstrated 16 , 17 , 19 , fiber installation in a standard plastic conduit does not preclude sufficient sensor coupling required for the detection of earthquakes, but the case of recording DAS data with repurposed telecommunications fiber is yet untested at regional scales. To explore this question with the Sacramento Dark Fiber DAS experiment we extract raw strain-rate waveforms for major global and regional earthquakes that occurred during the continuous recording interval (Fig. 5 ). We again use the linear quiet portion of the array shown in Fig. 1B and process the data by averaging 100 seismic traces (200 m section) and applying a bandpass filter to isolate the appropriate earthquake signals. To plot the raw strain-rate data in a more familiar unit we multiplied the data by a reference length equal to the gauge length (10 m) to convert to a unit that is proportional to velocity. We observed broadband DAS sensitivity to ground motion from earthquakes of varying magnitudes (M4.4–M8.1) and distances (100–7757 km). For example, in the case of the M7.5 Honduras event there is clear evidence of short period body waves and longer period surface waves over the two hour window following the origin. Figure 5 Example earthquakes recorded by the Sacramento Dark Fiber DAS array. The recorded data are plotted as strain-rate after multiplying by the gauge length (10 m) to convert to units proportional to velocity (1e-6 m/s), and have been averaged over 100 m of linear fiber length (50 traces) and then bandpass filtered in the 0.1–0.4 Hz range for regional events, and 0.01–0.1 Hz for teleseisms. Events are sorted by increasing epicentral distance from Sacramento. Earthquake amplitudes for the Peru and Honduras events are scaled by the factors in parentheses. Full size image While long period sensitivity is a major limitation of many inertial seismic sensors (e.g., accelerometers, short-period geophones, smartphone sensors), the long period response of DAS is currently a topic of active research with only limited available data 24 ; teleseismic earth motion (strains near 1 × 10 −8 ), for example, may be dominated by thermal expansion of the fiber-optic cable (strains on the order of 1 × 10 −6 ) depending on the frequency studied as well as the depth, composition, and condition of the fiber-optic cable and conduit. Recent studies 24 , 56 have used shallow hydrogeologic pump tests in a well with a fiber-optic cable to show that DAS has sensitivity to 9.4 × 10 −3 Hz (period = 1080 seconds) oscillations in strain induced by the variable confining pressure, presumably due to Poisson effects. This subject is complicated by the known directionality of DAS cables 57 , 58 , which for the horizontal geometry of telecommunications dark fiber cables is theoretically insensitive to vertically-incident compressional motion (P-waves). To explore the long period sensitivity of the Dark Fiber DAS array to teleseismic events, we extract raw strain-rate seismograms from the largest earthquake recorded during the experiment, the M8.1 2017-Sep-08 Chiapas, Mexico earthquake (Fig. 6 ). We observe broadband dispersive surface waves with strong energy at periods from 50–100 seconds. P-wave signal amplitude is lower than S-wave amplitude, perhaps because the sensor has minimal sensitivity to compressional particle motions for waves with incidence angles approaching 0° with respect to vertical (i.e. perpendicular to the fiber “broadside arrivals”). Nonetheless, the arrival times of major seismic phases are detected because of free surface scattering. Incidence angles of seismic phases are given in Fig. 6a . Differences in incidence angle also likely affect recorded amplitude, and appear to result in more coherent surface wave arrivals across the array (incidence angle = 0°) and as much as a 0.5 second delay time across the record section. Figure 6 Teleseismic DAS recording of the M8.1 Chiapas, Mexico 2017-Sep-08 earthquake. ( a ) Seismic data for [black trace] one location and [red and blue] all locations from 0.0–7.6 km at a 2 m spacing (4001 traces total); top right inset shows surface waves arriving at the [black] south and [pink] north end locations of the array (backazimuth 120°), bottom left inset shows body waves arriving coincidently at both locations. A two-corner, zerophase, f = 0.01–0.5 Hz bandpass filter was applied. ( b ) Stacking 400 m or 200 consecutive DAS channels, color-coded by the bandpass filter applied to emphasize the broadband observation (1–100 seconds). Gray background traces show the single trace recording for cases that make a significant difference. Each of the traces is normalized to peak amplitude. Full size image Telecommunications cables are commonly routed along railways, roads, and through high noise urban areas. We find that major regional earthquakes ( M ≈ 4) generate ground motions on the Sacramento array that have equal or lesser amplitude than local moving vehicles, however anthropogenic seismic signals typically are dominant in a higher frequency band (5–30 Hz). Figure 7 shows how the seismic signals from two different regional earthquakes (M4.22 Geysers 2018-Jan-18 and M4.38 Berkeley 2018-Jan-04) are easily discriminated from the local noise field based on spectral content. Higher frequencies of interest for local microearthquake analysis (f ≤ 50 Hz) will not always separate in this way. Figure 7 ( a ) Locations and focal mechanisms of the M4.2 2018-Jan-18 Geysers (red) and M4.4 2018-Jan-04 Berkeley (blue) earthquakes, which occurred approximately 100 km from the Sacramento Dark Fiber DAS array (black line). ( b , c ) Raw and lowpass filtered DAS strain-rate waveforms for these events averaged over 100 m (50 channels) at the yellow circle position shown in ( a ) (channel 4975 +/− 50 channels). Note the similarity between seismic and non-seismic signal amplitudes and the differences in frequency content. Full size image The two earthquake records shown in Fig. 7 appear very different despite having been generated in similar sized ruptures and traveled similar distances to Sacramento. This may be due to source rupture depth differences (z = 2.4 km for The Geysers, z = 12 km for Berkeley), or the major differences in geologic structure along the raypath, but could also be the result of strong DAS axial sensitivity to energy in the direction of the fiber axis. We observe larger recorded amplitudes for shear waves from Berkeley than from The Geysers, but larger recorded amplitude for Rayleigh waves from The Geysers because of the more favorably oriented polarization. We hypothesize that the method of installation (direct-burial, single conduit, conduit inside a larger conduit, conduit attached to infrastructure) has a significant effect on DAS recorded ground motion (see Fig. 8 ). The fiber-optic cable itself (gel-filled, aramid wrapped vs. loose-tube, polyethelene-jacketed vs. steel-armored, polyethelene vs. steel exterior) has each been shown to have only a small effect on recording quality at high frequencies 26 . Figure 8 ( a ) Illustration of different installation geometries. ( b ) Earthquake (M4.2 Geysers 2018-Jan-18) trace comparison for each installation mode at Sacramento – trenched conduit (green), cased conduit (blue), attached conduit (red); strain-rate data are stacked over 100 m and filtered (BP 0.5–2 Hz n 4 p 2). ( c ) Normalized Fourier amplitude spectra for the waveforms shown in b. Full size image Installation information for the Sacramento Dark Fiber DAS array provides clues as to the heterogeneity of fiber-soil coupling across our experimental profile. Cable installation occurred in 1999–2000. Most of the fiber was pulled through one of 12 high-density polyethelene (HDPE) conduits (ID = 3.5–4 cm, wall thickness = 0.5 cm) that were buried together in a trench at 1–1.5 m and backfilled with soil before installing the fiber cable inside. Each fiber cable contains 84 gel-filled, loose-tube Corning LEAF fibers that are polythelene jacketed and steel-armored. The DAS data were recorded using a single 9/125 μ m single-mode fiber from one of these cables. In a few locations, trenching was not possible so directional boring was used to install a large casing conduit (ID = 20–25 cm, wall thickness = 0.4 cm), inside of which the 12 smaller conduits were pulled. Depth of boring varied between one meter and a few meters when navigating around various culverts, sections of road and railway, and other obstacles. In some instances the casing was not required, or a steel casing may have been used. A third mode of installation used for approximately 300 m of the dark fiber array involved attaching a 20–25 cm diameter steel casing directly to the elevated rail line where it crosses a section of protected wetlands, the Sacramento Bypass Wildlife Area. Inside this attached conduit, the 12 HDPE conduits were installed as the boring method described above. Figure 8 shows DAS strain-rate earthquake waveforms (BP 0.5–2 Hz n 4 p 2) and normalized Fourier amplitude spectra for the M4.2 Geysers 2018-Jan-18 event stacked over 100 m of each of the three install modes. Any phase shifts between traces are due to these install locations being separated by as much as 7 km along the array. The conduit and cased conduit data show very similar seismic wave response to the ground motion centered in the f = 0.1–10 Hz range. Seismic signal amplitudes are observed to be on order with the optical noise at f ≥ 100 Hz. Data from attached section are noisier in a narrow frequency band centered on 12 Hz ± 3 Hz, perhaps caused by interaction of the incident seismic energy with the infrastructure and/or tube waves traveling in the attached conduit at air velocity. The trenched conduit shows a broader spectral response to near-surface scattering into surface waves, while the cased conduit is relatively insensitive to it. We should note that three installation conditions discussed in this study are certainly not a comprehensive survey. A large variety of techniques are used for fiber installation, ranging from direct cable burial to installation on utility poles; the impact on DAS recording for many have yet to be evaluated. Discussion While the focus of our study was the specific utilization of installed telecom fiber probed by DAS for seismic sensing, our static imaging and monitoring results are consistent with and rely on broad advances in the field of ambient noise seismology applied to the near-surface. Beyond the foundational studies cited previously 38 , a variety of recent projects have utilized ambient noise approaches to probe hydrologic cycles 7 , 59 , 60 , 61 and aquifer structure 62 although typically using a sparse network of stations. Such studies have typically relied on the microseisms band as a noise source (0.1–1 Hz) and hence are observing averaged velocity perturbations over significant vertical extent, often to km depths. Despite this, such approaches have yielded convincing correlations with environmental parameters such as groundwater level 59 although often for larger perturbations (10 s of m perturbations) in comparison to our study. In contrast, our experiment is densely sampled in space and utilizes infrastructure noise in the 0.5–18 Hz range; this enables adoption of surface wave inversion approaches utilized in the geotechnical community 46 , 47 , 63 and the monitoring of very shallow features with some degree of depth resolution. Unfortunately, the minimal precipitation and small perturbations (1 m) in groundwater level which occurred during our study period precluded effective observation of hydrologic variations. Despite this fact, the static structure observed was consistent with our available ground truth dataset. As mentioned previously, the ambient noise aspect of our study was also greatly enhanced by the broadband seismic signal generated by rail traffic co-linear to our measurement fiber. Recent seismic interferometry examples from the 2014 Belen experiment 64 which utilized a more classical sensor array (4.5 Hz geophones) confirms the utility of rail noise as an imaging source for both P as well as S-wave imaging. These observations bode well for future dark fiber exploitation; since many telecom fiber installations use railway right-of-ways for installation, judicious selection of fiber paths can exploit this powerful source of seismic energy. Our experiment also provide an unusual case study of large (12k+ channel) array recording with dense receiver spacing (2 m). The only comparably sized experiments (5000+ sensors) are massive nodal deployments, which have recently been leveraged for similar event detection 34 , 65 and ambient noise 66 , 67 applications. Outside of the validation datasets presented previously, prior studies in close proximity to our site are not extensive. Geotechnical evaluations, conducted in support of levee evaluation and construction broadly agree with our conclusions. At a location near our fiber profile, a past study evaluating levee safety in West Sacramento 68 reported an average V s 30 of 234 m/s which is broadly consistent with our surface wave inversion values of (210–280 m/s), a NEHRP class D soil environment. As demonstrated in the prior sections, DAS-based seismic measurements acquired using dark fiber can provide a wealth of information relevant to near-surface seismic property estimation, hydrologic state, and natural seismicity. Measurements using dark fiber also have advantages in a host of situations ranging from marine to urban scenarios where classical seismic networks are challenging to execute. An obvious strength of dark fiber DAS deployments, demonstrated in this study, is the potential to record data across long (10 s of km) transects at high spatial resolution without any required sensor installation or power source. Chains of such deployments could be utilized to provide true basin scale sensing; the Sacramento Basin’s central width (120 km) could be spanned using only 4 independent interrogation units and existing dark fiber resources, providing an unprecedented sensing resource. While basin scale hydrogeophysical monitoring studies using point sensors and ambient noise have been recently conducted 7 , the spatial resolution of such investigation is typically on the order of km due to sparse sensor distribution. Recent advances in large N processing approaches 32 also offer strategies for leveraging dense arrays for detecting small seismic events. A second advantage of utilizing dark fiber for seismic measurements is dense non-invasive coverage in urban areas where diverse deployment and permitting environments challenge classical acquisition strategies. The present study provides coverage spanning urban (Sacramento), suburban, and rural zones without the typical landowner permission effort, permitting, and survey work required for deployment in occupied areas. While not demonstrated in this study, dark fiber can also be utilized for observations in the transition zone and offshore domains, areas where almost no measurements exist at present due to the high cost of tethered marine observatories including seismometers. Offshore DAS measurements would be particularly useful for improving the hypocenter accuracy for small events occurring on marine faults and earthquake early warning in subduction zones. Limitations of using DAS and dark fiber for offshore observations include (a) distance constraints for DAS measurements on single mode fiber for existing optical chains, currently in the range of 30–40 km, (b) the considerably higher cost for dedicated use of fibers in transoceanic cables, (c) the lower density of offshore cable routes which reduces coverage. Having said this, offshore dark fiber recording provides a clear future opportunity to extend the domain of seismological measurements into previously uninstrumented regions. Despite these opportunities, challenges exist to fully exploit these conceptually novel sensing networks. Extremely large data volumes are among the most pressing, although solvable, problems; at maximum acquisition rates, a single interrogator can generate upwards of 20 TB/day. Combining these large N deployments with the long time periods required for ambient noise processing and monitoring yields raw data volumes that exceed the capacity of the computational infrastructure available to most researchers. The array in this study, which included 12000 channels sampled at 500 Hz, generated 128 TB of raw data in the first 3 months of operation and approximately 0.3 PB of raw data when the system was demobilized. Volumes of this size require careful consideration of data transport requirements, storage, archiving, and automated processing to be successfully utilized. Fortunately, on-going efforts to solve I/O and computational barriers in ambient noise studies 69 , 70 provide a path to potentially handle the much larger datasets generated by dark fiber studies. Methods DAS system installation The Silixa iDAS unit utilized in this study was installed on a vibration isolated table located in the West Sacramento PoP. Vibration isolation consisted of a 18 × 24′′ Nexus Breadboard (Thorlabs) with passive Sorbothane feet (Thorlabs) placed on a durometer 70 Sorbothane sheet. This assembly was placed on a rack shelf within the PoP cage where the utilized dark fiber was terminated in a standard fiber-optic patch panel with SC-UPC connection. Connection to the Silixa iDAS was made using an SC-UPC/SC-APC single-mode patch cable. An optical time-domain reflectometer (OTDR) was used to evaluate fiber integrity prior to recording. An OTDR trace measured a total loss of 20.8 dB over the full fiber length of 101 km at 1550 nm, or an average loss of 0.2059 dB/km. Details of data collection, computing, and processing infrastructure As mentioned in the main manuscript, the data was collected at 500 Hz in the iDAS native format in the form of raw 1 minute records at 2 m spatial sampling. Files were written to a local USB3-connected 8 or 16 TB external hard drive. To maintain continuity of the dataset, hard drives were replaced on a weekly or biweekly basis and manually transferred from West Sacramento CA to Berkeley CA, where the data were uploaded to a local RAID storage server using the Globus protocol ( , last accessed: 2018-05-21) from a networked data transfer node. The storage server was linked to five RAID6 disk arrays and the full dataset was striped to improve performance. Primary processing was carried out on a 32-core GPU server connected to the storage server via a fast GB switch. Final results were visualized using a combination of MATLAB (Mathworks) and the ObsPy package. Processing Framework & Parameters for Ambient Noise Analysis As mentioned in the primary manuscript, we adopted the ambient noise surface wave processing and inversion approach detailed in a prior DAS study 26 with the steps shown in Supplemental Fig. S1 . The overall workflow was initially derived from ambient noise analysis strategies developed over the last decade in the crustal imaging community 41 . Ambient noise interferometry was performed on sequential 120 m long subsections of the array to provide a combination of sufficient spectral resolution and a useful spatial resolution short enough for 1D analysis during inversion. In each subsection, the southernmost channel was treated as the virtual source and cross-correlated with the remainder of the channels. Results shown in the ambient noise study are comprised exclusively of processing minutes when a train was passing near the selected array section; the train was the most energetic as well as broadband ambient noise source and selection of these records allowed for efficient processing. Train passes were identified for each section by scanning trace windowed RMS amplitude on the raw records. One minute records where the train was approaching or departing from a section were tagged for ambient noise processing. The train schedule was variable (2–6 passes per day) so each epoch utilized 40 minutes of train noise for analysis, independent of the number of days required to accumulate this stack. These N = 40 stacks were typically generated over approximately 10 days. After selection and to prepare the raw records for noise correlation, static offsets and linear trends were removed, followed by temporal decimation down to a coarser sampling rate of 8 ms. Next, a temporal normalization with running-absolute-mean was applied over a 0.5-second running window. The frequency content of the data between 0.5 Hz and 18 Hz was then balanced with a spectral whitening step. Finally, in each of the 1-minute records, the noise of the virtual-source channel was cross-correlated with the rest of the channels’ records to form a common virtual-source gather. To achieve good SNR with a minimal stack count, phase-weighted stacking ( ν = 0.5) was used and a stack count of 40 was sufficient for reaching temporal stability. Prior studies 71 document the utility of phase weighted stacks in ambient noise analysis. A slant stack was then applied to the stacked common virtual-shot gathers to transform the data from space-time domain to frequency-velocity (dispersion) domain. The input to all inversions were experimental multimodal dispersion curves extracted from the frequency-velocity domain images. Given the very large number of datasets generated including 100 s of sub-arrays over 10 s of epochs, automation, rather than hand-picking, was a necessity. A simple algorithm was developed to pick the dispersion curves by 2D scans for local maximums with a lower threshold. As can be seen from the example in Fig. 2c , the presence of strong modes besides the fundamental forced adoption of a multimodal inversion approach. One weakness of the auto-picking approach is handling of the low frequency ends of the dispersion curves; in these cases, the auto-picks often are biased to high velocities which can distort (increase) the velocity of the bounding 1/2 space. As discussed in prior work 26 , mode-labeling can present a challenge for this class of DAS dataset hence we adopted an inversion approach which did not require explicit mode numbering. For our objective function in the inversion, we utilized a recently developed formulation developed by Maraschini and collaborators 46 , 47 . Their novel approach, which we refer to as the Haskell-Thomson determinant method, searches for models that can minimize the determinant of a model-predicted propagator matrix whose frequency and velocity terms are replaced with the experimental dispersion curves. We refer the interested reader to prior descriptions of the algorithm 46 , 47 and a prior example of applying it to DAS data 26 . The advantage of this approach is that (a) the objective function can be very efficiently evaluated without root finding, thus allowing global search, and (b) non-labeled multimodal dispersion data can be utilized as an input. Because of the nonlinear nature of the problem and the low computational costs of the Haskell-Thomson determinant method, Monte Carlo sampling was used 46 as part of search and model selection. We adopted a sparse parameterization for the search problem assuming a four layer model for each array subsection and solved for V s and layer thickness for each location. For each 1D inversion, a Monte Carlo pool size of 1 × 10 6 models was used. Search bounds of the model parameters are shown in Table 1 . Each model parameter was assumed to be uniformly distributed in the sampling process within the bounds, with the restriction that all models were required to have increasing V s with depth, likely a valid assumption in this geology. As can be seen in Table 1 and Fig. 2d , the bounds are not tight and allow effective model exploration. All models were ranked by L1 misfit; the optimum models were selected for interpretation. Each set of the inversions took 1.5 minutes on 24 cores (2.3 GHz Intel Xeon processors). Table 1 Upper and lower bounds in Monte Carlo sampling of the inversion variables. Full size table Processing Framework & Parameters for Earthquake Analysis We identified earthquake records in the continuous raw DAS dataset using catalogued origin times and approximate travel times to Sacramento for the 1-D iasp91 Earth velocity model. 1-minute duration earthquake records were merged together, and then, if desired, stacked with a mean average of traces over a specified length (100 m = 50 traces averaged and plotted at the midpoint), prior to the application of a specified bandpass filter to remove unwanted or uninteresting signals. In Fig. 5a , the iasp91 model was again used to calculate the phase arrival angles with respect to vertical. Figure 5b shows one stacking effect. Data on Reference Wells Data on the reference groundwater monitoring well discussed (W1) was acquired from the CASGEM database. As mentioned in the text, the well is referred to as the Sac Bypass Shallow Well (State well ID 09N04E20-N001M and CASGEM ID 25619). The well is located in Yolo County at 38.6062 N, −121.5602 W with a surface elevation of 6.55 m (21.49 ft) above sea level; the well is associated with the Yolo County Water Resources Association (WRA) with measurements conducted by the CA Department of Water Resources (DWR). The well is completed with slotted PVC and a sandpack over a 6.1 m (20 ft) interval from 24.4 to 30.5 m below ground surface. Water table depths were measured manually at an irregular intervals (every 1 to 3 months) with a reported accuracy of 0.3 cm (0.01 ft). Data Availability Due to the very large size of this dataset, only decimated raw components or processed subsections are available upon request.

In traditional seismology, researchers studying how the earth moves in the moments before, during, and after an earthquake rely on sensors that cost tens of thousands of dollars to make and install underground. And because of the expense and labor involved, only a few seismic sensors have been installed throughout remote areas of California, making it hard to understand the impacts of future earthquakes as well as small earthquakes occurring on unmapped faults. Now researchers at the U.S. Department of Energy's Lawrence Berkeley National Laboratory (Berkeley Lab) have figured out a way to overcome these hurdles by turning parts of a 13,000-mile-long testbed of "dark fiber," unused fiber-optic cable owned by the DOE Energy Sciences Network (ESnet), into a highly sensitive seismic activity sensor that could potentially augment the performance of earthquake early warning systems currently being developed in the western United States. The study detailing the work—the first to employ a large regional network as an earthquake sensor—was published this week in Nature's Scientific Reports. Shaking up seismology with dark fiber According to Jonathan Ajo-Franklin, a staff scientist in Berkeley Lab's Earth and Environmental Sciences Area who led the study, there are approximately 10 million kilometers of fiber-optic cable around the world, and about 10 percent of that consists of dark fiber. The Ajo-Franklin group has been working toward this type of experiment for several years. In a 2017 study, they installed a fiber-optic cable in a shallow trench in Richmond, California, and demonstrated that a new sensing technology called distributed acoustic sensing (DAS) could be used for imaging of the shallow subsurface. DAS is a technology that measures seismic wavefields by shooting short laser pulses across the length of the fiber. In a follow-up study, they and a group of collaborators demonstrated for the first time that fiber-optic cables could be used as sensors for detecting earthquakes. The current study uses the same DAS technique, but instead of deploying their own fiber-optic cable, the researchers ran their experiments on a 20-mile segment of the 13,000-mile-long ESnet Dark Fiber Testbed that extends from West Sacramento to Woodland, California. "To further verify our results from the 2017 study, we knew we would need to run the DAS tests on an actual dark fiber network," said Ajo-Franklin, who also heads Berkeley Lab's Geophysics Department. "When Jonathan approached me about using our Dark Fiber Testbed, I didn't even know it was possible" to use a network as a sensor, said Inder Monga, Executive Director of ESnet and director of the Scientific Networking Division at Berkeley Lab. "No one had done this work before. But the possibilities were tremendous, so I said, 'Sure, let's do this!" Chris Tracy from ESnet worked closely with the researchers to figure out the logistics of implementation. Telecommunications company CenturyLink provided fiber installation information. By coupling DAS technology with dark fiber, Berkeley Lab researchers were able to detect both local and distant earthquakes, from Berkeley to Gilroy, California, to Chiapas, Mexico. Credit: Ajo-Franklin/Berkeley Lab Because the ESnet Testbed has regional coverage, the researchers were able to monitor seismic activity and environmental noise with finer detail than previous studies. "The coverage of the ESnet Dark Fiber Testbed provided us with subsurface images at a higher resolution and larger scale than would have been possible with a traditional sensor network," said co-author Verónica Rodríguez Tribaldos, a postdoctoral researcher in Ajo-Franklin's lab. "Conventional seismic networks often employ only a few dozen sensors spaced apart by several kilometers to cover an area this large, but with the ESnet Testbed and DAS, we have 10,000 sensors in a line with a two-meter spacing. This means that with just one fiber-optic cable you can gather very detailed information about soil structure over several months." Digging deep for data underground After seven months of using DAS to record data through the ESnet Dark Fiber Testbed, the researchers proved that the benefits of using a commercial fiber are manifold. "Just by listening for 40 minutes, this technology has the potential to do about 10 different things at once. We were able to pick up very low frequency waves from distant earthquakes as well as the higher frequencies generated by nearby vehicles," said Ajo-Franklin. The technology allowed the researchers to tell the difference between a car or moving train versus an earthquake, and to detect both local and distant earthquakes, from Berkeley to Gilroy to Chiapas, Mexico. The technology can also be used to characterize soil quality, provide information on aquifers, and be integrated into geotechnical studies, he added. With such a detailed picture of the subsurface, the technology has potential for use in time-lapse studies of soil properties, said Rodríguez Tribaldos. For example, in environmental monitoring, this tool could be used to detect long-term groundwater changes, the melting of permafrost, or the hydrological changes involved in landslide hazards. The current study's findings also suggest that researchers may no longer have to choose between data quality and cost. "Cell phone sensors are inexpensive and tell us when a large earthquake happens nearby, but they will not be able to record the fine vibrations of the planet," said co-author Nate Lindsey, a UC Berkeley graduate student who led the field work and earthquake analysis for the 2017 study. "In this study, we showed that inexpensive fiber-optics pick up those small ground motions with surprising quality." With 300 terabytes of raw data collected for the study, the researchers have been challenged to find ways to effectively manage and process the "fire hose" of seismic information. Ajo-Franklin expressed hope to one day build a seismology data portal that couples ESnet as a sensor and data transfer mechanism, with analysis and long-term data storage managed by Berkeley Lab's supercomputing facility, NERSC (National Energy Research Scientific Computing Center). Monga added that even though the Dark Fiber Testbed will soon be lit for the next generation of ESnet, dubbed "ESnet 6," there may be sections that could be used for seismology. "Although it was completely unexpected that ESnet—a transatlantic network dedicated for research—could be used as a seismic sensor, it fits perfectly within our mission," he said. "At ESnet, we want to enable scientific discovery unconstrained by geography."

10.1038/s41598-018-36675-8

Earth

Prediction of large earthquake probability improved

Isabel Serra et al. Deviation from power law of the global seismic moment distribution, Scientific Reports (2017). DOI: 10.1038/srep40045 Journal information: Scientific Reports

http://dx.doi.org/10.1038/srep40045

https://phys.org/news/2017-01-large-earthquake-probability.html

Abstract The distribution of seismic moment is of capital interest to evaluate earthquake hazard, in particular regarding the most extreme events. We make use of likelihood-ratio tests to compare the simple Gutenberg-Richter power-law (PL) distribution with two statistical models that incorporate an exponential tail, the so-called tapered Gutenberg-Richter (Tap) and the truncated gamma, when fitted to the global CMT earthquake catalog. Although the Tap distribution does not introduce any significant improvement of fit respect the PL, the truncated gamma does. Simulated samples of this distribution, with parameters β = 0.68 and m c = 9.15 and reshuffled in order to mimic the time occurrence of the order statistics of the empirical data, are able to explain the temporal heterogeneity of global seismicity both before and after the great Sumatra-Andaman earthquake of 2004. Introduction The Gutenberg-Richter (GR) law is not only of fundamental importance in statistical seismology 1 but also a cornerstone of non-linear geophysics 2 and complex-systems science 3 . It simply states that, for a given region, the magnitudes of earthquakes follow an exponential probability distribution. As the (scalar) seismic moment is an exponential function of magnitude, when the GR law is expressed in terms of the former variable, it translates into a power-law distribution 4 , 5 , i.e., with M seismic moment, f ( M ) its probability density, (fulfilling ), the sign “ ∝ ” denoting proportionality, and the exponent 1 + β taking values close to 1.65. This simple description provides rather good fits of available data in many cases 6 , 7 , 8 , 9 , with, remarkably, only one free parameter, β . A totally equivalent characterization of the distribution uses the survivor function (or complementary cumulative distribution), defined as for which the GR power law takes the form S ( M ) ∝ 1/ M β . The power-law distribution has important physical implications, as it suggests an origin from a critical branching process or a self-organized-critical state 3 , 10 , 11 . Nevertheless, it presents also some conceptual difficulties, due to the fact that the mean value 〈 M 〉 provided by the distribution turns out to be infinite 4 , 12 . These elementary considerations imply that the GR law cannot be naively extended to arbitrarily large values of M , and one needs to introduce additional parameters to describe the tail of the distribution, coming presumably from finite-size effects. However, a big problem is that the change from power law to a faster decay seems to take place at the highest values of M that have been observed, for which the statistics are very poor 13 . Kagan 7 has enumerated the requirements that an extension of the GR law should fulfil; in particular, he considered, among other: (i) the so called tapered (Tap) Gutenberg-Richter distribution (also called Kagan distribution 14 ), with a survivor function given by and (ii) the (left-) truncated gamma (TrG) distribution, for which the density is Note that both expressions have essentially the same functional form, but the former refers to the survivor function and the later to the density. As f ( M ) = − dS ( M )/ dM , differentiation of S tap ( M ) in (i) shows the difference between both distributions. In both cases, parameter θ represents a crossover value of seismic moment, signalling a transition from power law to exponential decay; so, θ gives the scale of the finite-size effects on the seismic moment. The corresponding value of (moment) magnitude (sometimes called corner magnitude) can be obtained from , when the seismic moment is measured in N · m 15 , 16 . Kagan 7 also argues that available seismic catalogs do not allow the reliable estimation of θ , except in the global case (or for large subsets of this case), in particular, he recommends the use of the centroid moment tensor (CMT) catalog 17 , 18 . From his analysis of global seismicity, and comparing the values of the likelihoods, Kagan 7 concludes that the tapered GR distribution gives a slightly better fit than the truncated gamma distribution, for which in addition the estimation procedure is more involving. In any case, the β −value seems to be universal (at variance with θ ), see also refs 9 , 19 and 20 . Nevertheless, the data analyzed by Kagan 7 , from 1977 to 1999, comprises a period of relatively low global seismic activity, with no event above magnitude 8.5; in contrast, the period 1950–1965 witnessed 7 of such events 21 . Starting with the great Sumatra-Andaman earthquake of 2004, and following since then with 5 more earthquakes with m ≥ 8.5 (up to the time of submitting this article), the current period seems to correspond to the past higher levels of activity. Main et al . 22 and Bell et al . 23 have re-examined the problem of the seismic moment distribution including recent global data (shallow events only). Using a Bayesian information criterion (BIC), Bell et al . 23 compare the plain GR power law with the tapered GR distribution, and conclude that, although the tapered GR gives a significantly better fit before the 2004 Sumatra event, the occurrence of this changes the balance of the BIC statistics, making the GR power law more suitable; that is, the power law is more parsimonious, or simply, is enough for describing global shallow seismicity when the recent mega-earthquakes are included in the data. Similar results have been published in ref. 24 . In the present paper we revisit the problem with more recent data, including also the truncated gamma distribution, using other statistical tools, and reaching somewhat different conclusions: when data includes periods of high seismic activity, indeed, the tapered GR distribution does not introduce any significant improvement with respect to the power law 23 , but the truncated gamma does. Data, Models and Maximum Likelihood Estimation As Main et al . 22 and Bell et al . 23 , we analyze the global CMT catalog 17 , 18 , in our case for the period between January 1, 1977 and October 31, 2013, with the values of the seismic moment converted into N · m (1 dyn · cm = 10 −7 N · m). We restrict to shallow events (depth <70 km) and, in order to avoid incompleteness, to magnitude m > 5.75 (equivalent to M > 5.3 · 10 17 N · m), as Main et al . 22 and Bell et al . 23 . This yields 6150 events. As statistical tools, we use maximum likelihood estimation (MLE) for fitting, and likelihood-ratio (LR) tests for comparison of different fits. Maximum likelihood estimation is the best-accepted method in order to fit probability distributions, as it yields estimators which are invariant under re-parameterizations, and which are asymptotically efficient for regular models, in particular for exponential families 25 (the three models under consideration here are regular, and the PL and the TrG belong to the exponential family). When maximum likelihood is used under a wrong model, what one finds is the closest model to the true distribution in terms of the Kullback-Leibler divergence 25 . Model selection tests based on the likelihood ratio have the advantage that the ratio is invariant with respect to changes of variables (if these are one-to-one 25 ). Moreover, for comparing the fit of models in pairs, LR test is preferable in front of the computation of differences in BIC or AIC (Akaike information criterion), as the test relies on the fact that the distribution of the LR is known, under a suitable null hypothesis, which provides a significance level (or level of risk) to its value. So, LR tests constitute probability-based model selection (in contrast to BIC and AIC). But note that the log-likelihood-ratio is equal to the difference of BIC or AIC when the number of parameters of the two models is the same. In order to perform MLE it is necessary to specify the densities of the distributions, including the normalization factors. In our case, all distributions are defined for M above the completeness threshold a , i.e., for M > a , being zero otherwise (as mentioned above, a is fixed to 5.3 × 10 17 N · m). For the power-law (PL) distribution (which yields the GR law for the distribution of M ) Eq. (1) reads with β > 0. For the tapered Gutenberg-Richter, with β > 0 and θ > 0. And for the left-truncated (and extended to β > 0) gamma distribution; with −∞ < β < ∞ and θ > 0, and with the upper incomplete gamma function, defined for z > 0 when γ < 0. We summarize the parameterization of the densities as f ( M ; Θ), where Θ = { β, θ } for the Tap and TrG distributions and Θ = β for the power law. Note that for the TrG distribution, it is clear that the exponent β is a shape parameter and θ is a scale parameter; in fact, these parameters play the same role in the Tap distribution, which turns out to be a mixture of two truncated gamma distributions, one with shape parameter β and the other with β − 1, but with common scale parameter θ . Exactly, (in our case, the contribution of the second TrG will be only about 0.14%). In contrast, the power law lacks a scale parameter. In all cases the completeness threshold a is a truncation parameter, but it is kept fixed and is not a free parameter, therefore. Other authors consider the upper truncated power-law distribution 13 , 26 , given by f ( M ; β, θ ) ∝ 1/ M 1+ β for a < M < θ , and zero otherwise; then θ becomes a truncation parameter. We disregard this model because such an abrupt truncation is unphysical 7 , because the occurrence of one single earthquake with size larger than the resulting value of θ invalidates the selected model, and because the fact that the support of the distribution involves the unknown parameter θ leads to a violation of the regularity conditions for which standard likelihood theory holds 25 . The knowledge of the probability densities allows the direct computation of the likelihood function as where M i are the N observational values of the seismic moment. Maximization of the likelihood function with respect the values of the parameters leads to the maximum-likelihood estimation of these parameters, with the value of the likelihood at its maximum. Note that the independence assumption that is implicit in the expression for L (Θ) arises in fact as the maximum-entropy solution when there is no information about dependence 27 . If the data cannot be considered independent, the MLE results will just describe a marginal distribution f ( M ; Θ) of the sample under consideration, and inference about the underlying population will not be possible, as the sample may be not representative of the population. In any case, the results of MLE for our three models are reported in Table 1 , and an illustration of the corresponding fits is provided in the Supplementary Information (SI) . Although the TrG model has the highest likelihood one has to perform a proper model comparison. Table 1 Maximum likelihood estimation of the parameters with their standard errors (s.e.) and maximum value of the log-likelihood function, l = ln when the PL, Tap, and TrG distributions are fitted to the seismic moment of shallow CMT earthquakes, using the whole data set ( N = 6150). Full size table Model Comparison A powerful method for comparison of pairs of models is the likelihood-ratio test, specially suitable when one model is nested within the other, which means that the first model is obtained as a special case of the second one. This is the case of the power-law distribution with respect to the other two distributions; indeed, the power law is nested both within the Tap and within the truncated gamma, as taking θ → ∞ in any of the two leads to the power-law distribution. This is easily seen taking into account that S tap ( M ) = ( a / M ) β e −( M−a )/ θ , or just performing the limit in the expression for f tap ( M ) above. For the truncated gamma distribution, when doing the θ → ∞ limit in f trg ( M ) one needs to use that, for γ < 0, z γ /Γ( γ, z ) → − γ when z → 0, see ref. 28 for γ ≠ −1, −2, … Given two probability distributions, 1 and 2, with 1 nested within 2, the likelihood-ratio test evaluates , where is the likelihood (at maximum) of the “bigger” or “full” model (either Tap or TrG) and corresponds to the nested or null model (power law in our case). Taking logarithms we get the log-likelihood-ratio with , where f i denotes the probability density function of the distribution j for every j = 1, 2, and the MLE corresponds to and . In order to compare the fit provided by the two distributions, it is necessary to characterize the distribution of . Let n 1 and n 2 be the number of free parameters in the models 1 and 2, respectively. In general, if the models are nested, and under the null hypothesis that the data comes from the simpler model, the probability distribution of the statistic in the limit N → ∞ is a chi-squared distribution with degrees of freedom equal to n 2 − n 1 > 0. So, for n 2 = n 1 + 1, with a level of risk equal to 0.05. Note that the chi-squared distribution provides a penalty for model complexity as the “range” or “scale” of the distribution is given directly by the number of the degrees of freedom. This likelihood-ratio test constitutes the best option to choose among models 1 and 2, in the sense that it has a convergence to its asymptotic distribution faster than any other test 29 . The null and alternative hypotheses correspond to accept model 1 or 2, respectively, although the acceptance of model 1 does not imply the rejection of 2, it is simply that the “full” model 2 does not bring any significant improvement with respect the simpler model 1, which is more parsimonious. On the other hand, when the nesting of distribution 1 within 2 takes place in such a way that the space of parameters of the former one lies within a boundary of the space of parameters of distribution 2, the approach just explained for the asymptotic distribution of is not valid 30 , 31 . This happens when testing both the Tap or the TrG distributions in front of the power-law distribution, as the θ →∞ limit of the latter corresponds to the boundary of the parameter space of the two other distributions, and then, what one should obtain for is a mixture of a chi-square and a Dirac delta function. Nevertheless, this latter result is also unapplicable in our case, as the power-law distribution does not fulfil the sufficient conditions stated in ref. 30 , due to the divergence of the second moment 32 . This illustrates part of the difficulties of performing proper model selection when fractal-like distributions are involved 33 . In order to obtain the distribution of and from there the p −values of the LR tests, we are left to the simulation of the null hypothesis. We advance that the results seem to indicate that the distribution of , for high percentiles, is close to chi-square with one degree of freedom, so that Eq. (10) is approximately valid, but we lack a theoretical support for this fact. Let us proceed, using this method, by comparing the performance of the power-law and Tap fits when applied to the global shallow seismic activity, for time windows starting always in 1977 and ending in the successive times indexed by the abscissa in Fig. 1(a) (as in ref. 23 ). The log-likelihood-ratio of these fits (times 2), is shown in the figure together with the critical region of the test. In agreement with Bell et al . 23 , we find that: (i) the power-law fit can be safely rejected in front of the Tap distribution for any time window ending between 1984 and before 2004; and (ii) the results change drastically after the occurrence of the great 2004 Sumatra earthquake, for which the power law cannot be rejected at the 0.05 level. So, for parsimony reasons, the power law becomes preferable in front of the Tap distribution for time windows ending later than 2004. The fact that, for these time windows, the Tap distribution cannot be distinguished from the power law is also in agreement with previous results showing that the contour lines in the likelihood maps of the Tap distribution are highly non-symmetric and may be unbounded for smaller levels of risk 7 , 24 , 34 . Figure 1: Results of the likelihood-ratio tests. The points (joined by lines) denote the value of the statistic for the empirical data. Lines show different percentiles of the distribution of for 10000 simulations of the power-law null hypothesis with the same number of data (dotted black lines: first, second, and third quartiles), including the critical value of the test (at level 0.05, dashed red line). The abscissa corresponds to the ending point of a time window starting always in Jan 1, 1977. Note that the year is considered a continuous variable (not a categorical variable), so, the time window ending on Dec 31, 2004 takes value 2004.99… ≃ 2005. ( a ) Tap distribution versus power law. ( b ) Truncated gamma versus power law. Full size image When we compare the power-law fit with the truncated gamma, using the same test, for the same data, the results are more significant, see Fig. 1(b) . The situation previous to 2004 is nearly the same, with an extremely poor performance of the power law; but after 2004, despite a big jump again in the value of the likelihood ratio, the power law remains non-acceptable, at the 0.05 level. It is only after the great Tohoku earthquake of 2011 that the p −value of the test enters slightly into the non-rejection region, but keeping values very close to the 0.05 limit. From here we conclude that, in order to find an alternative to the power-law distribution, the truncated gamma distribution is a better option than the Tap distribution, as it is more clearly distinguishable from the power law (for this particular data). At this point, a direct comparison between these two distributions (Tap and TrG) seems pertinent. In this case we may use the likelihood-ratio test of Vuong for non-nested models 35 , 36 . As the number of parameters is the same for both models, their log-likelihood-ratio coincides with the difference in BIC or AIC, but the LR procedure incorporates a statistical test which specifies the distribution of the statistic under consideration. Unfortunately, the results are inconclusive, as no significance difference shows up. This is not surprising if one considers that the LR test for non-nested models is less powerful than the LR test for nested models used above. In order to check the possible influence of the different heterogeneous populations present in global seismicity, associated to different tectonic zones, we have separately analyzed subduction zones, similarly as done in ref. 37 , using Flinn-Engdahl’s regionalization 6 . The results for the LR tests are qualitatively the same, with the main difference that the values of l trg − l pl become somewhat smaller (not shown); nevertheless, as long as a time window of several years is considered, the power-law hypothesis can always be rejected except after the Tohoku earthquake. The resulting MLE parameters for the TrG are and ( N · m) for N = 4067 events. Then, the slightly larger value of with respect the global case ( Table 1 ) makes the power law a bit harder to reject. Simulated Data with Temporal Reshuffling As we have seen, in contrast to the Tap, the TrG distribution does bring an improvement with respect the PL, so, we concentrate on further comparisions between TrG and PL. With the purpose of gaining further insight, we simulate random samples following the truncated gamma distribution, with the parameters and obtained from MLE of the complete dataset ( Table 1 ), with the same truncation parameter a and number of points ( N = 6150) also. To avoid that the conclusions depend on the time correlation of magnitudes in the empirical data, we reshuffle the simulated data in such a way that the temporal occurrence of the order statistics in the seismic moment is the same as for the empirical data; in other words, the largest simulated event is assigned to take place at the time of the 2011 Tohoku earthquake (the largest of the CMT catalog 23 ), the second largest at the time of the 2004 Sumatra event, and so on. In this way, we model earthquake seismic moments as arising from a gamma distribution with fixed parameters, with occurrence times given by the empirical times, and with practically the same seismic-moment correlations as the empirical data. We simulate 1000 datasets with N = 6150 each. The results, summarized in Fig. 2 using boxplots 38 , show that the behaviour of the empirical data is not atypical in comparison with this gamma modelling. In nearly all time windows the empirical data lies in between the first and third quartile of the simulated data, although before 2004 the empirical values are close to the third quartile whereas after 2004 they lay just below the median. This leads us to compute the statistics of the jump in the log-likelihood-ratio between 2004 and 2005. The estimated probability of having a jump larger than the empirical value is around 4.5%, which is not far from what one could accept from the gamma modelling explained above. Thus, a TrG distribution, with fixed parameters, is able to reproduce the empirical findings, if the peculiar time ordering of magnitude of the real events is taken into account. Notice also that, although the simulated data come from a TrG distribution, they are not distinguishable from a power law for about half of the simulations of the last time windows, as the critical region is close to the median indicated by the boxplots. Figure 2: Comparison of the empirical values of the statistic 2 = 2 (points with lines, shown also in Fig. 1 ) with those resulting from 1000 simulations of the TrG distribution (boxplots) using the final parameters of Table 1 (i.e., β = 0.681 and m c = 9.15). The 95th percentile of the boxplots is also shown, in continuous red. Simulated seismic moments are reshuffled as explained in the text to make the comparison possible. The agreement between empirical data and simulations is very remarkable. The red dashed line is the same as in Fig. 1 . Remember that the central lines in the boxplots represent the three quartiles of the distribution of . Full size image We can also compare the evolution of the estimated parameters for the empirical dataset and for the reshuffled TrG simulations, with a good agreement again, see Fig. 3 . There, it is clear that although the exponent β reaches very stable values relatively soon (around 1990), the scale parameter θ (equivalent to m c ) is largely unstable, and the occurrence of the biggest events makes its value increase. Figure 3: Comparison between empirical estimated parameters and for the TrG distribution (points with lines) and the estimations for the 1000 simulations of Fig. 2 (i.e., TrG with β = 0.681 and m c = 9.15 with temporal reshuffling, boxplots). The different stability of both parameters is apparent, as well as the similarity between data and simulations. ( a ) . ( b ) . Full size image As a complementary control we invert the situation, simulating 1000 synthetic power-law datasets with β = 0.685 ( Table 1 ), a = 5.3 × 10 17 N · m, and N = 6150, for which the same time reshuffling is performed, in such a way that the order of the order statistics is the same. In this case, the results of the simulations lead, on average, to much smaller values of the log-ratio in comparison with the empirical data, which corresponds to the limit of rejection for many time windows, see Fig. 4 . So, a power-law distribution with temporal reshuffling cannot account for the empirical results as clearly as a truncated gamma distribution. Doing the same with a Tap distribution one finds something in between, see SI. Figure 4: As Fig. 2 , but simulating a power law with parameter β = 0.685 ( Table 1 ) instead of a TrG distribution. The reshuffling is also as in Fig. 2 , as explained in the text. The simulations lead, on average, to values of the likelihood ratio smaller than the empirical ones. Note that the difference with Fig. 1(b) is that there (i) there is no reshuffling and (ii) the value of β in the simulations corresponds to the obtained for each time window. Full size image Discussion Testing different statistical models for the distribution of seismic moment of global shallow seismicity (using the CMT catalog) we have found that, in contrast to the Tap distribution, the truncated gamma brings significant improvement with respect to the power law. Moreover, in order to reproduce the time evolution of the statistical results, it suffices that independent seismic moments following a truncated gamma distribution with fixed parameters β = 0.68 and m c = 9.15 are reshuffled so that the peculiar empirical time sequence of magnitudes is maintained (note that after reshuffling independence is broken). So, despite the fact that the future occurrence of more and larger mega-earthquakes could significantly change the value of parameter m c 13 , the current value is enough to explain the available data. Although ref. 13 claims that no less than 45,000 events are necessary for the reliable estimation of m c , our simulations with 6150 events indicate otherwise, see for instance the last boxplot for the estimation of m c in Fig. 3 , which yields a mean value of 9.11, with a standard deviation of 0.24, totally consistent with the results in Table 1 . We conclude that the fundamental problem in the estimation of m c is not the number of available data but the temporal heterogeneity of the seismic moment distribution. We have also found, with a similar reshuffling procedure, that a power-law distribution cannot account for the empirical findings. Direct comparison of Figs 2 and 4 shows how the TrG distribution outperforms the power law. Additionally, it would be very interesting to investigate if the high values of the likelihood ratio attained before the 2004 Sumatra event could be employed to detect the end of periods of low global seismic activity. Certainly, more case studies would be necessary for that purpose. As extra arguments in favour of the truncated gamma distribution in front of the tapered GR, we can bring not statistical evidence but physical plausibility and statistical optimality. On the one hand, the former distribution can be justified as coming from a branching process that is slightly below its critical point 12 , 39 . Further reasons that may support the truncated gamma are that this arises (i) as the maximum entropy outcome under the constrains of fixed (arithmetic) mean and fixed geometric mean of the seismic moment 40 ; (ii) as the closest to the power law, in terms of the Kullback-Leibler divergence, when the mean seismic moment is fixed 41 ; and (iii) as a stable distribution under a fragmentation process with a power-law transition rate 41 . We are not aware of similar theoretical support in favour of the Tap distribution. On the other hand, it is straightforward to check that the truncated gamma belongs to the exponential family 25 , in contrast to the Tap distribution. And it is well known that estimators in the exponential family achieve the Cramér-Rao lower bound for any sample size, in contrast to other regular models, where the bound is only achieved asymptotically. Additional Information How to cite this article : Serra, I. and Corral, Á. Deviation from power law of the global seismic moment distribution. Sci. Rep. 7 , 40045; doi: 10.1038/srep40045 (2017). Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Researchers of the Mathematics Research Centre (CRM) and the UAB have developed a mathematical law to explain the size distribution of earthquakes, even in the cases of large-scale earthquakes such as those which occurred in Sumatra (2004) and in Japan (2011). The probability of an earthquake occurring exponentially decreases as its magnitude value increases. Fortunately, mild earthquakes are more probable than devastatingly large ones. This relation between probability and earthquake magnitude follows a mathematical curve called the Gutenberg-Richter law, and helps seismologists predict the probabilities of an earthquake of a specific magnitude occurring in some part of the planet. The law however lacks the necessary tools to describe extreme situations. For example, although the probability of an earthquake being of the magnitude of 12 is zero, since technically this would imply the earth breaking in half, the mathematics of the Gutenberg-Richter law do not consider impossible a 14-magnitude earthquake. "The limitations of the law are determined by the fact that the Earth is finite, and the law describes ideal systems, in a planet with an infinite surface", explains Isabel Serra, first author of the article, researcher at CRM and affiliate lecturer of the UAB Department of Mathematics. To overcome these shortages, researchers studied a small modification in the Gutenberg-Richter law, a term which modified the curve precisely in the area in which probabilities were the smallest. "This modification has important practical effects when estimating the risks or evaluating possible economic losses. Preparing for a catastrophe where the losses could be, in the worst of the cases, very high in value, is not the same as not being able to calculate an estimated maximum value", clarifies co-author Álvaro Corral, researcher at the Mathematics Research Centre and the UAB Department of Mathematics. Obtaining the mathematical curve which best fits the registered data on earthquakes is not an easy task when dealing with large tremors. From 1950 to 2003 there were only seven earthquakes measuring higher than 8.5 on the Richter scale and since 2004 there have only been six. Although we are now in a more active period following the Sumatra earthquake, there are very few cases and that makes it statistically a poorer period. Thus, the mathematical treatment of the problem becomes much more complex than when there is an abundance of data. For Corral, "this is where the role of mathematics is fundamental to complement the research of seismologists and guarantee the accuracy of the studies". According to the researcher, the approach currently used to analyse seismic risk is not fully correct and, in fact, there are many risk maps which are downright incorrect, "which is what happened with the Tohoku earthquake of 2011, where the area contained an under-dimensioned risk". "Our approach has corrected some things, but we are still far from being able to give correct results in specific regions", Corral continues. The mathematical expression of the law at the seismic moment, proposed by Serra and Corral, meets all the conditions needed to determine both the probability of smaller earthquakes and of large ones, by adjusting itself to the most recent and extreme cases of Tohoku, in Japan (2011) and Sumatra, in Indonesia (2004); as well as to determine negligible probabilities for earthquakes of disproportionate magnitudes. The derived Gutenberg-Richter law has also been used to begin to explore its applications in the financial world. Isabel Serra worked in this field before beginning to study earthquakes mathematically. "The risk assessment of a firm's economic losses is a subject insurance companies take very seriously, and the behaviour is similar: the probability of suffering losses decreases in accordance with the increase in volume of losses, according to a law that is similar to that of Gutenberg-Richter, but there are limit values which these laws do not take into consideration, since no matter how big the amount, the probability of losses of that amount never results in zero" Serra explains. "That makes the 'expected value of losses' enormous. To solve this, changes would have to be made to the law similar to those we introduced to the law on earthquakes".

10.1038/srep40045

Medicine

Researchers may have found the missing link between Alzheimer's and vascular disease

Annie J. Lee et al, FMNL2 regulates gliovascular interactions and is associated with vascular risk factors and cerebrovascular pathology in Alzheimer's disease, Acta Neuropathologica (2022). DOI: 10.1007/s00401-022-02431-6

https://dx.doi.org/10.1007/s00401-022-02431-6

https://medicalxpress.com/news/2022-05-link-alzheimer-vascular-disease.html

Abstract Alzheimer’s disease (AD) has been associated with cardiovascular and cerebrovascular risk factors (CVRFs) during middle age and later and is frequently accompanied by cerebrovascular pathology at death. An interaction between CVRFs and genetic variants might explain the pathogenesis. Genome-wide, gene by CVRF interaction analyses for AD, in 6568 patients and 8101 controls identified FMNL2 ( p = 6.6 × 10 –7 ). A significant increase in FMNL2 expression was observed in the brains of patients with brain infarcts and AD pathology and was associated with amyloid and phosphorylated tau deposition. FMNL2 was also prominent in astroglia in AD among those with cerebrovascular pathology. Amyloid toxicity in zebrafish increased fmnl2a expression in astroglia with detachment of astroglial end feet from blood vessels. Knockdown of fmnl2a prevented gliovascular remodeling, reduced microglial activity and enhanced amyloidosis. APP/PS1dE9 AD mice also displayed increased Fmnl2 expression and reduced the gliovascular contacts independent of the gliotic response. Based on this work, we propose that FMNL2 regulates pathology-dependent plasticity of the blood–brain-barrier by controlling gliovascular interactions and stimulating the clearance of extracellular aggregates. Therefore, in AD cerebrovascular risk factors promote cerebrovascular pathology which in turn, interacts with FMNL2 altering the normal astroglial-vascular mechanisms underlying the clearance of amyloid and tau increasing their deposition in brain. Working on a manuscript? Avoid the common mistakes Introduction Alzheimer’s disease (AD) affects more than 6.2 million people in the United States and approximately 24 million worldwide. Neuropathological studies indicate that in AD, the hallmark findings of neuritic plaques and neurofibrillary tangles can frequently be accompanied by varying degrees of cerebrovascular disease in up to 70% of patients [ 4 , 16 , 44 , 74 ]. Amyloid β in blood vessels, as in cerebral amyloid angiopathy, reduces cerebral blood flow and is present in most patients with AD. However, as cardio and cerebrovascular risk factors (CVRFs) increase in the elderly, accompanied by inflammation, cytokine release, endothelial dysfunction and arterial stiffening in brain [ 81 ]. Cholesterol laden macrophages accumulate in vessel walls also decreasing blood flow. The increase in atherosclerosis in the intracerebral arteries [ 55 ] and capillaries leads to microinfarcts in the hippocampus contributing to cognitive decline [ 34 ]. The relationship between CVRFs such as hypertension [ 78 ], body mass index [ 33 , 57 ], diabetes [ 22 , 59 ] and coronary heart disease [ 42 , 65 ] and AD is well known, but there has been limited mechanistic evidence directly linking these vascular risk factors in AD to the presence of ischemic microvascular pathology. Each of these vascular factors has the capacity to impair the blood–brain barrier and glio-vascular units. Arterial pulses and flow are required for glymphatic clearance of molecules including amyloid β [ 42 , 65 , 86 ]. Jagust and colleagues argued that cerebrovascular disease contributed to AD by perturbing the amyloid β pathway in addition to causing neurodegeneration [ 38 ]. However, the effects reported in epidemiological studies of the association between CVRFs and AD are inconsistent, increasing AD risk in some studies [ 86 ], but in others showing protection against AD [ 64 , 86 ]. The relationship between vascular risk factors and cerebrovascular pathology in AD could simply be the result of aging, the stage of the disease or a coincidental occurrence. An alternative explanation is that an unidentified interaction between genetic variants and vascular risk factors leading to cerebrovascular pathology in AD brain may contribute to disease pathogenesis. APOE , CLU , ABCA7 , and SORL1 , typically associated with immune mechanisms in AD, are also part of the lipid metabolism pathway providing evidence for a putative molecular relationship. In fact, ABCA7 and SORL1 have been associated with brain infarcts [ 31 ]. Using summary statistics from multiple genome-wide associations studies (GWAS), a recent study detected evidence of pleiotropy between vascular risk factors and AD [ 17 ]. Another investigation [ 88 ] found shared genetic contribution to AD and small vessel disease. Despite the possible associations between genes, vascular factors and AD, there has been no unbiased genome-wide study of genes or genetic loci to investigate the interaction between cerebrovascular risk factors, cerebrovascular pathology, and AD. Previously, we observed that the cumulative burden of vascular risk factors increased the association with AD risk [ 58 , 70 ] compared with a single risk factor. Therefore, in this investigation, a dimensionality reduction of four common cardiovascular risk factors frequently associated with cerebrovascular disease and with AD: hypertension, heart disease, diabetes, and body mass index (BMI), was employed to create a vascular risk factor burden score. To augment the sample size, we included data from five different cohorts representing different ethnic groups, facilitating the interaction analyses. Materials and methods Clinical studies Participants Participants were from the following studies: Washington Heights–Inwood Columbia Aging Project (WHICAP), Estudio Familiar de Influencia Genetica en Alzheimer (EFIGA), the National Alzheimer’s Coordinating Center (NACC), and the Religious Orders Study and Rush Memory and Aging Project (ROSMAP) (Table 1 ). WHICAP is a multiethnic, prospective, community-based cohort study of aging and dementia in Medicare recipients 65 years and older residing in northern Manhattan. A detailed description of the study was previously published [ 83 ]. Participants are initially recruited as nondemented elderly and were non-Hispanic white, African American or Caribbean Hispanic. A consensus diagnosis was derived for each participant by experienced clinicians based on NINCDS‐ADRDA criteria for possible, probable, or definite AD (moderate or high likelihood of neuropathological criteria for AD) [ 61 ]. Recruitment for the EFIGA began in 1998, to study the genetic architecture of AD in the Caribbean Hispanic population. Patients with familial AD were recruited and if a sibling of the proband had dementia, all other living siblings and available relatives underwent evaluation. Cases were defined as any individual meeting NINCDS-ADRDA criteria for probable or possible AD. Details of the study have previously been reported [ 90 ]. NACC collects, organizes, and maintains phenotype information from National Institute on Aging (NIA) Alzheimer’s Disease Centers (ADCs). Genetic data were provided from wave 1–10 of the ADC genotyping. The Uniform Data Set (UDS) is made up of standardized clinical evaluations and diagnoses along with demographic information. Details on the UDS and NACC database can be found at . ROSMAP are two harmonized longitudinal studies enrolling older adults without dementia: The Religious Orders Study and the Memory and Aging Project (ROSMAP). Details of the ROSMAP studies have been described elsewhere [ 9 , 27 ]. Only participants 65 years of age or older, that had one or more in person clinical assessments and data concerning four vascular risk factors: any heart disease, hypertension, BMI, and diabetes were included. Table 1 Participant demographics included Full size table Cerebrovascular Risk Factors Score Self-reported data from each participant were recorded as a binary indicator (Yes-Ever Had or No-Never Had) for heart disease, hypertension, and diabetes. The self-reported of vascular risk factors has reasonable reliability and validity [ 45 ]. Details on heart conditions varied between groups, so report of any heart disease qualified as presence of heart disease in this study. A quantitative variable was recorded for body mass index (BMI) from the last visit. The PCAMix package [ 20 ] computes principal components (PCs) in a mixture of quantitative and qualitative data, and was used to summarize the cardiovascular risk factors into one summary score by computing PCs from the four cerebrovascular variables. The goal of the dimensionality reduction was to capture the greatest amount of variance accounted for by the four cerebrovascular risk factors, and thus each participant’s values from the first principal component from each of the cohorts was used as their vascular risk factor score (CVRF score). Genotyping Genotyping was performed on different platforms separately for each study cohort. Data from all cohorts underwent quality control (QC). For African American and white, non-Hispanics WHICAP participants, variants with missing rate greater than 5%, out of Hardy–Weinberg equilibrium ( p < 1 × 10 –6 ), or with less than 1% minor allele frequency (MAF) were removed. Participants were excluded if they were missing more than 2% of variants present in the overall cohort. Following quality control (QC), participants were imputed separately based on self-reported ethnicity. Imputation was performed using the HRC r1.1.2016 reference panel and SHAPEIT phasing on the Michigan Imputation Server [ 26 ]. Data from the WHICAP Caribbean Hispanics and EFIGA Caribbean Hispanics underwent QC and imputation in seven batches that included participants from both studies, consistent with the genotyping batch. Variants with missing rate greater than 5%, MAF ≤ 1%, out of Hardy–Weinberg equilibrium ( p < 1 × 10 –6 ) were removed. Samples with missing call rate greater than 5% were excluded. SHAPEIT2 [ 28 ] was used for phasing and IMPUTE2 was used for imputation with the HRC r1.1.2016 reference panel. Quality control and imputation details for the ROSMAP have been reported previously [4]. Briefly, variants missing in more than 5%, MAF ≤ 1%, out of Hardy–Weinberg equilibrium ( p > 0.001), and a p value of mishap test < 1 × 10 –9 were removed. Imputation was performed using the HRC r1.1.2016 reference panel of Caucasian ancestry and Eagle v2.3 phasing. ADC genotype data from waves 1–10 underwent QC by the Alzheimer’s Disease Genetic Consortium as described previously [ 52 ]. Imputation was performed on the Michigan Imputation Server individually for each of the cohorts including all ethnicities of the Haplotype Reference Consortium (HRC) 1.1 reference panel. EAGLE was used for phasing and Minimac3 was used for imputation. NACC and ROSMAP cohorts included only participants that self-reported as non-Hispanic White. For all cohorts, post-imputation genotype data were filtered for imputation quality ( R 2 > 0.8) and minor allele frequency (> 1%). KING[ 60 ] was used to perform multidimensional scaling (MDS) to identify population substructure within each cohort. Participants more than six standard deviations from the mean within cohort in the first three calculated MDS components were removed from analysis. Statistical Analyses Study design and analyses are described in Supplementary Fig. 1, online resource. Differences in allele frequency were present across ethnic groups, thus we performed genome-wide gene-CVRF score interaction analysis independently in each cohort and summarized the results in a meta-analysis to improve analysis power and results interpretability. Gene-based tests were performed using the adaptive gene-environment interaction (aGE) test [ 20 ]. Independently, we validated the results using an alternative method implemented in the gene-environment set association test (GESAT) [ 56 ]. Both tests allowed for gene-based analyses with an interaction between the vascular risk factor variable and multiple variants within a gene. GESAT uses a variance component test while, aGE combines variance-component and burden tests to maximize power across broader association patterns. We observed that aGE controls the type I error rate in the presence of many neutral variants, and the power is resilient to the inclusion of neutral variants [ 20 ]. Therefore, we used the aGE to test for the genes that interacted with CVRFs to alter AD risk. Additionally, we used GESAT to confirm the robustness of statistically significant results reported from the aGE. Covariates included in the model were age at diagnosis for AD and age at last visit for the unaffected participants, sex, and the first three principal components to adjust for population substructure. We analyzed genes with a minimum of two variants with allele frequency greater than 0.01 in the gene-based test. The aGE R package and GESAT functions in the iSKAT R package were performed with default settings. Both tests only export p value of testing the gene-CVRF score interaction term. We combined the results in a meta-analysis using weighted sum of Z -scores method through METAL [ 96 ]. The p value of the interaction term in each cohort was first converted into signed Z -score and then calculated as the weighted mean of the Z -scores, with square-root of the sample size from each cohort as weights. Individual SNPs were tested for CVRF interactions within each gene when a significant gene by vascular risk factors interaction was found. The SNP by CVRF interaction test was performed in each cohort for all SNPs within candidate genes by fitting a logistic regression for AD as the outcome and testing for the interaction term of SNP by CVRF score. The interaction tests were adjusted for the main genetic effect and the CVRF score as covariates in all models, in addition to age, sex, and the first three principal components. The models were tested with additive coding for the SNPs. For models used in the NACC cohort, batch effects were included as covariates because genotyping from ten waves were imputed separately. The results were then combined in a meta-analysis using inverse-variance weighted average method through METAL [ 96 ] by calculating the weighted mean of the SNP-CVRF interaction effect sizes, with the inverse variance from each cohort as weights. Results were shown for SNPs and genes present in at least 10,000 participants. Statistical significance was set at p = 5 × 10 –6 for the genome-wide gene-based tests and used the false discovery rate (FDR) adjusted p value or Bonferroni correction for the number of genes tested. RNA expression analysis As shown in Table 3 , we performed association analyses of pathologically diagnosed AD using the multi-region brain transcriptomes from ROSMAP [ 7 , 8 , 9 ] and a replication analysis was performed using the transcriptomes from Mount Sinai Medical Center [ 94 ] and Mayo Clinic Brain Banks [ 2 ]. A logistic regression was fitted for pathological AD as the outcome while adjusting for age, sex, and processing factors. The bulk RNA-sequencing (RNA-Seq) transcriptomic profiles in ROSMAP were accessed from three brain regions—the dorsolateral prefrontal cortex (DLPFC) in 1,092 individuals, the posterior cingulate cortex (PCC) in 661 individuals and the anterior caudate (AC) in 731 individuals. For the MSBB tissues, the transcriptomic profiles were accessed from four brain regions—the frontal pole (BM10) in 214 individuals, the superior temporal gyrus (BM22) in 191 individuals, the parahippocampal gyrus (BM36) in 161 individuals, the inferior frontal gyrus (BM44) in 186 individuals. For the Mayo Clinic brain tissues, the transcriptomic profiles were accessed from the temporal cortex (TCX) in 261 individuals and the cerebellum (CBE) in 262 individuals. RNA sequencing data in ROSMAP was generated in multiple batches from different sequencing centers and experimental protocols. RNA sequencing of dorsolateral prefrontal cortex (DLPFC) was first done in 10 batches for 739 subjects at the Broad Institute. Subsequently, 124 subjects in 2 batches were sequenced at the New York Genome Center. Moreover, 229 samples in a single batch were sequenced at the Rush Alzheimer’s Disease Center. A detailed description of the study was previously published [ 101 ]. RNA sequencing of anterior caudate (AC) was done in 76 subjects in a single batch at the Broad Institute using the Illumina TruSeq method, as described above in DLPFC. Subsequently, 655 subjects in 2 batches were sequenced at the New York Genome Center using the same experimental protocols described in DLPFC. RNA sequencing of posterior cingulate cortex (PCC) was done in 79 subjects in a single batch at the Broad Institute using the Illumina TruSeq method. Subsequently, 487 subjects in 2 batches were sequenced at the New York Genome Center and 95 subjects in a single batch were sequenced at the Rush Alzheimer’s Disease Center. The experimental protocols of each sequencing center are same as described in DLPFC. Each brain region was pre-processed separately. We used gene-level transcription values derived from the RNA sequencing data. The counts values were used to measure the gene expression levels and subsequently normalized using Trimmed means of M-values (TMM) to create a frozen dataset that are available on Synapse (Synapse: syn25741873). Outlier samples were removed based on quantified expression profiles and lowly expressed genes with a median Counts less than 10 were filtered out to reduce the influence of technical noise. A linear regression was fitted for each transcript with the log2-transformed normalized values as outcome adjusted for age, sex, and processing factors, and used residual from the transcript as our final transcription dataset. The residuals represent a quantitative trait capturing variability in transcripts outcome not captured by known demographics or technical factors. Technical factors for DLPFC include batch, library size, percentage of aligned reads, percentage of coding bases, percentage of intergenic bases, percentage of ribosomal bases, percentage of UTR base, percentage of duplication, median 3 prime bias, median 5 prime to 3 prime bias, median CV coverage, pmi, and study index of ROS or MAP. Technical factors for AC and PCC are very similar. A detailed description of outlier removal, normalization, and calculation of the residuals in MSBB and Mayo are described elsewhere [ 93 ]. We assumed that differences in candidate gene expression would be associated with brain infarcts. A logistic regression was fitted for brain infarcts as the outcome and tested for the candidate gene expression while adjusting for age, sex, and processing factors. Details of neuropathological evaluation and qualifying determinants of infarcts in the ROSMAP dataset have been reported previously [ 75 ]. We performed association analyses of pathologically diagnosed AD with expression level of candidate gene across cells of the dorsolateral prefrontal cortex in 24 individuals from the ROSMAP cohort. A logistic regression was fitted for pathological AD as the outcome while adjusting for age, sex. A detailed description of the ROSMAP single nucleus RNA sequencing data are described elsewhere [ 3 ]. Causal mediation analysis To quantify the involvement of gene expression in aging-related brain pathology in AD, we investigated whether the effect of gene expression on cerebrovascular pathology in AD brain would be mediated by AD specific pathology. We also investigated whether the effect of the AD pathology in brain infarcts would be mediated by candidate gene expression. We used a causal mediation analysis aimed at identifying whether the candidate expression resulted from amyloid or phosphorylated tau deposition or the reverse. In the mediation analysis, the mediated effect refers to an indirect effect of the exposure on outcome through a mediator. Direct effect refers to the effect of the exposure on the outcome after adjusting for the mediator. Total effect refers to the total effect of the exposure on the outcome, obtained by combining direct and indirect effects. Causal mediation modeling was performed using the R package mediation and confidence intervals were obtained by nonparametric bootstrap procedure with 1000 resamples [ 85 ]. In parallel, we performed cell type specific analyses of the expression of our gene of interest. Human brain sections and immunohistochemistry Human brain sections from BA9 prefrontal cortex were also obtained from the New York Brain Bank at Columbia University (Table S14) and immunohistochemical staining for FMNL2 and GFAP were performed as described [ 92 ] with the following modifications: heat-induced antigen retrieval was performed with pressure cooker, and concomitant biotin-based immunodetection was performed. Nine random images per patient from the sections were acquired with identical acquisition parameters to control brains and Arivis-based quantification was performed (script available upon request). The analyses were based on the overlapping surface area of GFAP and FMNL2 that is normalized to the total GFAP surface area. Automated masking was quality controlled manually; images were excluded from the analyses if machine-predicted surfaces show substantial difference to manual selections. Analyses were performed in blind fashion: sample IDs were revealed after immunohistochemical stainings and quantifications for individual samples. Statistical comparisons were performed with Dunn’s Kruskal–Wallis test and linear mixed effects model. Mouse studies Ethics statement and animal maintenance All animal studies followed European animal regulations and were approved by the appropriate authority (Landesdirektion Sachsen, Germany) under license number TVV87/2016. To limit suffering and overall animal numbers, animals were treated with extraordinary caution. Mice were maintained on a 12-h alternating light and dark cycle, with free access to conventional mouse diet (chow) and water. Animals were housed in groups in normal ventilated cages, and after the surgeries, they were maintained in individual cages. Fixed gender mice aged between 52 and 54 weeks were used for analysis. Jackson Laboratories (Bar Harbor, ME, USA) provided the B6.Cg-Tg(APP695)3DboTG(PSEN1dE) mice, which were kept as a heterozygous breeding colony. Stereotaxic injection for mouse brain injury The brain tissue injury was induced by stereotaxic injection. The mouse was anesthetized with a combination of oxygen and isoflurane (49:1) (Baxter–HDG9623) flow during the procedure and put on a pre-warmed heat-pad to avoid hypothermia. Ear bars were used to keep the head immobile, and a protective ointment was applied to the eyes to prevent cornea dehydration. An analgesic was administered subcutaneously before to the surgery to reduce any potential pain afterward. The injury was caused in the right hemisphere, coordinates were ± 1.6 mm mediolateral, − 1.9 mm anterior–posterior, and − 1.9 mm dorsoventral from the Bregma, where the PBS was dispensed at 200 nL/min speed. The left hemisphere was used as a control for the analysis. The capillary was progressively retracted after the injection, followed by the ear being released. Mouse brain was analyzed 3 days after the injury via immunohistochemistry. Tissue preparation An overdose of Ketamine/Xylazin (0.25 mL per 25 g of body weight) was used to euthanize the mice, which were then transcardially perfused with NaCl (0.9 percent w/v) and 4% paraformaldehyde (PFA). Brains were extracted and post-fixed in 4% PFA overnight at 4 °C. Brains were placed in a 30% sucrose solution for 2–3 days to cryopreserve the fixed tissue. A sliding microtome (Leica SM2010) chilled with dry ice was used to cut coronal sections with a thickness of 40 μm. Free-floating sections were collected in six series and kept at − 20 °C in cryoprotection solution (CPS; 25% ethylene glycol, 25% glycerol in 0.1 M phosphate buffer pH 7.4). For immunohistochemistry, every sixth section series of each brain was used. Immunohistochemistry Free-floating sections were rinsed three times in PBS before being blocked in a 10% Donkey Serum, 0.3% Tween 20, and 1 × PBS solution for 1 h at room temperature. Primary antibodies were diluted in PBS containing 3% donkey serum and 0.3% Tween-20 and sections were treated overnight at 4 °C. After 3 washes with PBS, secondary antibody conjugated with a chosen fluorophore was then incubated for four hours at room temperature. After a quick wash, samples are incubated for 15 min in 4,6-diamidino-2-phenylindole (DAPI) diluted in PBS. Additional washing steps were performed, and samples were mounted on glass slides. Imaging Fluorescence images were acquired using a spinning disc Zeiss Axio Observer.Z1 microscope (Oberkochen, Germany) equipped with ZEN software (version blue edition, v3.2, company, Carl Zeiss, Jena, Germany). Zebrafish studies Ethics statement Animal experiments were carried out in accordance with the animal experimentation permits of Referate 24 (Veterinärwesen, Pharmazie, und GMP) of the state administration office of Saxony, Germany (Landesdirektion Sachsen), the ethical commission of TU Dresden (Kommission für Tierversuche), and the Institutional Animal Care and Use Committee (IACUC) at Columbia University (protocol number AC-AABN3554). Zebrafish handling and maintenance was according to the provided guidelines [ 1 , 32 , 51 , 79 ] and EU Directive 2010/63 Article 33 and Annex III (permit numbers: TVV-35/2016, TVV-52/2015, TVV31/2019, and TVV39/2020). For zebrafish studies, 8–12 months old wild type AB strain, Tg( her4.1 :GFP) [ 100 ] and Tg( kdrl: GFP) [ 40 ] reporter fish of both genders were used. In every experimental set, animals from the same fish clutch were randomly distributed for each experimental condition. Amyloid-β42 (Aβ42) and morpholino injection Cerebroventricular microinjection of Amyloid-β42 peptides in adult zebrafish brain were performed as previously described [ 13 ]. Morpholinos were injected to embryos and adult zebrafish brains as described [ 11 , 49 ]. For morpholino experiment, 10 μM concentration of control morpholinos, control morpholinos with Aβ42 or fmnl2a + fmnl2b morpholinos with Aβ42 were injected to adults. For embryos, 2 ng of morpholinos were injected. See Table S15 for more information on the reagents. Histological preparation of zebrafish tissue and immunohistochemistry Euthanasia and tissue preparation were performed as described [ 13 ]. 12-µm thick cryo-sections were prepared from these brain samples using a cryostat and collected onto glass slides which were then stored at − 20 °C. Immunohistochemistry was performed as described [ 13 ]. Imaging, quantifications, and statistical analyses Images were acquired using ZEN software on a Zeiss fluorescent microscope with ApoTome or Zeiss Spinning Disc microscope. Images were analyzed using ZEN or Arivis Image processing software. For glial-vasculature interaction, automated quantification pipeline for at least 50% overlap between GFP (vasculature) and S100β (glia) was developed in Arivis software (script available upon request). Quantifications were performed on z -stacks images obtained from the telencephalon sections (region between caudal olfactory bulb until the rostral optic tectum). 6–8 histological sections per animal were used for stereological analyses and quantifications. For colocalization in Fig. 2 , ImageJ version 2.1.0/1.53c was used with default settings. A two-tailed Student’s t -test was performed for comparison of two experimental groups, one-way ANOVA with Tukey’s multiple comparison test were performed for comparison of multiple experimental groups. Statistical analyses were performed in GraphPad Prism software. Bars in the graphs indicate the mean values and 95% CI. p values less than 0.05 were considered significant. Significance is indicated by ∗( p < 0.05), ∗∗( p < 0.01), ∗∗∗( p < 0.001) or n.s. (not significant, p > 0.05). No sample set was excluded from the analyses unless the histological sections were damaged severely during the acquisition of the sections (constitutes less than 5% of all sections analyzed). Results Genetic analyses Genome-wide array data were available from 10,287 non-Hispanic whites, 3404 Caribbean Hispanics and 978 African–Americans (Table 1 ). The majority of the participants were women (65%) and had a history of hypertension (70%). A history of heart disease, hypertension and diabetes and measured BMI in the different cohorts were modestly correlated (spearman correlation coefficient ranging from 0.03 to 0.26, Supplementary Fig. 2. Principal component (PC) analysis was performed to create the CVRF score, summarizing the joint effect of the four risk factors for each cohort. The first principal component was used to represent the CVRF score, Supplementary Fig. 1 shows the contribution of each risk factor to the score. In all cohorts, diabetes and hypertension were highly correlated and most strongly influenced the CVRF score. The percentage of variance explained by principal component 1 for each ethnic group were: African–Americans 34.7%, non-Hispanics Whites 35.5%, Caribbean Hispanics 38.6%. Variance explained by the first four principal components and contributions of the risk factors to each component is also described in Table S1. In a gene by CVRF interaction analysis of AD, three genes met the false discovery rate (FDR) adjusted p value of 0.05 using the adaptive gene-environment interaction (aGE) test (Table 2 ). One gene located on chromosome 2, Formin-like protein 2 ( FMNL2 ), also met the Bonferroni corrected p value threshold for significance ( p = 6.6 × 10 –7 ). Cohort-level results showed nominal significance in all ethnic groups except in African-Americans. QQ plots for each cohort are shown in Supplementary Fig. 4. There was no evidence of inflation in any group. This FMNL2 gene by CVRF score interaction was subsequently validated using gene-environment set association test (GESAT) ( p = 7.7 × 10 –7 ). Other nominally significant genes are reported in Table S2. As an additional measure of FMNL2 interaction with CVRF, we repeated the interaction test by constructing the CVRF as the sum of first two principal components. FMNL2- CVRF interaction was still robustly associated with AD ( p = 3.47 × 10 –4 ). Analyzing each risk factor of the CVRF separately, FMNL2 interacts most strongly with history of hypertension ( p = 7.53 × 10 –4 ) and BMI (2.51 × 10 –3 ) (Table S3). Table 2 Gene-CVRF score interaction results in GWAS: Top ten genes interacted with the CVRF score to modify AD risk, sorted by meta-analysis p value Full size table Meta-analysis of single nucleotide variants (SNP) within the FMNL2 identified 130 nominally significant SNPs ( p < 0.05) present in at least three cohorts and interacting with the CVRF score (Table S4). Two individual SNPs in FMNL2 met the Bonferroni corrected p value in the interaction model rs57223657 ( p = 3.61 × 10 –4 ) and rs6760139 ( p = 3.63 × 10 –4 ). Random-effects model in meta-analysis was used as comparison and rs57223657 stayed similar ( p = 3.61 × 10 –4 ). Main genetic effect of rs57223657 was significant when adjusted for SNP–CVRF interaction effect and the CVRF score ( p = 0.004). FMNL2 expression in autopsy brains We assumed that any gene or genes interacting with vascular risk factors to modify AD risk would be directly related to molecular pathways in AD and cerebrovascular disease. Using autopsied brain tissue from the ROSMAP with multi-region brain transcriptomes, higher expression of FMNL2 was associated with pathological AD in all three regions sampled: the dorsolateral prefrontal cortex ( p = 0.004), the posterior cingulate cortex ( p = 0.001) and in the anterior caudate ( p = 0.012). In the ROSMAP cohort a small number of participants were considered cognitively normal but had autopsy findings confirming hallmarks of AD in brain. As a sensitivity analysis, we tested the association of FMNL2 expression by restricting the analysis to individuals with both clinical and pathological AD compared them with healthy individuals without ante and post-mortem AD. The FMNL2 association with AD was even stronger ( p = 7.89E−08; Table S5). Available data from the Mount Sinai Brain Bank, also showed increase FMNL2 expression in the parahippocampal gyrus ( p = 0.0004) in pathological AD, and in data from the Mayo Clinical Brain Bank, FMNL2 expression was increased in the temporal cortex ( p = 0.011) in pathological AD (Table 3 ). Table 3 Association of FMNL2 expression with pathologic diagnosis of AD using the multi-region brain transcriptomes from three cohorts Full size table Using data from ROSMAP, we found that an increased expression of FMNL2 in AD brain was associated with Aβ (adjusted β = 0.188, p = 0.034) and tau deposition (adjusted β = 0.415, p = 0.011). The expression level of FMNL2 was also significantly higher in those with gross chronic infarcts in cortex compared to those without infarctions (adjusted for age, sex, and processing factors β = 1.024, p FDR = 0.003; Fig. 1 a; Table 4 ). Restricted to pathological AD, FMNL2 expression was higher in those with, compared to those without, gross chronic infarcts in cortex (adjusted β = 0.792, p = 0.025). A mediation model suggested that FMNL2 expression mediates the association of Amyloid-β and phosphorylated tau deposition with brain infarcts, although no direct association was observed between them. The mediation analysis suggested an indirect effect, in which Aβ or tau deposition increased FMNL2 expression, which was then associated with brain infarcts (mediation effect: p = 0.002 amyloid; p < 2 × 10 –16 tau; Table S6). Of interest, the ante-mortem CVRF score was also modestly increased in those with gross chronic infarcts (detected post-mortem) compared to no infarcts ( β = 0.21, p = 0.036). Fig. 1 FMNL2 expression level by brain infarcts. a Distribution of expression level of FMNL2 in those with gross chronic infarction in cortex compared to those without infarction was statistically significant (adjusted for age, sex, and processing factors b = 1.024, p FDR = 0.003). b Distribution of FMNL2 expression in those with pathological AD and gross chronic infarction in cortex compared to those with pathological AD and no infarction, in the dorsolateral prefrontal cortex from ROSMAP was statistically significant (adjusted b = 0.792, p = 0.025). A logistic regression was fitted for the brain infarcts as outcome and FMNL2 expression level (log2-tranformed normalized transcripts per million values) as exposure adjusted for age, sex, and processing factors. For visualization purposes, the residual of FMNL2 expression level (y-axis) was used to represent the residual of transcripts from fitting the linear regression on the FMNL2 expression level adjusted for age, sex, and processing factors, capturing variability in transcripts outcome not captured by known demographics or processing factors Full size image Table 4 Association of expression level of FMNL2 with brain infarcts in all brain samples in the dorsolateral prefrontal cortex from ROSMAP Full size table Rs77136812 in FMNL2 was among 14 SNPs that were correlated with FMNL2 expression ( β = 0.118, p = 0.027) (Table S7). In addition, SNP rs4664586 was associated with hypomethylation of several CPG sites surrounding the FMNL2 gene, the strongest having a p value of 6.75 × 10 –7 , indicated that one of these variants might also have regulatory effects on expression (Table S8). We then tested the cell-specific expression of FMNL2 using data from single nucleus RNA-sequencing of the frontal cortex in 24 individuals from the ROSMAP cohort. FMNL2 was highly expressed ( p = 0.024) in a subset of astrocytes enriched in interferon response genes, derived from AD brains compared to brain tissue from non-demented individuals. Zebrafish model To determine the putative evolution of FMNL2 expression changes in pathological AD, we used a previously described in vivo zebrafish model of amyloidosis [ 11 , 12 , 13 , 24 ]. The zebrafish model of amyloid β pathology is an acute model where injection of amyloid peptides leads to an intracellular aggregation/polymerization of amyloid peptides (Supplementary Fig. 5), which lead to phenotypes reminiscent of the alterations in mammalian systems (e.g., cell death, synaptic degeneration, inflammation, glial hypertrophy and cognitive impairment) [ 12 , 13 , 14 ]. This model allows to mechanistically investigate the effects of amyloidosis in a vertebrate brain within a short time, and the molecular understanding from zebrafish model is relevant to mammalian models and human AD [ 48 , 50 , 66 , 69 , 77 ]. Based on our existing zebrafish single cell sequencing data, we determined that fmnl2a, a zebrafish ortholog of human FMNL2, is expressed in her4.1 -positive glial cells (Fig. 2 a). To confirm this, we analyzed the bulk RNA-seq data from sorted her4.1 -positive glial in transgenic reporter line Tg( her4.1 :GFP), which marked the astroglial cells in adult zebrafish in control and amyloid-β42-injected brains. We found that amyloid deposition significantly increased the expression of fmnl2a in astroglia (Fig. 2 a) . Fmnl2 is localized to the cell bodies and the projections of the astroglia in the healthy adult zebrafish brain as determined by immunohistochemistry with an antibody cross-reactive to FMNL2 protein (Fig. 2 c, c′). Amyloid β increased the FMNL2 reactivity levels in the astroglia consistent with the gliotic hypertrophy (Fig. 2 d, d′). FMNL2 and her.1-driven GFP expression strongly correlates, indicating that these proteins co-localize in astroglial cells (Fig. 2 g, h). We found that the other ortholog of human FMNL2 gene in zebrafish, fmnl2b, is primarily expressed in neuronal populations as determined by single cell sequencing (Supplementary Fig. 6a). We determined the specificity of the FMNL2 antibody in zebrafish by performing the antibody staining only with the secondary antibody, which resulted in no signal in glial extensions and punctate staining in the parenchyma (Supplementary Fig. 6b), indicating that Fmnl2 immunohistochemistry in zebrafish is detecting true signals. Neuronal and glial Fmnl2 immunoreactivity can be distinguished in neurons and glia by the signal in the radial extensions (glial Fmnl2) and parenchymal perinuclear staining (neuronal Fmnl2) (Supplementary Fig. 6c). Fig. 2 Expression of fmnl2a in zebrafish brain overlaps with astroglia. a Singe cell sequencing tSNE plots for her4.1 and fmnl2a in adult zebrafish telencephalon. b Amyloid toxicity assay and bulk RNA sequencing results for fmnl2a gene expression . c , d Double immunohistochemical staining (dIHCs) for her4.1: GFP (astroglia) and Fmnl2 in control ( c ) and amyloid-injected ( d ) brains with DAPI counterstain. c′ , d′ individual fluorescent channels for Fmnl2. e , f Higher magnification image showing overlapping Fmnl2 and her4.1-driven GFP. g , h Double channel colocalization analyses of e and f . R Pearson correlation coefficient, ρ Spearmann’s rank coefficient. Scale bars as indicated Full size image The radially elongated endfeet of astroglia (S100β) in zebrafish telencephalon coalesce at ventrolateral blood vessels (ZO-1) (Fig. 3 a–a′′), which displayed colocalized astroglial endfeet and tight junctions of the blood vessels (Fig. 3 b, b′). This interaction was verified using a transgenic zebrafish reporter line Tg( kdrl :GFP), which marked the endothelial cells of the blood vessels and S100β immunolabeling that marked the astrocytic endfeet (Fig. 3 c, c′). To determine if these gliovascular interactions are affected in amyloid pathology, we performed immunohistochemical staining in amyloid β-injected animals and found that gliovascular interactions were less pronounced compared to the control (Fig. 3 d, d′). To determine this change in a quantitative manner, we developed an automated image analyses pipeline that determines the colocalization of glial endfeet and blood vessels (Fig. 3 e) and found that the surface area of gliovascular interactions reduced significantly upon amyloid β (Fig. 3 f; Table S9). This result suggest that astroglial-vascular contacts undergo a relaxation upon amyloid β toxicity. Fig. 3 Fmnl2 is required for remodeling of gliovascular interactions. a dIHCs for S100β (astroglia) and ZO-1 (tight junctions/vessels) with DAPI counterstain. a ′, a ′′ Individual channels for the inset in a . b Higher magnification view of dorsoventral gliovascular (GV) junctions. b ′ Close-up of inset in b . Arrows indicate overlaps. c , d dIHCs for kdrl:GFP (blood vessels) and S100β in control ( c ) and amyloid-injected ( d ) brains with DAPI counterstain. c ′, d ′ DAPI omitted from c and d . e Representative snapshot of the automated quantification of the surface of gliovascular interactions. f Quantification results of GV interactions. g , h dIHCs for kdrl:GFP and S100β in amyloid injected brains that were co-injected with control morpholino ( g ) and fmnl2a/b morpholino ( h ) with DAPI counterstain. g ′, h ′ DAPI omitted from g and h . i Close-up image of blood vessels interacting with glial endfeet. j Quantification results of GV interactions. Yellow arrows in c ′, d ′, g ′ and h ′ indicate exemplary gliovascular interactions. Scale bars as indicated Full size image To determine the role of Fmnl2 in the dynamic regulation of gliovascular interactions and amyloid deposition, we knocked down Fmnl2 activity by cerebroventricular microinjection of fmnl2 -targeting morpholinos, which were previously verified in zebrafish [ 91 ] and we also determined to be effective for reducing Fmnl2 expression in zebrafish (Supplementary Fig. 6d). Cerebroventricular microinjection targets mainly the ventricular cells and therefore would effectively knock-down Fmnl2 in astroglia. We found that relaxation of gliovascular interactions upon amyloid toxicity was negatively affected by fmnl2 knock-down in zebrafish brains (Fig. 3 g–j; Table S10), which suggested that gliovascular remodeling in response to amyloid toxicity requires Fmnl2 function. Defects in blood–brain barrier integrity are associated with several chronic diseases. An increase in blood–brain barrier permeability, required for the passage of toxic aggregates from brain to blood as well as for immune cells to infiltrate the brain to attack the toxic entities at earlier stages of disease progression [ 25 ], suggests that stage-specific regulation of gliovascular interactions modulates the cerebrovascular pathology in AD. To determine whether the microglial interactions with the blood vessels are affected by Fmnl2 function, we performed immunohistochemical staining for L-plastin (microglial marker) in relation to the blood vessels (Fig. 4 a–f). We found that while amyloid β injection increased the microglia-vasculature contact compared to control animals, and fmnl2 knock-down reduced the extent of this contact and the number of activated microglia (Fig. 4 g; Table S11). These results suggest that the remodeling of the gliovascular interactions by fmnl2 may be a critical factor for microglial dynamics or leukocyte extravasation from the blood vessels. Fig. 4 Fmnl2 knock-down reduces the number of activated microglia and alters the amyloid load. a – f dIHCs for L-plastin (microglia) and kdrl:GFP with DAPI counterstain on control morpholino ( a , b ), Aβ42+ control morpholino ( c , d ), and Aβ42+ fmnl2a/b morpholino ( e , f ) injected brains. a ′, b ′, c ′ L-plastin channel alone. g Quantification graph for the number of activated microglia in the indicated conditions. h High magnification images from three blood vessels in Aβ42+ control morpholino-injected zebrafish brains (left column) and Aβ42+ fmnl2a/b morpholino-injected brains (right column). Scale bars as indicated Full size image Microglial activity and vascular dynamics are important for clearance mechanisms in the brain. If Fmnl2 knock-down resulted in reduced number of activated microglia, it may have an impact on amyloid clearance as well. Therefore, we analyzed whether Fmnl2 knock-down could alter the clearance and vascular deposition of amyloid β. After acute amyloidosis through microinjection of amyloid, we performed immunostaining for amyloid β, and analyzed the extent of amyloid aggregation near the blood vessels with and without fmnl2 knock-down (Fig. 4 h). We observed that fmnl2 knock-down zebrafish brains showed higher levels of amyloid aggregation around the blood vessels compared to control animals, which could be due to impaired amyloid clearance because of reduced number of activated microglia (Fig. 4 h). These findings suggest that relaxed gliovascular interactions through the activity of Fmnl2 is required for amyloid clearance through bloodstream and/or interaction of the immune cells with the vasculature, but not for blood vessel development (Supplementary Fig. 6d-f; Tables S9 and S12). Mouse model of AD recapitulates gliovascular remodeling and upregulation of Fmnl2 To determine whether chronic amyloid pathology in mammalian models of AD would alter Fmnl2 expression and gliovascular remodeling, we utilized a well-established AD model APP/PS1dE9 [ 39 ] (Fig. 5 ). Compared to age-matched control mice (12 months old) (Fig. 5 a–d), APP/PS1dE9 mice shows extensive amyloid plaques (Fig. 5 e, f), gliosis (Fig. 5 g) and elevated levels of Fmnl2 (Fig. 5 h), supporting our previous findings. To determine whether gliovascular interactions are also altered in mice upon AD pathology, we performed triple immunostaining for Gfap (astroglia), Fmnl2 and Cd31 (blood vessels) (Fig. 5 i–p). We observed that concomitant to elevated levels of Fmnl2 and gliosis, astroglial interactions with the blood vessels are reduced with pathology (Fig. 5 i, m). Fmnl2 staining was quality controlled by excluding the primary antibody, which confirmed that the signal for Fmnl2 is specific (Supplementary Fig. 7). These results indicate that in acute or chronic amyloid pathology models in zebrafish and mouse, upregulation of Fmnl2 and reduction in gliovascular interactions are consistent. Fig. 5 Upregulated Fmnl2 and altered gliovascular interactions observed in chronic APP/PS1dE9 AD model in mice. a , h Triple immunohistochemical stainings (tIHC) for Gfap (astroglia), 4G8 (amyloid plaques) and Fmnl2 with DAPI counterstain in control ( a , d ) and APP/PS1dE9 mice ( e – h ) at 12 months of age. Panels are from cerebral cortex. Black-white panels are individual fluorescence channels of the composite images in a and e . i–p tIHC for Gfap, Cd31 (blood vessel) and Fmnl2 with DAPI counterstain in control ( i – l ) and APP/PS1dE9 mice ( m – p ) at 12 months of age. Panels are from cerebral cortex. Black-white panels are individual fluorescence channels of the composite images in i and m . Scale bars as indicated Full size image AD pathology is strongly associated with a gliotic response in astroglia, and the gliovascular remodeling response could be a generic outcome due to gliosis. To test whether Fmnl2 upregulation and relaxation of the gliovascular interactions are dependent on gliosis, we performed traumatic injury in the cerebral cortex of healthy mouse brains (Supplementary Fig. 8). Compared to control mice, injured brains showed extensive gliotic response; yet the gliovascular interactions and the levels of Fmnl2 were not significantly altered (Supplementary Fig. 8). These results suggest that the remodeling of gliovascular interactions and concomitant upregulation of Fmnl2 could be a specific response to amyloid pathology. Validation in post-mortem human brains Functional studies imply that vascular risk factors would increase the expression of FMNL2 in humans with AD. To test this hypothesis, we performed immunohistochemical staining in post-mortem human brains for FMNL2 and GFAP (Fig. 6 ) (control, AD, Primary age-related Tauopathy [PART] with or without cardiovascular pathology, Tables S12 and S13). We observed that in control brains, FMNL2 expression was punctate and scarce around the blood vessels (Fig. 6 a, b), while in early onset AD and in AD with severe atherosclerosis, there was significant FMNL2 expression that delineated the blood vessels (Fig. 6 c–f). In these brains, astroglia were reactive and in some regions glial endfeet were detached from the vessel. In patients with primary age-related tauopathy, FMNL2 localizes around the blood vessels albeit less prominent than in AD (Fig. 6 g, h). The tight associations of astroglia to the blood vessels inversely correlate with FMNL2 expression (Fig. 6 i). Quantification of FMNL2 expression in control and patients showed a significant increase in FMNL2 in astroglial cells in disease conditions (Fig. 6 j), supporting the findings that FMNL2 is a response to AD and is involved in regulating the gliovascular interactions. Fig. 6 FMNL2 expression is upregulated at gliovascular junctions in AD patients. a – h dIHC for FMNL2 and GFAP in control ( a , b ), early onset AD patient ( c , d ), AD patient with severe atherosclerosis ( e , f ), and primary age-related Tauopathy (PART) with mild atherosclerosis patient ( g , h ). Black and white images are individual fluorescent channels of the leftmost panels for FMNL2 and GFAP. Rightmost panels ( b , d , f , h ) are higher-magnification composite images around a representative blood vessel. i Two examples of astroglia (GFAP+, green) contacting blood vessels (bv), and their expression of FMNL2 (violet). Arrows indicate contact with blood vessels. In AD, astroglia express FMNL2 and its endfeet is retracted from the blood vessel, while in controls tight association is observed. Black-white panels indicate individual fluorescence channels. j Arivis-based automated image quantification pipeline and the quantification graph for FMNL2 expression in human brains. Comparisons are for every diseased brain to control. with Dunn's Kruskal–Wallis test. Scale bars as indicated Full size image Discussion A major effort to identify genes and loci associated with AD risk has been underway for several decades [ 41 , 54 ]. Many of these genes have been replicated and confirmed in other studies [ 15 , 53 ]. The focus has now shifted to fully understand mechanisms by which variants in genes lead to disease. Epidemiological and neuropathological research has established an intertwined relationship between AD and cerebrovascular disease and its antecedents. Yet, our understanding of the course of this inter-dependent relationship is unclear. To probe the interaction between CVRFs, cerebrovascular disease and AD based on the genotype–phenotype relationship, we used data from five cohorts of different ancestries to perform an unbiased genome-wide association study of the interaction effects of CVRFs and AD genetic variants. We found that FMNL2 , formin-like protein 2, significantly interacted with the CVRF score to modify AD risk. We also found that the association between the CVRF scores and AD differed by SNP genotypes within FMNL2. FMNL2 brain expression was increased in pathological AD, with Aβ and tau deposition and with brain infarcts independently. The mediation analysis allowed us to hypothesize that the effects of FMNL2 expression on pathological AD were related to accumulation of Aβ and tau deposition. Furthermore, we found that FMNL2 to be highly expressed in astrocytes, a key cell type of interest in AD. SNPs in FMNL2 interacting with CVRF are also correlated with gene expression and hypo-methylation levels suggesting a regulatory effect of common polymorphisms leading to increased FMNL2 expression. A strength of the current study is the inclusion of three ethnic groups, which is particularly important in the study of genetics and cerebrovascular disease, for two reasons. First, specific variants associated with disease within a gene can differ between ethnic groups based on linkage disequilibrium (LD) structure. Identifying significant SNPs through a meta-analysis across cohorts and collapsing SNPs within gene for gene-based analyses increases the power of discovery and generalizability of the results. Second, there are differences in the frequency of cerebrovascular disease and associated CVRFs between non-Hispanic Whites, African–Americans, and Caribbean Hispanics. Clinically, these results could have important implications for personalized treatment plans, suggesting more aggressive treatments of cardiovascular and cerebrovascular risk factors in persons with specified genotypes. Third, we bring together data from three species—human, mouse, and zebrafish—and demonstrate a validation-oriented functional genomics pipeline to analyze the biological mechanisms of action for increasing number of AD-related genes [ 6 ] and their evolutionary conservation. Finally, we also propose zebrafish as a useful experimental model for amyloid β pathology, an important aspect of AD. Previous studies of gene-environment interactions in AD have focused on candidate genes, stemming from prior genome-wide studies of AD or cerebrovascular disease (e.g., [ 19 , 95 ]). Here we focused on a genome-wide, unbiased interaction approach to identify novel genes that interact with CVRFs in AD. Hypertension [ 29 , 87 ], type II diabetes [ 89 ] and obesity [ 5 ] have each been associated with AD, cerebrovascular disease and alterations in the blood–brain-barrier integrity. The resulting microvascular dysfunction increases the permeability of the blood–brain barrier, oxidative stress, activation of immune mechanisms and leakage of proteins. Additionally, vascular defects generate hypoxic conditions that enhance the toxic protein aggregation [ 73 , 80 ]. Therefore, these changes increase the demand for waste removal mechanisms in the brain. How do variants in FMNL2 contribute to the cerebrovascular pathology in AD? FMNL2 encodes a formin-related protein. Formins are important in regulating actin and microtubules, with cellular effects and consequences in cell morphology, cytoskeletal organization and cell–cell contact [ 30 ] and are implicated in AD pathology [ 35 , 68 ]. Based on our results here, we propose that FMNL2 regulates cell–cell interactions between glia and vasculature, and the clearance of extracellular aggregates. Astroglial endfeet form a perivascular space that is involved in arterial pulsation-dependent clearance [ 10 ]. The size of the astroglia-regulated interstitial space correlates with the drainage efficiency of the brain, for instance sleep-dependent adrenergic signaling expands the glympathic space [ 76 , 98 ]. However, with a higher load of extracellular protein aggregation, more efficient mechanisms such as the clearance through blood–brain barrier [ 84 ], where major amyloid-binding efflux receptor complexes including CLU, LRP1/2 and VLDLR are present [ 97 ], is required. FMNL2 -mediated gliovascular remodeling could be a mechanism to potentiate the clearance through blood–brain barrier. Notwithstanding the elevated protein aggregation, clearance mechanisms become less efficient with aging and the efflux receptor expression increases [ 67 ]. FMNL2 -mediated enlargement of perivascular space can serve to facilitate both glympathic and blood–brain barrier clearance mechanisms. Recent findings confirm our hypothesis that perivascular remodeling is a critical response to amyloid pathology [ 21 , 46 ] and could act as the interface between CVRFs and AD. A recent single cell transcriptomics analyses also confirmed our findings that FMNL2 is upregulated in astroglial cell populations in human AD brains [ 37 ]. Additionally, detached gliovascular interactions can lead to immune cell infiltration to the brain from the blood and may help the initial attempts of amyloid clearance [ 18 , 36 ], which is supported by our findings that fmnl2 knock-down reduces the microglia that contact the blood vessels in zebrafish. The long-term consequence of gliovascular detachment is unclear [ 47 , 82 ], as breached blood–brain-barrier could inadvertently lead to exacerbation of the disease pathology. Additionally, whether the gliovascular endfeet retraction is a consequence of a generic gliotic response or is part of a disease-associated mechanism is debated. Our results showed that amyloid aggregation but not the traumatic injury induced FMNL2 expression, suggesting that FMNL2-regulated astroglial endfeet remodeling could be preferentially related to AD pathophysiology. A recent study demonstrating that selective focal ablation of astroglia endfeet around the blood vessels does not cause blood–brain-barrier disruption and gliovascular contacts are efficiently replaced [ 62 ] supports our findings. In this study, we experimentally showed that (1) FMNL2 is expressed in astroglial cells, (2) FMNL2 is upregulated with amyloid deposition in zebrafish, mouse, and human brains, (3) gliovascular interactions loosen with the disease pathology in three species, (4) this response requires FMNL2, and (5) loss of FMNL2 function leads to amyloid deposition and reduced immune response. Post-mortem human brain immunohistochemical staining also confirmed that compared to control individuals, patients with AD have a significant increase in the expression of FMNL2 near the glial endfeet at the blood vessels and in astrocytes that appear hypertrophic and reactive. We also observed that glial end feet were arrayed around the blood vessels in controls while this arrayed structure is impaired and show a dilated appearance concomitant to prominent FMNL2 expression. Based on these results, we propose that FMNL2 functions in astroglial cells to regulate the extent of gliovascular interactions and toxic protein clearance from the brain on demand. The role of FMNL2 may also be required in other proteinopathies such as progressive age-related tauopathy, and FMNL2 is also upregulated and delineated in blood vessel structures. Gliovascular contacts are important components of the blood–brain barrier, and they partake in sustained homeostasis and clearance of amyloid β [ 99 , 102 ]. The role of FMNL2 in amyloid clearance is supported by the findings where the knock-down of FMNL2 protein resulted in the defective sorting into late endosomes and lysosomes [ 43 ]. Abnormalities in the endosomal-lysosomal network are known to be important to the mechanism of neurodegeneration and AD [ 63 ]. We do not exclude that FMNL2 could function in other cell types such as pericytes around the blood vessels to regulate the clearance and drainage mechanisms as well as to control the immune response either by altering microglia–blood vessel interactions or extravasation of immune cells into the brain. Future studies aiming at elucidating disease stage and cell-specific roles of FMNL2 will be instrumental to further delineate the role of this protein and to propose new drug candidates. The animal models that we used in this study—zebrafish and mouse—reflect the acute and chronic pathological changes and cellular response to amyloidosis. We recognize that the complex late-stage cognitive decline in AD may not be fully represented in these model systems. However, the post-mortem human tissue analyses, AD cohort studies and single cell datasets are relevant to the cognitive decline in AD through clinical (e.g., cognitive assessment scores) and neuropathological (e.g., Braak staging and histological analyses) aspects. Our hypothesis favors for an FMNL2-related pathological mechanism that is likely to initiate during the early course of the disease and extend toward the later stages. In all our investigations, the dynamics of FMNL2 was consistent either in acute zebrafish model, chronic mouse model or post-mortem human brains with documented clinical and pathological AD. This suggests that the biological role of FMNL2 and the cellular interactions between neuro-glio-vascular niche cells could be an early response to the disease and extends toward later stages even when cognitive decline manifests. The relationship of clearance to the cognitive decline and brain atrophy is yet to be investigated. FMNL2 is most highly expressed in the brain. In a study of ischemic stroke in young individuals ( n = 1816), one SNP was found to be associated (rs2304556 p = 1.18 × 10 –7 ) [ 23 ]. Interestingly, we found that increased FMNL2 expression (ENSG00000157827) was associated with brain infarcts and AD independently, suggesting that this gene may be involved in a shared molecular mechanism with cerebrovascular disease. The SNP with the highest significant association in our study, rs57223657, is a point mutation in the third intron of FMNL2 gene, where a promoter regulatory flank region resides. Intronic regions that contain gene regulatory, or enhancer functions are known to be critical determinants of gene expression and function [ 71 , 72 ]. Thus, it is possible that regulatory variants of FMNL2 may also alter gene function and the extent of the AD pathology at least through the regulation of gliovascular interactions. The multidisciplinary work reported here demonstrates the unique involvement of genes, FMNL2 , and antecedent CVRFs in AD with human GWAS and in vivo animal model data. While this study has several strengths, we recognize that there are also some limitations. First, we relied on self-report of three of the four vascular risk factors. While this approach is reliable and valid [ 45 ] it is also possible some individuals responded incorrectly. However, we did measure the fourth, BMI, directly. Second, we did not consider the possible effects of medications used to treat vascular risk factors, such as hypertension or diabetes. We were unable to define more deeply some of the cardiovascular diseases. In some groups, investigators only reported “any heart disease”, while others included information on the specific heart condition. We used a transient knock-down of FMNL2 function in an acute amyloidosis zebrafish model, and a model of chronic amyloid pathology in mouse. A gene editing approach coupled to chronic models to mimic human FMNL2 SNP variants will help to capture long-term effects of FMNL2 on vascular risk factors and AD pathology in the zebrafish or mouse experimental models. Taken together, this investigation implies that hypertension, diabetes, heart disease and obesity are more likely to have acquired cerebrovascular disease in the brain and are at higher risk for AD. In these patients, FMNL2 expression resulting from the vascular risk factors and the resulting cerebrovascular pathology regulate the ongoing deposition of amyloid and tau proteins by altering the normal interplay between glia and the vasculature and ultimately the clearance of extracellular aggregates (Supplementary Fig. 9). Our results also indicate that FMNL2 expression is upregulated in the presence of CVRFs among individuals who develop AD, and that having both conditions augment the effect of FMNL2 on AD pathology. The sequence of events—whether CVRF or AD comes first—will be better addressed with longitudinal experimental models for both diseases, yet our mediation analyses and functional experiments suggest that CVRF potentiates the AD pathology. Expanding on the relationship of genes and loci identified here, along with expanded basic molecular work will move us closer to conceptualizing the etiological interaction between AD and cerebrovascular disease and hopefully identifying therapeutic targets for AD. Data availability Genetic data for all cohorts is available at the National Institute on Aging Genetics of Alzheimer’s Disease Data Storage Site (NIAGADS; ). Primary clinical data for WHICAP and EFIGA data are available at . Clinical data from NACC data is accessible on request at . Clinical data from ROSMAP data is accessible on request at . For the RNA sequencing data, gene-level transcript values in counts and normalized expression data from ROSMAP are available on the AMP-AD Knowledge Portal at Synapse ( ) (Synapse: syn25741873). The residual expressions from gene transcripts are available at Synapse (ROSMAP: syn25741873, MSBB: syn8485027, and Mayo: syn8466826). The ROSMAP single nucleus RNA sequencing data and the description of the data are available at Synapse (Synapse ID: syn16780177). Expression datasets for zebrafish are publicly available as NCBI GEO accessions GSE74326 and GSE161834. Code availability We used publicly available statistical software to perform analyses in the Methods section. Code to perform analyses in this manuscript are available from the author upon request (A.J.L.).

For more than 20 years, scientists have known that people with hypertension, diabetes, high cholesterol, or obesity have a higher likelihood of developing Alzheimer's disease. The conditions can all affect the brain, damaging blood vessels and leading to strokes. But the connection between vascular disease in the brain and Alzheimer's has remained unexplained despite the intense efforts of researchers. Now, a study published in Acta Neuropathologica and led by researchers at Columbia University's Vagelos College of Physicians and Surgeons has uncovered a possible mechanism. The study found that a gene called FMNL2 links cerebrovascular disease and Alzheimer's and suggests changes in FMNL2 activity caused by cerebrovascular disease prevent the efficient clearance of toxic proteins from the brain, eventually leading to Alzheimer's disease. The finding could lead to a way to prevent Alzheimer's in people with hypertension, diabetes, obesity, or heart disease. "Not only do we have a gene, but we have a potential mechanism," says senior author Richard Mayeux, MD, chair of neurology at Columbia and NewYork-Presbyterian/Columbia University Irving Medical Center. "People have been trying to figure this out for a couple of decades, and I think we have our foot in the door now. We feel there must be other genes involved and that we've just scratched the surface." Mayeux and his colleagues found FMNL2 in a genome-wide hunt designed to uncover genes associated with both vascular risk factors and Alzheimer's disease. The search involved five groups of patients representing different ethnic groups. One gene, FMNL2, stood out during the analysis. But what role it could possibly play was unclear. That's when Caghan Kizil, Ph.D., a visiting associate professor at Columbia, leveraged his expertise with zebrafish as a model organism for Alzheimer's disease. FMNL2 and the blood-brain barrier "We had this gene, FMNL2, that was lying at the interface between Alzheimer's disease in the brain and cerebrovascular risk factors," says Kizil. "So we had an idea that FMNL2 might operate in the blood-brain barrier, where brain cells meet the vasculature." The blood-brain barrier is a semi-permeable, highly controlled border between capillaries and brain tissue that serves as a defense against disease-causing pathogens and toxins in the blood. Astrocytes, a specialized type of brain cell, compose and maintain the structure of the blood-brain barrier by forming a protective sheath around the blood vessel. This astrocyte sheath needs to loosen for the clearance of toxic amyloid—the aggregates of proteins that accumulate in the brain and lead to Alzheimer's disease. The zebrafish model confirmed the presence of FMNL2 in the astrocyte sheath, which retracted its grip on the blood vessel once toxic proteins were injected into the brain, presumably to allow for clearance. When Kizil and his colleagues blocked the function of FMNL2, this retraction did not occur, preventing clearance of amyloid from the brain. The same process was then confirmed using transgenic mice with Alzheimer's disease. The same process may also occur in the human brain. The researchers studied postmortem human brains and found increased expression of FMNL2 in people with Alzheimer's disease, along with breach of the blood-brain barrier and retraction of the astrocytes. Based on these findings, the researchers propose that FMNL2 opens the blood-brain-barrier—by controlling its astrocytes—and promotes the clearance of extracellular aggregates from the brain. And that cerebrovascular disease, by interacting with FMNL2, reduces the clearance of amyloid in the brain. The team is currently in the process of investigating other genes that could be involved in the interplay between Alzheimer's and cerebrovascular disease, which, along with FMNL2, could provide future approaches for drug development.

10.1007/s00401-022-02431-6

Medicine

Study uncovers the molecular events by which popular antidepressants work

Revathy U. Chottekalapanda et al. AP-1 controls the p11-dependent antidepressant response, Molecular Psychiatry (2020). DOI: 10.1038/s41380-020-0767-8 Journal information: Molecular Psychiatry

http://dx.doi.org/10.1038/s41380-020-0767-8

https://medicalxpress.com/news/2020-08-uncovers-molecular-events-popular-antidepressants.html

Abstract Selective serotonin reuptake inhibitors (SSRIs) are the most widely prescribed drugs for mood disorders. While the mechanism of SSRI action is still unknown, SSRIs are thought to exert therapeutic effects by elevating extracellular serotonin levels in the brain, and remodel the structural and functional alterations dysregulated during depression. To determine their precise mode of action, we tested whether such neuroadaptive processes are modulated by regulation of specific gene expression programs. Here we identify a transcriptional program regulated by activator protein-1 (AP-1) complex, formed by c-Fos and c-Jun that is selectively activated prior to the onset of the chronic SSRI response. The AP-1 transcriptional program modulates the expression of key neuronal remodeling genes, including S100a10 (p11), linking neuronal plasticity to the antidepressant response. We find that AP-1 function is required for the antidepressant effect in vivo. Furthermore, we demonstrate how neurochemical pathways of BDNF and FGF2, through the MAPK, PI3K, and JNK cascades, regulate AP-1 function to mediate the beneficial effects of the antidepressant response. Here we put forth a sequential molecular network to track the antidepressant response and provide a new avenue that could be used to accelerate or potentiate antidepressant responses by triggering neuroplasticity. Introduction Major depressive disorder (MDD) is a disabling psychiatric disorder with diverse etiology, and is a leading cause of mortality and morbidity worldwide [ 1 ]. The symptoms of depression are varied, and include cognitive, motivational, emotional, and physiological changes [ 2 ]. Evidence for molecular and cellular alterations in depression includes reduction in neuroplastic properties, reduced levels of neurotrophins, and decreased neurogenesis [ 3 , 4 , 5 ]. Clinical imaging and postmortem studies indicate functional and structural alterations in the limbic brain regions, prefrontal cortex (PFC), hippocampus, and amygdala [ 4 , 6 ] in depressed patients. Many psychopharmacological agents are currently used for the treatment of depression, among which the selective serotonin reuptake inhibitors (SSRIs) are the most effective and widely prescribed [ 7 ]. However, clinical studies show that only two-thirds of patients respond to treatment, and that treatment has a delayed onset of action [ 8 ]. The therapeutic response to chronic antidepressant treatment is thought to be mediated by neuroadaptive changes in specific neuronal networks, which collectively reverses the dysregulation caused by depression [ 3 ]. There is therefore a pressing need to elucidate the identity of molecular pathways that mediate long-term treatment and antidepressant efficacy. Gene expression regulation represents a general mechanism through which mature neuronal circuits control their physiology and behavior [ 9 , 10 ]. Several transcription factors have been identified as the first responders to extracellular signals. Such factors include the immediate early genes (IEGs) such as c-Fos, FosB, Fra1, Fra2, c-Jun, Junb, Jund, Egr1, Nr4a1 , and Npas4 . These IEGs initiate and maintain gene expression profiles of crucial effector proteins, causing sustainable changes in the structure and function of mature synapses and thereby promoting behavioral plasticity. So far, among the IEGs, Creb1 and Egr1 have been previously shown to be associated with the antidepressant response [ 11 , 12 , 13 ]. Additionally, activation of cAMP, PKA, CAMK, and BDNF signaling molecules have been implicated in the chronic antidepressant response [ 14 , 15 , 16 ], likely through the regulation of physiological processes, including neuroplasticity, neuroprotection, and neurogenesis. However, how drug treatment is specifically coupled to transcription and how the signaling molecules are activated in response to chronic antidepressant administration are still unclear. Within the corticolimbic network [ 17 ], we chose to focus on the cortex, as the integral role of this brain-region in the regulation of behavior and the control of stress reactivity has been well characterized in patients and animal models [ 18 , 19 , 20 , 21 ]. Additionally, as the primary role of the serotonin-dependent function of the cortical circuit in the effective treatment of depression has been well established, studying the cortex would allow us to further delineate the molecular mechanisms regulating the complex response to antidepressants. In this study, we investigated which of the IEGs are activated in the cortex by chronic treatment with an SSRI, fluoxetine, and we elucidated the target genes regulated by these factors. Furthermore, we addressed whether the IEGs and their target genes contribute directly to the behavioral response. Our findings reveal activation of a network of molecules that are sequentially linked together to provide a robust antidepressant response. Method details Treatments, transfections, and DNA constructs For chronic drug and JNK inhibitor treatment, BALB/cJ mice (Jackson Laboratories) were housed two per cage and fluoxetine hydrochloride (Sigma) at a dose of (0.167 mg/mL) was administered in drinking water in 1% saccharine solution to mask the taste of the drug. Saccharine alone was given to the vehicle-treated animals. Mice were treated on average for 28 days and replaced with fresh solution every 3 days. On average, the fluoxetine-receiving mice drank approximately ∼ 3–4 mL a day, somewhat less compared with the control mice (that received 1% saccharine), which drank approximately ∼ 5–7 mL a day presumably due to the taste of the drug. The drinking volume of the fluoxetine-treated mice eventually normalized to that of the control mice. The fluoxetine-treated mice thus received 16–23 mg/kg/day of fluoxetine, an effective dose that is known to produce an antidepressant response in different strains of mice [ 22 ]. BALB/cJ mice were used as these mice are inherently anxious and show a robust antidepressant response. The S100a10-EGFP/Rpl10a ES691 mice of C57BL/6 background also produces a robust antidepressant response to fluoxetine treatment. For JNK inhibitor treatment in vivo, 16 mg/kg was injected intraperitoneally (i.p.) to the BALB/cJ mice on days 4, 6, 8, 10 of fluoxetine treatment to block JNK function to block c-Jun phosphorylation. For growth factor stimulation experiments in PC12/TrkB cells and primary mixed cortical neurons, growth factors BDNF (50 ng/mL) [ 23 , 24 ], FGF2 (50 ng/mL) [ 25 ], EGF (100 ng/mL) [ 26 ], IGF (100 ng/mL) [ 27 , 28 ], NGF (100 ng/mL) [ 29 ], VEGF (100 ng/mL) [ 30 ], BMP4 (100 ng/mL) [ 31 ], TGFβ (100 ng/mL) [ 32 ], Bicuculline (50 μM) [ 33 ], and KCl (55 mM) [ 34 ], all from Sigma, were acutely applied onto cells. Samples were collected at 2 h to assess c-Fos or c-Jun expression, and samples were collected at 12 and 24 h for p11 expression in PC12-TrkB cells and primary mixed cortical neurons, respectively. For the BDNF- and FGF2-stimulated time course experiments, samples were collected at 2, 6, 12, 24, and 48 h after an acute application. The experiments were conducted in a 12-well plate with at least three biological replicates. For inhibitor experiments in primary cortical neuronal cultures , we tested which pathways affect basal c-Fos and c-Jun transcription. We applied inhibitors to block the various molecules in the tyrosine kinase pathway. To test the BDNF- and FGF2-inducible c-Fos and c-Jun transcription, we treated cells with inhibitors 30 min before stimulation. We used inhibitor concentrations that were previously known to induce neurogenic and neurotrophic effects: TrkB (K252a, 1 μM) [ 35 ], MAP kinase kinase, MEK1 and MEK2 (U0126, 10 μM, EMD Millipore) [ 36 ], PI3K (LY294002, 50 μM) [ 35 ], PLCγ (U73122, 10 μM) [ 37 ], p38 MAPK (SB203850, 20 μM) [ 38 ], and JNK (SP600125, 20 μM) [ 39 ]. All inhibitors were bought from Sigma unless otherwise mentioned. The experiments were conducted in a 12-well plate with at least three biological replicates. For transfections of transcription factor small interfering RNA (siRNA ), experiments were done in PC12-TrkB cells. To identify the transcription factor regulating p11, inhibition of transcription factors were done upon treatment with two pre-validated silencer select siRNAs (Thermo Fischer Scientific) for each transcription factor according to the manufacturer’s instructions. siRNAs were transfected using Lipofectamine RNAiMAX reagent (Thermo Fischer Scientific) according to the manufacturer’s instructions, and expression levels of the transcription factor and S100a10 was analyzed by quantitative PCR, 48 h after transfection. siRNA efficciency was calculated based on the downregulation of the transcription factor by its specific siRNA. The experiments were conducted in a 12-well plate with at least three biological replicates. The siRNAs used are Atf3 , s129666, s129668, Foxo1 , s136654, s136655, Lrrfip1 , s173268, s173270, c-Myc , s128068, s128069, Nfkb , s135615, s135616, Stat5a s128672, s128673, Egr1 , s127689, s127690, Ets1 , s127719, s127721, Junb , s127982, s127983, Jund , 201434, 201435, Fosl1 , s217983, s129817, Fosl2 , 197390, 197391, Bhlhe40 , s135199, s135201, Yy1 , s128676, s128677, Sp1 , s128429, s128430, Srf , s180122, s180123, Creb1 , s135438, s135439, Crem , s130285, s234984, Stat3 , s129048, s129047, c-Fos s66197, s66198, c-Jun , s68563, and s201552. Luciferase reporter assays Assays were performed in N2A cells to identify the functional S100a10 promoter using the Dual-Luciferase Reporter Assay system (Promega). Exon 1 of S100a10 with varying lengths of 3′UTR region were cloned upstream of a firefly luciferase reporter construct in a mammalian expression vector pGL4.11 using the SLIC method (Li M 2007). We identified the functional S100a10 promoter reflected by robust luciferase activity with the coordinates, chromosome3: 93554373–93555181, based on mouse genome assembly, mm10. Next, we identified three c-Fos and c-Jun binding motifs within this region, and mutated the three sites named mutation1, mutation2, and mutation3 and deleted mutation3. The mutated and deleted firefly luciferase constructs were transfected into confluent Neuro-2a (N2a) cells grown in 12-well culture plates using lipofectamine LTX with Plus reagent (Thermo Fischer Scientific). The renilla luciferase gene was cotransfected as a control for transfection and expression efficiency. Cells were harvested and lysed 48 h after transfection and cell lysates were assayed for firefly and renilla luciferase activity according to the manufacturer’s instructions. Firefly luciferase activity was normalized to renilla luciferase activity for each cell culture well and plotted as activity relative to control transfections. At least four biological replicates were used. For dual transfections, c-Fos and c-Jun siRNA was first transfected using RNAiMAX and then 12 h later, DNA constructs were transfected using lipofectamine LTX with Plus reagent. The cells were harvested 36 h later, when the siRNA effect is maximum. The sequences of all plasmids were verified by sequencing and restriction enzyme digestion. Western blotting For tissue analysis, mice were anesthetized with CO 2 and decapitated, and cortex was rapidly dissected, frozen in liquid nitrogen, and stored at −80 °C until further processing. For cell analysis, cells were scraped and collected, rinsed with PBS, flash frozen in liquid nitrogen, and stored at −80 °C. For both tissues and cell pellet, samples were sonicated at 4 °C in lysis buffer containing 20 mM Tris pH 7.5, 150 mM NaCl, 1% SDS supplemented with protease inhibitor (Roche) and phosphatase inhibitor (Roche), and boiled for 10 min. The protein concentration was determined using a BCA protein assay kit (Thermo Fisher Scientific) according to the manufacturer’s instructions. Protein samples were diluted in equal volume of 2× LDS sample buffer (Invitrogen) and supplemented with DTT to a final concentration of 200 mM (Sigma). Twenty micrograms of protein samples were separated on 4–12% Bolt Bis-Tris precast denaturing gels (Invitrogen) and transferred onto PVDF membranes and blocked with 5% milk in TBS–0.1% Tween (TBST) solution for 1 h at room temperature. Membranes were probed with primary antibodies diluted in 5% milk–TBST solution overnight at 4 °C. Membranes were then washed and probed with horseradish-peroxidase-conjugated anti-mouse (Thermo Fischer Scientific, 31460, 1:10,000), anti-rabbit (Thermo Fischer Scientific 31430, 1:10,000), or anti-goat antibody (Jackson Immunoresearch, 305-035-003, 1:10,000) for 1 h at room temperature. Membranes were developed using Pierce Western blotting substrate (Thermo Fischer Sceintific, 32106) and exposed on film. Exposed films were scanned, and protein bands were quantified using ImageJ Software (NIH, USA). Protein quantities were normalized using GAPDH. All values were plotted relative to control/untreated samples. Antisera and antibodies against the following proteins were purchased from the indicated sources: p11 (AF2377, R&D systems, 1:500), GAPDH (Mab374, EMD Millipore 1:1000), total c-Fos (4384, Cell Signaling Technology, 1:300), phospho c-Fos (5348S, Cell Signaling Technology, 1:250), c-Jun (9165S, Cell Signaling Technology, 1:500), phospho c-Jun (3270S, Cell Signaling Technology, 1:500). Chromatin immunoprecipitation Antibodies against c-Fos (sc-52, Santa Cruz Biotechnology), c-Jun (9165S, Cell Signaling Technology), and rabbit IgG antibody (C15410206, Diagenode) were bound to Protein-G magnetic beads (Diagenode, kch-818-220) for 2 h, at 4 °C (30 μL Protein-G magnetic beads were incubated with 5 μg of Jun antibody, 10 μg of Fos antibody, and 1 μg of IgG antibody. A total of eight frontal cortices from four S100a10-EGFP/Rpl10a ES691 mouse brains from 9-day vehicle- and fluoxetine-treated animals were pooled. They were briefly washed with ice-cold 1× PBS and 1 mM MgCl 2 . Tissue was transferred to a dounce homogenizer and buffer containing 1% formaldehyde, 50 mM HEPES-KOH, pH 7.5, 100 mM NaCl, 1 mM EDTA, 0.5 mM EGTA was added and fixed for 10 min end-to-end rotation. Formaldehyde was then quenched by adding 0.125 mM of glycine for 5 min. Halt-protease and phosphatase inhibiter (Thermo Fischer Scientific and Roche) and chromatin immunoprecipitation (ChIP) cross-link Gold (Diagenode) were added to the buffer before homogenization. Throughout the protocol buffers were supplemented with protease and phosphatase inhibitors. Lysate was then subjected to two wash steps with the same 1× PBS and 1 mM MgCl 2 at 1350 g at 4 °C for 5 min. Each sample was washed with lysis buffer 1 containing 50 mM HEPES-KOH, pH 7.5, 140 mM NaCl, 1 mM EDTA, 10% glycerol, 0.2% NP-40, 0.2% Triton X-100 after incubating for 10 min at 4 °C. The pellet was then washed with lysis buffer 2 containing 10 mM Tris-HCl, pH 8.0, 140 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, after incubating at RT for 10 min. The pellet was resuspended in 1 mL shearing buffer at 1 × 10 7 cells, containing 10 mM Tris-HCl, pH 8.0, 140 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 0.1% deoxycholate, 0.2% N-lauroylsarcosine, and 0.2% SDS. The samples were sonicated in Covaris settings duty cycle 10% and 175 cycles/burst for 3 min to achieve fragments of sizes ranging from 200 to 500 bp. To each sample, 0.2% Triton X-100 was added and spun at 20,000 g @ 4 °C for 15 min. The supernatant was collected; 1% of the input sample of each ChIP reaction was collected. The sheared chromatin was immunoprecipitated with washed antibody-bound beads at 4 °C overnight. The antibody coupled beads were washed 2× with low-salt buffer containing 50 mM HEPES-KOH, pH 7.6, 100 mM LiCl, 0.5 mM EDTA, 0.2% NP-40, 0.7% deoxycholate, and then washed 2× with 50 mM HEPES-KOH, pH 7.6, 100 mM LiCl, 0.5 mM EDTA, 0.2% NP-40, 0.7% deoxycholate. Then the beads were washed 2× with 10 mM Tris-HCl, pH 8.0, 50 mM NaCl, 1 mM EDTA, and eluted with 50 mM Tris-HCl, pH 8.0, 10 mM EDTA, 10% SDS buffer and incubated at 65 °C for 1 h. The isolated chromatin and the input samples were immediately cleaned with the IPure kit (Diagenode). The unbound fraction of the immunoprecipitation reaction was used to validate the fragmentation on a 1% agarose gel. Purification with the IPure kit was carried out according to the manufacturer’s instructions with the addition of Proteinase K treatment (Roche) for 1 h at 55 °C after the de-crosslinking step. Immune-enriched chromatin was further purified with phenol:chloroform:isoamyl alcohol (25:24:1) and concentrated by ethanol precipitation. ChIP-sequencing libraries were prepared using the Ovation Ultralow system V2 (Nugen). The samples were further validated, and processed for single reads 75 bp sequencing on the Illumina HiSeq 2500 platform. Genome-wide sequencing The ChIP-seq library samples were further validated, and processed for single read 75 bp sequencing on the Illumina HiSeq 2500 platform. FastQC was invoked for the sequencing quality control. Good-quality reads were aligned to mouse genome (mm10) with bowtie v1.0.1 (parameter “--best –strata” for keeping only the best hit, default parameters otherwise). Peak regions were called with MACS2 (parameters “-g mm -q 0.01”, Minimum FDR cut-off 0.01 for peak detection, default parameters otherwise) and reads were extended by 200 bp to account for the size of the fragment isolated by the ChIP reaction. Three-way comparisons were performed to identify locations of differential enrichment in the genome between two conditions. This was done by using the tool bdgdiff in MACS2 (parameters “-g 60 -l 120” and “--d1/--d2 according to the callpeak output, default parameters otherwise). IGVTools (IGV Version 2.3.52) were used to convert the pileup peak files into binary tdf files for viewing in IGV. Possible promoter and closest gene of the differential enriched peaks were annotated by R package ChIPpeakAnno; peaks overlapped 5kbp upstream or downstream of a TSS were annotated as possible promoters. We invoke MEME (Version 4.12.0 with parameter -dna -nmotifs 16 -p 8, default parameters otherwise) to discover the motifs with the sequences from the peak regions. RNA purification and quantitative PCR Mice were anesthetized with CO 2 and decapitated. The PFC or the whole cortex were rapidly dissected, frozen in liquid nitrogen, and stored at −80 °C. RNA extraction from frozen samples was performed using the Trizol/chloroform technique according to the manufacturer’s instructions (Thermo Fischer Scientific). After extraction, RNA was precipitated overnight at −80 °C in isopropanol with 0.15 M sodium acetate and Glycoblue (Ambion, Austin, TX), washed twice with 80% ethanol, air-dried, and resuspended in nuclease-free water. Purified samples were analyzed using a Nanodrop 1000 spectrophotometer (Thermo Fischer Scientific) in order to assess mRNA quantity and quality. cDNA was prepared from DNase-treated total RNA using the High Capacity RNA-to-cDNA kit (Thermo Fischer Scientific). Relative gene expression of the cDNA was assayed by qRT-PCR using pre-designed recommended Taqman gene expression assays from Applied Biosystems (ABI) following the manufacturer’s recommendations. Cycle counts for mRNA quantification were normalized to Gapdh . Relative expression (ΔCt) and quantification (RQ = 2 –ΔΔC) for each mRNA were calculated using the ΔΔCt method as suggested and the graphs were plotted. Calculation of standard deviation (SDΔCt = (SDtarget2 + SDref2)1/2) and error bars (RQ1 = 2 (−(ΔΔCt+SDΔΔCt) ) and RQ2 = 2 (−(ΔΔCt-SDΔΔCt)) ) was performed according to ABI technical literature Part Number 4371095 Rev B. For the control samples, the RQ values are close to 1 but not exactly 1, because the ΔΔCt was obtained after substracting the ΔCt of each control sample from the average ΔCt of the biological replicates. The following pre-designed TaqMan gene expression assays from Applied Biosystems (ABI) were used. The probe names start with Mm for mouse probes (used to test in mouse brain samples and in primary mixed cortical neurons), and Rn for Rat probes (used to test in rat PC12/TrkB cells). Creb1 , Mm00501607_m1, Rn00578826_m1; Crem , Mm04336053_g1, Rn04338541_m1; Egr1 , Mm00656724_m1, Rn00561138_m1; c-Fos , Mm00487425_m1, Rn00487426_g1; Fosl2 , Mm00484442_m1, Rn00564121_m1; Fosl1 , Mm04207958_m1, Rn00564121_m1; Fosb , Mm00500401_m1, Rn00500401_m1; c-Jun , Mm00495062_s1, Rn99999045_s1; Junb , Mm04243546_s1 Rn00572994_s1; Jund , Mm04208316_s1, Rn00824678_s1; S100a10 , Mm00501457_m1, Rn01409218_m1; Bdnf , Mm04230607_s1, Rn02531967_s1; Fgf2 , Mm00433287_m1, Rn00570809_m1; Egf , Mm00438696_m1, Rn00563336_m1; Igf , Mm00439560_m1, Rn00710306_m1; Ngf , Mm00443039_m1, Rn01533872_m1; Vegf , Mm00437306_m1, Rn01511602_m1; Tgfβ , Mm01178820_m1, Rn00572010_m1; Gapdh , Mm_99999915_g1, Rn99999916_s1; Srsf5 , Mm00833629_g1; Slc1a2 , Mm01275814_m1; Sirt1 , Mm01168521_m1; Glul , Mm00725701_s1; Glo1 , Mm00844954_s1; Crhr1 , Mm00432670_m1; Adrb1 , Mm00431701_s1; Abcb1 , Mm00440736_m1. Behavior testing All behavioral studies were carried out and analyzed with the experimenter blind to the treatment group. Genotypes were decoded after data were processed and analyzed. Procedures were performed as described previously: open-field test (OFT) [ 40 ]; novelty suppressed feeding (NSF) [ 41 ]; and tail suspension test (TST) [ 24 ]. Cohorts of chronic fluoxetine-treated mice were subjected to multiple behavioral testing from a less to worse invasive nature of the tests in the following order: OFT, TST, and NSF. Quantification and statistical analysis Statistical details of each experiment are included in the figure legends. Briefly, for two group comparisons, we used two-tailed unpaired Student’s t -test. For multiple group comparisons, we used one-way or two-way ANOVAs and corrections were applied using the appropriate post hoc test. In all experiments, P < 0.05 was considered significant. Bar graphs show mean values and the error bars for bar plots are standard error of the mean (±SEM). N represents the number of mice for behavior experiments and biological replicates for cell culture experiments. For behavior experiments, all studies were carried out and analyzed with the experimenter blind to the treatment group. Results AP-1 transcription is stimulated by chronic fluoxetine treatment To test the hypothesis that IEGs implement specific gene expression programs to mediate an antidepressant response, we aimed to first determine the kinetics of expression of transcription factors in response to treatment with chronic fluoxetine. To do so, we used BALB/cJ mice as they are inherently anxious and produce a robust behavioral response to chronic fluoxetine treatment [ 22 ]. We administered fluoxetine orally in drinking water for 28 days to mimic the treatment in humans. The timeline of the behavioral and biochemical experiments is shown in a schematic diagram in Fig. 1a . At various intervals (2, 5, 9, 14, 21, and 28 days of treatment), the PFC was dissected for biochemical analysis and behavioral experiments were performed as indicated. Fig. 1: Identification of c-Fos and c-Jun upregulation during chronic fluoxetine treatment. a A schematic diagram showing the timeline for biochemistry and behavior experiments during chronic SSRI fluoxetine (Flx) treatment. Fluoxetine was administered orally to BALB/cJ mice for 28 days. Start of treatment indicated as Day 1 (red arrow). Behavioral analysis (black arrows) using open-field test (OFT) was performed on day 14 (Supplementary Fig. S1 ), tail suspension test (TST) on day 16, and novelty suppressed feeding test (NSF) on day 18. Biochemistry was done after harvesting the mouse prefrontal cortex (PFC) on 2, 5, 9, 14, 21, and 28 days of treatment (blue arrows). The onset and maintenance of behavioral response is shaded in gray. b TST. c NSF. d – k Quantitative-polymerase chain reaction (qPCR) to measure the levels of transcription factor mRNA Creb1 ( d ), Crem ( e ), Egr1 ( f ), c-Fos ( g ), Fosl2 ( h ), c-Jun ( i ), Junb ( j ), Jund ( k ). l A schematic diagram describing the formation of the AP-1 complex. Extracellular signals activate c-Fos and c-Jun mRNA transcription and protein expression. The proteins get phosphorylated by kinases, forming a stable heterodimeric AP-1 complex, thereby binding to target DNA and controlling their transcription. m A schematic diagram depicting c-Fos and c-Jun regulation by chronic fluoxetine treatment. Statistical analysis was performed between vehicle- (1% saacharine in drinking water) and fluoxetine-treated samples using two-tailed unpaired Student’s t -test; n = 6 for biochemistry and n = 11-14 for behavioral experiments. Data are mean ± SEM; * P ≤ 0.05, ** P ≤ 0.01, *** P ≤ 0.005. Full size image The response to treatment was assessed by using well-described behavioral paradigms for depressive behavior, including TST and NSF [ 42 ]. The fluoxetine-treated mice exhibited reduced immobility in TST ( P < 0.01, Fig. 1b ) and reduced latency to bite food in NSF ( P < 0.04, Fig. 1c ). We observed no effects of the treatment on the locomotor activity of the animals using open-field test (OFT), confirming that the performance of the animals in the TST and NSF tests were not confounded by the overall changes in animal behavior (Supplementary Fig. S1 ). Next, to identify transcriptional programs that are initiated in response to antidepressants, we analyzed the gene expression changes of selected families of IEGs in the PFC after days 2, 5, 9, 14, 21, and 28 of fluoxetine treatment. Analysis of the mRNA levels of CREB, Fos, and Jun family IEGs ( Creb1 , Crem , Egr1 , Fosl1 , Fosl2 , c-Fos , Fosb , c-Jun , Junb , and Jund) was performed by quantitative PCR (qPCR) (Fig. 1d–k ). Among all the transcription factors tested, c-Fos was the most strongly induced gene (Fig. 1g ). Surprisingly, c-Fos mRNA expression increased to ∼ 1.5-fold that of the vehicle controls at 9 days of treatment and peaked at ∼ 3.5-fold at 21 days of treatment. Fosl1 and Fosb mRNA levels were below the threshold for reliable quantification and therefore excluded from analysis. We also observed a statistically significant induction of Creb1 and Egr1 , factors that have been previously implicated in neuronal plasticity and neuropsychiatric disorders [ 11 , 12 , 43 ]. As c-Fos is the earliest induced transcription factor in response to fluoxetine and as we wanted to determine the function of molecules that precede initiation of the behavioral response, we decided to focus on c-Fos regulation. In order for c-Fos protein to exert its function, it must be phosphorylated and bound to a member of the Jun protein family, thereby forming a stable functional dimer termed the AP-1 complex (Fig. 1l ) [ 44 , 45 , 46 , 47 ]. The Jun proteins can function either as a homodimer or as a heterodimer by interacting with members of the Fos family. We found that c-Jun is the most likely binding partner of c-Fos based on the coordinated induction of c-Fos and c-Jun between 9 and 21 days of treatment (Fig. 1g, i ). In summary, we show that c-Fos is the most induced gene among all IEGs in reponse to chronic fluoxetine treatment, and c-Jun is the likely binding partner of c-Fos , together forming the AP-1 complex (Fig. 1l, m ). AP-1 controls the expression of neuronal remodeling genes in response to chronic fluoxetine treatment The AP-1 complex (Jun–Jun and Jun–Fos dimers) functions as a transcriptional regulator by binding to a common 12- O -tetradecanoylphorbhol-12-acetate(TPA)-responsive element (TRE) palindromic sequence, TGAC/GTCA [ 48 ]. We hypothesized that the target genes regulated by the Jun–Fos complex may be directly linked to the initiation of the behavioral response to fluoxetine. Hence, we aimed to characterize the targets of the AP-1 transcriptional program using c-Fos and c-Jun genome-wide chromatin immunoprecipitation paired with high-throughput sequencing (ChIP-seq). Mice were treated with fluoxetine for 9 days, the earliest time point at which c-Fos and c-Jun transcription was upregulated (Fig. 1g ). Comparison of the target sites between untreated and treated samples from the frontal cortex revealed a number of binding sites in response to drug treatment for both c-Fos and c-Jun (Fig. 2a–g ). Analysis of genome-wide binding sites bound by c-Fos and c-Jun at ∓ 4 kb of the transcription start site (TSS) showed predominant binding to promoter regions of target genes (Fig. 2a, e ). In general, we observed c-Fos occupancy on fewer target sites in the vehicle-treated animals reflecting low c-Fos endogenous expression, and c-Fos occupancy on many target sites in response to fluoxetine reflecting robust c-Fos induction (Fig. 2a, b ). In contrast, we observed c-Jun occupancy on many target sites in the vehicle-treated mice, indicating high c-Jun endogenous levels, and an even further increase in binding to target sites was observed in response to fluoxetine (Fig. 2e, f ). Fig. 2: Structural and synaptic plasticity genes are the targets of AP-1 complex in response to chronic fluoxetine treatment. a , e Aggregate plot of genome-wide ChIP-seq signals for c-Fos and c Jun in the mouse frontal cortex before and after 9 days of Flx treatment at −/+ 4 kb of the transcription start site (TSS). b , f Venn Diagram showing the overlap in c-Fos and c-Jun binding targets between vehicle- and Flx-treated animals. c , g Position-weight matrix of the c-Fos and c-Jun binding motif identified using an MEME de novo motif search for all significant peaks genome wide. d , h Ingenuity pathway analysis (IPA) to identify significant canonical pathways ( y -axis), which are associated with c-Fos and c-Jun target genes ( x -axis), displaying the −log P value cut-off set to 1.3, calculated by right-tailed Fischer’s exact test. For full gene list, see Supplementary Tables S1 and S2 and for profiles of depression-associated genes, see Supplementary Fig. S2 ). i Venn Diagram showing the overlap between c-Fos and c-Jun bound target sites. For full gene list, see Supplementary Table S3 . j Downstream effects analysis examining the genes in the dataset affecting a particular disease or function. k Ingenuity pathway analysis (IPA) to identify significant canonical pathways ( y -axis), associated with c-Fos and c-Jun overlapping sites ( x -axis), displaying the −log P value cut-off set to 1.3, calculated by right-tailed Fischer’s exact test. l Genome browser tracks showing the representative profiles of genes bound by both c-Fos (red trace) and c-Jun (blue trace), compared to input samples (green trace), treated with vehicle or fluoxetine (Flx), represented by – and + symbols. A schematic diagram showing the intron–exon structure of each target gene, gene name, and TSS (black arrow to indicate the direction of gene transcription). ChIP peaks near the TSS are shaded in gray. The y- axis signal intensity labels are indicated in parenthesis. Full size image Next, we identified the significantly enriched motifs in all the c-Fos and c-Jun bound target genes by performing a de novo motif search. We found that the consensus TRE sequence was highly enriched for both c-Fos and c-Jun target sites confirming direct binding of AP-1 to DNA (Fig. 2c, g ). We then annotated these target genes and used Ingenuity Pathway Analysis (IPA) to determine association of these genes with specific canonical pathways. We observed that many of the c-Fos target genes regulate pathways including CREB, PKA, neurotrophin signaling, synaptic long-term potentiation, circadian rhythm signaling, and synaptic plasticity (Fig. 2d , Supplementary Table S1 ). Only the pathways with the most significant −log ( P value) cut-off of 1.3 are shown. As expected, we found that the c-Jun target genes regulated some of the same pathways as c-Fos target genes, but in addition also bound to genes that regulated structural plasticity, such as Rho, Rac signaling, and actin cytoskeleton signaling (Fig. 2h , Supplementary Table S2 ). The role of these regulatory proteins in synapse development and plasticity has been well established [ 49 , 50 , 51 ]. As c-Jun is known to have broader DNA-binding specificity through homodimer binding or heterodimer formation with other Fos proteins, we analyzed the target genes that showed coincidental binding for both c-Fos and c-Jun. We identified a large overlap between the c-Fos and c-Jun bound genes and identified 1781 commonly regulated binding sites. (Fig. 2i , Supplementary Table S3 ). Using this list of c-Fos/c-Jun overlapping targets, we carried out downstream effects analysis using IPA to find whether genes in this dataset would affect a particular biological process or disease. Overall we identified pathways affecting neuronal morphology, remodeling, and homeostasis as indicated in Fig. 2j . We also determined the association of these genes with canonical pathways, and, as expected observed a number of commonly regulated pathways (Fig. 2k ). These results indicate that c-Fos and c-Jun work together as a complex to function as a transcriptional regulator modulating expression of crucial genes essential for the fluoxetine response. Next we tested whether any of the human MDD-associated genes based on GWAS studies [ 52 , 53 ] are targets of c-Fos and c-Jun regulation. Interestingly, we found that both c-Fos and c-Jun are bound to many depression-associated genes described in humans. These include Srsf5, Srf, Slc1a2, Sirt1, Rab4b, Rab3a, Glul, Abcb1b, Glo1, Crhr1, Creb1, Calm2, Calm1, Bdnf, Adrb1, and S100a10 (Supplementary Fig. S2 ). Among the above depression-associated genes, Bdnf and S100a10 (p11) stand out as they are both essential for mediating the antidepressant response in the cerebral cortex and hippocampus, and play a major role in mood and depressive disorders [ 54 , 55 ]. These results demonstrate that AP-1 regulated target genes have links to human depression and antidepressant response. Among the c-Fos and c-Jun targets, transcription factors were the most upregulated group across the various categories (Fig. 2j ), and representative ChIP-binding profiles of target genes are shown in Fig. 2l . These include transcription factors Srf, Egr1, Arc, Junb, Npas4, Jund, Creb1 , and other effector genes , Dusp6, Trib1 , and Nptx2 in response to fluoxetine. Taken together, these findings indicate that the specific induction of AP-1 by antidepressant treatment initiates a transcriptional program that modulates the expression of neuronal plasticity genes which then leads to an antidepressant response. These results prompted us to characterize the pathways that regulate AP-1 transcription and activity. Regulation of AP-1 transcription and activity We show that c-Fos and c-Jun transcription is stimulated by chronic fluoxetine only after 9 days of chronic treatment. This prompted us to investigate and identify the factors that regulate c-Fos and c-Jun transcription. The expression of c-Fos and c-Jun are known to be readily induced by growth factors, neurotransmitters, and electrical stimulation [ 45 , 56 , 57 ]. Therefore, it is possible that growth factor signaling or neuronal activity in response to antidepressant treatment promotes c-Fos and c-Jun expression. To determine which factors regulate c-Fos and c-Jun transcription, we established an in vitro primary cortical culture system that allowed for efficient and simultaneous screening of multiple growth factors that could directly stimulate c-Fos and c-Jun mRNA expression. This cell culture system comprises predominantly cortical neurons and other supporting cell types, and is widely used to study physiological properties of neurons [ 58 ]. To capture the peak induction of c-Fos and c-Jun transcription, we measured their levels 2 h after stimulation (as reported previously [ 59 ]) with a variety of growth factors (BDNF, FGF2, EGF, IGF, NGF, VEGF, BMP4, and TGFβ) (Fig. 3a ). We also tested neuronal activity modulators such as bicuculline (GABA-A receptor antagonist) and potassium chloride (KCl) that induces depolarization in cultured neurons. c-Fos mRNA expression was stimulated by BDNF ( ∼ 20-fold), FGF2 ( ∼ 4-fold), and EGF ( ∼ 1.8-fold) (Fig. 3a ). c-Fos mRNA expression was significantly stimulated by KCl ( ∼ 60-fold), as previously reported [ 45 , 56 ]. c-Jun mRNA expression was similarly induced by KCl, BDNF, FGF2 (all ∼ 2-fold), and EGF ( ∼ 1.2-fold). The other factors showed no effects. Fig. 3: Regulation of c-Fos and c-Jun mRNA and protein. a Stimulation of c-Fos and c-Jun mRNA in response to acute application of factors, BDNF, FGF2, EGF, IGF, NGF, VEGF, Bmp4, TGFβ, bicuculline, and KCl in DIV 7 primary cortical cultures at 12 h after stimulation by qPCR. b – d Determination of basal ( b ), BDNF ( c ), or FGF2 ( d ) dependent c-Fos and c-Jun mRNA expression in response to pharmacological inhibition of the receptor tyrosine kinase pathways in DIV 7 primary cortical cultures by qPCR. Inhibitors for TrkB (K252a), MAPK (U0126, MEK1/2 inhibitor), PI3K (LY294002), PLCγ (U73122), p38 MAPK (SB203850), JNK (SP600125) were used. The figure legends have an “inh” for inhibitor (e.g., inhTrkB for inhibitor of TrkB). e Western blots showing the effect of kinase inhibition on BDNF-induced c-Fos and c-Jun phosphorylation. GAPDH loading control blots are shown. f A schematic diagram summarizing the BDNF- and FGF2-inducible c-Fos and c-Jun regulation in primary cortical cultures. Statistical analysis was done using one-way ANOVA and corrections for multiple comparisons were performed using post hoc Bonferroni test. Comparisons were made between vehicle- and growth factor-treated samples in a , and between growth factor-treated samples with and without inhibitors in b – d ; n = 6. In c and d , + symbol indicates comparison between vehicle- and BDNF- or FGF-treated samples. Data are mean ± SEM; * P ≤ 0.05, ** P ≤ 0.01, *** P ≤ 0.005. Dashed lines in a and b indicate fold change of the control sample, and dashed lines in c and d indicate fold change of the BDNF- and FGF2-induced sample. The fold change values for the control vehicle samples are normalized to 1. Full size image Having demonstrated that BDNF and FGF2 predominantly regulate c-Fos and c-Jun transcription, we next addressed which specific signaling cascade(s) induces AP-1 transcription and activity. Growth factors act by binding to their cognate receptors and activating MAPK signaling cascades that include extracellular signal-regulated kinases (ERKs), stress-activated protein kinases (JNK/SAPK), and p38 family of kinases (P38 MAPK), all of which stimulate c-Fos and c-Jun gene expression [ 60 ]. Furthermore, the Fos and Jun proteins are functionally regulated via phosphorylation by various kinases. The phosphorylation of c-Fos by ERK and RSK (p90 ribosomal S6 kinase) [ 61 ] and the phosphorylation of c-Jun by MAPK and JNK [ 62 , 63 ] are known to activate their target gene regulation. Using the primary cortical culture system, we screened pharmacological inhibitors of these kinases and assessed the level of c-Fos and c-Jun transcription in the absence (Fig. 3b ) and presence of BDNF (Fig. 3c ) and FGF2 (Fig. 3d ). Endogenous expression of c-Fos and c-Jun transcription was strongly attenuated by TrkB (BDNF receptor) and MAPK inhibition, and moderately by p38 MAPK inhibition. No effects were observed upon inhibition of the PLCγ, PI3K, or JNK pathway (Fig. 3c ). In the presence of BDNF and FGF2-stimulation, c-Fos and c-Jun induction was attenuated by TrkB and MAPK kinase inhibition as well as by JNK inhibition (Fig. 3c, d ). No effects were observed upon inhibition of other kinases. To determine which kinases phosphorylate c-Fos and c-Jun proteins, we repeated the inhibitor experiment as above, and tested the levels of phosphorylated c-Fos and c-Jun proteins. Phosphorylation of c-Fos was reduced by MAPK or PI3K inhibition, and c-Jun phosphorylation was reduced by MAPK or JNK inhibition (Fig. 3e ). Taken together, these data indicate that BDNF and FGF2 signals regulate both c-Fos and c-Jun transcription via the MAPK and JNK pathways. In addition, the kinases MAPK and PI3K phosphorylate c-Fos protein, and the kinases MAPK and JNK phosphorylate c-Jun protein (Fig. 3f ), which leads to the formation of the functional AP-1 complex. Specificity of AP-1 transcriptional regulation To ensure that AP-1 is a specific transcription factor that is relevant for the chronic fluoxetine response, we chose to study in detail the regulation of one of the AP-1 target genes— S100a10 . S100a10 expression is downregulated during depressive disorders, is essential for the antidepressant response, and shows enriched expression in specific cell types in brain regions relevant for depression [ 55 ]. We also found S100a10 as a top target gene of AP-1 regulation in our ChIP-seq experiments. Hence characterizing the transcriptional regulation of S100a10 would further contribute to the understanding of the antidepressant response. We aimed to determine whether AP-1 is the only transcription factor that regulates S100a10 transcription, or if there is a need for other additional factors which we may have missed during our analysis. To pursue this goal, we first identified the S100a10 promoter in silico. Based on the ChIP-seq profile, bioinformatic analysis, and search for canonical promoter elements [ 64 ], we determined that the exon 1 region of the S100a10 gene exhibits all the characteristics of a promoter (Supplementary Fig. S3a ). Also, our ChIP-seq data revealed that c-Fos and c-Jun bind to this region of the S100a10 gene. To identify the functional promoter, we cloned varying lengths of the S100a10 regulatory/promoter region upstream of a luciferase gene reporter, which was used to measure promoter activity (Supplementary Fig. S3b ). We transfected these constructs into Neuro-2a (N2a) cells and evaluated the luciferase activity of various constructs. We pinpointed the core functional promoter region of S100a10 , which is located on chromosome3: 93554373–93555181(based on mouse chromosome assembly, mm10). Next, we searched for a suitable cell line that would enable robust stimulation of AP-1 by BDNF and FGF2, and, in addition, would allow for efficient and targeted transcription factor knockdown following siRNA inhibition of candidate transcription factors. We identified the rat PC12-TrkB cell line (PC12 cells that stably express the BDNF receptor, TrkB) as a suitable system to study growth factor signaling. These cells demonstrated robust stimulation of AP-1 expression in response to BDNF and FGF2 (Fig. 4a, b ). To evaluate the dynamics of expression of these molecules in the PC12-TrkB system, we measured the kinetics of c-Fos , c-Jun , and S100a10 mRNA by qPCR. c-Fos and c-Jun mRNA were rapidly stimulated within 30 min of BDNF or FGF2 application, with a peak induction at 2 h as observed in primary cortical cultures. In contrast, S100a10 mRNA expression was first observed at 2 h (when c-Fos and c-Jun expression were at their peak), and the highest expression was observed at 24 h after treatment. (Fig. 4a, b ). To further validate the expression kinetics, we analyzed their protein levels after BDNF stimulation. The expression dynamics of p11 ( S100a10 ) protein and the total and phosphorylated forms of c-Fos and c-Jun were congruent to the mRNA levels, as illustrated in Fig. 4c and quantification of the blots shown in Fig. 4d . We observed the initial induction of S100a10 mRNA (2 h) at the same time we detected the phosphorylated forms of c-Fos and c-Jun (when they are able to form the AP-1 complex) (Fig. 4c ). Fig. 4: AP-1 specifically regulates the basal and inducible transcription of S100a10 . a , b Analysis of c-Fos , c-Jun , and S100a10 mRNA expression by qPCR at 2, 6, 12, 24, and 48 h of acute BDNF application (red arrow) and acute FGF2 application (orange arrow) in PC12/TrkB cells. c Western blots showing the relative kinetics of p11 ( S100a10 ) protein, AP-1 (total and phosphorylated forms of c-Fos and c-Jun), and GAPDH (loading control) after 2, 6, 12, 24, and 48 h of acute BDNF application. d Quantification of the blots from c . e qPCR of basal, BDNF- and FGF2-inducible S100a10 mRNA expression, when treated with specific siRNAs for c-Fos , c-Jun , or both. f Western blots of c-Fos, c-Jun, and p11 ( S100a10 ) protein, untreated (basal) or stimulated with BDNF (induced), and treated with specific siRNAs for c-Fos , c-Jun , or both. g Effect of c-Fos/c-Jun siRNA on S100a10 promoter activity measured by luciferase reporter gene assay (the predicted S100a10 promoter sequence is shown in Supplementary Fig. S3a and the identification of the functional promoter is shown in Supplementary Fig. S3b ). h The consensus AP-1-binding motif is shown on the top. The binding of c-Fos and c-Jun to the promoter region of S100a10 gene from the ChIP-seq experiment is shown below. The S100a10 exon 1 promoter sequence comprising potential binding sites for the AP-1 TRE consensus motif were mutated and are denoted as mut1, mut2, and mut3 as shown in the lower panel. The mut3 sequence was deleted and denoted as Δmut3. i S100a10 promoter activity was measured using the luciferase reporter assay to test the effects of AP-1 site mutations in mouse N2A cells. For samples in a, b , and d , comparisons were made between untreated and treated conditions ( n = 3) for all samples. Statistical analysis was done using two-way ANOVA to test effects of BDNF or FGF2 treatment and time; corrections for multiple comparisons were performed by running post hoc Tukey’s multiple comparisons test. For e and i , statistical analysis was done using one-way ANOVA and corrections for multiple comparisons were performed using post hoc Bonferroni test. In a and b , P value significance indicated by symbols # for c-Jun , + for c-Fos , * for S100a10 . In e , comparisons were made between each scrambled siRNA control and the three siRNA treatments (control, BDNF-, and FGF2-induced). The + symbol represent t -test between vehicle- versus BDNF- or FGF2-treated samples. In i , comparisons were made between empty pGL4 vector and S100a10 luciferase promoter construct (+); and between S100a10 promotor construct and the three AP-1 site mutations (*). In g , two-tailed unpaired Student’s t -test was performed between the WT S100a10 promotor construct and the mutant pGL4 constructs ( n = 4). In g and i , the + symbol represent t -test between vehicle versus BDNF- or FGF2-treated samples. Data are mean ± SEM; * P ≤ 0.05, ** P ≤ 0.01, *** P ≤ 0.005. Full size image Having identified the peak time of induction of S100a10 at 24 h, we next tested which of the transcription factors, when silenced by specific small-interfering RNAs (siRNAs), affected S100a10 transcription. We tested 22 candidate transcription factors, all of which have potential to bind to the S100a10 promoter region. The siRNA inhibition efficiency for each factor and their effect on modulating S100a10 mRNA expression are indicated in Table 1 . Only siRNA against c-Fos , c-Jun , or both decreased S100a10 mRNA (Table 1 , Fig. 4e , gray bars), with an siRNA efficiency of 55% for c-Fos and 67% for c-Jun . The siRNA inhibition of Jund or Junb had no effect. These results further strengthen our previous findings confirming that c-Jun is indeed the authentic binding partner for c-Fos, and that the AP-1 complex specifically regulates S100a10 transcription. Table 1 Identification of transcription factor(s) regulating p11 by RNAi. Full size table In contrast, siRNA inhibition of factors Bhlhe40 , Crem , Fosl2 , Stat3 , Sp1 , and Srf upregulated S100a10 expression (Table 1 , Supplementary Fig. S4a–f ), indicating that they are potential repressors. Next we tested whether both BDNF- and FGF2-dependent induction of S100a10 requires AP-1 activity. We observed a decrease in BDNF-inducible (red bars in Fig. 4e ) and FGF2-inducible (orange bars in Fig. 4e ) S100a10 mRNA when c-Fos , c-Jun , or both were silenced. To confirm these results, we also measured the protein levels of c-Fos, c-Jun, and p11. We demonstrate that the p11 protein was also downregulated by silencing of c-Fos, c-Jun, or both (Fig. 4f ). Our results confirm that AP-1 activity is necessary for both the BDNF- and FGF2-dependent regulation of S100a10 expression. Having identified the specific regulation of S100a10 by AP-1 and determined the potential binding of c-Fos and c-Jun to the S100a10 promoter region by ChIP-seq, we analyzed if S100a10 promoter activity would be affected by silencing AP-1 activity using specific siRNAs for both c-Fos and c-Jun (AP-1 siRNA). The S100a10 luciferase promoter construct was transfected into mouse Neuro-2a (N2a) cells in the presence and absence of AP-1 siRNA (Fig. 4g ). Strikingly, AP-1 depletion resulted in decreased luciferase activity. This indicated that the binding site for AP-1 is indeed located within the identified functional promoter sequence. Hence we searched for the consensus AP-1 motif (TGAC/GTCA) within S100a10 exon 1 and identified three potential binding sites (Fig. 4h ). We then mutated these sites to block AP-1 binding and called them as mutation1 (mut1), mutation2 (mut2), mutation3 (mut3), and deleted mut3 (Δmut3) (Fig. 4h ). Mut1 and mut2 contain AP-1-binding motifs oriented in the sense direction, whereas mut3 contains an AP-1-binding motif oriented in the antisense direction. Normalized luciferase reporter activity was measured after transfection of mutants: mut1, mut2, mut3, and Δmut3. Reduced luciferase activity was only observed when the AP-1 consensus motif in mut3 was mutated or deleted (Fig. 4i ), confirming that the functional AP-1-binding site is within the S100a10 promoter sequence. Therefore, our results indicate that the specific binding of AP-1 to the mut3 site within the S100a10 promoter is likely critical for implementing a robust antidepressant response. AP-1 activity is essential for the antidepressant response in vivo Finally, we examined whether inhibition of AP-1 function would affect fluoxetine efficacy in vivo. We chose to block JNK as our previous experiments showed that JNK function is essential for the inducible expression of c-Fos and c-Jun (Fig. 3c, d ), whereas MAPK function is necessary for both basal and inducible expression; and JNK selectively phosphorylates c-Jun (Fig. 3e ). We used the JNK inhibitor, SP600125, that was previously shown to selectively block JNK 1, 2, 3, and prevent c-Jun phosphorylation when injected into mice intraperitoneally (i.p) at a concentration of 16 mg/kg [ 65 ]. We chose to use this concentration of 16 mg/kg rather than 30 or 50 mg/kg [ 66 , 67 ] to avoid any stress-inducing effects. We used a milder treatment paradigm by injecting the inhibitor on days 4, 6, 8, and 10 of fluoxetine treatment, only when JNK function is likely essential for the antidepressant response (Fig. 5a ). We had four groups of animals: Vehicle; Vehicle Fluoxetine; Vehicle inhJNK; Fluoxetine inhJNK; and we powered this study with a large sample size ( n = 10–18) to overcome influence of variability. We observed that the JNK inhibitor had no effect on the body weight of the mice and the animals drank about 3–6 mL of fluoxetine per day (Supplementary Fig. S5a ). The JNK inhibitor would likely inhibit cJun phosphorylation and block the formation of the AP-1 complex. Therefore, the low dose and the mild JNK-inhibitor treatment, while effective at attenuating the onset of the fluoxetine response, did not produce long-term stress-like behavioral effects. We utilized NSF and TST tests to measure the behavioral effects of antidepressant treatment in the four groups of mice as described in Fig. 5a . We observed a robust antidepressant response to fluoxetine reflected by reduced tail suspension immobility ( P < 0.003) and reduced latency to bite food ( P < 0.003) in the fluoxetine-treated animals compared with the vehicle controls (Fig. 5b, c ). Strikingly, we observed a blunted fluoxetine response in the presence of JNK inhibitor, shown by comparing the fluoxetine-treated group with fluoxetine/JNK inhibitor-treated animals in both the behavioral tests of TST ( P < 0.02) and NSF ( P < 0.01). The injection of the JNK inhibitor itself on vehicle-treated animals did not affect their behavior (Fig. 5b, c ). We observed no effects of the inhibitor on the locomotor activity of these animals as shown by the open-field test (Supplementary Fig. S5b ). In addition, we also determined the expression levels of a representative set of depression-associated AP-1 target genes. We observed that several of the target genes were modulated by fluoxetine and this regulation was partially reversed by the JNK-specific inhibitor treatment (Supplementary Fig. S5c ). These data indicate that the JNK inhibitor was effective in blocking target gene expression in the mouse cortex that is relevant for the response. Together, our results demonstrate that the JNK pathway, which we have shown is specifically required for inducing AP-1 activity, is a critical component to activate specific molecules essential for mediating the antidepressant response in vivo. Fig. 5: Blocking AP-1 function by inhibiting JNK activity attenuates the antidepressant response. a A Schematic diagram showing the timeline of chronic Flx treatment that was initiated on day 1 (red arrow). The JNK inhibitor, SP600125, was injected intraperitoneally (i.p) on days 4, 6, 8, and 10 (blue arrows) of treatment, and behavioral tests (black arrows) of OFT, TST, and NSF were performed on days 14, 17, and 37 of treatment, respectively. b Four groups of mice ( n = 12–18) for vehicle-treated, Flx-treated, vehicle/JNK inhibitor-treated (inhJNK), and Flx/JNK inhibitor-treated were tested for their immobility in TST. c Four groups of mice ( n = 10–18 per group) were subjected to NSF test to measure their latency to bite food in a novel environment. The drinking water consumption and OFT data are shown in Supplementary Fig. S5a and b respectively. A few of the AP-1 target genes affected by inhJNK treatment are shown in Supplementary Fig. S5c . d A schematic diagram illustrating the molecular programs and the sequence of signaling pathways that are activated during chronic antidepressant response is shown. The dotted line between fluoxetine and BDNF/FGF2 indicate that SSRIs are known to stimulate BDNF/FGF2 signaling. Statistical comparisons were made using two-way ANOVA to test effects of flx and inhibitor-treatment and were corrected for multiple comparisons by running a post hoc Tukey’s multiple comparisons test. Data are mean ± SEM; * P ≤ 0.05, ** P ≤ 0.01, *** P ≤ 0.005. Full size image In summary, our findings reveal activation of a molecular cascade pertaining to S100a10 regulation during chronic fluoxetine response. The cascade involves the activation of BDNF and FGF2 growth factor signaling, which in turn stimulates MAPK, JNK, and PI3K intracellular pathways to activate an AP-1-driven transcriptional program that regulates neuroplasticity-associated effector genes, which ultimately produce the antidepressant response (Fig. 5d ). Altogether, our results highlight the need for temporal activation of select molecules and the significance of sequential signaling events during antidepressant response. Discussion The molecular mechanisms underlying the delayed onset of action of antidepressant drugs is highly debated in the field and is not clearly understood. Both animal and human research has provided supporting evidence that chronic stress and neuropsychiatric disorders have deleterious effects on the brain, both structurally and functionally [ 5 ]. Emerging evidence on the role of neuroplasticity and its correlation with behavioral improvement in humans [ 68 ] and in mouse models [ 5 , 69 ] explains the time lag needed to reorganize and remodel the synaptic morphology changes and neural networks disrupted during depression. However, the molecular mechanism is not clearly understood. Here we performed a detailed study to characterize the molecular response to fluoxetine, an SSRI that is widely prescribed for the treatment of several neuropsychiatric disorders. By looking for the early transcription changes during the response, we identified the activation of a selective AP-1 transcriptional program that precedes the onset of the behavioral response in rodents [ 3 , 70 ], and clinical efficacy in humans [ 71 ]. Importantly, a sudden drop in suicidal rate and ideation in humans at 9 days after treatment has been reported [ 72 ]. These findings indicate that there is a functional relationship between the genes that are induced at this time point and the behavioral response. Interestingly, we identified effector molecules of the AP-1 transcriptional program that particularly regulate the expression of neuronal remodeling and plasticity-inducing genes, many with known links to depression and antidepressant responses such as S100a10 . Additionally, mice with a brain-specific deletion of c-Fos and c-Jun show defects in synaptic plasticity and axonal regeneration respectively [ 73 , 74 ]. Our findings demonstrate that the onset of the AP-1 transcriptional program links neuronal plasticity to the antidepressant response. However, we cannot rule out the involvement of other transcription factors, microRNAs and RNA-binding proteins that possibly contribute to the observed antidepressant response, as we have mapped the antidepressant pathway by focusing on factors controlling S100a10 transcription. The delay in the stimulation of c-Fos and c-Jun transcription is intriguing as AP-1 transcription is shown to be regulated within 30 min of stimulation by growth factors, neurotransmitters, and electrical stimulation [ 45 , 56 , 57 ]. In general, AP-1 function and activity is shown to be regulated by (1) the composition of the Fos:Jun complex (Fos:Jun heterodimer or Jun:Jun homodimer) [ 45 ]; (2) by AP-1 binding to transcription factors that regulate gene activation [ 75 ] or gene repression [ 76 ]; (3) by the ability of AP-1 to affect chromatin accessibility [ 77 ]; (4) by AP-1 binding to enhancers [ 78 ]. Here we demonstrate that neuronal activity (stimulation by KCl), BDNF and FGF2, two well established growth factors necessary for antidepressant action [ 54 , 79 ], are the strongest stimulators of c-Fos and c-Jun transcription in cortical cultures. As sustained neuronal activity is known to induce BDNF-mediated TrkB signaling [ 80 ], and as AP-1 is part of the BDNF-positive feedback loop [ 23 ], it is likely that BDNF and FGF2 are the rate-limiting factors during the antidepressant response. The molecular steps mediating the interaction between immediate increase in serotonin and BDNF activity during chronic fluoxetine treatment is not completely characterized. However, bidirectional regulation between these two signaling systems and their distinct neuronal functions in survival, neurogenesis, and synaptic plasticity has been well documented [ 81 , 82 , 83 , 84 ]. Also, S100a10 -expressing corticostriatal neurons in the cerebral cortex have been shown to exhibit distinct serotonin responses during stress and fluoxetine response indicating the essential role of these neurons in antidepressant action [ 85 ]. Moreover, evidence for AP-1 regulation by serotonin signaling [ 86 , 87 , 88 , 89 ], further substantiates the role of the signaling network comprising serotonin, BDNF, AP-1, and S100a10 in the initiation of the antidepressant response. Our studies identifying the role of the MAPK, PI3K, and JNK cascades during BDNF- and FGF2-dependent regulation of the AP-1 complex, together with their reported function in mediating structural and synaptic plasticity function(s) confirm a significant role for these cascades during the antidepressant response. Such neuroadaptive properties have been documented for BDNF [ 90 , 91 , 92 ]; FGF2 [ 93 ]; MAPK [ 94 , 95 , 96 ]; PI3K [ 97 ]; JNK [ 98 ]; c-Jun [ 43 ]; and c-Fos [ 9 ]. Interestingly, JNK −/− mice have disorganized cortical layers and impaired dendritic architecture due to disrupted microtubule integrity [ 99 ]. Consistent with these observations, we have identified Rho, Rac, and cytoskeletal signaling molecules as targets of AP-1 during the fluoxetine response. Although not much is known about the specific function of S100a10 in neuronal plasticity, its role in mediating BDNF-dependent structural plasticity has recently been demonstrated [ 100 ]. These results further substantiate and favor the idea that AP-1 regulated genes operate in various molecular and cellular remodeling pathways mediating neuroplasticity. We establish that AP-1 function is essential for the fluoxetine response. We chose to inhibit the JNK pathway based on key observations. First, our data showed that the JNK pathway was the only pathway that was essential for the inducible expression of c-Fos and c-Jun in response to BDNF and FGF2-stimulation, whereas the MAPK and PI3K pathways were required for both basal and inducible expression. Also, the role of MAPK in the induction of depressive behavior has been documented before [ 96 ]. Second, our experiments showed that JNK is the kinase that predominantly phosphorylates c-Jun, which would then promote DNA-binding and regulation of target effector molecules [ 63 ]. Third, our experiments showed that the substrate of JNK, c-Jun, is induced in response to BDNF and FGF2 in vitro and during chronic fluoxetine treatment in vivo. These observations indicate that the JNK function becomes essential under conditions that require c-Fos and c-Jun induction and Jun phosphorylation, potentially during chronic fluoxetine response. Our results show that an attenuation in the fluoxetine response occurs when the activity of JNK is inhibited, thus validating that AP-1 function is necessary for providing behavioral response to fluoxetine in mice. Although we show evidence that both c-Fos and c-Jun are necessary to regulate the transcription of genes essential for the antidepressant response, we cannot rule out the effects of the AP-1 complex formed by the Jun-Jun homodimer. We show that the transcriptional regulation of the target genes by the AP-1 complex is selective. By characterizing the regulation of one of the AP-1 target genes, S100a10 , we demonstrate that the AP-1 complex comprised of c-Fos and c-Jun (and not any of the other members of the Jun-Fos protein family) is the only transcriptional activator, among 22 other transcription factors tested. Furthermore, we identified the AP-1 binding site within the S100a10 promoter, and proved with nucleotide resolution that this binding site is functionally active, and is likely important for the antidepressant response. In addition, we also identified transcriptional repressors (Bhlhe40, Crem, Fosl2, Stat3, Sp1, and SRF) of S100a10 gene that may a play a role in the regulation of basal- versus inducible transcription. These findings reveal that the activators and repressors of S100a10 transcription, together, potentially regulate the cell-type-specific expression of S100a10 and its signal-dependent induction—both of which are potentially important for the antidepressant response. In summary, we have identified AP-1 target genes with known links to human MDD based on GWAS data, indicating that these molecules could be used as potential biomarkers for predicting antidepressant responses. We have linked a number of molecules including growth factors, signaling cascades, and a fully integrated gene expression program that function consecutively to activate neuronal remodeling pathways to provide the antidepressant response. Importantly the molecules, FGF2 [ 101 ], BDNF [ 102 ], c-Fos [ 103 ], and p11 [ 104 ], are also induced by exercise and enrichment reflecting their function in mediating the homeostatic and neuroplasticity mechanisms in the brain. More evidence for this function comes from the observed upregulation of the molecules, BDNF [ 105 ] and p11 [ 55 ], following treatment with multiple classes of chemical antidepressants or brain stimulation. Future studies to unravel the mechanisms that trigger the activation of these molecules will therefore help design novel strategies for treating all neurological disorders that benefit from improving neuroplasticity.

Some highly effective medications also happen to be highly mysterious. Such is the case with the antidepressant drugs known as selective serotonin reuptake inhibitors, or SSRIs: They are the most common treatment for major depression and have been around for more than 40 years, yet scientists still do not know exactly how they work. Nor is it known why only two out of every three patients respond to SSRI treatment, or why it typically takes several weeks for the drugs to take effect—a significant shortcoming when you're dealing with a disabling mood disorder that can lead to impaired sleep, loss of appetite, and even suicide. New research by a team of Rockefeller scientists helps elucidate how SSRIs combat depression. Their work, published in Molecular Psychiatry, could one day make it possible to predict who will respond to SSRIs and who will not, and to reduce the amount of time it takes for the drugs to act. Brain Teasers Major depression—also known as clinical depression—is firmly rooted in biology and biochemistry. The brains of people who suffer from the disease show low levels of certain neurotransmitters, the chemical messengers that allow neurons to communicate with one another. And studies have linked depression to changes in brain volume and impaired neural circuitry. Scientists have long known that SSRIs rapidly increase the available amount of the neurotransmitter serotonin, leading to changes that go well beyond brain chemistry: Research suggests the drugs help reverse the neurological damage associated with depression by boosting the brain's innate ability to repair and remodel itself, a characteristic known as plasticity. Yet the molecular details of how SSRIs work their magic remains a mystery. Researchers in the late Paul Greengard's Laboratory of Molecular and Cellular Neuroscience set out to trace the chain of molecular events triggered by one of the most widely prescribed and effective SSRIs: fluoxetine, also known as Prozac. In particular, they wanted to see if they could tie the effects of the drug to specific changes in gene expression. Led by senior research associate Revathy Chottekalapanda, the team treated mice with fluoxetine for 28 days, then measured the animals' biochemical and behavioral responses to the drug. They conducted these experiments in a mouse strain prone to anxiety that offered several advantages including the fact that "improvement in anxiety, in addition to depressive-like behavior, is a good measure of antidepressant drug response," Chottekalapanda says. Using a combination of behavioral tests and real-time RNA analysis, the researchers were able to monitor changes in the animals' behavior and gene expression as they responded to the drug. Things started to get interesting on the ninth day of treatment: The activity of a gene called c-Fos began to increase markedly, and by day 14 the mice were showing telltale behavioral changes—they were moving around more, for example, and took an interest in food even after being moved to a new environment. The timing was remarkable, Chottekalapanda says, because it aligned with well-established milestones in the treatment of human patients. "For example, the rate of suicide drops after nine days of treatment, and people tend to feel better after two to three weeks," she says. As it turns out, c-Fos helps create a so-called transcription factor, AP-1, which activates specific genes by binding to their DNA. The sudden appearance of these molecules therefore raised several new questions: Which genes does AP-1 activate? What triggers the production of AP-1 in the first place? And how does this whole sequence of events ultimately beat back depression? Domino Effect Chottekalapanda and her colleagues began by analyzing the mice's cortex, a brain region previously shown to be essential for the antidepressant response, looking for changes in genes and DNA-binding proteins to which AP-1 might possibly bind. Focusing especially on the nine-day mark, they found changes in mouse genes whose human counterparts have been linked to depression and antidepressant drug response. Walking the cat backward, the team was able to identify specific molecules known as growth factors that spur the manufacture AP-1 itself, and the pathways through which they act. Taken together, these results painted a detailed picture of how fluoxetine and other SSRIs work. First, the drugs ramp up the amount of serotonin available in the brain. This triggers a molecular chain reaction that ultimately makes brain cells increase their AP-1 production—an effect that only begins to take off on day nine of treatment. AP-1 then switches on several genes that promote neuronal plasticity and remodeling, allowing the brain to reverse the neurological damage associated with depression. After two to three weeks, the regenerative effects of those changes can be seen—and felt. "For the first time, we were able to put a number of molecular actors together at the crime scene in a time- and sequence-specific manner," says Chottekalapanda. To confirm their molecular model of SSRI response, Chottekalapanda and her team gave mice treated with fluoxetine an additional substance designed to block one of the pathways necessary to the production of AP-1. The results were striking: When the researchers prevented the mice from producing AP-1, the effects of the drug were severely blunted. Moreover, gene-expression analysis showed that blocking the formation of AP-1 partially reversed the activation of some of the genes responsible for the antidepressant response. One Word: Plasticity The implications of the team's findings could be far-reaching. For example, the genes that AP-1 targets could be used as biomarkers to predict whether a given patient will respond to SSRI treatment. And the cast of molecular characters involved in fluoxetine response could potentially be targeted with drugs to improve the efficacy of antidepressant treatment, perhaps even reducing the amount of time it takes for SSRIs to kick in. Towards that end, Chottekalapanda is already conducting experiments to clarify the precise reasons for the nine-day delay in AP-1 production. She would also like to know whether these molecular players are mutated or inactive in people who don't respond to SSRIs, and to discover precisely how the genes that AP-1 regulates in response to fluoxetine would promote neuronal plasticity. That last piece of the puzzle might prove to be especially important. Depression is not the only disorder that could potentially be remedied by enlisting the brain's innate healing abilities: Chottekalapanda suspects that the same plasticity-promoting pathways that are activated by antidepressants such as fluoxetine could potentially be used to treat other neurological and neurodegenerative conditions such as Alzheimer's and Parkinson's disease. "If we can figure out what they do, we can potentially develop treatments to reverse the damage involved not only with depression but also with other neurological disorders," she says.

10.1038/s41380-020-0767-8

Medicine

Cutting off melanoma's escape routes

Prudence Donovan et al. Endovascular progenitors infiltrate melanomas and differentiate towards a variety of vascular beds promoting tumor metastasis, Nature Communications (2018). DOI: 10.1038/s41467-018-07961-w Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-018-07961-w

https://medicalxpress.com/news/2019-01-melanoma-routes.html

Abstract Tumor vascularization is a hallmark of cancer central to disease progression and metastasis. Current anti-angiogenic therapies have limited success prompting the need to better understand the cellular origin of tumor vessels. Using fate-mapping analysis of endothelial cell populations in melanoma, we report the very early infiltration of endovascular progenitors (EVP) in growing tumors. These cells harbored self-renewal and reactivated the expression of SOX18 transcription factor, initiating a vasculogenic process as single cells, progressing towards a transit amplifying stage and ultimately differentiating into more mature endothelial phenotypes that comprised arterial, venous and lymphatic subtypes within the core of the tumor. Molecular profiling by RNA sequencing of purified endothelial fractions characterized EVPs as quiescent progenitors remodeling the extracellular matrix with significant paracrine activity promoting growth. Functionally, EVPs did not rely on VEGF-A signaling whereas endothelial-specific loss of Rbpj depleted the population and strongly inhibited metastasis. The understanding of endothelial heterogeneity opens new avenues for more effective anti-vascular therapies in cancer. Introduction Tumor vascularization is one of the hallmarks of cancer. It is classically proposed that for tumor progression to occur new blood vessels must form that will allow the provision of oxygen and nutrients, however these vessels also provide beneficial roles by allowing immune cells and drug delivery to inhibit tumor growth 1 . Furthermore, blood vessels have angiocrine capacity supporting directly the growth of tumors through the production of cytokines and growth factors 2 . They have also been proposed to facilitate tumor spread through the blood or lymphatic vasculature 3 . More recently, it has also been proposed that tumor vessels behaving abnormally contribute to the hypoxic environment and hence maintain tumor cells in an invasive state 4 . Beyond its detrimental role, tumor vasculature is an essential component of drug and immune cell delivery to the tumor. Overall these properties have prompted numerous attempts at normalizing abnormal blood vessel formation in the context of cancer rather than outright abrogating tumor vessels 5 Tumors are vascularized through a variety of modalities but rely predominantly on angiogenesis where VEGF family members play a crucial role. However, anti-VEGF therapy has failed in many indications to reduce tumor size, spread, or vascularization 5 . Although many of the molecular factors that drive tumor vascularization are well known and established, much less is known about the cellular origin of vessel network in tumors. In particular, it is often unclear which vascular bed or which cells are at the source of newly formed vessels in primary tumors. Past studies have proposed the existence of endothelial progenitors driving tumor vascularization 6 . Contrasting early studies, the hematopoietic, bone marrow derived, and circulating nature of this endothelial progenitors has been refuted 7 and it has been clarified that these cells are of a resident endothelial nature 8 . Although informative, many of these studies relied on single markers or cell transfers as opposed to cell fate mapping of endogenous progenitors as well as a functional rather than a marker-based definition of progenitors 9 , 10 . We have recently reported in a variety of vascular beds in mice 11 and humans 12 that the adult endothelium is heterogeneous and is composed of 3 distinct populations: an endovascular progenitor (EVP), a transit amplifying (TA), and a differentiated (D) population. In the present study, we have extensively examined the formation of tumor vasculature and show that very early upon inoculation, tumors are infiltrated by SOX18 expressing EVP cells that originate from arterial or venous but not from lymphatic beds. These EVP cells give rise to TA and D cells that form mostly venous/arterial capillaries but also lymphatics through the combined contribution of multiple clones of endothelial cells. At the functional level, only EVP cells have colony forming and transplantation capacity. The molecular characterization of EVPs shows significant differences with TA and D cells. Critically for a therapeutic perspective, anti-VEGF-A therapy did not affect EVP cells. On the other hand, conditional ablation of RBPJ, a direct protein interactor of SOX18 13 , 14 , dramatically reduced EVP cells and resulted in the abrogation of metastases providing perspectives for anti-vascular therapy of cancer by targeting the EVP population. Results Functional and molecular heterogeneity of tumor vasculature To explore the cellular origin of vessels in a growing tumor and establish the kinetics of vessel assembly, we undertook orthotopic delivery of B16-F0 melanoma cells intradermally. Cdh5-Cre ER RosaYFP mice were used to label all endothelial cells with YFP using tamoxifen injection for 5 consecutive days. Subsequently, mice were inoculated intradermally with B16-F0 tumor cells. Upon injection and development, tumors were visible from 5–7 days macroscopically and could be easily distinguished from surrounding tissues. We first examined the heterogeneity of endothelial cells based on variation in cell surface markers 11 . Dissection, single cell suspension, and analysis of tumors with minimum contamination by surrounding cells using multi-color flow cytometry allowed determining positive and negative staining for every marker using the fluorescence minus one (FMO) method (Supplementary Figure 1A ). In the absence of tamoxifen injection no YFP could be observed (not shown). Using a live gate, doublets were discounted and hematopoietic cells were identified using the lineage cocktail (Lin). Among Lin-CD34+ cells, most cells (>98%) were YFP+ demonstrating their endothelial origin (Fig. 1a ). Among these VE-cadherin expressing cells, three populations could be easily identified based on their levels of CD31 and VEGFR2 identified as EVP (CD31 lo VEGFR2 lo ), TA (CD31 int , VEGFR2 lo ), and D (CD31 hi VEGFR2 hi ) 11 . We also ensured that these Lin-CD34+ populations did not express the hematopoietic progenitor marker c-kit (Supplementary Figure 1B ). We further ensured the validity of our analysis using spontaneous melanomas developed on Tyr::Nras; Cdk4r24c 15 transgenic mice as well as in Lewis Lung Cancer (LLC) and EO771 breast cancer tumors and observed similar endothelial heterogeneity (Supplementary Figure 1C & D ). Fig. 1 Melanoma endothelium in heterogeneous. a Schematic diagram demonstrating experimental set up using vascular lineage tracing Cdh5-Cre ER RosaYFP mice. Flow cytometry plots showing cells dissociated from B16-F0 tumors harbor distinct CD34 positive, lineage (LIN) negative populations (red gate) as determined using strict fluorescence-minus-one (FMO) analysis. >98% of CD34+LIN- cells are YFP+. Three distinct populations were observed based on CD31 and VEGFR2 expression in tumors (from left to right: EVP, TA, and D) amongst CD34+LIN-YFP+cells ( n = 4). b Schematic diagram demonstrating the isolation of GFP+ EVP, TA, and D from tumors, which were subsequently re-transplanted in a 1:1 ratio with B16-F0 cells into a wild-type host. c , d Flow cytometry plots showing only GFP+ EVP cells re-transplanted were able to persist and engraft in secondary tumors. TA and D cells inoculated in secondary tumors could not be recovered 14 days later (** p < 0.01 vs TA and D; Mann–Whitney T -Test) ( n = 6). e Immunofluorescence images of only GFP+ EVP cells (white arrows) engrafting and surviving transplantation with B16-F0 (scale bar represents 100 µm). f Delivery of B16-F0 cells with GFP+ EVP resulted in larger tumors (increase in weight by 34% compared to B16-F0 alone), whereas tumor cells delivered with D cells alone resulted in smaller tumors (decrease in weight by 33% compared to B16-F0 alone) (* p < 0.05; Mann–Whitney T -Test). Results presented as mean ± SEM. EVP endovascular progenitor, TA transit amplifying, D definitive differentiated, CON control (no GFP cells) Full size image We next evaluated the functional differences between the three populations. Here, we injected B16-F0 cells in CAG-GFP transgenic mice ubiquitously expressing GFP (Fig. 1b ). After 15 days, tumors were removed and GFP+ EVP, TA, and D populations were flow-sorted, mixed with B16-F0 tumor cells, and re-implanted in secondary non-transgenic recipients. After 14 days, secondary tumors were removed, weighed, and subjected to flow cytometry to identify GFP+ endothelial cells. Only EVP cells had a capacity to survive and repopulate, whereas TA and D cells inoculated in secondary tumors could not be recovered 14 days later (** p < 0.01 vs TA and D) (Fig. 1c, d ) and visualized in tumor sections (Fig. 1e ). These cells maintained their endothelial characteristics as they expressed CD34 and were devoid of any hematopoietic lineage marker. Furthermore, delivery of B16-F0 cells mixed with purified EVP resulted in significantly larger tumors (increase in weight by 34% compared to B16-F0 alone) collected 7 days after transplant, whereas tumor cells delivered with D cells alone resulted in smaller tumors (decrease in weight by 33% compared to B16-F0 alone, * p < 0.05, Mann–Whitney T -Test) (Fig. 1f ). Overall, these findings confirm that vascular endothelial cells are heterogeneous in their cell surface molecular markers as well as in their self-renewal and engraftment potential. Lineage relationship between endothelial populations We next asked whether these three populations derived from one another and used fate mapping analysis to establish their lineage relationship over time. Cdh5-Cre ER RosaYFP mice were inoculated with B16-F0 melanoma cells at D0 and injected with a single dose of tamoxifen on D3 (Fig. 2a ). Tumors were collected on subsequent days to establish the fate of endothelial cells labeled. On D5, flow cytometry allowed to clearly identify YFP+ CD34+ cells that did not express any hematopoietic markers (lineage negative) within tumors. Most of these cells were CD31 lo VEGFR2 lo and corresponded to EVP and a few TA cells. At this early time point, no D cell could be identified (Fig. 2a ) and tumors sections revealed that YFP+ EVP cells were mostly single cells that only expressed CD34 but not CD31 or VEGFR2 (Fig. 2b , Supplementary Figure 2A ). At D10, however, many cells were identified as D with some TA cells. At this time point, no obvious YFP positive EVP cell could be found suggesting that the progenitor pool labeled at D3 had undergone differentiation within the tumors. Finally by D15, D cells were the only population in abundance that was remaining from the cells that were stained on D3, with only small numbers of EVP and TA observed. At D10 and D15, YFP+ cells were organized in vascular structures infiltrating the tumors (Fig. 2b ). Overall these findings clarify the transition from EVP to TA to D in a 7–12 days period and demonstrate the lineage relationship between the three populations. Fig. 2 Lineage tracing reveals the kinetics of the endothelial hierarchy in tumors. a Schematic diagram of the Cdh5-Cre ER RosaYFP lineage tracing model employed with B16-F0 melanoma cells injected at day 0 (D0), with mice receiving tamoxifen (Tam) 3-days post tumor inoculation. Flow cytometry plots demonstrate that D5 post tumor inoculation, only EVP and TA amongst CD34+ LIN-YFP+ cells can be observed. At D10 EVP are absent and by D15 only mature D cells can be observed ( n = 5). b Immunofluorescence at D5 shows individual YFP+ foci. At D10 and 15 entire YFP+ vessel structures can be observed. c Schematic diagram of the Sox18-Cre ER RosaYFP lineage tracing model employed with B16-F0 melanoma cells injected at day 0 (D0), with mice receiving tamoxifen (Tam) 3-days post tumor inoculation. Flow cytometry plots demonstrate that D5 post tumor inoculation, only EVP and TA amongst CD34+ LIN-YFP+ cells can be observed. At D10 EVP are absent and by D15 only mature D cells can be observed ( n = 5). d Immunofluorescence at D5 shows individual YFP+ foci. At D10 and 15 entire YFP+ vessel structures can be observed. Scale bar represents 50 and 150 μm, respectively. Results presented as mean ± SEM. EVP endovascular progenitor, TA transit amplifying, D definitive differentiated Full size image We next examined if the EVP were infiltrating the tumor and most importantly if they were originating from a non-endothelial origin. We therefore used pulse chase of tamoxifen studies in the Cdh5-Cre ER RosaYFP mice over an extended period of time as previously reported 16 (Supplementary Figure 2B ). Tamoxifen was initially provided for 5 consecutive days and the normal back skin was processed to assess the endothelial compartment. Greater than 99% of all Lin-CD34+ cells were YFP+, demonstrating the entire labeling of the vasculature (Supplementary Figure 2C ). Of interest, the normal skin vasculature was also organized in EVP, TA, and D compartments. Next, after a similar pulse of tamoxifen, mice were left for a latency period of 14 days before being inoculated subcutaneously with B16-F0 cells. The tumors were then collected 10 days later, or 24 days after their final tamoxifen injection. Within the tumors greater than 99% of all Lin-CD34+ cells remained YFP+, clarifying that the endothelial population was only originating from the vasculature and not being diluted by other unlabeled populations (Supplementary Figure 2D ). Given the rapid development of these tumors that on D5 are devoid of any visible blood vessels, our findings suggest EVPs gain the center of the tumor through infiltration from surrounding vessels. Sox18 re-expression EVP progenitors in tumors In our previous work, we have shown the critical importance of Sox18 re-expression in EVPs 11 . Indeed, Sox18 is a gene involved in embryonic vascular development. Its expression in the vasculature is lost at adult age in endothelial cells except in situations of new vessel formation such as wounds 17 or tumors 18 . We performed similar lineage tracing using Sox18-Cre ROSA-YFP mice (Fig. 2c ). When induced with tamoxifen at D3 post tumors inoculation, the first YFP+ population identified at D5 was the EVP. At this time point we observed very few TA, no D cells, and EVP cells once again in this different model consisted in single cells (Fig. 2c, d ). D populations could be observed at D10–D15 post-tumor inoculation. This clearly implies that EVPs initially express Sox18 and then give rise to TA and D cells through complete differentiation. This additional model clearly reinforces the observation of transitions between EVP, TA, and D. We next asked whether the initial input of EVPs was sufficient to drive tumor vascularization or alternatively whether additional EVPs would enter the same process of differentiation. We therefore used Sox18-Cre ROSA-YFP mice inoculated with B16-F0 tumors at D0 and induced them with tamoxifen at D3 but also at D8 and collected tumors at D12 (Supplementary Figure 3 ). The additional stimulation of Sox18- induced Cre activity at D8, demonstrated the staining of additional EVP cells that could not be observed by a single injection on D3 alone. This result suggests that there is an ongoing process by which EVP cells arise to sustain a continuous makeup of the tumor vasculature. EVPs contribute to both arterial and venous vascular beds Having established that in tumors only EVPs have self-renewal potential and differentiate into TA and D cells, we next asked the capacity of these cells to give rise to all possible vascular beds. On D5, EVP cells labeled with YFP were essentially single cells infiltrating the bulk of tumor cells. This was observed for both Sox18 and Cdh5 -driven YFP labeling (Fig. 3 ). They did not express any specific vascular bed marker. By D10 and D15, YFP cells had formed entire vessel network. The majority of vessels expressed the venous capillary marker endomucin, and about a third of YFP+ vessels displayed arterial markers such as DLL4 or Sox17 (not shown). These observations were made in both models of lineage tracing showing the potential of Cdh5 or Sox18 expressing EVPs to give rise to these structures. Overall, the quantification of lineage tracing experiments shows that Sox18 and Cdh5 expressing cells seem both to give rise to the same proportions of capillary, veins, and arteries suggesting that all EVPs activate the expression of Sox18 prior to vessel formation. Fig. 3 EVP contribution to arterial and venous beds. Representative micrographs of tumor sections taken from Cdh5-Cre ER RosaYFP and Sox18-Cre ER RosaYFP lineage tracing models. At day 5 YFP+ cells did not display any arterial or venous markers. By day 10 and 15 YFP+ cells could be co-localized with arterial marker DLL4 and venous marker endomucin ( n = 5). Results presented as mean ± SEM. Scale bar represents 150 μm Full size image EVPs contribute to lymphatic vascular beds Tumor-induced neo-lymphangiogenesis has long been thought to exclusively derive from pre-existing lymphatic vessels 3 or from myeloid cells infiltrating the tumor vascular 19 . In both Cdh5 and Sox18 -driven lineage tracing models, a discrete population of YFP+ lymphatic vessels could be observed by D15 within the center of the tumor (Fig. 4a ). This was observed using both Lyve-1 and Podoplanin lymphatic markers. The rarity of this event upon quantification reaching upon 6–8% of YFP+ vessels (Fig. 4b ) prompted us to examine whether lymphatics had a separate progenitor. We performed lineage tracing experiments using Prox1Cre tdtomato (Fig. 4c ). On tumor sections obtained on D5 post tumor inoculation (2 days after tamoxifen delivery), the large majority of Prox1+ cells co-localized with Lyve1 and Podoplanin expressing lymphatics in the periphery of the tumor (Fig. 4d ). Flow cytometry on D5 showed that Prox1 (tdTomato) expressing cells were not Lin-CD34+ suggesting that EVPs within tumors at this time point were not derived from the Prox1 expressing lymphatic vascular bed (Fig. 4e ). This further suggests that EVPs contribute to lymphatics through the expression of Sox18 likely similar to the series of embryonic events that where this transcription factor induce lymphatic endothelial cell fate in venous progenitor cells 20 . However, this remains a small fraction of tumor lymphatics. Overall, these findings show that Sox18 expressing EVPs derive from venous or arterial but not lymphatic beds and contribute to most of the venous capillaries and some of the arterial and lymphatic vessels within tumors. Fig. 4 EVPs contribute to lymphatic vascular beds but do not originate from lymphatics. a , b Representative micrographs of tumor sections taken from Cdh5-Cre ER RosaYFP and Sox18-Cre ER RosaYFP lineage tracing models. At day 15 YFP+ vessels co-localized with lymphatic marker Lyve1 (white arrow) and Podoplanin (PDPN). Green arrows represent Lyve1+ vessels that are not YFP+ ( n = 5). c Schematic diagram of the Prox1-Cre ER tdTomato lineage tracing model employed with B16-F0 melanoma cells injected at day 0 (D0), with mice receiving tamoxifen (Tam) 3-days post tumor inoculation ( n = 5). d Lyve1+PDPN+ vessels co-localized with Tomato+ (Prox1) vessels. e Flow cytometry plots demonstrating that Prox1+CD34+LIN- cells do not contribute to the endothelial hierarchy. Results presented as mean ± SEM. Scale bar represents 150 μm. EVP endovascular progenitor, TA transit amplifying, D definitive differentiated Full size image EVP from multi-clonal populations derive tumor vessels We next wondered how single EVP cells observed on D5 could give rise to entire vessels of arterial or venous nature. In particular, we asked if this was related to a single clone growing into a given vessel or if multiple EVPs would compose a single vessel. We therefore performed multicolor lineage tracing using a Cdh5-Cre ER Rainbow reporter system as previously reported by us 21 , 22 . In this model, each cell upon Cre recombinase activation undergoes a random recombination of one or two Rainbow 1.0 cassette (Fig. 5a ) allowing the random expression of one out of 5 different colors. Cdh5-Cre ER Rainbow mice were injected with a low dose (0.3 mg) of tamoxifen at D3 post B16-F0 tumor inoculation. Five days post tumor inoculation, multiple EVP clones could be identified within the tumor (Fig. 5b ). There was no specific clustering of colors in EVP cells within each tumor suggesting probably the migration of multiple cells stained at D3 rather than the proliferation of a single clone. By D10, both venous capillaries and arterial vessels seemed composed of more than one color clearly suggesting the contribution of multiple clones to different sections of each vessel within the tumor. However, within each section of vessels labeled with the same color and deemed to be clonal, multiple cells could be identified. This suggests that EVP cells enter a differentiation process as well as some proliferation before joining within the resulting clone to form a vascular structure. Fig. 5 Clonality of the tumor vasculature. a Schematic diagram of the Cdh5-Cre ER Rainbow lineage tracing model employed with B16-F0 melanoma cells injected at day 0 (D0), with mice receiving tamoxifen (Tam) 3-days post tumor inoculation ( n = 5). b At day 5 multi-individual colored clones could be observed that were presumed to be EVP. By day 10 entire arterial (DLL4) and venous (endomucin-higher magnification) vessels were composed of multi colors demonstrating polyclonality. Scale bar represents 50–150 μm. EVP endovascular progenitors Full size image Molecular characterization of endothelial populations Having established the functional hierarchy between EVP, TA, and D endothelial populations and their contribution to venous and arterial beds within the tumor at clonal level, we next aimed to characterize their molecular profiles. CAG-GFP transgenic mice were injected with B16-F0 tumors. At D15 tumors were collected, GFP+ EVP, TA, and D cells sorted and subjected to RNA sequencing ( n = 5 per population). The three populations could be separated on PC plots based on their gene expression levels (Fig. 6a ). Unsupervised clustering clearly distinguished two groups: node 1 consisted in D cells only whereas node 2 gathered EVP and TA. This clearly shows that the TA cells are more closely related to EVP. Within node 2 however EVP and TA could be distinguished based on the same clustering (Fig. 6b ). Based on this unsupervised clustering, differential gene expression between EVP and D was therefore deemed to be the most important feature. We identified 3064 differentially expressed genes ( p < 0.05 after multiple testing correction). Reflecting the quality of the cell sorting, Pecam1 (74×) and K dr (VEGFR2, 40×) were largely overexpressed in D compared to EVP as expected (Fig. 6c , Table 1 and Supplementary Figure 4A–F ). Fig. 6 Endothelial hierarchy RNA sequencing and gene expression analysis. a , b Principal component analysis and hierarchical clustering demonstrating the distinct clustering of each population segregated from each other ( n = 5). c Differentially expressed genes (* p < 0.05 after multiple testing correction) were identified between EVP and D populations. Differentiated endothelial markers Pecam , Kdr , and Nos3 were all significantly upregulated in D compared to EVP (*** p < 0.001 vs EVP). Results presented as mean ± SEM. EVP endovascular progenitor, D definitive differentiated Full size image Table 1 Differentially expressed genes Full size table Gene expression reflects state of differentiation In general D cells overexpressed many classical endothelial markers such as Nos3 (enos, 74×), Vwf (14×), Ets1 (9×), Ets2 (10×), Gata2 (6×), Fli1 (22×), Esam (47×), Tie1 (60×), Cldn5 (111×). As expected from our previous work, D cells also overexpressed genes from the SoxF family ( Sox18 , 52×; Sox17 , 75×; Sox7 , 64×). However, as shown above, the expression of Sox18 despite being higher in D cells, was initiated in EVP cells. In contrast, EVP cells showed some different characteristics. EVP cells displayed mobility with the overexpression of many matrix metalloproteases (MMP; 6 to 293 fold, respectively) and a high capacity to remodel the extracellular matrix (dermatopontin and decorin, >600×; Col3a1 , 415×; Col1a2 , 294×; Has1 , 258×) potentially explaining their ability to infiltrate tumors as single cells. Finally, EVP cells expressed genes classically involved in stem cell function ( Aldh1a1 , 790×; Sox9 18×) as well as quiescence ( Nfatc4 , 43×; Il33 , 409×; Cdkn1c , 12×). Looking at significant differences between TA and EVP cells, we identified 580 differentially expressed genes ( p < 0.05). Once again Pecam1 and Kdr were more highly expressed in TA cells. In many aspects, TA cells had an intermediate gene expression profile between EVP and D cells. Overall, these findings reflect the remarkable heterogeneity within the endothelium and support the hierarchy from progenitor to differentiated cells. Overall, when performing pathway analysis (DAVID), extra-cellular matrix–receptor interaction and focal adhesion were the top pathways reflecting the activity of EVP cells. PI3kinase and cytokine receptor signaling were also reflecting some of the key signaling in these cells (Supplementary Figure 4A–F ). Progenitor cells have significant angiocrine capacity EVP cells overexpressed many Wnt ligands ( Wnt9a , 9×; Wnt5a , 18×; Wnt10b , 23×; Wnt16 , 200×; Wnt11 , 395×; Wnt2 , 489×), FGFs ( Fgf16 , 8×; Fgf23 , 8×; Fgf10 , 9×; Fgf11 , 11×; Fgf21 , 11×; Fgf2 , 13×; Fgf18 , 76×; Fgf7 , 165×), Hgf (15×), Vegfd (163×), Vegfa (9×), Tgfb3 (14×) and Pdgfc (20×). EVPs seemed also able to respond to a range of extracellular growth factors through the expression of Pdgfra (205×) and Pdgfrb (23×), Egfr (106×), Tgfbr2 (5×), Tgfbr3 (13×) as well as to a range of cytokines through the expression of Flt3 (22×), Il6ra (13×), Il11ra1 (11×), and Il1rl1 (ST2, IL33 receptor; 181×). In particular, cytokine receptor signaling seemed particularly activated given the expression of its downstream targets ( Socs1 and Socs3 , 5×; Ifi205 , 305×; Irf4 , Irf6 and Irf7 , 5–30×; Pim1 , 6×). Acquisition of arterial vs venous identity is a late event We next wondered whether progenitors were predetermined to give rise to a specific vascular bed. We therefore looked for arterial vs venous or lymphatic markers in EVP, TA, or D cells. Arterial markers such as Dll4 , Sox17 , EphrinB2 ( Efnb2 ), Notch1 and Notch4 , Gja5 and Gja4 were all overexpressed in D cells with little difference in the level of expression between EVP and TA populations suggesting a late differentiation (*** p < 0.001; Mann–Whitney T -Test). Regarding venous markers such as COUP-TFII ( Nr2f2 ), Ephb4 , Nrp2 , or endomucin ( Emcn ), once again the level of expression was higher in D vs EVP suggesting a late differentiation. However, there was also a significant increase in expression in the TA population suggesting the initiation of venous differentiation during the TA stage compared to EVP (* p < 0.05; ** p < 0.01; *** p < 0.001 vs EVP; Mann–Whitney T -Test, Supplementary Figure 5A ). EVPs share common molecular signature in aorta and tumors Finally, we have previously compared EVP to D cells in normal aorta 11 and wondered to what extent the tumor and aorta data would overlap. We compared genes differentially expressed between EVP and D in the aorta vs in tumors and found a significant overlap (44% of aorta DEGs and 38% of tumor DEGs, Venn diagram, Supplementary Figure 5B ). Genes reflecting stem cell function ( Sox9 , Aldh1a1 , Aldh1a2 ), quiescence ( Nfatc4 , Il33 , Cdn1c ) as well as many of the receptors and paracrine function described above had been also identified in the aorta (Supplementary Figure 5B ). We next compared the level of differential expression by plotting the fold change for overlapping genes in aorta vs tumors. Although there was a clear correlation in fold change between EVP and D in both aorta and tumors, some genes seemed to behave differently in the two situations. We therefore looked at genes that would specifically be overexpressed in EVP cells in the context of tumors and not aorta and identified 138 genes significantly overexpressed (at least 2 fold, p < 0.05, arrow in Supplementary Figure 5B ). In the “tumor only” setting, the most differentially upregulated genes in EVP were related to its paracrine function such as Gdf6 , a member of the TGFβa agonist, Wnt16 , epiregulin , and neuroregulin . Anti-VEGF-A therapy does not target EVP cells Having established the important role of EVP cells in tumor vasculature and its gene expression signature, we next wondered about strategies to target this population to reduce both tumor vascularization, growth, and metastasis. Anti-VEGF-A therapy is the main FDA approved anti-angiogenic treatment for a range of solid tumors. However, it has failed in many indications to reduce tumor size, spread or vascularization and a variety of mechanisms have been proposed to explain resistance to this treatment 5 . We injected C57Bl/6 mice with B16-F0 melanomas and started therapy with anti-VEGF-A vs isotype control from D3 to D15 (Supplementary Figure 6A ). Tumors were collected at D15 and analyzed for the relative frequency of endothelial populations: there was no significant reduction or change in the EVP or D population (Supplementary Figure 6B & C ). This clearly shows that anti-VEGF-A therapy does not affect the pool of EVP cells available. There was no effect of anti-VEGF-A treatment on tumor vasculature in animals that received anti-VEGF-A therapy compared to control (Supplementary Figure 6D ). Conditional ablation of Notch signaling alters EVP We have previously reported that active Notch signaling is essential in progenitor quiescence 23 . Moreover, RBPJ is a direct interactor with Sox18 suggesting that this pathway would be critical in EVPs. However, it has also been established that Notch signaling allows reduction of VEGFR2 receptors in sprouting angiogenesis forming stalk cells 24 . In our gene expression analysis, Notch signaling elements were overexpressed in D cells ( Notch4 , 47×; Notch1 , 7×; Dll4 , 73×; Dll1 , 24×; Jag2 , 17×) resulting in increased signaling via Hey1 (13×), whereas in EVP cells other receptors were overexpressed ( Notch3 , 6×; Notch2 , 16×; Rbpj , 4×) and signaling was mostly driven by Heyl (7×) suggesting that different ligand-receptors were involved in Notch signaling in different populations (Supplementary Figure 4E ). Given the indication that Rbpj as well as Heyl were overexpressed in EVP cells, we examined the consequences of endothelial-specific ablation of canonical Notch signaling. We used Rbpj fl/fl /Cdh5-Cre ER RosaYFP ( Rbpj eKO ) mice and Rbpj fl/fl RosaYFP ( Rbpj WT-Cre negative ) controls (Fig. 7a, b ). Animals were injected with tamoxifen for 10 days to fully ablate Rbpj in endothelial cells and they were inoculated with HCMel12 tumors intradermally. HCMel12 is a murine (C57Bl/6) derived melanoma cell line obtained carrying transgenic Hgf and Cdk4r24c mutation. Upon in vivo passaging this cell line is highly metastatic 25 . After 10 days, primary cutaneous tumors were excised and mice were kept alive to progress to metastasis (Fig. 7a ). Primary tumors in Rbpj eKO mice had dramatically smaller numbers of EVP cells whereas D cells were unchanged as compared to Rbpj WT controls (Fig. 7b, c ). In accordance, CD31 staining of primary tumor sections revealed only minimal difference between the two groups, whereas Lyve1 and Podoplanin staining of lymphatics was significantly increased in the absence of Rbpj (Fig. 7d ). Of note in a B16-F0 model, primary tumors remained of smaller weight in Rbpj eKO mice compared to Rbpj WT (Supplementary Figure 7 ; ** p < 0.01; Mann–Whitney T -Test). At D28, 2 weeks after primary HCMel12 tumor excision, animals were sacrificed and examined for metastasis microscopically in the lung and the liver. We could not observe any significant metastasis in RBPJ eKO animals whereas 50–90% of Rbpj WT mice had identifiable tumors in lung and/or liver (Fig. 7e ). Fig. 7 Vascular-specific knock-out of Rbpj results in EVP depletion and reduced metastasis. a Schematic diagram of the Rbpj fl/fl /Cdh5-Cre ER RosaYFP ( Rbpj eKO ) and Rbpj fl/fl RosaYFP ( Rbpj WT-Cre Negative ) lineage tracing model employed, with mice receiving tamoxifen (Tam) for 10 consecutive days prior to tumor inoculation with metastatic HcMEL12 melanoma cells injected at day 0 (D0). Tumors were harvested at D28 ( n = 5). b , c Flow cytometry plots demonstrate that EVP cells are entirely absent from Rbpj eKO mice compared to Rbpj WT (*** p < 0.001 vs Rbpj eKO ; Mann–Whitney T -Test). No change was observed in the percentage of D cells present from either group. d No difference was observed in CD31+ vessel surface area between the groups. Significantly more Lyve1+ and Podoplanin+ vessels were observed in the Rbpj eKO mice compared to Rbpj WT (** p < 0.01 vs Rbpj WT ; Mann–Whitney T -Test). e Histological analysis of lungs and liver showed vast metastatic nodules in the Rbpj WT compared to no visible nodules in the Rbpj eKO mice (* p < 0.05; *** p < 0.001 vs Rbpj eKO , Mann–Whitney T -Test). Results presented as mean ± SEM. Scale bar represents 150 μm. EVP endovascular progenitor, TA transit amplifying, D definitive differentiated Full size image Discussion Vascularization of tumors is a hallmark of cancer and has been shown as an important step in cancer progression and metastasis. Despite major progress in understanding key molecular pathways involved in angiogenesis, it remains unclear where tumor vasculature originates from and to what extent it has a supportive role in cancer progression. Here, we show that melanoma tumors from very early stages are infiltrated by a population of EVP. These cells initiate a vasculogenic process as a single cell and progress towards a TA stage and finally differentiate into arterial and venous capillaries within the central core tumor area. Of interest, a small fraction of these capillaries also gives rise to lymphatics although EVPs are clearly not derived from this vascular bed. At the molecular level, these progenitors are endothelial in origin as they express VE-Cadherin and reactivate the expression of the developmental transcription factor Sox18 . RNA sequencing on sorted cells as well as functional characterization clearly differentiates EVPs from the later stages of progression towards differentiation. In particular, EVPs harbor significant paracrine activity that allows considering them as a reliable cellular target for therapy. Unlike anti-VEGF-A therapy, conditional ablation of Rbpj in the endothelium significantly reduced the EVP population in tumors and strongly inhibited metastasis in a model of melanoma. Modes of new vessel formation have been described during development and include vasculogenesis and angiogenesis among others. Although our study does not address the contribution of angiogenesis, it clearly demonstrates the existence of a vasculogenic process within the tumor center that is devoid of any vessel. In this avascular area, single cells were tracked through their expression of either VE-Cadherin or Sox18 and shown to form entire blood vessels as already reported by us in skin wounds. This was further reinforced by multicolor lineage tracing showing that individual arterial or venous capillaries are formed by the juxtaposition of multiple clones and therefore unlikely to result from angiogenesis where a limited number of stalk cells would contribute to the clonal progression of vessel branching. In this setting, the vascularization process seemed to follow steps occurring during embryonic development through the expression of Sox18 but also the formation of lymphatics from cells that originate from different (Prox1 negative) vascular beds. Indeed adult lymphatic vessel formation has been shown to derive exclusively from existing lymphatics unlike the contribution of EVPs within the center of the tumor. Our findings cannot confirm or exclude if similar processes also occur at the periphery of the tumor. Indeed many vessels of blood or lymphatic origin surround the tumor. Given that the tumor periphery is not avascular, vasculogenesis is therefore unlikely in this area. Progenitor activity of EVPs was exclusive and shown through their colony forming capacity as well as engraftment potential 8 . These properties could not be observed in TA and D populations. This strongly argues against the hypothesis that EVP, TA, and D are simply different stages of activation of otherwise similar endothelial cells. In the latter scenario where no hierarchy exists, one would expect equal engraftment and colony forming capacity. The observed functional differences suggest strongly that the cells differ intrinsically. This is strongly supported by the gene expression studies reported here clearly showing major transcriptomic changes from EVP to TA to D. Our findings highlight the transition from an endothelial cell with active expression of many mesenchymal genes important for mobility towards an endothelial cell that is adherent and expresses all of the endothelial differentiation markers including arterial and venous markers. We have previously shown a similar gene expression pattern within the aorta. The current findings reinforce the existence of a core gene expression signature that defines EVPs which include a capacity to remodel the extracellular matrix and a core set of “stem cell” genes and quiescence genes. Moreover, our findings add to lines of evidence supporting the concept of progenitors in the endothelium as defined by their function, self-renewal, plasticity, and lineage, rather than cell surface markers 9 , 10 . This hierarchy described in both mouse 11 and human 12 tissues allows a more precise definition of such progenitors. As such the current findings might differ from past studies where endothelial progenitors were defined based on a few surface markers without functional validation, an approach prone to contamination by hematopoietic or other cell lineages. The impact of anti-angiogenic therapy on cancer progression is difficult to predict. There is even uncertainty on the goal of antiangiogenic therapy. Is it to normalize vessels to allow immune cell infiltration and drug delivery or is it to abrogate all vessel to prevent tumor growth and spread? We here also observe that EVPs have significant angiocrine activity producing many growth factors, morphogens, and cytokines that might affect tumor growth and spread. In support of this finding, inoculation of tumor cells with EVP compared to D cells resulted in significantly larger tumors. In such experiments, it is difficult to expect a large effect as the host is replete with EVPs and therefore the addition of external EVPs might not constitute a large advantage. The depletion of EVP cells through ablation of canonical Notch signaling as already observed in wounds and in vitro clearly showed the importance of these progenitors in tumor spread. Of note this did not significantly affect tumor CD31+ vessels. Previous studies of Rbpj depletion in the context of tumors have also reported an increase in immune cell infiltrate 26 . Interestingly, excision of tumors and follow-up showed clearly that this strategy was valid in neo-adjuvant settings. Of interest, in our hands the use of anti-VEGF-A therapy did not provide a similar protection and did not deplete EVP cells. Indeed anti-VEGF-A is unlikely to target EVP cells given their low level of expression of VEGF receptors. Given the demonstrated continuous input of progenitors we propose here that EVPs are probably a major therapeutic target for cancer control. In conclusion, tumor vascularization recapitulates mechanism observed during embryonic development via the activation and expression of sox18 in specific progenitors of endothelial origin that form arterial and venous capillaries but also lymphatics. The hierarchy between progenitor, TA, and differentiated cell can be tracked by fate tracing as well as gene expression studies. Our findings strongly suggest that targeting EVPs is a valid strategy in controlling the progression and spread of cancer. Methods Animals All mice were treated in accordance with University of Queensland ethics approvals and guidelines for care of experimental animals. Both males and females (8–14 weeks of age; genders housed separately) were used for this study. C57BL/6 mice (WT) were obtained from the Animal Resources Centre (Perth, Western Australia). For lineage tracing experiments, Sox18-Cre ERt2 and Cdh5-Cre ERt2 were crossed with ROSA lox YFP lox . The resultant double transgenic offspring were named Sox18-Cre ER RosaYFP (8–12 weeks old) and Cdh5-Cre ER RosaYFP . For endothelial-specific knockout of the gene Rbpj , Rbpj fl-fl mice were crossed with Cdh5-Cre ER RosaYFP to create the resultant triple-transgenic Rbpj fl/fl /Cdh5-Cre ER RosaYFP . For polyclonal lineage tracing, Cdh5-Cre ERt2 were crossed with the Rainbow (CAG-Brainbow 1.0) mice to create the resultant Cdh5-Cre ER Rainbow . Prox1-Cre ERt2 were crossed with tDTomato reporter mice to create the resultant Prox1-Cre ER tDTomato . Ubiquitous CAG-GFP mice were used for all stem cell transplant experiments. All lineage tracing experiments were conducted using Tamoxifen (Sigma-Aldrich, MI, USA) made up in 90% corn oil (Sigma-Aldrich) and 10% ethanol, with each mouse receiving a 2 mg dose per intraperitoneal injection (100 µL of 20 mg/ml solution). Mice being treated with either anti-mouse VEGF-A (BioLegend (#512808), CA, USA) or Isotype control (BioLegend (#400533)) were given a dose of 100 µg every 4 days via intraperitoneal injections. Tumor cell culture B16-F0 cells were cultured using RPMI 1640 supplemented with 10% fetal bovine serum (FBS). Cells were passaged every 4 days, with 5 × 10 5 cells in saline (300 µL) were injected subcutaneously on each flank of mice for in vivo tumor studies. Metastatic melanoma cell line HcMel12 cells were obtained as a single cell suspension following 3 subsequent in vivo passaging (in vivo tumors were passaged after 14 days of growth). 2 × 10 5 cells in saline (300 µL) were injected subcutaneously for in vivo tumor studies. Tissue processing of murine B16-F0 and HcMel12 tumors Tissues were collected for ex vivo analyses at defined end points tumors (D5, D10, and D15). Tumors were first digested for 20 min at 37 °C in 1 mg/ml collagenase I (Gibco, Life Technologies, NY, USA), 1 mg/ml dispase (Gibco, Life Technologies, NY, USA), 150 μg/ml DNase-I (Sigma-Aldrich, St. Louis, MO, USA) before passing the suspension through a 70 μM cell strainer. Lineage+ cells were then depleted from tumor cell suspensions via MACS® cell separation according to the manufacturer’s instructions (Miltenyi Biotech, Cologne, Germany). Cell number and viability for each sample was assessed using 0.4% Trypan blue solution and a hemocytometer. Single cell suspensions were then used for flow sorting or analysis by flow cytometry. For transplantation studies, tumors were grown in CAG-GFP+ mice and GFP+ EVP, TA, and D cells were isolated from harvested tumors. Flow cytometry and FACS Dissociated single cells in PBS/BSA/EDTA were then incubated with various antibody combinations for multi-parameter flow acquisition and analysis. A Gallios™ flow cytometer was used for sample acquisition, while unbiased data analyses were performed with Kaluza® analysis software (Beckman-Coulter, Miami, FL, USA). FACS was performed by using a FACSaria cell sorter (Becton Dickinson, Franklin Lakes, NJ, USA). Extreme care was taken during cell sorting to ensure only “singlets” were being gated and any potential “doublets” were being gated out. Cell populations were collected in 5 ml polypropylene tubes containing 100% fetal calf serum (FCS). The following combinations of antibodies were used to assess the endothelial hierarchy populations: Rat anti-mouse VEGFR2 PE (1:50), Sca-1 (1:50), c-KIT (1:50), CD31 PE-Cy7 (1:100), and CD34 Alexa647 (1:50) (Becton Dickinson, NJ, USA), Rat anti-mouse Lineage cocktail BV450 (1:50) (BioLegend), Rat anti-mouse CD144 FITC (1:50) (eBioscience). FMO was used to delineate negative gating for each antibody. RNA extraction RNA was extracted from FACS sorted EVP, TA, and D cells using a QIAGEN mini kit (Qiagen, Valencia, CA) according to the manufacturer’s instructions. RNA quality and concentration was assessed using A260 nm/A280 nm spectroscopy on the Nanodrop ND-1000 (Thermo-scientific, Langenselbold, Germany). 5–100 ng of RNA was used for cDNA synthesis using the Superscript III Reverse Transcription Kit (Invitrogen, Mount Waverley, Australia). Library preparation, RNASeq, and data analysis RNA-Seq libraries were prepared from purified total RNA using a modified Smart-Seq2 protocol 27 . 2 ng of purified total RNA (0.4 ng/µL) was combined with 1 µL of 10 µM oligo-dT primer (/5Biosg/AAGCAGTGGTATCAACGCAGAGTACT30VN; Integrated DNA Technologies) and 1 µL of dNTP mix (10 mM each; Invitrogen, y02256), then the protocol was continued as described (ref. 2 ). Briefly, the RNA was reverse transcribed with the Smart-Seq2 TSO (/5Biosg/AAGCAGTGGTATCAACGCAGAGTACATrGrGrG, Integrated DNA Technologies), followed by 12 cycles of PCR amplification to obtain enough cDNA to prepare a library. Volumes of reagents were scaled accordingly to maintain final concentration ratios as in the original protocol, except for the PCR preamplification where the Smart-Seq2 ISPCR primer (/5Biosg/AAGCAGTGGTATCAACGCAGAGT; Integrated DNA Technologies) was added to a final concentration of 0.25 µM. 0.5 ng of cDNA was prepped into a library using the Nextera XT DNA sample preparation kit (Illumina, FC-131-1096), with 12 cycles of PCR used to amplify the final library. The final Nextera XT libraries were quantified on the Agilent Bioanalyzer with the High Sensitivity DNA kit (Agilent Technologies, 4067–4626). Libraries were pooled in equimolar ratios, and the pool was quantified by qPCR using the KAPA Library Quantification Kit—Illumina/Universal (KAPA Biosystems, KK4824) in combination with the Life Technologies Viia 7 real time PCR instrument. Sequencing was performed using the Illumina NextSeq500 (NextSeq control software v2.0.2/Real Time Analysis v2.4.11). The library pool was diluted and denatured according to the standard NextSeq protocol (Document # 15048776 v02), and sequenced to generate paired-end 76 bp reads using a 150 cycle NextSeq500/550 High Output reagent Kit v2 (Catalog # FC-404-2002). After sequencing, fastq files were generated using bcl2fastq2 (v2.17). Library preparation and sequencing was performed at the Institute for Molecular Bioscience Sequencing Facility (University of Queensland). Data analysis The RNA-seq reads were mapped to the Mus musculus mm10 genome with STAR version 2.5.3a 28 using Ensembl annotation (GRCm38, release 91). Reads were quantified with HTSeqcount version 0.6.1 29 using the attribute gene_id from the Ensembl GTF file GRCm38 release 91 as feature ID. Read count normalization and differential gene expression analysis was performed using DESeq2 version 1.10.1 30 . Immunofluorescence and imaging Dissected tumors were fixed for 2 h in 4% PFA. The fixative was removed with 3× washes of 1× PBS (Amresco, Solon, OH, USA). Tissues were subsequently infused with sucrose before cryo-embedding. For staining of specific antigens, cryo-sections were permeabilized in 0.5% Triton-X-100 (Chem Supply, Gillman, South Australia) before blocking with 20% normal goat serum. For this study, primary antibodies used included rat anti-mouse CD31 (1:50), rat anti-mouse VEGFR2 (1:50), rat anti-mouse CD34 (1:50) (all from Becton Dickinson); rabbit anti-mouse LYVE-1 (1:100), rabbit anti-mouse Dll4 (1:250), rabbit anti-mouse endomucin (1:100) (Abcam, MA, USA) and hamster anti-mouse Podoplanin (1:100) (AngioBio, CA, USA). Excess and unbound antibody was then removed with 3 × 5 min washes in a solution containing 1× PBS/0.1% Tween-20 (Amresco, Solon, OH, USA). Secondary antibodies conjugated with Alexa-Fluor 568 or 647 (Invitrogen, Carlsbad, CA, USA) were used for fluorescence detection. Briefly, sections were incubated with secondary antibodies for 40 min at room temperature. Excess antibody was removed by 3× washes in PBS/0.1% Tween-20. Nuclear staining was revealed in specimens mounted with ProLong® Gold mounting media containing DAPI (Invitrogen, Carlsbad, CA, USA). Confocal images were acquired with a Zeiss LSM 710 microscope equipped with Argon 561-10 nm DPSS and 633 nm HeNe lasers, and a 405-30 nm diode. Images were obtained at 10× and 20×. Immunofluorescence vessel quantification was conducted using ImageJ (NIH). Vascular beds were located and imaged as described above. The region of interest was kept consistent, and vascular beds were overlaid with the YFP channel and quantified as a percentage overlap standardized to YFP coverage. Statistical analysis All statistical analyses were performed using GraphPad Prism v5c software. Data were analyzed using the following tests: Mann–Whitney (for non-parametric data), T -tests 1-way or 2-way ANOVA with Bonferroni correction for parametric data. A p -value < 0.05 was considered significant. Reporting summary Further information on experimental design is available in the Nature Research Reporting Summary linked to this article. Data availability All sequencing data that support the findings of this study have been deposited in the NCBI Gene Expression Omnibus (GEO) and are accessible through the GEO Series accession number GSE114528 .

Stopping melanoma from spreading to other parts of the body might be as simple as cutting off the blood supply to the cancer, according to researchers. Scientists from The University of Queensland's Diamantina Institute have discovered stem cells which form blood vessels in tumours, and have identified how to 'switch the cells off'. Professor Kiarash Khosrotehrani (below) said the study's findings had enormous implications for cancer patients. "Blood vessels are vital because tumours can't grow without them – they feed the tumours and allow the cancer to spread," Professor Khosrotehrani said. "If you get rid of these stem cells, then the blood vessels don't form and the tumours don't grow or spread to other locations." Professor Khosrotehrani said being able to block blood vessel development could be useful in treating recently diagnosed patients as it may help to prevent the cancer from spreading at an early stage. "This idea has been around for a while, but it has proven difficult to achieve because blood vessel formation is a fundamental mechanism by which our body responds to injury," he said. "Directly targeting the stem cells that form these blood vessels is a new approach that could make the difference." The research team will test the ability of a compound to stop these stem cells from forming blood vessels, in a study supported by National Health and Medical Research Council (NHMRC) funding. Researcher Dr. Jatin Patel said melanoma's ability to quickly spread from the skin to other parts of the body was what made it so deadly. "We know that before tumours spread to places like lymph nodes or lungs, the body starts growing extra blood vessels in these areas – almost as if preparing special 'niches' for the cancer," Dr. Patel said. "Our next study will focus on blocking the development of these niches. "If the body doesn't prepare them, then the cancer won't grow there."

10.1038/s41467-018-07961-w

Medicine

Dead probiotic strain shown to reduce harmful, aging-related inflammation

Shaohua Wang et al, Lipoteichoic acid from the cell wall of a heat killed Lactobacillus paracasei D3-5 ameliorates aging-related leaky gut, inflammation and improves physical and cognitive functions: from C. elegans to mice, GeroScience (2019). DOI: 10.1007/s11357-019-00137-4

http://dx.doi.org/10.1007/s11357-019-00137-4

https://medicalxpress.com/news/2019-12-dead-probiotic-strain-shown-aging-related.html

Abstract Increased inflammation associated with leaky gut is a major risk factor for morbidity and mortality in older adults; however, successful preventive and therapeutic strategies against these conditions are not available. In this study, we demonstrate that a human-origin Lactobacillus paracasei D3-5 strain (D3-5), even in the non-viable form, extends life span of Caenorhabditis elegans . In addition, feeding of heat-killed D3-5 to old mice (> 79 weeks) prevents high- fat diet-induced metabolic dysfunctions, decreases leaky gut and inflammation, and improves physical and cognitive functions. D3-5 feeding significantly increases mucin production, and proportionately, the abundance of mucin-degrading bacteria Akkermansia muciniphila also increases. Mechanistically, we show that the lipoteichoic acid (LTA), a cell wall component of D3-5, enhances mucin ( Muc2 ) expression by modulating TLR-2/p38-MAPK/NF-kB pathway, which in turn reduces age-related leaky gut and inflammation. The findings indicate that the D3-5 and its LTA can prevent/treat age-related leaky gut and inflammation. Access provided by DEAL DE / Springer Compact Clearingstelle Uni Freiburg _ Working on a manuscript? Avoid the common mistakes Introduction The world’s population is rapidly aging with a concomitant increase in aging-related comorbidities like obesity, diabetes, cardiovascular diseases, cancer, infections, and cognitive decline (Hodes et al. 2016 ; Jaul and Barron 2017 ; Tabloski 2004 ). The mechanism(s) of how aging contributes to these health ailments are still unclear, but low-grade inflammation is often found to be higher in older adults and is a major risk factor for mortality and morbidity (Franceschi and Campisi 2014 ). The pathways leading to low-grade inflammation are not completely understood, but increased gut permeability (“leaky gut”) can allow non-selective diffusion of microbial and dietary antigens from the gut lumen into the lymphatic and blood circulation, which in turn promotes inflammatory responses locally (gut mucosa) and systemically (Konig et al. 2016 ; Ulluwishewa et al. 2011 ). Thus, effective strategies to reduce leaky gut and consequent inflammation are needed. Emerging evidence suggests that detrimental perturbations in gut microbiome (dysbiosis) are associated with leaky gut, inflammation, and poor health outcomes in older adults. Gut permeability is tightly controlled by intestinal barriers including tight junction proteins and a mucus layer (Capaldo et al. 2017 ). The mucus layer functions as a physical barrier to separate gut microbes and host cells and protects from the abnormal leakage and invasion of pathogens and antigens into the circulation (Konig et al. 2016 ). The intestinal mucus layer covers the epithelial cells by making a viscoelastic gel layer, which is synthesized by goblet cells (Cornick et al. 2015 ). Intestinal goblet cells are interspersed between epithelial cells on the villi of intestinal lumen and constantly secrete mucins to form the mucus layer (Johansson 2012 ). Mucin 2 ( Muc2 ) is the major mucin secreted from intestinal goblet cells and is known to protect leaky gut and invasion of microbes (Johansson 2012 ; Vemuri et al. 2018 ). Muc2 knockout (KO) mice develop colitis with a dramatically inflammatory gut mucosa (Van der Sluis et al. 2006 ) associated with a marked decrease in mucus layer’s thickness, thereby highlighting the importance of Muc2 in maintaining gut integrity and inflammation. A thin mucus layer may allow gut microbes to invade intestinal epithelial cells and to activate inflammatory responses in immune cells (Allaire et al. 2018 ). The mucus layer’s thickness is also dramatically reduced in the gut of older adults, which is also linked with increased leaky gut and inflammation (Elderman et al. 2017 ). Given the importance of the mucus layer in protection against microbial translocation, development of gut modulators to promote gut mucin production may be beneficial to ameliorate aging-related leaky gut and inflammation. We hypothesize that probiotics can be used to modulate gut microbiome and ameliorate aging-related leaky gut and inflammation. Probiotics are live microorganisms that extend health benefits to the host/consumers upon administering in sufficient amounts (Hill et al. 2014 ). The most commonly used probiotics are non-spore-forming, gram-positive, lactic acid-producing bacteria such as specific Lactobacillus and Bifidobacterium strains that have been shown to have health-promoting effects against several human diseases (Azad et al. 2018 ; Landete et al. 2017 ; O'Toole et al. 2017 ). However, the precise mechanisms by which probiotics exert their benefits in aging remain largely unknown. Also, probiotics are thought to exert their beneficial effects as live cells; however, whether dead probiotics can also have beneficial effects is not well-known. Given the fact that commonly used probiotics like lactobacilli and bifidobacteria are Gram-positive, it has been hypothesized that bacterial molecules, such as bacterial cell wall constituents including peptidoglycan (PG) and lipoteichoic acid (LTA) and bacterial cytoplasm components including proteins and bacterial DNA, can have specific biological activities (Мokrozub et al. 2015 ). Here, we demonstrate that LTA from the cell wall of a human-origin probiotic bacterium, L. paracasei strain D3-5 (called D3-5 hereafter), exhibits potent activity to stimulate mucin production and reduces aging-related leaky gut and inflammation, which in turn improves physical and cognitive functions in older mice. Mechanistically, LTA from D3-5 stimulated Muc2 expression in goblet cells via activation of toll-like receptor 2/p38-MAPK (TLR-2/p38-MAPK)/NF-κB pathway. Our results highlight the potential of a dead probiotic strain and its cell wall-derived LTA to prevent and/or treat aging-related leaky gut and inflammation and to improve overall cognitive and physical functions in older adults. Results Dead L. paracasei D3-5 feeding extends lifespan in Caenorhabditis ( C.) elegans The onset of aging is associated with decline in muscle mass, and in physical and cognitive functions. Wild-type C. elegans N2 is a widely used animal model in several anti-aging screening studies (Mack et al. 2019 ; Tissenbaum 2015b ). In a previous study (Nagpal et al. 2018b ), we isolated six probiotic Lactobacillus strains (viz. L. plantarum SK9, L. plantarum D6-2, L. paracasei D3-5, L. paracasei D10.4, L. rhamnosus D4-4, and L. rhamnosus D7-4) from healthy infants. Here, to investigate if these strains displayed anti-aging benefits, we screened them using the wild-type C. elegans N2 (Tissenbaum 2015a ). Since longevity assay of C. elegans required addition of antibiotics and antimycotics to prevent contamination in long-term, all the bacteria used in C. elegans survival assay were considered killed by antibiotics. To control for the wide differences of bacterial genera and species as well as to prove the strain-dependent effects, we included two different strains but from the same genera and species. Only two strains of dead probiotics, L. paracasei D3-5 and L. plantarum SK9, exhibited beneficial effects of extending lifespan and preserving better physical function when fed to wild-type C. elegans N2. Of these, D3-5 exhibited the strongest effects (Fig. 1a, b ; Supplementary Figure S1 a). Interestingly, D3-5 feeding declined muscle mass with significantly low rate compared to their control ( E. coli OP50 strain)-fed and species mate L. paracasei 10-4-fed groups (Fig. 1c, d ), suggesting that decline in physical function is associated with higher muscle mass in D3-5 fed C. elegans . Body length was shorter in the worms with lower survival including D10-4 fed worms (Fig. 1c ; Supplementary Figure S1 b). To determine whether D3-5 feeding increased muscle mass or persevered the decline of muscle in older worms, we calculated (i) the ratio of muscle mass vs body length (which gives similar measure of BMI in humans) and (ii) the longitudinal muscle mass decline by subtracting muscle mass on day 7 from day 11, within the same groups. First analysis showed no significant differences in the ratio of muscle mass/body length in worms fed with D3-5 and D10-4 compared to E. coli OP50 fed control worms at the adult age of day 7 (Fig. 1c, e ), whereas second analysis demonstrated that muscle mass decline was significantly less in worms fed with D3-5 compared to E. coli OP-50 and D10-4 (Fig. 1d–f ; Supplementary Figure S1 c). These results suggest that D3-5 feeding preserved higher muscle mass at the older stage (day 11), without increasing it an adult age (day 7). No significant differences were observed in food intake (as measured by pharyngeal pumping rate) among all the groups (Supplementary Figure S1 d). Other strains, regardless of genera and species, showed no significant changes in the lifespan of C. elegans compared to control OP50-fed worms. These results indicate that (i) certain dead probiotics are beneficial to aging-related ailments, in terms of extending life-span and preserving better physical function and muscle mass, and (ii) these beneficial effects are strain-specific, as not all probiotics from the same species/genera exhibit beneficial effects against aging. Fig. 1 Feeding dead probiotics extends lifespan and improves physical function and muscle mass in C. elegans . a Selective strains of probiotics lactobacilli extended life-span of wild-type C. elegans N2 compared to E. coli OP50 strain-fed controls. b D3-5 feeding reduced physical function (movement) decline of C. elegans compared to their control OP50- and other lactobacilli strains-fed groups. c – e It also maintained higher body length (at day 7) ( c ), muscle mass (indicated by GFP-labelled MAH19 strain of C. elegans at × 4 magnification) at day 11 ( d ), and muscle mass/body length ratio at days 7 and 11 ( e ). f Muscle mass decline was prohibited with D3-5 feeding compared to their control OP50- and D10-4 fed groups. Values are mean of 150–195 in each group ± SEM (standard error of mean; error bars) done three times in five replicates (10–17 worms each replicate at each batch). Values with * p < 0.05, ** p < 0.001 are statistically significant. ns: not significant Full size image D3-5 feeding prevents HFD-induced metabolic dysfunctions in older mice As D3-5 feeding prolonged the lifespan of C. elegans , we investigated its effects in older obese mice (> 79 weeks or > 18 months old that are equivalent to > 65–70 years old humans) that were fed with HFD to induce metabolic dysfunction. Ten weeks of D3-5 feeding prevented the development of glucose intolerance and insulin resistance (measured by oral glucose tolerance test [GTT] and insulin tolerance test [ITT], respectively) in HFD-fed older mice compared to their age, gender, and diet-matched counterparts (Fig. 2a, b ). Body weight, fasting, and fed blood glucose levels and energy expenditure were not significantly changed after D3-5 treatment compared to control older mice (Supplementary Figure S2 a-j). Further, feeding of D3-5 significantly reduced hepatic steatosis or non-alcoholic fatty liver disease (NAFLD) indicated by fat accumulation in the liver and decreased crown-like structures (indicator of inflammation) in the white adipose tissue (WAT) (Fig. 2c, d ; Supplementary Figure S3 a). WAT adipocyte size was also significantly reduced in D3-5 fed older mice compared to controls (Fig. 2e ; supplementary Figure S3 b). Overall, these results demonstrate that the D3-5 feeding prevented HFD-induced glucose intolerance, insulin resistance, hepatic steatosis/NAFLD, and inflammation in WAT with small adipocytes, indicating that D3-5 feeding ameliorates detrimental metabolic effects of HFD in older mice. Fig. 2 Feeding dead probiotics prevents HFD-induced metabolic dysfunctions in older mice. a , b Feeding of dead L. paracasei D3-5 prevented high fat diet (HFD)-induced glucose intolerance ( a ) and insulin resistance in older mice ( b ). c – e D3-5 feeding also prevented HFD-induced hepatic steatosis (fat accumulation in the liver) (upper panel— c ) and reduced crown like structures (indicators of inflammation) (lower panel— c , d ) and reduced adipocyte size of white adipose tissue ( e ) of older mice. Values are mean of n = 5–8 mice per group ± SEM (error bars). Values with * p < 0.05, ** p < 0.001 are statistically significant. ns: not significant Full size image D3-5 improves physical and cognitive functions, with reduced leaky gut and inflammation in older obese mice Obese older adults commonly have poorer physical function, increased anxiety, and decreased cognition (learning and memory) (Dahl et al. 2013 ; DeJesus et al. 2016 ; Virta et al. 2013 ). Therefore, we tested the effects of D3-5 feeding on these ailments in older HFD-fed mice. We found that the older obese mice fed with D3-5 maintained better physical function, as evidenced by increased ambulatory and total motor activity (indicator of walking) compared to age- and gender-matched controls (Fig. 3a, b ). In addition, other physical function indicator, i.e., hanging time on inclined screen (a measure of muscle strength), was significantly improved in D3-5-fed older mice (Fig. 3c ). D3-5 feeding also significantly decreased anxiety (an another common ailment in older people) (Borta and Schwarting 2005 ) as indicated by (i) decreased rate of rearing and (ii) increased time spending in the center during open field test (Fig. 3d–g ). We also found that the D3-5 feeding significantly increased cognitive function in terms of enhanced learning improvements during Morris Water Maze test in older mice compared to their non-treated controls (Fig. 3h ). Because leaky gut and inflammation are often greater in older humans and mice and are major risk factors for poor health outcomes in older adults (Nagpal et al. 2018a ), we next assessed the gut permeability and inflammation in these mice. We found that D3-5 feeding significantly decreased gut permeability (a marker of leaky gut) as assessed by reduced diffusion of FITC-dextran (3–5 kDa) from gut to the blood (Fig. 3i ), and this effect was associated with lower expression of inflammatory genes including IL-6 and IL-1β in the colon of D3-5-fed older obese mice compared to controls (Fig. 3j, k ). Together, these results indicate that D3-5 feeding led to better physical and cognitive functions and reduced anxiety/depression which was also associated with decreased gut permeability and inflammation in older HFD-fed obese mice. Fig. 3 D3-5 feeding improves aging-related ailments like physical function, muscle strength, depression/anxiety, learning-memory, leaky gut, and inflammation in older obese mice. a , b Physical function [measured as ambulatory activity ( a ) and total activity ( b )], muscle strength [measured as inclining on tilted screen ( c )], anxiety/depression behavior [measured as rearing ( d ) and time spent in center during open field test ( e – g )], and learning-memory behavior [measured during Morris Water Maze test ( h )] were significantly enhanced in D3-5 fed older obese mice compared to their controls. Leaky gut [measured by appearance of FITC in blood leaked from gut ( i )] and mRNA expression of inflammatory markers (like IL-6 and IL-1β) were significantly decreased in D3-5 fed older obese mice compared to their age-sex-matched controls ( j , k ). Values are mean of n = 5–8 mice per group ± SEM (error bars). Values with * p < 0.05, ** p < 0.001 are statistically significant. ns: not significant Full size image D3-5 beneficially modulates gut microbiome in older obese mice We did not expect the feeding of dead D3-5 to have a major impact on gut microbiome, because of its dead form. Instead, we found that the gut microbiome signature was significantly different in older mice fed with D3-5 compared to their controls (Fig. 4 ). The β-diversity was dramatically different in D3-5 fed older mice compared to their controls (Fig. 4a ), without significant changes in α-diversity indices (Supplementary Figure S4 a–d). However, microbial signatures were distinctly clustered in D3-5 compared to their controls (Supplementary Figure S4 e). Further, D3-5 feeding significantly decreased the abundance of metabolically detrimental bacterial phyla including Firmicutes while increasing the abundance of metabolically beneficial Verrucomicrobia phyla and Verrucomibiaceae family (Fig. 4b, c ). Feeding with D3-5 increased the abundance of Verrucomicrobiaceae , Verrucomicrobiales , Verrucomicrobiae , and Akkermansia species, while decreasing the ratios of Actinobacteria , Adlercreutzia , Coriobacteriales , and Coriobacteriia (Fig. 4d–f ; Supplementary Figure S4 f). The abundance of common mucin-degrading bacterial species Akkermansia muciniphila was significantly increased in D3-5-fed older mice compared to their controls (Fig. 4g ). Fig. 4 D3-5 feeding results in favorable changes in gut microbiome. a Principal coordinate analysis (PCoA) showing the β-diversity clustering of gut microbiome from older obese mice fed with D3-5 and the control group. Major changes in bacterial phyla ( b ), families ( c ), and genera ( d ) in mice fed with D3-5 compared to their controls. e Abundance of Akkermansia was significantly increased in D3-5 fed mice compared to controls. f Cladograms of linear discrimination analysis (LDA) effect size (LEfSe) demonstrating the clustering of gut microbiome in D3-5-fed older mice compared to the controls. g Abundance of A. muciniphila was significantly increased in D3-5 treated older obese mice compared to their non-treated controls. Values are mean of n = 5 mice per group ± SEM (error bars). Values with * p < 0.05, ** p < 0.001 are statistically significant. ns: not significant Full size image D3-5 administration increases mucin production To explain the dramatic increase of A. muciniphila by D3-5 feeding, we reasoned that mucin (food source for A. muciniphila ) (Zhou 2017 ) levels may have increased in the gut of D3-5-fed older mice, leading to the increase in A. muciniphila as a feedback mechanism. Muc2 (a major isoform of mucin abundantly expressed in intestine) expression was significantly increased in duodenum, ileum, and colon of D3-5 fed older obese mice compared to their controls (Fig. 5a ). D3-5 feeding also significantly increased the expression of mucin protein in colon along with increased mucin secretion in the feces compared to their controls (Fig. 5b, c ). Mucin is secreted from goblet cells (Pelaseyed et al. 2014 ), and the goblet cell mass was significantly higher (34.6%) in D3-5 fed mice compared to their non-treated controls (Fig. 5d, e ), suggesting that the observed increase in mucin production could be explained by a larger goblet cell mass. Fig. 5 D3-5 treatment enhances mucin production and goblet cell mass in the gut of older obese mice. a Real-time PCR results indicated that the expression of mucin 2 ( Muc2 ) mRNA was significantly higher in the different regions of the intestine of D3-5 fed older mice compared to control mice. b , c Expression [measured by Western blotting in colon ( b )] and production [measured using ELISA in feces ( c )] of mucin was significantly higher in older obese mice fed with D3-5 compared to controls. d , e AB/PAS staining depicting more intestinal goblet cell mass in ileum ( d ) and their quantitative numbers ( e ) was significant in D3-5 fed older obese mouse gut compared to their controls. f – j Transplantation of fecal microbiome from older mice fed with D3-5 and their controls to gut-cleansed recipient mice also showed enhanced expression of Muc2 mRNA ( f ) and protein ( g ), marginal increase in fecal mucin ( h ), and goblet cell mass numbers ( i , j ). Real-time PCR reaction was performed in triplicates and repeated at least two to three times. Values are mean of n = 5–8 mice per group ± SEM (error bars). Values with * p < 0.05, ** p < 0.001 are statistically significant. ns: not significant Full size image To further explore whether D3-5-induced gut microbiome changes contribute in the modulation of goblet cell and mucin biology, we transplanted gut microbiome from D3-5 fed (FMT-D3-5) and their control mice (FMT-Control) to gut cleaned (GC; using antibiotics and polyethylene glycol [PEG] protocol) mice (Wrzosek et al. 2018 ). Muc2 mRNA and mucin protein as well as goblet cell mass were significantly increased in the gut of FMT-D3-5 recipient GC mice compared to FMT-control recipient GC mice (Fig. 5f–j ). These results indicate that D3-5-induced gut microbiome modulation contributes in promoting goblet cell mass and mucin production. Lipoteichoic acid (LTA) derived from the cell wall of D3-5 increases mucin production In order to discover the specific cellular components of dead D3-5 which enhance the mucin production, we first fractionated cell wall and cytoplasmic components and subsequently treated CMT93 cells (a mouse goblet cell line) with these components. Muc2 mRNA expression and mucin content (labelled with PAS staining) were increased only in the cell wall fraction of D3-5-treated cells (Fig. 6a, b ). To further determine which D3-5 cell wall constituents that enhanced mucin production, we fractionated the D3-5 cell wall into the peptidoglycan and LTA fractions, and found that the LTA fraction-treated CMT93 goblet cells exhibited significantly higher expression of Muc2 mRNA and mucin/PAS staining, suggesting that the LTA from the D3-5 cell wall activates mucin production from goblet cells (Fig. 6c, d ). Interestingly, the LTA derived from L. paracasei D3-5 has only stimulated Muc2 mRNA expression in CMT-93 cells, while LTA derived from L. paracasei D10-4 demonstrated no effects (Fig. 7a ), suggesting that LTA derived from D3-5 has unique biological activity to stimulate Muc2 mRNA expression in goblet cells. Fig. 6 Cell wall-derived lipoteichoic acid (LTA) of D3-5 enhances mucin production. a , b Cell wall fraction of D3-5 containing LTA increased mRNA expression of Muc2 ( a ) and production of mucin (PAS staining indicated in blue) in the mouse intestinal goblet cell line CMT93 cells ( b ). c , d D3-5 cell wall derived LTA (200 μg/ml) increased both Muc2 mRNA ( c ) and mucin production ( d ) compared to peptidoglycan (PGN)-enriched fraction. All the assays were performed in triplicate and were repeated at least two to three times. Values are mean of replicates and repeated batches ± SEM (error bars). Statistically significant ( p values < 0.05) was obtained using Student’s t test and/or ANOVA. Values with * p < 0.05, are statistically significant. ns: not significant Full size image Fig. 7 LTA induces mucin via TLR2 signaling. a Treatment of mouse intestinal goblet CMT-93 cells with 200 μg/ml LTA derived from D3.5 cell wall increased Muc2 mRNA expression, while LTA derived from D10.4 cell wall shows no effects. b mRNA expression of Muc2 was significantly increased in D3.5 LTA-treated CMT93 goblet cells, while this induction disappeared in TLR2 inhibitor (CU CPT22; 8 μM)-treated cells. c Mucin protein expression (measured by Western blot) in the colon was also increased in 5-day D3-5 and LTA-treated C57BL/6J (B6) young mice, while these effects of D3-5 and LTA treatment disappeared in age- and gender-matched TLR2 KO mice. d – g Similarly, goblet cell mass (AB/PAS staining) and numbers ( d , e ) were increased in D3-5 and LTA-treated B6 control mice, while these effects were not seen in TLR2 KO mice ( f , g ). Values are mean of n = 5–8 mice per group ± SEM (error bars). Values with * p < 0.05, ** p < 0.001 are statistically significant. ns: not significant Full size image LTA activates TLR-2/p38-MAPK signaling to promote Muc2 expression and suppress NFκB to reduce inflammation To further establish how D3-5-derived LTA increases Muc2 expression and mucin production, we hypothesized that LTA activates the toll-like receptor-2 (TLR-2) signaling and enhances Muc2 expression in intestinal goblet cells. LTA treatment-stimulated increase in Muc2 mRNA in CMT93 cells was significantly abolished in the presence of TLR2 inhibitor (CU CPT22) (Fig. 7b ), suggesting that TLR2 signaling mediates the effects of LTA to enhance the expression of Muc2 in CMT93 goblet cells. In addition, oral administration of D3-5 and its cell wall-derived LTA for 1 week significantly increased Muc2 expression as well as enhanced goblet cell mass in the mouse gut. However, such effects of D3-5 and its cell wall LTA failed in TLR2 knockout (KO) mice (Fig. 7c–g ), further indicating that LTA required TLR2 signaling activation to enhance goblet cell mass and mucin production. LTA also significantly reduced the inflammatory markers, i.e., IL-6, IL-1β, and TNF-α in the intestine of C57BL/6J mice; however, these effects were disappeared in the TLR2 KO mice (Fig. 8a–f ), suggesting that LTA-reduced inflammation also depends on intact TLR2 signaling. These results suggest that LTA stimulated TLR2 signaling to enhance goblet cell mass and mucin production, which in turn reduced leaky gut and intestinal inflammation. Fig. 8 D3-5 cell wall-derived LTA activates TLR2-p38 MAPK signaling and suppresses NF-κB signaling. a – f D3-5 and its cell wall-derived LTA feeding significantly decreased mRNA expression of pro-inflammatory genes like IL-6, IL-1ββ and TNF-α in the colon of normal B6 (C57BL/6J) mice ( a – c ), while these effects are abolished in the TLR2 KO mouse colons ( d – f ). g Western blot analyses show that D3-5 and its cell wall-derived LTA increased abundance of phospho-p38-MAPK, and decreased NF-κB, with no changes in p38-MAPK, Phospho-JNK, and JNK in the colons of older mice and B6 young mice. Such effects disappeared in TLR2 KO mice. Values are mean of n = 5–8 mice per group ± SEM (error bars). Values with * p < 0.05, ** p < 0.001 are statistically significant. ns not significant Full size image To further discover which signaling mediators of the TLR2 pathway participate in response to LTA, we found that the levels of phosphorylated p38 MAPK proteins were significantly increased in the intestines of both old and young C57BL/6J mice, while this activation was abolished in TLR2 KO mice (Fig. 8g ). However, total p38 MAPK as well as phospho- and total-JNK proteins did not significantly change upon LTA treatment (Fig. 8g ), suggesting that p38-MAPK is activated downstream of TLR2, without impacting JNK signaling. The protein levels of NF-κB, a master regulator of inflammatory pathways, were significantly decreased in LTA-treated wild-type mice, but not in TLR2 KO mice (Fig. 8g ). Together, these data show that LTA activation of TLR2-p38 MAPK signaling enhanced mucin production, suppressed NF-κB signaling, and reduced inflammatory cytokines. Discussion Increasing prevalence of age-related ailments including decline in physical and cognitive functions and increased predisposition to obesity, diabetes, cardiovascular diseases, and cancer is significantly associated with dysbiotic gut microbiome, leaky gut, and inflammation (Buford 2017 ; Nagpal et al. 2018a ; Shimizu 2018 ). Herein, we demonstrated that a dead probiotic strain L. paracasei D3-5 isolated from infant gut enhanced life-span and maintained better physical function and muscle mass in C. elegans and prevented HFD-induced metabolic dysfunctions, leaky gut, and inflammation in older mice. Mechanistically, D3-5 cell wall-derived LTA enhanced goblet cells and mucin production by activating TLR2-p38 MAPK signaling and reducing inflammation by inhibiting NF-κB resulting in decreased expression of the pro-inflammatory cytokines IL-6, IL-1β, and TNF-α (Fig. 9 ). Fig. 9 Proposed model of how dead L. paracasei D3-5 and its cell wall-derived LTA ameliorate leaky gut and inflammation, and promotes healthy aging. Mechanistically, D3-5-derived LTA activates TLR2/p38-MAPK signaling resulting in increased mucin production to suppress leaky gut and inhibit NF-κB to reduce inflammation Full size image Although the precise mechanisms underlying increased inflammation in older adults remain elusive, increased gut epithelial permeability in patients presenting leaky gut is common in older adults and can serve as a major trigger for intestinal and systemic inflammation (Stehle Jr et al. 2012 ). Emerging evidence indicates that gut microbiome dysbiosis is associated with poor health outcomes in older adults and may contribute to the development of leaky gut and inflammation (Nagpal et al. 2018a ). The abundance of certain detrimental Gram-negative bacteria are often increased in the older gut (Schiffrin et al. 2010 ), elevating the levels of pro-inflammatory molecule, i.e., LPS (Lipopolysaccharide), an endotoxin and major constituent of Gram-negative bacterial cell wall. LPS can diffuse in leaky gut conditions, causing endotoxemia and inflammation in local tissues as well as systemic inflammation (Mu et al. 2017 ). Therefore, gut microbiome modulators that are able to (i) reduce the abundance of Gram-negative bacteria and (ii) improve the gut barrier integrity by enhanced mucus layer thickness and expression of tight junctions could reduce leaky gut and inflammation (Kelly et al. 2015 ). Probiotics can be ideal candidates for this purpose because common probiotics are Gram-positive and therefore can balance the growth of Gram-negative bacteria in older gut, thereby reducing the LPS load and the associated inflammation. In addition, specific strains of probiotics have been shown to induce mucin production. For example, bifidobacteria increase mucus production and protect from HFD-induced microbiota dysbiosis and obesity (Schroeder et al. 2018 ), and Lactobacillus GG upregulates Muc2 gene expression in intestinal epithelial cells (Wang et al. 2015 ). However, the impact of probiotics in age-related leaky gut, inflammation, and mucin biology remains elusive. Also, specific probiotics are known for their beneficial effects in several human diseases including diabetes, obesity, cardiovascular diseases, and cancer (Azad et al. 2018 ; Landete et al. 2017 ), but their impact on longevity is not well defined. In this study, we demonstrated that only a few selected human-origin probiotic strains exhibit beneficial effects in extending the lifespan of C. elegans . Among the six selected probiotic strains (Nagpal et al. 2018b ), only two strains of lactobacilli like L. plantarum SK9 and L. paracasei D3-5 fed C. elegans show beneficial effects to extend life-span. D3-5 feeding showed the highest extension in life-span and hence was selected for further studies. These results indicate that not all probiotics have anti-aging effects and that these effects are highly strain-specific. Interestingly, D3-5 feeding also enhanced physical function and maintained higher muscle mass in older C. elegans , suggesting that the probiotic D3-5 prevented the aging-related muscle mass decline. The C. elegans longevity assay protocols require addition of antibiotics and antifungal to prevent contamination of the worm media. Thus, probiotics fed to C. elegans were considered and confirmed dead. That led us to discover that the dead D3-5 has anti-aging effects including increased life-span of C. elegans and preserved muscle mass and physical functions. These findings are unusually significant in several ways: (i) they establish a new paradigm that dead probiotics can also have beneficial effects, in contrary to the old paradigm suggesting that probiotics are needed to be alive to exhibit beneficial effects; (ii) dead probiotics eliminate the risk of leakage of live bacteria systemically in patients with severe leaky gut, thus increasing the risk of bacteremia/sepsis; and (iii) dead probiotics may offer an advantage for food industry applications as they can easily be supplemented in several food lines/products not amenable for live probiotics, such as beverages, baked products, and others. Obesity and metabolic dysfunctions are common in older adults with poor health outcomes, and HFD-feeding induces these ailments much faster in older mice compared to younger ones (Dominguez and Barbagallo 2016 ; Nunes-Souza et al. 2016 ). We found that D3-5 feeding to older mice prevented the HFD-induced glucose intolerance, insulin resistance, hepatic steatosis (lipid accumulation in liver), and low-grade inflammation in adipose tissue. D3-5-fed older mice maintained better physical and cognitive functions with reduced leaky gut and inflammation, indicating that the anti-aging effects observed with D3-5 in C. elegans were translated to mammals, like older obese mice. Intriguingly, we found that the D3-5 also significantly modulated the gut microbiome composition in these older obese mice, as it considerably enhanced the abundance of a mucin-degrading bacterium A. muciniphila . Indeed, A. muciniphila is known to have beneficial effects against obesity, diabetes (Dao et al. 2016 ; Everard et al. 2013 ), and other human diseases along with anti-inflammatory effects (Naito et al. 2018 ; Ottman et al. 2017 ). We hypothesized that the increased abundance of A. muciniphila might be due to increased mucin production in the gut of D3-5 fed mice. Because mucin is a major food source for A. muciniphila , increased mucin production may enhance the growth of this bacteria. Interestingly enough, the intestine of D3-5 fed mice had significantly increased mucin production along with higher Muc2 expression and goblet cell mass. The fecal microbiome transplantation studies demonstrated that the D3-5 modulated gut microbiome enhances goblet cell mass and mucin production. However, these changes might be due to the presence of D3-5 and its cellular ingredients in the donor feces which might have been transferred to the recipient mice gut and showing these residual effects. Using biochemical fractionation, we demonstrated that the D3-5’s cell wall-derived LTA is the key component responsible for enhancing Muc2 expression and mucin production from goblet cells. The LTA derived from L. paracasei D3-5 cell wall has this unique biological activity to stimulate Muc2 expression, while LTA of same species but different strain of probiotics like L. paracasei D10-4 has no effect on Muc2 mRNA expression. We further investigated the mechanism(s) by which LTA can stimulate goblet cells to produce mucin. We also found that LTA action depends on TLR2 signaling to induce Muc2 expression. Activation of TLR2 signaling further recruits and initiates downstream signaling cascade by phosphorylation of p38 MAPK (Kawai and Akira 2010 ; Ribeiro et al. 2010 ). Accordingly, D3-5 and LTA treatment increased phospho-p38-MAPK levels in the mouse intestine, indicating that LTA activates TLR2 and p38-MAPK signaling, which in turn increases Muc2 expression and goblet cell mass. Interestingly, our results also demonstrated that NF-κB protein levels were significantly decreased. Suppression of NF-κB signaling is known to diminish the expression of proinflammatory cytokines such as IL-6, IL-1β, and TNF-α (Tak and Firestein 2001 ). Thus, our results indicate that LTA downregulates NF-κB-mediated proinflammatory responses in the intestinal tissue milieu. Our future studies are planned to comprehend beneficial effects of LTA against aging-related leaky gut and inflammation, and determine the structural uniqueness and molecular mechanism(s), using cell culture, C. elegans , and mouse models. Altogether, our results demonstrated that a newly isolated human-origin probiotic strain D3-5 is beneficial to enhance longevity and ameliorate aging-related health ailments such as metabolic dysfunctions, leaky gut, and inflammation. D3-5 cell wall-derived LTA activated TLR2-p38 MAPK signaling that promoted Muc2 expression and mucin production from goblet cells thereby reducing the leaky gut and suppressing the NF-κB signaling eventually reducing the inflammation. These results demonstrated that the D3-5 and its cell wall-derived LTA could be used as a biotherapy to prevent and/or treat age-related gut microbiome dysbiosis, leaky gut, and inflammation in older adults. Methods C. elegans culture and longevity assay A wild-type strain of C. elegans N2 (Tissenbaum 2015a ) and a muscle-specific GFP-labelled strain MAH19 (Kumsta and Hansen 2012 ) (procured from CGC, University of Minnesota) were cultured by standard procedures on NGM complete agar media and E. coli OP50 (Rea et al. 2005 ). The life span screening assay was carried out using liquid medium in 96-well plates according to a protocol described elsewhere (Solis and Petrascheck 2011 ) at room temperature (21–22 °C). In brief, 10–17 age-synchronized C. elegans (L1 larva) were cultured in each well with S-complete media containing Ampicillin, Carbenicillin, and Amphotericin B (for inhibiting bacterial and fungal contamination) and 100 μg/mL fluorodeoxyuridine (FUDR; for blocking fertilization) in 96-well plates. C. elegans were fed (0.3 × 10 8 cfu/mL) with six human-origin probiotic Lactobacillus strains ( L. paracasei D3-5, L. paracasei D10-4, L. rhamnosus D4-4, L. rhamnosus D7-4, L. plantarum SK9, and L. plantarum D6-2) and were compared with E. coli OP50-fed controls. Fifty to 85 worms in five wells were used for each treatment, and the numbers of live worms were counted daily on the basis of body movement. Food intake was measured by counting the number of pharyngeal contractions per minute (pumping rate) using a microscope. The fraction of animals alive was expressed as the number of animals displaying body movement per total number of animals. Body length of the worms was measured using ImageJ tools. This assay was repeated three times using 10–17 worms in five replicates each time, accounting for 150–195 worms in each group for cumulative data presented in figures. For the movement assay, C. elegans were picked into M9 buffer on a glass slide. The number of left- and right-ward body bends within 10 s was counted. One left-ward and one right-ward bend counted equal to one stroke. The number of strokes within 10 s was multiplied by six to calculate the movement rate per min (60 s). The counting of body strokes was repeated five times for each animal, and 10 animals were tested for each treatment, which accounted 50 measures per group. To determine the movement decline on 7, 11, and 14 days of age in worms, we subtracted data from 7th day movement activities. To determine food intake, rate of pharyngeal pumping within 1 min was counted, while worms kept on S-complete media containing bacteria (food) in 6-well plates at 20 °C under microscope without any disturbances. This assay was repeated five times for each animal, and 10 animals were tested for each group. Mice Older male C57BL/6J mice (78–80-week-old; equivalent to > 65 years of human age) were purchased from the Jackson Laboratory and were randomized in two groups: (1) control and (2) D3-5 ( n = 5–8 animals in each group), after 2 weeks of acclimatization. Group 2 were fed with heat killed (70 °C incubation for 2 h and confirmed their death on MRS agar cultivation) D3-5 by supplementing the equivalent of 10 9 colony forming unit (cfu)/ml in drinking water, while control animals were fed with normal drinking water. Drinking water was freshly changed every day. Both groups of animals were fed with 60% kcal of fat high-fat diet (HFD; Research Diets). Body weight and food intake were measured weekly. Glucose and insulin tolerance tests, metabolic phenotyping, physical activity measures, behavior test, and gut permeability assays were performed as described below. Fecal microbiome transplant and gut cleansed (GC) mice were prepared a modified method of Wrzosek et al. ( 2018 ). GC mice were fed with antibiotics cocktail (containing ampicillin 1 g/l, metronidazole 1 g/l, neomycin 1 g/l, and vancomycin 0.5 g/l and 3 g/l sweetener) in drinking water for 4 days, followed by orally gavage four doses of PEG (425 g/l, 200 μl/mouse) at 20 min interval on day 5 (this procedure depletes > 95% live microbial counts in the feces). Four hours after the last PEG dose, mice were administered with freshly prepared cecal microbiome slurry suspension (200 μl/mouse). Microbiome slurry was prepared using 500 mg of cecal contents dissolved in 5 ml of reduced PBS (phosphate buffer supplemented with 0.1% Resazurin ( w / v ) and 0.05% l -cysteine-HCl) under anaerobic condition (Kang et al. 2017 ). Control mice (called CCM-control) received microbiome from older-control mice, while CCM-D3-5 mice received microbiome from older-D3-5-treated mice, and continued for additional 3 days to enrich donor microbiome in recipient GC mice. After 1 week, GC animals were euthanized and tissues were collected and used for histology and gene and protein expression assays (using RT-PCR and western blots, respectively). C57BL/6 wild-type and TLR2 KO mice (6–8 week-old) were randomized the mice into (1) control, (2) D3-5, and (3) LTA groups. Group 2 was fed with dead D3-5 bacteria (as described above), and group 3 was fed with LTA (6 mg/mL) based on the body weight (4 ml/kg) through oral gavage for 5 days. Control animals were fed with normal drinking water in the same way. Body weight, food and water intake, and feces were collected before, during, and after feeding period, and all the animals were then euthanized and tissues were collected for histological and gene expression analyses. All the animal experiments and procedures were approved by the Institutional Animal Care and Use Committee (IACUC) of Wake Forest School of Medicine. Glucose and insulin tolerance tests Glucose tolerance test (GTT) and insulin tolerance test (ITT) were performed in fasted (10–12 h for GTT and 4–6 h for ITT) mice by giving oral gavage of glucose (2.5 g/kg body weight) and intraperitoneal injection with insulin (0.75 U/kg body weight), respectively. Blood glucose was measured from tail puncture before (0 min) and 15, 30, 60, and 120 min after administration using Truetrack® glucometer (Ahmadi et al. 2019 ; Yadav et al. 2011 ) Muscular endurance (incline screen) Treadmill and muscular endurance tests were performed in blinded groups, as described before (Silverstein et al. 2015 ). For four-limb testing, mice were suspended by all four limbs onto a high wire mesh. Timer recording started from the mesh was inverted and stopped when the mouse fell onto the padding below. Behavioral measures Open field test to determine anxiety and depression was performed in a quiet and dim room. Mice were placed individually into the center of a clean and sanitized white cylindrical tank (30-cm height × 50-cm diameter), and mice movements were recorded using camera for 5 min. The measures of behavior parameters such as average moving speed, total distance travelled, percentage distance covered in the center, and percentage time spent in the center were assessed using a ToxTrac tool (Rodriguez et al. 2018 ). Spatial learning and memory abilities were assessed using Morris Water Maze test (Nunez 2008 ). Tests were carried out in a clean and sanitized white cylindrical tank (30-cm height × 50-cm diameter) filled with tap water (25 °C) about 1 in below an island-like rescue platform placed at the center of the tank. During the pre-training, each mouse was trained three times in different directions. The water maze testing was performed at the same time next day with the same condition. Mice underwent 12 trials from four different directions. Time to reach the platform was recorded. All mice groups were blinded for the assessors. Gut permeability assay Intestinal/gut permeability was performed by giving oral gavage of FITC-dextran (3–5 kDa; 1 g/kg body weight) to mice pre-fasted for 4–6 h, and measuring the appearance of FITC fluorescence (excitation at 485 nm and emission at 520 nm) in serum with reference to the FITC standard curve, as earlier described by us (Ahmadi et al. 2019 ). Histological analyses Tissues from liver and intestine and white adipose tissue were collected, washed with PBS, and fixed in 10% formalin. Sections (0.5 μm thickness) were stained with hematoxylin and eosin (H&E) and imaged with AmScope microscope on × 20 magnification. Adipocyte size and distribution were assessed in ~ 480 random adipocytes from each mouse with ImageJ software. Crown-like structures (CLS)—indicative of inflammation in white adipose tissue (Apovian et al. 2008 ; Bagarolli et al. 2017 ; Ebke et al. 2014 ; Wijetunge et al. 2019 )—were measured by counting the total number of CLS proportionate to the total number of adipocytes in each section. To analyze the goblet cell mass, sections were stained with Alcian Blue/PAS kit (Newcomer Supply, Middleton, WI) and goblet cells (blue dots under the pink background); numbers were counted by a person blinded for the groups. Real-time PCR Total RNA was extracted from cells or tissues using RNeasy and was reverse transcribed using High-Capacity cDNA reverse transcription kit. The cDNA was used to quantify the expression of Muc2 , IL-6, IL-1β, and TNF-α using TaqMan Gene Expression Assays for real-time PCR. 18S rRNA was used as an internal control. Relative gene expression was calculated using ΔΔ CT method and presented as relative fold change. All the assays were performed in triplicate and repeated minimum two to three times. Values presented as mean of replicates and combination of repeated batches. Western blot Total proteins from cells and tissues were extracted using homogenization lysis buffer as earlier described by us (Yadav et al. 2011 ). Proteins were resolved by SDS-PAGE electrophoresis and transferred to PVDF membrane for Western blotting. Membranes were developed with primary antibodies (anti-Mucin 2, phospho-p38 MAPK, p38-MAPK, phospho-SAPK/JNK, SAPK/JNK, and NF-kB p65 from Cell Signaling) followed by secondary antibody and developing with chemiluminiscent ECL kit and imaged on PXi with the GeneSys software. Tubulin was used as the loading control. Fecal mucin Fecal mucin content was determined using an ELISA kit (Cosmo BioCo. Ltd) to quantify mucin within the feces in triplicate from each mice (Crowther and Wetmore 1987 ). Gut microbiome analyses The gut microbiota composition was analyzed using 16S rRNA sequencing with MiSeq Illumina platform and analyzed using bioinformatics pipelines, as per our previously described protocols (Ahmadi et al. 2019 ; Nagpal et al. 2018b ). In brief, genomic DNA was extracted from ~ 200 mg of mice feces ( n = 5 in each group) using the Qiagen DNA Stool Mini Kit (Qiagen, CA, USA). Primers 515 F (barcoded) and 806 R were used to amplify the V4 region of bacterial 16S rDNA. After being purified and quantified with AMPure® magnetic purification beads and Qubit-3 fluorimeter, respectively, equal amounts (8 pM) of the amplicons were applied for sequencing on an Illumina MiSeq sequencer (using Miseq reagent kit v3). The sequences were de-multiplexed, quality filtered, clustered, and analyzed with the Quantitative Insights into Microbial Ecology (QIIME) and R-based analytical tools (Navas-Molina et al. 2013 ). Linear discriminatory analysis (LDA) effect size (LEfSe) was used to identify unique bacterial taxa that drive differences between different groups (Ahmadi et al. 2019 ; Nagpal et al. 2018b ). Microbiome analyses were performed from n = 5 animals from each group. Cell wall, cytoplasm, peptidoglycan, and lipoteichoic acid preparation We prepared bacterial cell wall and cytoplasmic fractions as described by Kim et al. ( 2002 ) with slight modifications. In brief, L. paracasei D3-5 was cultivated in MRS at 37 °C up to log phase, and cell pellets were disrupted with a high pressure homogenizer (EmulsiFlex®-C3, AVESTIN, Inc., Canada) in citrate buffer. After removing cell debris, the supernatants were ultra-centrifuged at 70,000 g for 30 min (Beckman). The resulting supernatant was designated as cytoplasmic fraction, while the pellet as cell wall fraction. Further, the peptidoglycan-wall teichoic acid (PGN-WTA) was further fractionated using SDS according to the protocol developed by Heß et al. ( 2017 ). In brief, 4% SDS was added to the supernatant and incubated at 100 °C for 30 min, followed by stirring overnight at room temperature. Then, this solution was centrifuged (30,000 g , 15 min), and the pellets were washed four times with citrate buffer and five times with ethanol, and lyophilized to form PGN-WTA powder. Lipoteichoic acid (LTA) from D3-5 cell wall fraction was also extracted according to Heß et al. ( 2017 ). L. paracasei D3-5 cells were disrupted as described above, and cell homogenates were stirred with an equal volume of butanol for 30 min. Aqueous and organic phases were separated through centrifugation at 21,000 g for 15 min. The aqueous phase containing LTA was collected and dialyzed (3.5 kDa cutoff membrane) against water which was changed every 24 h for 5 days, and the resultant LTA solution inside the dialysis membrane was lyophilized to get the LTA powder. Cell culture and treatment The mouse mucin-secreting goblet cells CMT-93 were grown in 12-well plate (for RNA extraction) and 6-well plate (for protein extraction) for 24 h; cells were treated with 1% ( v / v ) of cell wall, cytoplasm, PGN, or LTA fractions derived from D3-5 bacterial cells. Cells were collected after 14-h treatment, and RNA and proteins were used for real-time PCR and Western blotting, respectively. TLR2 inhibitor (CU CPT22, 8 μM) was used to determine whether LTA action is involving TLR2 signaling in mucin production. Mucin production in cells was visualized using PAS Kit (Sigma), following the manufacturer’s protocols. Statistical analyses All the assays were performed at least two to three times with three to five replicates at each time and/or n = 5–8 animals in each group, and the values presented in graphs/tables are mean ± standard error of means. Statistical differences among groups/treatments were analyzed using two-tailed unpaired Student’s t test and/or ANOVA. Bacterial diversity and OTU abundance between groups were compared using Kruskal-Wallis test followed by Dunn’s post hoc analysis, and heatmaps were created in R statistical software package (version 3.4.3; ). P values with < 0.05 were considered statistically significant.

Scientists at Wake Forest School of Medicine have identified a dead probiotic that reduces age-related leaky gut in older mice. The study is published in the journal GeroScience. But what exactly is leaky gut and what does a probiotic—dead or alive—have to do with it? Some research has indicated that leaky gut, in which microbes and bacteria in the gut leak into the blood stream through holes or cracks in the intestinal lining, causes an increase in low-grade inflammation, and these conditions are common in older people. This resulting inflammation is thought to play a role in the development of many age-related diseases, such as diabetes, obesity, cancer, cardiovascular disease and decline in physical and cognitive functions. "We know that probiotics are instrumental in maintaining a healthy gut and preventing leakage, but there isn't much data available to pinpoint which ones work and which ones don't," said Hariom Yadav, Ph.D., assistant professor of molecular biology at Wake Forest School of Medicine and principal investigator of the study. "Determining the strain that is most effective at reducing leaky gut and inflammation would help us target more effective strategies to address the problem, and help explain why probiotics work for some people but not others." In the study, Yadav's team first screened eight strains of human-origin probiotics in roundworms, a commonly used model with a short lifespan of 11 to 20 days. They discovered that a strain of Lactobacillus paracasei (D3-5), even in the non-viable or heat-killed form, extended the roundworms' life span. They then tested their initial findings in mice. The results showed that feeding heat-killed D3-5 to older mice prevented high fat diet-induced metabolic dysfunctions, decreased leaky gut and inflammation, and improved physical and cognitive functions. "Not only did we determine which probiotic strain was the most effective in preventing leaky gut and inflammation, we also showed that the dead version of that probiotic had the same benefits," Yadav said. "This is the first-of-its kind study to show that a component (lipoteichoic acid) from the cell wall of a dead probiotic induced changes in the gut microbiome and mucin production, thereby reducing leaky gut and inflammation in elderly mice. "We think our findings could be very useful to the food and supplement industries because dead probiotics have the potential to be more stable and have a longer shelf life than live probiotics." Yadav has filed a provisional patent application for D3-5.

10.1007/s11357-019-00137-4

Nano

Nanotechnology transforms molecular beams into functional nano-devices with controlled atomic architectures

dx.doi.org/10.1038/nature11999 Journal information: Nature

http://dx.doi.org/10.1038/nature11999

https://phys.org/news/2013-04-nanotechnology-molecular-functional-nano-devices-atomic.html

Abstract The incorporation of impurities during the growth of nanowires from the vapour phase alters their basic properties substantially, and this process is critical in an extended range of emerging nanometre-scale technologies 1 , 2 , 3 , 4 . In particular, achieving precise control of the behaviour of group III and group V dopants has been a crucial step in the development of silicon (Si) nanowire-based devices 5 , 6 , 7 . Recently 8 , 9 , 10 , 11 it has been demonstrated that the use of aluminium (Al) as a growth catalyst, instead of the usual gold, also yields an effective p-type doping, thereby enabling a novel and efficient route to functionalizing Si nanowires. Besides the technological implications, this self-doping implies the detachment of Al from the catalyst and its injection into the growing nanowire, involving atomic-scale processes that are crucial for the fundamental understanding of the catalytic assembly of nanowires. Here we present an atomic-level, quantitative study of this phenomenon of catalyst dissolution by three-dimensional atom-by-atom mapping of individual Al-catalysed Si nanowires using highly focused ultraviolet-laser-assisted atom-probe tomography. Although the observed incorporation of the catalyst atoms into nanowires exceeds by orders of magnitude the equilibrium solid solubility 12 and solid-solution concentrations in known non-equilibrium processes 13 , 14 , the Al impurities are found to be homogeneously distributed in the nanowire and do not form precipitates or clusters. As well as the anticipated effect on the electrical properties, this kinetics-driven colossal injection also has direct implications for nanowire morphology. We discuss the observed strong deviation from equilibrium using a model of solute trapping at step edges, and identify the key growth parameters behind this phenomenon on the basis of a kinetic model of step-flow growth of nanowires. The control of this phenomenon provides opportunities to create a new class of nanoscale devices by precisely tailoring the shape and composition of metal-catalysed nanowires. Main The nanowires investigated in this study were grown on Si(111) in an ultrahigh-vacuum chemical deposition system 10 , 15 . Their growth was accomplished with Al-Si nanoparticles, which act catalytically as the energetically favoured sites for vapour-phase reactant (SiH 4 ) adsorption and (when saturated) as the nucleation sites for crystallization and one-dimensional growth 16 . The resulting nanowires are aligned along the [111] direction and display morphological characteristics that are sensitive to the growth temperature ( Supplementary Figs 1 and 2 ). Growth at the highest temperature (470 °C) yields tapered nanowires with an average height and base diameter of 475 ± 30 nm and 94 ± 8 nm, respectively ( Supplementary Fig. 1 ). Figure 1a displays a high-resolution transmission electron microscopy (TEM) micrograph of an Al-catalysed Si nanowire near the interface with the catalyst. This micrograph demonstrates that both nanowire and catalyst nanoparticle are single crystals with a coherent interface between them. However, the close atomic numbers of 13 Al and 14 Si make them indistinguishable in high-resolution TEM images of nanowires. Energy-dispersive X-ray analysis (EDX) in an aberration-corrected TEM did not find a significant Al signal in Si nanowires grown at the lowest temperature (410 °C). This indicates that the Al concentration is less than the EDX detection limit, which is typically ∼ 0.5 atomic per cent (at.%). For a rigorous analysis, we used atom-probe tomography (APT), which has a superior detection sensitivity of less than ten atomic parts per million (p.p.m.), with atomic-scale spatial resolution 17 . Pulsed evaporation of individual atoms was achieved using a focused picosecond ultraviolet (wavelength λ = 355 nm) laser. The use of a highly focused ultraviolet laser beam decreases local heating, thereby improving the detection limits and mass resolving power ( m /Δ m , where m is the mass of an isotope) 18 . Figure 1: Structure and three-dimensional map of Al-catalysed Si nanowires. a , High-resolution cross-sectional TEM image (left) displaying the interface between the catalyst particle and the nanowire. The right panel exhibits a close-up image of the interface (top), the fast Fourier transform of the image (middle), and the corresponding colour-filtered image (bottom, Al and Si regions correspond to red and blue regions, respectively) indicating that the interface is epitaxial. b , Three-dimensional APT atom-by-atom map of a nanowire grown at 410 °C. For the sake of clarity, only a limited number of atoms is displayed (2.5 × 10 4 atoms of each element). Inset, a cross-sectional TEM image of an identical Si nanowire (scale bar, 40 nm). c , Si 50 at.% isoconcentration surface of an 80-nm-long segment of a nanowire determined by analysing a three-dimensional atom-probe tomographic reconstruction: left, side view; right, top view. PowerPoint slide Full size image Figure 1b shows a three-dimensional atom-by-atom map of a single Al-catalysed Si nanowire covered by a protective nickel (Ni) layer. The inset in Fig. 1b displays a cross-sectional TEM image of an identical nanowire. Figure 1c shows two projections of an 80-nm-long segment of a Si nanowire delineated by an isoconcentration surface drawn at the 50 at.% Si level. The 50 at.% Al isoconcentration surface of the catalyst nanoparticle is displayed in Fig. 2a . Axial and radial concentration profiles of Al and Si, taken from the top 10 nm of the nanoparticle, are presented in Fig. 2b and c , respectively. Both concentration profiles demonstrate Si segregation close to the nanoparticle’s surface, possibly as a result of Si expulsion during post-growth cooling or simply due to oxidation on exposure to air. The average Si concentration measured at the core of the nanoparticle is in the range of 2.7 ± 0.9 to 3.1 ± 0.4 at.%, which is nearly twice the solubility of Si in Al at the eutectic temperature 19 . This indicates that Si solubility in Al increases in undercooled catalyst. Figure 2d displays the three-dimensional distribution of Al atoms detected in an as-grown Si nanowire. The average Al solid concentration ( ) measured for Si nanowires grown at 410 °C is ∼ (2.0 ± 0.5) × 10 20 atom cm −3 ( ∼ 0.4 at.%). Strikingly, this is about four orders of magnitude greater than the extrapolated equilibrium solubility of Al in Si (ref. 12 ), which corresponds to an increase of two to three orders of magnitude compared to Al-assisted recrystallization 13 or solid-phase epitaxy 14 at a similar temperature. Notwithstanding this colossal Al concentration, clusters or precipitates of Al do not form. To verify this quantitatively, we statistically analysed the spatial distribution of Al atoms ( Fig. 2e ) and performed partial Al–Al radial distribution function analyses. This analysis shows that the concentration frequency distribution is not different from a binomial distribution with the same mean at 99% confidence, and thus confirms that Al does not form precipitates or clusters in Si nanowires. Figure 2: APT data obtained for a catalyst nanoparticle and a Si nanowire. a , Al 50 at.% isoconcentration surface of a catalyst nanoparticle. b , c , Axial ( b ) and radial ( c ) concentration profiles of Al and Si in the catalyst nanoparticle; error bars, ±1 s.d. calculated for the analysed volume containing 2.9 × 10 5 Al atoms and 8 × 10 3 Si atoms. d , Three-dimensional distribution of Al atoms in a Si nanowire grown at 410 °C. e , Results of statistical analysis of Al concentration frequency (histogram) and of binomial analysis (solid line). The binomial analysis was performed by sampling blocks containing 200 atoms. PowerPoint slide Full size image Figure 3a displays a set of radial concentration profiles measured in 8-nm-diameter cylinders in different regions across the nanowire grown at 410 °C, perpendicular to the growth direction ( Fig. 3a insets). The estimated mean value varies between 0.39 and 0.45 at.%. The profiles display random fluctuations indicative of a random distribution of Al in Si nanowires. This is also consistent with the Al axial concentration profiles ( Fig. 3b ). Indeed, averaged over a cylindrical volume remains practically the same regardless of the diameter of the selected volume, thereby confirming the uniform distribution of Al. Analysis of in Si nanowires grown at a higher temperature (470 °C) demonstrates a significantly stronger incorporation of Al, leading to concentrations sufficiently large to be detected by EDX ( Fig. 3c ). Interestingly, the Al radial concentration displays flat profiles, indicating a uniform distribution of Al in the nanowire. Surface segregation of Al would have led to ‘U’-shaped profiles. Similarly, the Al axial concentration profile indicates a uniform concentration ( ∼ 4.3 at.%) along the nanowire. We note the ∼ 10-fold increase in when compared to the nanowires grown at 410 °C. Figure 3: Al concentration profiles in individual nanowires. a , Al radial concentration profiles (left) measured in different regions of a Si nanowire. Each concentration profile is taken from a cylindrical segment having a diameter of 8 nm, as indicated in the images (right). b , Al axial concentration profiles (left) measured for cylindrical segments (right) with a diameter of 10 nm (top) and 20 nm (bottom). The horizontal dashed line in all concentration profiles denotes the average concentration, which is given as blue numerals; the shaded band is the uncertainty interval; and error bars in a and b represent ±1s.d. The number of Al atoms in the sampling volumes varies between ∼ 2,000 and 4,500. c , EDX Al concentration profiles: left and middle plots, radial concentration profiles near the base and middle of the Si nanowire, respectively; right plot, axial concentration profile. Error bars, approximately ±0.5 at.%, which is the sensitivity of the instrument. Insets, the corresponding TEM images with EDX scan lines (dashed lines). PowerPoint slide Full size image It is generally recognized that nanowire tapering is induced by vapour–solid growth on sidewall surfaces 8 and/or catalyst diffusion from one nanowire to another 20 . As demonstrated here, this is not the case for Al-catalysed nanowires, for which the observed incorporation of catalyst atoms during growth is the key element determining nanowire morphology. It is noteworthy that the uniform Al radial concentration profile ( Fig. 3c ) rules out vapour–solid deposition on the nanowire sidewalls as a possible mechanism for nanowire tapering. Specifically, if homoepitaxy did take place on the nanowire’s sidewalls, the overgrown layer would have been Al-free, because Al diffusion is negligible under these growth conditions 21 . The strong incorporation of Al catalyst atoms into the nanowire that we observe should involve a continuous dissolution of the catalyst throughout the growth process, with direct implications for nanowire morphology ( Supplementary Information , section D). For instance, at 470 °C, the total volume of Al atoms incorporated in the completely grown Si nanowire is ∼ 5.5 × 10 4 nm 3 , which is significantly larger than the volume of the catalyst at the end of growth ( ∼ 1.3 × 10 4 nm 3 ). This means that there is more Al in the Si nanowire than in the nanoparticle at the end of the growth process—that is, the catalyst has lost more than 80% of its initial volume. This leads to a continuous reduction of the diameter of the catalyst nanoparticle during growth, and thus to nanowire tapering. We note, however, that the catalyst’s volume loss is very small ( ∼ 6%) at 410 °C, which explains the absence of tapering at this temperature ( Supplementary Fig. 2 ). The observed uniform Al concentration distributions suggest that the dissolution of Al atoms and their injection into the growing nanowire occurs at a constant rate during the growth process. Moreover, the fact that Al displays uniform radial and axial concentration profiles suggests the existence of a single atomic pathway behind this colossal incorporation. We anticipate that surface effects, capillarity and related nanoscale stresses will shift the boundaries in the phase diagram of the catalyst–nanowire system in comparison to those of a bulk system 22 . However, by taking into account these thermodynamic considerations and including the nanowire tapering ( Supplementary Information , section E), we found that the calculated solid composition of the nanowire is always smaller than the equilibrium content, which is of the order of 1 atomic p.p.m. at the growth temperature of our Si nanowires ( Supplementary Fig. 11 ). This stands in sharp contrast to the observed excess concentrations (that is, concentrations above the equilibrium value) of orders of magnitude in our Si nanowires, which demonstrates that kinetic factors must be responsible for the incorporation of Al atoms. Supersaturations of several orders of magnitude occur as a result of deviations from local equilibrium in ultra-rapid solidification of a melt 23 , 24 , 25 . This phenomenon (so-called solute trapping) has been observed for a variety of solutes in Si for growth rates of the order of metres per second (refs 24 , 25 ). In general, solute trapping is quantified in terms of a partition coefficient, k , which is the ratio of the solute concentration in the solvent to its concentration in the melt at the solid–liquid interface, as a function of the equilibrium partition coefficient ( k e ), the average interface velocity ( v min ), and a characteristic velocity known as diffusive velocity ( v D ) 25 . Because nanowires grow by successive addition of bilayers through step flow 26 , the average interface velocity is defined as v min = b / τ , where b = 0.3 nm is the thickness of a Si(111) bilayer and τ is the total time needed for one bilayer to grow ( τ includes the time needed to reach the supersaturation (the incubation time, τ I ) and the time between the end of the incubation period (that is, the beginning of nucleation) and the complete growth of the bilayer, τ BL ). Note that the incubation time is a nanoscale phenomenon, not expected for macroscopic systems 26 . Under our growth conditions, v min equals 0.4 and 0.2 nm s −1 at 470 and 410 °C, respectively. Intriguingly, from studies of rapid solidification of melts 25 , one can infer that solute trapping should not occur for systems with an interface velocity of the order of nanometres per second, and thus the anticipated amount of Al in Si nanowires should not exceed the equilibrium concentration. We note that the influence of temperature on the rate of the thermally activated jumps is not sufficient to explain the observed incorporation of Al at exceedingly low interface velocities ( Supplementary Information , section F). Current solute trapping models 25 involve the assumption that atomic jumps occur over the entire infinite liquid–solid interface, and solute atoms are trapped only after the passage of the next layer across the nanowire surface. Although this mechanistic picture describes the behaviour in bulk systems very well 25 , it cannot be extended to nanowire growth, which is characterized by a relatively large delay between two successive bilayers. During this incubation time, there is no driving force for atoms to leave the catalyst and thus the exchange (that is, the process of atomic jumps) should stop until the system reaches supersaturation. It is also important to note that in step-flow growth, atoms at the step edges have few nearest neighbours, which provides them with more steric freedom for an exchange process. Thus, it is reasonable to assume that atomic jumps are more favourable at the step edges than at terraces. This suggests that the impurity atoms become frozen in the solid immediately after the formation of the next row of atoms at the step edge. Therefore, the time, τ e , during which the local exchange is possible corresponds to the average time needed to add one row, and can be expressed as τ e = a / v step , where a is equivalent to the width of one row of atoms and v step is the average step velocity. This latter can be expressed as a function of nanowire radius R NW : v step = 2 R NW / τ BL . Thus, at fixed radius, τ e = τ BL ( a /2 R NW ) = τ BL / n , where n = 2 R NW / a ≫ 1. The parameter n depends only on step speed ( Supplementary Information , section F). Experimentally, we can only determine the average time between the completion of one bilayer and the completion of the next, τ = τ I + τ BL . By taking f = τ BL / τ , the exchange time can be written as τ e = ( f / n ) τ . The transport of impurity atoms between the two sides of the interface for 0 < t < τ e can be described on the basis of chemical rate theory and mass conservation 27 , 28 . Assuming that the step growth is random rather than periodic 29 ( Supplementary Information section F), the partition coefficient can be expressed as: where D i is the coefficient of interdiffusion at the interface ( v D = D i / b ). Figure 4a displays the evolution of the calculated k in ( f ; n ) space for nanowires grown at 470 °C ( τ = 0.75 s and k e = 2.6 × 10 −6 ) using v D = 6.7 nm s −1 obtained from earlier solute trapping models ( Supplementary Information section F). We note that any pair ( f ; n ) (except around (1;1)) predicts Al trapping at levels always significantly higher than the measured value ( k = 5.6 × 10 −2 ). At these trapping levels, nanowire growth cannot occur. Obviously, diffusive velocities of the order of nanometres per second cannot describe the observed Al injection in nanowires. The complete map of v D values fitting the measured k values is displayed in Fig. 4b . Note that the expression for τ BL can be deduced from the model above: . The extrapolation of the correlation between the equilibrium partition coefficient, k e , and v D obtained in laser resolidification 25 to our experimental conditions yields a diffusive velocity of ∼ 2 × 10 11 nm s −1 at 470 °C ( k e = 2.6 × 10 −6 ), corresponding to τ BL ≈ 7 ns. Interestingly, the time between two layers is ∼ 0.75 s, thus suggesting that the incubation time is very much larger than the time needed for the step to grow across the whole nanowire diameter, in qualitative agreement with in situ TEM observations 8 , 26 , 29 , 30 . Similarly, at the lowest growth temperature ( k e = 6.4 × 10 −7 ), the estimated τ BL is ∼ 18 ns. A reduction in growth temperature is anticipated to affect not only the rate of the thermally activated atomic jumps, but also other growth parameters that are critical for the incorporation of the catalyst atoms, as discussed below. Figure 4: Calculations of partition coefficient and diffusive velocity. a , The calculated partition coefficient, k (on colour scale and labelled on contours), for nanowires grown at 470 °C using a diffusive velocity ( v D ) of 6.7 nm s −1 obtained from an early solute trapping model 25 . Axes show f (given by τ BL / τ ) and n (given by τ BL / τ e ); see text for details. Inset, the calculated k near ( f ; n ) = (1;1). b , The map in ( f ; n ) space of diffusive velocities, v D (on colour scale and labelled on contours), fitting the measured partition coefficient at 470 °C ( k = 5.6 × 10 −2 ). It is important to indicate that only low values of f have a physical meaning in nanowire growth as the incubation time is much larger than the growth time of a single bilayer (that is, τ BL ≪ τ I , corresponding to f ≪ 1). Note the logarithmic scales. In both a and b , v min is taken as 0.4 nm s −1 , which is the average interface velocity corresponding to nanowires grown at 470 °C. PowerPoint slide Full size image Using a step-flow growth model that takes catalyst dissolution into consideration, we derived an expression for the Al content in the nanowire as a function of growth parameters ( Supplementary Information section G). Note, however, that the elucidation of the kinetically controlled atomic processes involved in the incorporation of Al would require a deep understanding of Al-induced Si surface reconstructions ( Supplementary Information section H). From our analysis, it emerges that the second derivative of the Gibbs free energy of the liquid evaluated at the equilibrium composition, , and the characteristic supersaturation at which the nucleation occurs deterministically, μ c , are the two important parameters that influence catalyst injection. The quantity can be modified by adjusting Si solubility, whereas high μ c implies a high nucleation barrier. In principle, both parameters can be engineered through control of the catalyst chemistry (by adding impurities, for instance), thus providing the possibility of tailoring the properties of nanowires. This flexibility in fabrication and functionalization of nanowires could be greatly enhanced owing to the availability of other catalysts that can act as dopants—for example, Bi (which yields n-type doping 31 , 32 ) or Ga (which gives p -type doping 33 )—if their injection during the growth of Si or other group IV nanowires can be controlled. Our observations and predictions provide motivation to pursue the synthesis and characterization of atomically controlled nanowires with potentially adjustable morphology and physical characteristics that might offer new routes for catalytic assembly of nanowire-based devices. Methods Summary The growth of Al-catalysed Si nanowires was accomplished by using monosilane, SiH 4 , (diluted to 5% in argon) as a precursor in the temperature range ∼ 400–470 °C (refs 10 , 15 ). The partial pressure of the monosilane was held below 0.15 mbar during growth. Although the temperature used is below the Al-Si macroscopic eutectic temperature, the growth here is believed to be via the vapour–liquid–solid mechanism involving undercooled Al-Si nanodroplets. The morphology of Al-Si nanowires was characterized using an FEI dual-beam Nanolab 600 scanning electron microscope, a Philips CM 20T TEM operating at 200 kV, a JEOL JEM-4010 TEM operating at 400 kV, and an aberration-corrected FEI TITAN 80-300 analytical scanning TEM operating at 300 kV, which yields a spatial resolution of about 100 pm in both modes (TEM and scanning TEM). Additionally, this TEM is also equipped with an EDX detector having a detection limit of about 0.5 at.%. We use an ultraviolet laser-assisted LEAP (local electrode atom probe) tomograph (LEAP 4000XSi, Cameca). The three-dimensional reconstructions and statistical analysis of concentration frequency were performed using Cameca’s IVAS program. For the preparation of nanowire LEAP specimens, we have developed a focused ion beam (FIB)-based damage-free methodology to attach individual nanowires to commercially available Si microtips ( Supplementary Information section B). The Si microtips were subsequently inserted into the LEAP tomograph’s ultrahigh-vacuum chamber and cooled to 60 K before pulsed ultraviolet-laser-assisted evaporation analyses. To confirm that the preparation and laser-assisted evaporation do not affect the accuracy of the analysis, a reference sample consisting of an Al-capped Si needle was also analysed under exactly the same conditions, and the results demonstrate that the needle is Al-free. The binomial analysis was performed by sampling blocks of 200 atoms. The entire volume was first divided into columns along the z axis, with a cross-section in x- and y- directions, so that the volume d x d y d z contains 200 atoms on average. To obtain blocks with exactly 200 atoms, each column with cross-section d x d y was then cut into sections with 200 atoms as d z is varied.

Semiconductor nanowires are quasi-one-dimensional nanomaterials that have sparked a surge of interest as one of the most powerful and versatile nanotechnological building blocks with actual or potential impact on nanoelectronics, photonics, electromechanics, environmentally friendly energy conversion, biosensing, and neuro-engineering technologies. Bottom-up synthesis of nanowires through metal-catalyzed vapor phase epitaxy is a very attractive process to generate high-quality nanowires thus providing an additional degree of freedom in design of innovative devices that extend beyond what is achievable with the current technologies. In this nano-fabrication process, nanowires grow through the condensation of atoms released from a molecular vapor (called precursors) at the surface of metallic nano-droplets. Gold is broadly used to form these nano-droplets. This self-assembly of nanowires takes place spontaneously at optimal temperature and vapor pressure and can be applied to synthesize any type of semiconductor nanowires. However, to functionalize these nanomaterials a precise introduction of impurities is central to tune their electronic and optical properties. For instance, the introduction of group III and V impurities in a silicon lattice is a crucial step for optimal design and performance of silicon nanowire technologies. The accurate control of this doping process remains an outstanding challenge that is increasingly complex as a result of the relentless drive toward device miniaturization and the emergence of novel nanoscale device architectures. In a recent development, a team of scientists from Polytechnique Montréal (Canada), Northwestern University (USA), and Max Planck Institute of Microstructure Physics (Germany) led by Professor Oussama Moutanabbir has made a fascinating discovery of a novel process to precisely functionalize nanowires. By using aluminum as a catalyst instead of the canonical gold, the team demonstrated that the growth of nanowires triggers a self-doping process involving the injection of aluminum atoms thus providing an efficient route to dope nanowires without the need of post-growth processing typically used in semiconductor industry. Besides the technological implications, this self-doping implies atomic scale processes that are crucial for the fundamental understanding of the catalytic assembly of nanowires. The scientists investigated this phenomenon at the atomistic-level using the emerging technique of highly focused ultraviolet laser-assisted atom-probe tomography to achieve three-dimensional atom-by-atom maps of individual nanowires. A new predictive theory of impurity injections was also developed to describe this self-doping phenomenon, which provides myriad opportunities to create entirely new class of nanoscale devices by precisely tailoring shape and composition of nanowires. The results of their breakthrough will be published in Nature.

dx.doi.org/10.1038/nature11999

Chemistry

Newly discovered form of carbon is graphene's 'superatomic' cousin

Elena Meirzadeh et al, A few-layer covalent network of fullerenes, Nature (2023). DOI: 10.1038/s41586-022-05401-w Journal information: Nature

https://dx.doi.org/10.1038/s41586-022-05401-w

https://phys.org/news/2023-01-newly-carbon-graphene-superatomic-cousin.html

Abstract The two natural allotropes of carbon, diamond and graphite, are extended networks of sp 3 -hybridized and sp 2 -hybridized atoms, respectively 1 . By mixing different hybridizations and geometries of carbon, one could conceptually construct countless synthetic allotropes. Here we introduce graphullerene, a two-dimensional crystalline polymer of C 60 that bridges the gulf between molecular and extended carbon materials. Its constituent fullerene subunits arrange hexagonally in a covalently interconnected molecular sheet. We report charge-neutral, purely carbon-based macroscopic crystals that are large enough to be mechanically exfoliated to produce molecularly thin flakes with clean interfaces—a critical requirement for the creation of heterostructures and optoelectronic devices 2 . The synthesis entails growing single crystals of layered polymeric (Mg 4 C 60 ) ∞ by chemical vapour transport and subsequently removing the magnesium with dilute acid. We explore the thermal conductivity of this material and find it to be much higher than that of molecular C 60 , which is a consequence of the in-plane covalent bonding. Furthermore, imaging few-layer graphullerene flakes using transmission electron microscopy and near-field nano-photoluminescence spectroscopy reveals the existence of moiré-like superlattices 3 . More broadly, the synthesis of extended carbon structures by polymerization of molecular precursors charts a clear path to the systematic design of materials for the construction of two-dimensional heterostructures with tunable optoelectronic properties. Main C 60 fullerene, the first synthetic carbon allotrope 4 , 5 , is a geometrically closed, polycyclic polymer composed solely of carbon atoms (Fig. 1a ). This polymer is infinite in the literal sense of not having any termini, but it is obviously quite finite in being the size of a normal, albeit large, molecule. Graphene, another allotrope of elemental carbon 6 , is also a polymer of carbon atoms, but in this case, the polymerization leads to a geometrically open result: infinite, two-dimensional (2D) sheets (Fig. 1b ). Here we disclose a 2D polymer of C 60 , which we synthesize by linking C 60 molecules into layered, graphene-like hexagonal sheets (Fig. 1c ). By analogy to graphene and graphite, we have dubbed this material graphullerene, and its three-dimensional van der Waals (vdW) solid, graphullerite. Fig. 1: Carbon allotropes. a , b , C 60 fullerene, a zero-dimensional molecular cage composed of 60 carbon atoms ( a ), and graphene, consisting of a single layer of atoms ( b ), both composed of three-coordinate carbon. c , Graphullerene, a molecular sheet of carbon assembled from covalently linked C 60 fullerene superatomic building blocks. Full size image Our chemical strategy to prepare graphullerene was inspired by a recent study 7 , which used the chemical vapour transport (CVT) approach to grow single crystals of metal-doped polyfullerides. First, we grow single crystals of magnesium (Mg)-doped polyfulleride—(Mg 4 C 60 ) ∞ . These polyfullerides are obtained by pressing a pellet of C 60 and Mg powder under an inert atmosphere, sealing it in a fused silica tube under vacuum, and placing it in a horizontal furnace with a temperature gradient (Fig. 2a ). Large, black, hexagonal crystals (hundreds of micrometres in lateral dimensions), with a metallic luster, are obtained at the cold end of the tube (Fig. 2b ). Fig. 2: Synthesis and crystal structures of (Mg 4 C 60 ) ∞ . a , Schematic of the CVT technique used for the growth of (Mg 4 C 60 ) ∞ single crystals. b , Optical micrograph of a single crystal. c , Crystal structure of (Mg 4 C 60 ) ∞ showing a top view of a single layer and a side view emphasizing the stacking of the layers along the a axis. The C 60 units within each layer are much closer to one another than they are in molecular C 60 crystals 5 . The closest C···C distance between two C 60 subunits (1.573(1) Å) is roughly half of that in molecular C 60 (3.116 Å). This very close spacing between the fullerenes is a direct reflection of covalent bonding between the molecules. d , The log of the conductance ( σ ) versus temperature ( T ) for a 70-nm-thick (Mg 4 C 60 ) ∞ device. A fit to a thermally activated (Arrhenius) model is given by the dashed green line. A typical device and corresponding four-terminal measurement scheme are shown in the inset. E a , activation energy; k B , Boltzmann constant; V SD , source-drain voltage; V xx , longitudinal voltage drop. Full size image Single-crystal X-ray diffraction (SCXRD) reveals that the crystals have a layered structure, and display a quasi-hexagonal lattice, with each C 60 forming eight covalent σ bonds to six neighbours within a molecular plane. Four of these make single connections between the C 60 molecules, and each of the other two pairs doubly connects the C 60 molecules (Fig. 2c ). The synthesis yields highly reduced sheets with four Mg counterions per fullerene. The counterions are closely associated with each individual layer (Fig. 2c ), and not shared between layers; hence, the layers are only weakly bonded to each other, predominantly through vdW interactions. Single crystals of (Mg 4 C 60 ) ∞ were also grown in a recent study 8 using a similar CVT approach. We fabricated mesoscopic devices to investigate the electrical transport properties of these highly reduced polymerized fullerene sheets. We produced thin bulk flakes of (Mg 4 C 60 ) ∞ by mechanical exfoliation and deposited gold contacts using the dry-transfer and high-resolution stencil mask technique (an approximately 70-nm-thick device is shown in the inset of Fig. 2d ). (Mg 4 C 60 ) ∞ exhibits thermally activated transport along the in-plane direction (Extended Data Fig. 1 ) with an activation energy of 121 meV, as calculated from fitting an Arrhenius thermal activation model (Fig. 2d ). One of the key benefits of 2D materials prepared by mechanical exfoliation of vdW crystals is their ultra-clean surfaces without counterions or contaminants. This is critical for many applications, and in particular for the assembly of heterostructures and optoelectronic devices 3 . To create a vdW C 60 polymer material that can be mechanically exfoliated, we attempted to remove the Mg from the (Mg 4 C 60 ) ∞ lattice by immersing the crystals in different aqueous acidic solutions, expecting Mg to form water-soluble salts with the conjugate bases. Suspending (Mg 4 C 60 ) ∞ in dilute aqueous solutions of acetic acid or nitric acid leaches out most of the Mg, yielding (Mg 0.5 C 60 ) ∞ , as determined by energy-dispersive X-ray spectroscopy (EDS). By suspending the (Mg 0.5 C 60 ) ∞ crystals in N -methylpyrrolidone at 180 °C, we completely remove the Mg counterions (Fig. 3a ). Upon examining the graphullerite crystals using scanning electron microscopy (SEM), we find that the crystals remain intact following Mg deintercalation (Fig. 3a , inset). With Mg taken out, the remaining material is entirely and purely carbon, yet it is not C 60 ; it is a vdW solid, graphullerite ((C 60 ) ∞ in the figures). We note that the lack of long-range registry of the covalent layers along the stacking direction, indicated by the broadening of the powder X-ray diffraction (PXRD) peaks (Extended Data Fig. 2 ), has thus far prevented structural determination using SCXRD. Fig. 3: Mg deintercalation and mechanical exfoliation to produce graphullerene. a , Elemental composition of single crystals before ((Mg 4 C 60 ) ∞ ) and after ((C 60 ) ∞ ) treatment with a dilute acetic acid solution, as determined by EDS. The absence of the oxygen peak in the spectrum of graphullerite implies that the observed oxygen peak in (Mg 4 C 60 ) ∞ crystals corresponds to oxidized Mg species and not the fullerene sheets. Inset: SEM image of a graphullerite crystal. b , Optical micrograph of mechanically exfoliated graphullerene. c , AFM image of the bilayer selected in red in b . Inset, height profile measured over the region defined by the white dashed box. A 2.4-nm step height characteristic of bilayer graphullerene is observed. d , TEM image of a few-layer graphullerene flake prepared by mechanical exfoliation on a TEM grid. Inset: selected area electron diffraction pattern from a 100 × 100-nm 2 region of the flake. A low-magnification TEM image of this flake is shown in Extended Data Fig. 7 . e , iFFT of the area selected in blue in d , which is magnified in the inset. The filtering removes the aperiodic pixel noise and represents the lattice superstructure, showing that each C 60 is connected to six neighbouring fullerenes in a molecular plane. f , iFFT image of the area selected in pink in d . The high degree of crystallinity can be observed in the Fourier transform (inset) and the electron diffraction pattern in d . Full size image Raman spectroscopy is a diagnostic probe of C 60 polymerization 9 , 10 , and when we compare the Raman spectrum of graphullerite to that of molecular C 60 , we find a splitting of the C 60 H g modes at 1,420 cm –1 and 1,560 cm –1 (Extended Data Fig. 3a ), which is attributed to the lower symmetry of polymerized C 60 . Furthermore, the A g (2) pentagonal pinch mode at 1,469 cm –1 , characteristic of molecular C 60 , is not observed in graphullerite. Quenching of the most intense A g (2) mode corroborates the high degree of polymerization 11 . An alternative interpretation of the Raman spectrum is that the A g (2) mode is shifted to a lower energy as a result of the polymerization 12 and overlaps with the broad H g (7) mode at 1,420 cm –1 . The Raman spectra of (Mg 4 C 60 ) ∞ and graphullerite show no significant differences, indicating that the covalent bonding between fullerene subunits is retained in graphullerite despite the complete removal of the Mg. Note that the H g (7) mode of bilayer graphullerene, obtained by mechanical exfoliation (described below), is slightly shifted to higher energy compared with bulk graphullerite (Extended Data Fig. 3a ). To test whether the Mg counterions, which constitute the scaffolding for the construction of graphullerite, are essential for the thermal stability of the structure, we performed differential scanning calorimetry (DSC) and thermogravimetric analysis (TGA) measurements on graphullerite and (Mg 4 C 60 ) ∞ crystals. The DSC shows no endothermic peak up to about 550 °C, and no mass loss occurs up to about 700 °C, as determined by TGA (Extended Data Fig. 4 ). PXRD measurements, however, show that although graphullerite crystals are structurally stable up to 400 °C, they depolymerize and crystallize as molecular C 60 when heated to 500 °C for 1 h (Extended Data Fig. 5 ). Furthermore, the characteristic A g (2) mode for molecular C 60 at 1,469 cm –1 appears in the Raman spectra of the annealed crystals (Extended Data Fig. 3b ). The absence of an endothermic peak in the DSC data suggests that depolymerization is a gradual process, hard to capture by calorimetry. In as much as there appears to be no interlayer C–C covalent links in graphullerite, we suspected that we could exfoliate the crystals down to a few layers, as in the case of graphite 6 . Indeed, mechanical exfoliation of graphullerite routinely produces uniform flakes as thin as bilayers with lateral dimensions on the order of tens of micrometres. Figure 3b,c shows the optical micrograph and atomic force microscopy (AFM) image of a bilayer. A recent study 8 obtained ionic monolayers of [(NBu 4 + ) 6 (C 60 6– )] ∞ from (Mg 4 C 60 ) ∞ crystals via cationic exchange of Mg with tetrabutylammonium (NBu 4 + ) cations, followed by solution exfoliation. The presence of counterions associated with the reduced sheets precludes the creation of clean, high-quality interfaces for the fabrication of optoelectronic devices and 2D heterostructures. This recent study 8 demonstrated that the ionic sheets can be neutralized using hydrogen peroxide; however, solvent residue resulting from solution exfoliation continues to be a problem in the field. In contrast, we present a simple preparation and the properties of the charge-neutral (Extended Data Fig. 6 ), mechanically exfoliatable graphullerene sheets. The ability to exfoliate thin flakes allows us to directly image graphullerene using high-resolution transmission electron microscopy (HR-TEM) (Fig. 3d and Extended Data Fig. 7 ). Selected area electron diffraction of a few-layer flake shows that, despite the lack of long-range registry along the stacking direction, the flakes are highly crystalline within a molecular plane (Fig. 3d , inset). The magnified HR-TEM image shows bright spheres, corresponding to the C 60 subunits, which are arranged in a hexagonal lattice (Fig. 3e , inset), as anticipated from the SCXRD structure of a single layer of (Mg 4 C 60 ) ∞ . Performing a bandpass-filtered inverse fast Fourier transform (iFFT) discards the lower frequencies and removes the aperiodic pixel noise. The Fourier-filtered image of the magnified area (Fig. 3e ) clearly shows the hexagonal morphology of the covalently interconnected fullerenes. A closer inspection of a roughly 2-μm 2 area reveals some variations in the periodic structure of the flakes, characteristic of moiré-like superstructures (Fig. 3f ). The presence of discreet peaks in the FFT (Fig. 3f , inset) indicates that the layers are single crystalline. Such variations in the periodic structure can be rationalized by considering mechanical strain resulting from ion diffusion during Mg deintercalation, which leads to local changes in the alignment of the layers with respect to each other. To elucidate the optoelectronic properties of this material, we measured the photoluminescence (PL) of molecular C 60 , bulk graphullerite and graphullerene bilayer. The PL spectral shape of graphullerite is significantly different from that of molecular C 60 , particularly in the high-energy region (Fig. 4a ). Furthermore, the PL peak of graphullerene bilayer is slightly blueshifted compared with that of bulk graphullerite, consistent with trends observed in other 2D materials 13 . In C 60 crystals, the fundamental transition between the conduction band minimum and valence band maximum is parity forbidden because of the icosahedral symmetry of the molecule. Therefore, the (0,0) vibronic transition (ground vibrational levels in both S 1 and S 0 states) corresponding to the bandgap energy (about 1.9 eV) is not observed 14 . Instead, the PL spectrum consists of redshifted (0,1) and (0,2) transitions. The emergence of the higher-energy PL in graphullerite and graphullerene suggests that the covalent polymerization leads to a large change in the electronic structure near the bandgap, making the parity-forbidden (0,0) transition allowed. A more detailed analysis of the PL spectra from (Mg 4 C 60 ) ∞ is in Extended Data Fig. 8 . Fig. 4: Photoluminescence and scanning near-field optical microscopy. a , PL spectra of molecular C 60 , bulk graphullerite and bilayer graphullerene (2L). Inset: the emission polarization dependence of the PL intensity for graphullerite. The highest emission is presumably along the b axis where the fullerenes are doubly connected, and the lowest intensity between two single σ bonds, 90° with respect to the b axis. b , Nano-PL image of a 12-nm-thick graphullerene flake. At each pixel, a full PL spectrum is acquired and integrated over the entire emission range (700–900 nm). Inset: AFM topography of the flake. c , High-resolution image of the region defined by the blue box in b , showing that the PL intensity varies on the roughly 50-nm length scale. The pixel size is 8 × 8 nm 2 . d , Topographic image of 315-nm-thick flake of graphullerite obtained using AFM. e , Near-field amplitude collecting at the fifth harmonic of the tip tapping frequency (S 5 ) of the flake shown in d using an incident laser wavelength of 780 nm. f , Sine of the associated near-field phase normalized to the SiO 2 substrate. Both near-field channels in e and f show oscillations emanating from the edge of the graphullerite flake characteristic of both waveguide and air modes. Full size image Polarization-resolved PL measurements show that graphullerite exhibits polarized emission: the PL intensity depends on the analyser angle (Fig. 4a , inset), reflecting the anisotropic structure of the b – c plane. We note that the presence of counterions in the structure is detrimental to its optical properties: the PL intensity of the crystals increases by more than two orders of magnitude following Mg deintercalation. These findings demonstrate the importance of the deintercalation process that yields a charge-neutral purely carbon-based material without counterions. Changing the atomic registry between vdW layers through twisting or lattice mismatch can modulate the optoelectronic properties of 2D materials. In particular, moiré engineering provides a powerful approach for tailoring new excitonic systems 15 . To test whether the formation of moiré-like superstructures in graphullerene (Fig. 3f ) affects the optical properties of the flakes, we performed near-field nano-photoluminescence (nano-PL) imaging 16 on a 12-nm-thick flake. We find that the lack of the long-range registry along the stacking direction, as deduced from the PXRD and HR-TEM, leads to variations in the optical response and formation of linear domains with different PL intensities (Fig. 4b,c ). Although in this study we did not twist the material intentionally, these findings suggest that tuning the optoelectronic properties of graphullerene is possible by the construction of heterostructures with different twist angles. Maps of the near-field optical response obtained using scattering-type scanning near-field optical microscopy (s-SNOM) further attest to the material quality. Figure 4d–f shows AFM and s-SNOM images with systematic oscillations in the near-field amplitude and phase (that is, near-field fringes) emanating from the edges of a thin graphullerite flake (thickness 315 nm). Such fringes in the near-field signal arise owing to interference between modes propagating in either free space near the crystal surface or within the crystal bulk (so-called air modes and waveguide modes, respectively) 17 . To observe either phenomenon, a low density of defects is required in the graphullerite flake to prevent decoherence of propagating light. Thus, the observation of near-field fringes corroborates the high quality of our crystals and suggests that it may be a viable platform for further study of 2D confined light. Finally, we elucidate the thermal conductivity ( k ) of graphullerite (Fig. 5a ). Vibrational thermal transport is strongly impacted by the strength of the interatomic and intermolecular interactions 18 . In fullerene-based materials, the modification of intermolecular interactions has been shown to impact the vibrational scattering mechanisms affecting k (refs. 19 , 20 , 21 , 22 ). Figure 5b compares the room temperature k of mechanically exfoliated, thin, bulk graphullerite flakes with that of molecular C 60 crystals. Unlike previous reports in which chemical modifications of C 60 resulted in reductions of k (refs. 18 , 19 , 20 , 22 ), the high in-plane order and the formation of intermolecular covalent bonds in graphullerite enhance thermal transport. Graphullerite shows a marked increase in k (2.7 W m –1 K –1 ) that is nearly one order of magnitude higher than that measured in molecular C 60 crystals (0.3 W m –1 K –1 ). We note that this value for graphullerite is an average of all directions, as our technique is equally sensitive to thermal transport in both in-plane and cross-plane directions 23 . Fig. 5: Thermal transport properties of graphullerite. a , Schematic of the thermal conductivity measurement. b , Experimentally measured thermal conductivities of a few-micrometres-thick graphullerite flake at room temperature (black open triangle) and a molecular C 60 single crystal (green open triangle). The error bars incorporate the standard deviation between measurements and propagated uncertainty of transducer properties and the specific heat of the sample. The filled symbols represent the non-equilibrium molecular dynamics predictions for the three crystallographic directions. c , Vibrational density of states of molecular C 60 and graphullerite showing additional modes throughout the vibrational spectrum for the 2D polymer phase. Full size image Molecular dynamics simulations provide fundamental insights into the vibrational thermal transport dynamics and the mechanisms that drive this order of magnitude increase in k . Figure 5b presents the molecular-dynamics-predicted k values for graphullerite along the three crystallographic directions between 100 K and 400 K. These simulations support the marked increase in k for the polymerized phase compared with that of molecular C 60 . More specifically, the in-plane k (across the covalently bonded sheets) (Fig. 2b ) is more than one order of magnitude higher than the out-of-plane k , which is similar to that of molecular C 60 crystals. The in-plane k of graphullerene is also highly anisotropic, with the b direction (along the [2 + 2] inter-fullerene bonds) demonstrating the highest thermal transport. A comparison between the calculated vibrational spectra of C 60 and graphullerite demonstrates that the 2D covalent bonding of C 60 introduces many new vibrational modes throughout the spectrum (Fig. 5c ). These modes provide additional pathways for heat conduction, which results in an overall increase in k for graphullerite compared with molecular C 60 . Fullerenes can be covalently bonded to one another by photoirradiation 24 , 25 , 26 , solid-state reaction with alkali metals 7 , 27 , 28 , 29 , 30 , 31 and high-temperature, high-pressure annealing 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 . However, the resulting polycrystalline materials have high defect concentrations, mixed stoichiometries and unreacted starting materials. Moreover, the small dimensions of the single crystals obtained so far 37 , 38 , 39 have limited the physical and structural characterization of polyfullerenes. Although previously reported polyfullerenes 40 generally revert to molecular C 60 when heated above 300 °C, graphullerite depolymerizes above 400 °C. Such stability results from the unique bonding architecture of graphullerite, which comprises both isolated C–C single bonds and pairs of inter-C 60 bonds that give cyclobutane-like functionality; other polyfullerenes show only the latter type of inter-C 60 bonding 37 , 38 . Perhaps the lower thermal stability of these other polyfullerenes is the result of a [2 + 2] cycloreversion, whereas the thermal stability of graphullerite is primarily due to the isolated C–C bonds. To realize the technological promises of 2D materials, it is critical to grow macroscopic single crystals that can provide high-quality macroscopic flakes with clean surfaces. We have presented a chemical strategy to grow a 2D polymer of C 60 as large, single crystals that are readily exfoliatable. The ability of graphullerite to withstand grinding, oxidation and treatment with acid highlights the strong in-plane covalent bonding between the fullerenes. Graphullerite vdW crystals are charge neutral and the exfoliated molecularly thin flakes have no residual counterions or impurities, providing a platform for the investigations of confined light, and the construction of quantum-materials-based devices 3 , 15 . This study also reveals that there is an entire family of higher- and lower-dimensional superatomic allotropes of carbon that may be chemically prepared and studied. Methods Synthesis Single crystals of (Mg 4 C 60 ) ∞ were grown by the CVT technique. C 60 (MTR, 99.9%, sublimed) and Mg powder (Sigma-Aldrich, 50 mesh, ≥99%) were combined in a Mg:C 60 6:1 molar ratio, ground into a homogeneous mixture, pressed into a pellet and placed in a quartz tube (7-mm inner diameter) in a nitrogen-filled glovebox. The tube was then sealed under vacuum and placed in a two-zone horizontal furnace with a temperature gradient of 500 °C to 600 °C for 24 h and then allowed to cool to room temperature. Highly air-sensitive polyfulleride crystals up to a few hundred micrometres in size grew at the cold end of the tube. Selected crystal structure data are presented in Extended Data Table 1 . We determined the Mg:C 60 ratio by both SCXRD and EDS and found that the Mg:C 60 atomic ratio varies between 3.6 and 4.4 for different crystals from the same growth batch. The lattice structure is the same for crystals within this composition range. We note that using finer Mg powder (Alfa Aesar, 325 mesh, 99.8%) without changing the other growth conditions, we obtain (Mg 2 C 60 ) ∞ crystals, similar to ref. 7 . Single crystals of pure C 60 were grown using a similar procedure but without added Mg powder. The crystal structure is as reported 5 . Optical microscopy Optical microscope images were taken on a Nikon ECLIPSE LV150N microscope. Single-crystal X-ray determination SCXRD data were collected using an Agilent SuperNova X-ray diffractometer configured in a four-circle kappa goniometer geometry and a mirror-monochromated microfocus Cu Kα radiation(1.54184 Å). Single crystals were mounted on a 150-μm MiTeGen MicroMount pin using Paratone N oil and cooled to 100 K with an Oxford Cryosystems nitrogen flow apparatus. X-ray intensities were measured with the Titan charge-coupled-device (CCD) detector placed about 34–42 mm from the sample. The data were processed with CrysAlisPro 41 (Oxford) and corrected for absorption. The structures were determined in OLEX2 version 1.5 42 using SHELXT 43 and refined using SHELXL 44 . Two of the Mg atoms were found to be disordered and modelled as occupying three possible positions each. Device fabrication and electronic transport measurements (Mg 4 C 60 ) ∞ crystals were mechanically exfoliated in a nitrogen-filled glovebox with oxygen and water levels <0.1 ppm, using Scotch-brand magic tape. The 285-nm silicon dioxide (SiO 2 )/silicon (Si) substrates (p doped) were cleaned with a low-power oxygen plasma etch for 5 min before transfer and heated to 100 °C for 2 min for the transfer process. Flakes of desired thicknesses were identified by optical contrast and transferred to the pre-cleaned 285-nm SiO 2 /Si substrates via a dry-transfer method; a cold polydimethylsiloxane (PDMS) stamp (−100 °C) was used to pick up and subsequently release the flakes on the designated spot at a higher temperature (−20 °C). A high-resolution stencil mask technique was used to fabricate rigorously air-free devices inside the glovebox. This process is free of solvents and polymers and helps to preserve the quality of the crystals, eliminating possible artefacts that might arise from conventional lithography processes using solvent-based methods and conducting paints usually applied as contacts for bulk crystals. Stencil masks were fabricated from silicon nitride (Si 3 N 4 )(500 nm)/Si(300 µm)/Si 3 N 4 (500 nm) wafers. At first, a combination of photolithography, reactive ion etching, and wet etching (potassium hydroxide solution) was used to create 200-µm and 500-µm square windows of Si 3 N 4 . Subsequently, the desired patterns of devices were written with photolithography on the windows. In the last step, we used reactive ion etching to remove Si 3 N 4 in the exposed patterned area. We used a micro-positioner to align and place the prefabricated Si 3 N 4 masks on transferred flakes. A small amount of vacuum grease (Apiezon H) was used to secure the mask onto the substrate. The device fabrication was completed by evaporating gold (Au) metal as electrical contacts. The final device was cut, mounted on a 16-pin chip carrier and wire-bonded inside the glovebox. Finally, the chip carrier was covered with glass and loaded into the cryostat for the measurements. Electronic transport measurements were performed in a Quantum Design Physical Property Measurement System. Electrical conductivity was measured with a combination of a Keithley 2400 source meter and an Agilent 34401A Digital Multimeter with a high internal impedance of more than 10 GΩ. Deintercalation (Mg 4 C 60 ) ∞ crystals were immersed in a 10% v/v aqueous solution of hydrochloric acid, acetic acid or nitric acid for 2 h. The supernatant was decanted, the crystals were rinsed thoroughly with deionized water and acetone, and dried in vacuo at room temperature. The Mg:C 60 ratio was determined by EDS. Hydrochloric acid only partially removed the Mg to give (Mg 3 C 60 ) ∞ , whereas suspending (Mg 4 C 60 ) ∞ in dilute aqueous solutions of acetic acid or nitric acid decreased the Mg:C 60 ratio to 0.5:1. Our hypothesis is that these chelating acids form Mg salts that are more easily removed from the structure than those produced with hydrochloric acid. To remove the remaining Mg from the crystals, we suspended (Mg 0.5 C 60 ) ∞ in N -methylpyrrolidone at 180 °C for 12 h. The supernatant was then decanted, the crystals were rinsed thoroughly with tetrahydrofuran and dried in vacuo at 80 °C for 12 h. Powder X-ray diffraction The crystals were ground and loaded on a zero-background Si sample holder. PXRD data were collected on a PANanlytical Aeris diffractometer housed in a glovebox. The PXRD pattern of graphullerite features two broad peaks at 9.3° and 19° (Extended Data Fig. 2 ). We assign these peaks to the (200) and (400) reflections based on a model constructed from the SCXRD structure of (Mg 4 C 60 ) ∞ , assuming preferred orientation with the basal plane parallel to the substrate (marked as red vertical lines in Extended Data Fig. 2 ). Such a preferred orientation is common for vdW materials. When compared with (Mg 4 C 60 ) ∞ , the (200) and (400) peaks of graphullerite are shifted to lower angles, suggesting that the diffusion of ions during deintercalation results in an expansion along the a axis—the stacking direction. The peak broadening is due to stacking faults and inhomogeneous strain as a result of ion diffusion during deintercalation 45 . Scanning electron microscopy and energy-dispersive X-ray spectroscopy SEM samples were prepared by placing single crystals on carbon tape. Scanning electron micrographs were collected using a ZEISS Sigma VP SEM. EDS was performed with a Bruker XFlash 6 | 30 attachment. Mechanical exfoliation Graphullerite crystals were mechanically exfoliated using Scotch-brand magic tape. The 285-nm SiO 2 /Si substrates (p doped) were cleaned with a low-power oxygen plasma etch for 5 min before transfer and heated to 100 °C for 2 min for the transfer process. Flakes of desired thicknesses were identified by optical contrast using an optical microscope. Atomic force microscopy AFM images were acquired in PeakForce QNM in scanning mode using a Bruker Dimension FastScan AFM under ambient conditions. Height profiles were analysed and extracted from the AFM images using Gwyddion software. Raman spectroscopy Raman measurements were performed on (Mg 4 C 60 ) ∞ and graphullerite, as well as on bilayer graphullerene prepared by mechanical exfoliation. The samples were sealed in a 1-cm cuvette inside a nitrogen-filled glovebox with oxygen and water levels less than 0.1 ppm and then placed on a home-built micro-Raman spectrometer. Light from a solid-state laser with wavelength λ = 532 nm reflects off a dichroic beamsplitter in a Nikon Eclipse Ti/U inverted microscope and is focused by a ×40, 0.6-numerical-aperture (NA) objective to an approximately 1-μm 2 spot on the sample. The backscattered light was collected by the same objective and passes through a 50-μm confocal pinhole before entering a 0.3-m Acton spectrometer where a 1,200 g mm −1 diffraction grating disperses it onto a PIXIS 400 CCD imaging detector. Typical laser powers were between 20 μW and 100 μW. Spectra were measured on at least two different exfoliated crystals for a particular sample, and in each case, two consecutive spectra were measured on the same spot with a 300-s exposure time to check for sample degradation. The spectral resolution is typically 3.7 cm –1 and was determined using the λ = 546 nm peak of a mercury calibration lamp. High-resolution transmission electron microscopy Graphullerite crystals were mechanically exfoliated using Scotch-brand magic tape, and transferred to a 200-mesh copper PELCO TEM grid with carbon support film using a PDMS stamp under ambient conditions. HR-TEM was performed on FEI Talos F200X. We determined the average centroid-to-centroid distance between neighbouring C 60 units (8.8 ± 0.3 Å) by applying an FFT to the TEM image (Fig. 3f , inset). This value is in good agreement with the corresponding distance within a molecular plane of (Mg 4 C 60 ) ∞ , as calculated from the SCXRD data (9.16 Å). Photoluminescence spectroscopy PL measurements were performed on single crystals of molecular C 60 and graphullerite, as well as on graphullerene bilayer prepared by mechanical exfoliation. The C 60 crystals were glued to BK7 glass substrates using silver paste plus (SPI Supplies). A continuous-wave laser ( λ = 450 nm) was used as an excitation light source. The laser was sent to a ×40 microscope objective and focused onto a sample in a cryostat. The PL was collected with the same objective. The PL spectrum was recorded using a liquid-nitrogen-cooled CCD equipped with a monochromator. Long-pass filters were used to cut the scattering of the excitation laser. The emission polarization dependence was measured by placing a linear polarizer in the detection path and rotating it. The sensitivity of the set-up and its polarization dependence owing to the rotation of the linear polarizer was calibrated using a standard white light. The PL spectra reported in this paper were taken under vacuum at room temperature. We observed a multiple-peak structure in the PL spectrum of (Mg 4 C 60 ) ∞ (Extended Data Fig. 8a ). We also confirmed that the PL intensity increases linearly with increasing excitation laser power (Extended Data Fig. 8b ). The observed spectral shape suggests that the optical transition in (Mg 4 C 60 ) ∞ is dominated by a strong vibronic nature, where the electronic exciton transition and the vibration are coupled to each other. Indeed, the multiple-peak structure is found to be well explained by considering the single-oscillator Franck–Condon model 46 . We performed the theoretical fitting according to the Franck–Condon analysis (equations (3) and (4) in ref. 46 ), where the PL spectral shape is described by the vibronic (0,0) transition energy E 0 , the frequency of the vibration involved ω (corresponding energy is ħω , where ħ is the reduced Planck constant), the Huang–Rhys factor S and the width of the vibronic transition w . The refractive index is assumed to be constant. The fitting result, shown with the solid line in Extended Data Fig. 8a , explains the experimental spectral shape. For this Franck–Condon analysis, we used the vibrational frequency ħω = 1,429 cm –1 (177 meV) because this T 1 u mode is known to give rise to the most intense vibronic PL sideband in C 60 -based materials 14 ; one can see the equally spaced PL peaks in Extended Data Fig. 8a , whose space is ħω = 1,429 cm –1 . The obtained Huang–Rhys factor of 2.7 indicates a strong coupling between the excitonic transition and the vibrational mode. The fitting also yielded the values E 0 = 2.02 eV and w = 0.166 eV (full-width at half-maximum). The fact that the PL of (Mg 4 C 60 ) ∞ is described by the Franck–Condon model suggests a large change in the electronic states near the bandgap compared with those of pristine C 60 crystals. In the pristine C 60 crystals, the fundamental transition between the conduction band minimum and valence band maximum is parity forbidden, thus the vibronic (0,0) transition is not observed and accordingly the Franck–Condon model is not applicable (Fig. 4a ). The PL results suggest that the covalent polymerization leads to a large change in the electronic structure near the bandgap, making the parity-forbidden (0,0) transition allowed. Near-field nano-photoluminescence imaging Graphullerite crystals were mechanically exfoliated onto a 70-nm Au/SiO 2 /Si substrate and heated to 100 °C for 2 min for the transfer process. Imaging of the nano-PL is performed at room temperature on the TRIOS near-field optical microscope from Horiba Scientific. A single-mode helium–neon laser (633 nm) is focused from the side onto the apex of an Au-coated AFM probe (OMNI-TERS-Au, AppNano) with a ×100, 0.7-NA long-working-distance objective (Mitutoyo). This excites a localized plasmon that enhances the local optical field, with a spatial extent of about 10 nm. The enhanced PL is collected through the same objective and passed to the LabRAM HR Evolution spectrometer (Horiba Scientific), and recorded with an electron multiplying CCD (Synapse II, Horiba Scientific). Hyperspectral imaging is conducted in ‘Spec-Top’ mode with ‘dual-spec’ used, a hybrid contact/non-contact scanning mode where the nano-PL is measured with the probe in contact with the sample, then retracted about 20 nm to measure the far-field background and move to next pixel, avoiding contact wear of the probe and sample. Excitation powers for the nano-PL images were 200 μW with exposure times of 150 ms. Scattering-type scanning near-field optical microscopy s-SNOM measurements were conducted using a commercial Neaspec system under ambient conditions using Arrow AFM probes with a nominal resonant frequency of f = 75 kHz. The sample was illuminated with a tunable continuous-wave laser from M-squared (titanium–sapphire module with an output range of 700–1,000 nm (SolsTiS)). The backscattered light was registered by pseudo-heterodyne interferometry and demodulated at the fifth harmonic of the tip tapping frequency to suppress background contribution to the detected signal. Thermal conductivity measurements We used two different pump–probe thermoreflectance techniques to measure the thermal conductivities of C 60 and graphullerite. For the C 60 crystals, we measured the thermal conductivity with time-domain thermoreflectance (TDTR), a pump–probe experiment that utilizes subpicosecond laser pulses to monitor the transient changes in reflectivity on the sample surface owing to changes in temperature induced from the pump pulses. In our measurements, we utilized TDTR in a two-tint configuration, described elsewhere 47 . Before TDTR measurements, we coated the surface of the samples with 80 nm of aluminium (Al) via electron-beam evaporation. We used an optical camera to focus the pump and probe spots on the surface of each isolated crystal that we test. At focus, the pump and probe 1/e 2 diameters were 4.5 μm and 4.5 μm, respectively. To measure thermal conductivity, we modulated the pump path at 8.4 MHz, and monitored the changes in reflectivity from the reflected probe pulses owing to temperature changes induced from the modulated pump heating event as a function of pump–probe delay time (Extended Data Fig. 9a ). We monitored the negative ratio of the in-phase to out-of-phase voltages recorded from the lock-in amplifier and fit this signal to the solution to the cylindrical heat equation for an 80-nm Al film on a semi-infinite C 60 substrate 48 . The high-frequency modulation and spot sizes ensure that this semi-infinite assumption is valid 49 . For our analysis, we assumed volumetric heat capacities for the Al and C 60 from previous literature, and a thermal conductivity for the Al film of 180 W m –1 K –1 as determined via the Wiedemann–Franz law applied to four-point probe resistivity measurements, and we set the Al/C 60 thermal boundary conductance to 100 MW m –2 K –1 , leaving the only free parameter in our TDTR fit as the thermal conductivity of C 60 . We have negligible sensitivity to the Al/C 60 thermal boundary conductance owing to the low thermal conductivity of the C 60 , so our assumption for this value does not impact our results. For graphullerite, we attempted TDTR measurements, but we were unable to obtain reliable and repeatable results owing to sample-to-sample variation and non-temperature related variations in our measured signals that arose during measurements. A major limitation in applying TDTR to graphullerite is that these crystals had a much smaller measurable cross-sectional dimension than the C 60 crystals (for example, most graphullerite samples were only a few micrometres in measurable surface area). This limited dimensionality restricted the number of individual crystals in which we could perform reliable TDTR tests. In addition, as our TDTR spot sizes were on the order of the dimension of the sample, we observed artefacts in our measured data that could be attributed to thermoelastic effects from the pump-induced pressure wave interacting with the sample dimensions 50 . Therefore, to alleviate these measurement issues, we use steady-state thermoreflectance (SSTR) 23 to measure the thermal conductivity of the graphullerite crystals. Similar to TDTR, SSTR is a pump–probe technique but SSTR utilizes continuous-wave lasers to measure the change in reflectivity on the surface of the sample as a function of change in pump power (Extended Data Fig. 9a ). We modulate the pump beam at a low frequency (in our case, 1,000 Hz) to ensure that the temperature gradients induced during pump heating reach steady state, and we measure this thermoreflectance during steady state, as recorded from the in-phase lock-in voltage, as a function of pump power. Given that the sample has reached steady-state conditions, the temperature response is directly related to the change in heat flux from the pump path via the cylindrically symmetric Fourier law, as detailed previously 23 , 51 . In our specific case, SSTR offers an advantage to measuring the thermal conductivity of graphullerite as we can focus on small spot sizes (1.8 mm and 1.4 mm 1/e 2 diameters for pump and probe, respectively), ensuring we are taking measurements in the middle of the crystals that normally we either could not measure or posed difficulty measuring with TDTR. In addition, as SSTR operates in the steady-state regime, we do not require an assumption of the heat capacity of the materials in our thermal model; given that graphullerite is a new crystal with an unknown heat capacity, this reduces our uncertainty in our determined thermal conductivity. Finally, our measurements are not subjected to the same strong thermoelastic conditions that we potentially observed when applying TDTR to these systems as the absorbed power density on the surface of the Al-coated graphullerite from the continuous-wave pump is orders of magnitude smaller than that from a short pulse with similar average power, and thus the thermally induced acoustic waves generated from the short pulse in TDTR that we posit could be impacting our measurements on these areal-confined samples are non-existent in SSTR. We note that we also attempted SSTR measurements on the C 60 sample that we measured with TDTR; however, owing to the much lower thermal conductivity of C 60 compared with the graphullerene (0.3 W m –1 K –1 for C 60 compared with 2.7 W m –1 K –1 for graphullerite), our SSTR measurements were much more sensitive to the thermal conductivity of the Al film compared with the C 60 . Given the ninefold increase in thermal conductivity of graphullerite compared with C 60 , the graphullerite SSTR measurements did not suffer from this same issue and were quite robust for measurements of the thermal conductivity of the relatively smaller crystallites. Non-equilibrium molecular dynamics simulations We performed non-equilibrium molecular dynamics (NEMD) simulations to predict the thermal conductivity of graphullerite described via the polymer consistent force-field (PCFF) 52 . The Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS) package was used to perform all simulations 53 . Throughout the simulations, a timestep of 0.5 fs was used. For the NEMD method, we established hot and cold baths at opposite ends of the computational domain that were extended along the heat flux direction. An equal amount of heat was added and subtracted from the hot and cold sides, respectively. The equations of motion of the atoms between the baths were integrated under the microcanonical ensemble (with the number of atoms, volume and energy held constant). A steady-state temperature gradient was established across the computational domain, which is averaged over 9 ns. As size effects can considerably reduce the NEMD-predicted thermal conductivity owing to scattering of longer wavelength vibrations at the walls and the baths along the heat flux direction, we performed simulations on multiple domain lengths. From these simulation results, the inverse of thermal conductivity (1/ k ) was compared with the inverse of the length of the simulation domain length (1/ d ) and a linear extrapolation to 1/ d → 0 corresponds to 1/ k ∞ , the inverse of which is the approximate thermal conductivity of the bulk structure 54 (Extended Data Fig. 9b ). Differential scanning calorimetry DSC traces were collected on a DSC Q2000 (TA Instruments) equipped with an RCS90 cooling accessory. Baseline calibration was performed with sapphire disks. The temperature and the cell constant were calibrated using an indium standard. Thermogravimetric analysis TGA traces were collected on a TGA Q500 (TA Instruments) under N 2 flow. Zeta-potential analysis The zeta potential of graphullerite was measured using a Malvern nano-ZS zeta-potential analyser. Microcrystalline crystals were suspended in isopropanol by sonication in an ultrasonic bath (fisherbrand CPX2800) for 1 h and then centrifuged for 15 min at 5,000 r.p.m. to separate the sediment from the colloidal suspension. Data availability The data that support the findings of this study are present in the paper and its Extended Data. The crystallographic data presented in this work are available through the Cambridge Crystallographic Data Centre (CCDC) referencing deposition no. 2151576 . Further data are available from the corresponding authors upon reasonable request.

Carbon in its myriad forms has long captivated the scientific community. Besides being the primary component of all organic life on earth, material forms of carbon have earned their fair share of breakthroughs. In 1996, the Nobel Prize in Chemistry went to the discoverers of fullerene, a superatomic symmetrical structure of 60 carbon atoms shaped like a soccer ball; in 2010, researchers working with an ultra-strong, atom-thin version of carbon, known as graphene, won the Nobel Prize in Physics. Today, in work published in Nature, researchers led by Columbia chemists Xavier Roy, Colin Nuckolls, and Michael Steigerwald with postdoc and first author Elena Meirzadeh have discovered a new version of carbon that sits somewhere in between fullerene and graphene: graphullerene. It's a new two-dimensional form of carbon made up of layers of linked fullerenes peeled into ultrathin thin flakes from a larger graphullerite crystal—just like how graphene is peeled from crystals of graphite (the same material found in pencils). "It is amazing to find a new form of carbon in the year 2022," said Nuckolls. "It also makes you realize that there is a whole family of materials that can be made in a similar way that will have new and unusual properties as a consequence of the information written into the superatomic building blocks." Meirzadeh, who synthesized the first crystals of graphullerite, referred to graphullerene as the superatomic "cousin" of graphene. Unlike graphene and most other two-dimensional materials that are made of repeating elements that are limited to specific bonding geometries and have specific properties as a result, graphullerene's superatomic structure makes it incredibly modular, she explained. With 60 carbon atoms in the ball to work with, fullerenes can theoretically be linked together in a number of different ways, each of which could yield different electronic, magnetic, and optical properties—this first version represents just one possible configuration, Roy said. It's a new way of thinking about structures and their properties as they grow, added Steigerwald. "For more than 30 years, researchers have had the notion that clusters of atoms will behave differently than the larger solids that they form," he said. "Here, we are making a solid out of an existing carbon superatom to see how that kind of organization will influence its properties. Would the new material behave like fullerene, or like something else?" The team set out to bond fullerenes molecules rather than individual carbon atoms into a layered, peelable crystal, in order to study its superatomic properties in two dimensions. Meirzadeh used a high-temperature solid-state synthesis technique involving a magnesium scaffold that was later removed—a process involving acid that, after a year spent working with air-sensitive crystals inside a glovebox, was a slightly nerve-wracking final step. "As chemists, we try things and don't always know what will happen. I thought it would fall apart, but it remained intact," she recalled. "Seeing an intact, pure carbon crystal that we could then easily exfoliate and study was a great surprise." Once the new material was made, Meirzadeh sent samples off to collaborators at Columbia and beyond for initial imaging and characterization. The battery of tests revealed a number of intriguing electrical, optical, and thermal properties. Like graphene, graphullerene can confine and polarize light, it can accept lots of extra electrons, and it can form tunable superlattice structures; these properties make it a promising material with potential applications in new kinds of optical and electronic devices. Compared to fullerenes, graphullerite crystals are shown to have a much higher thermal conductivity, a result of the strong covalent bonds within each graphullerene sheet. Thermal conductivity helps dissipate heat, an important consideration when building devices. The work is a starting point for the team to explore the potential of graphullerene. From a chemistry perspective, they plan to tweak and tune its modular properties and introduce new structures, while collaborators will look more closely at what happens when graphullerene sheets are combined with different kinds of two-dimensional materials studied at Columbia to see what other secrets carbon conceals. "The discoveries of both graphene and fullerenes were incredibly impactful," said Roy. Fullerenes, for example, are used to build organic photovoltaic cells and in medicine as contrast agents for MRI and X-ray imaging and to deliver drug therapies. The more recently discovered graphene, an extremely light yet strong material with numerous unique properties, is being actively explored for potential uses in electronics, energy applications, and more. "Now we've combined them together into this new form of carbon. We don't know exactly what will come out of this work, but it will be very exciting to explore," he said.

10.1038/s41586-022-05401-w

Earth

Cutting HFCs to cool the Earth

Pallav Purohit, Achieving Paris climate goals calls for increasing ambition of the Kigali Amendment, Nature Climate Change (2022). DOI: 10.1038/s41558-022-01310-y Journal information: Nature Climate Change

http://dx.doi.org/10.1038/s41558-022-01310-y

https://phys.org/news/2022-03-hfcs-cool-earth.html

Abstract Hydrofluorocarbon emissions have increased rapidly and are managed by the Kigali Amendment to the Montreal Protocol. Yet the current ambition is not consistent with the 1.5 °C Paris Agreement goal. Here, we draw on the Montreal Protocol start-and-strengthen approach to show that accelerated phase-down under the Kigali Amendment could result in additional reductions of 72% in 2050, increasing chances of staying below 1.5 °C throughout this century. Main Hydrofluorocarbon (HFC) refrigerants are factory-made chemicals produced for use in refrigeration, air-conditioning, insulating foams, fire extinguishers, solvents and aerosol propellants. Since their introduction, emissions of HFCs have grown rapidly as they are the primary replacement for ozone-depleting substances (ODSs) currently managed under the Montreal Protocol (MP) 1 , 2 , 3 . HFCs do not contain ozone-destroying chlorine or bromine atoms but are powerful GHGs and account for ~1.5% of global anthropogenic GHG emissions 2 . Without any controls, HFC emissions are expected to double by 2030 and nearly quadruple by 2050, compared to 2015 levels 3 , 4 . In 2016, HFCs were included in the Kigali Amendment (KA) to the MP—a result of an international consensus that HFCs could be most effectively controlled through the phase-down of their production and consumption under the MP (ref. 5 ), complementary to mitigation under the UN Framework Convention on Climate Change (UNFCCC). The MP process accumulated experience and expertise to ensure a fast and efficient phase-down of HFCs, which are in the same family of gases, have similar chemical properties and are used in the same sectors as the ODSs that they are replacing. The MP also uses a ‘start and strengthen’ approach wherein controlled substances are phased out in an orderly and transparent schedule which is regularly evaluated and strengthened, through amendments, as markets innovate and adjust (Extended Data Fig. 1 ; ref. 6 ). Furthermore, unlike the Paris Agreement to the UNFCCC, the MP and its amendments are legally binding for countries that ratify them. KA is a global agreement (in force since 1 January 2019) to phase-down consumption of HFCs of 80–85% by the late 2040s (Supplementary Table 1 ). Unlike previous MP amendments, which managed ODSs, the KA is primarily a climate treaty, therefore it is appropriate to evaluate the sufficiency of its ambition on the basis of its consistency with climate mitigation targets. The 2015 Paris Agreement established an ambitious target of limiting global mean temperature rise this century to well below 2 °C, preferably to 1.5 °C compared to pre-industrial levels, but did so in the context of broader international goals of sustainable development and poverty eradication. The 1.5 °C-consistent scenarios used in the IPCC Special Report on Global Warming of 1.5 °C (SR1.5) include a 75–80% reduction in HFC emissions by 2050, compared to 2010 levels 7 , along with deep and simultaneous reductions of CO 2 and all non-CO 2 climate-forcing emissions. A recent IIASA study 8 used the Greenhouse Gas and Air Pollution Interactions and Synergies (GAINS) model 9 framework to develop a range of long-term scenarios for HFC emissions under varying degrees of stringency in climate policy and assessed co-benefits in the form of electricity savings and associated reductions in GHG and air pollutant emissions. Full compliance with the KA (Fig. 1a ) is expected to achieve a 56% reduction in HFC emissions by 2050, compared to 2010 levels 8 . Hence, it will not achieve the 75–80% threshold set by 1.5 °C-consistent scenarios 7 . Full compliance with the KA phase-down schedule is estimated to avoid 0.2–0.4 °C additional warming by the end of this century 3 , 10 which is significant but insufficient to achieve a 1.5 °C-consistent pathway 7 . Despite the widely recognized success of the MP for phasing out ODSs faster 11 and at a lower cost than originally assumed, some observers question whether the HFC-reduction process under KA is taking place quickly enough to adequately address the urgency of the climate crisis 12 . Considering the role HFC mitigation plays in 1.5 °C-consistent scenarios 7 , enhancing the ambition of mitigation efforts by all parties to the MP is called for. In this study, we develop a series of alternative HFC phase-down scenarios (Fig. 1 ) consistent with the MP’s history and experience of a gradual increase in ambition. Fig. 1: HFC phase-down schedule of the Article 5 and non-Article 5 parties to the MP in KA and alternative scenarios. Article 5 parties are divided into two groups: Group 1, most of the Article 5 parties; and Group 2, Bahrain, India, Iran, Iraq, Kuwait, Oman, Pakistan, Qatar, Saudi Arabia and the United Arab Emirates. Group 2 (dashed black line) has a later freeze and phase-down steps compared with Group 1 (solid yellow line). Several non-Article 5 parties with dashed red line (Belarus, the Russian Federation, Kazakhstan, Tajikistan and Uzbekistan) have a different formulation for the calculation of baseline and have different initial phase-down steps (that is, the first two steps) from the other non-Article 5 parties with solid blue line. a , KA HFC phase-down schedules. b , Article 5 Groups 1 and 2 alignment scenario. c , 95% scenario. d , Combined scenario. e , Accelerated combined scenario. Full size image KA defines HFC phase-down schedules for four different party groups. The first group (Article 5 Group 1) includes 136 primarily developing countries that make up all Article 5 countries with the exception of ten countries characterized by high ambient air temperatures forming a second group (Article 5 Group 2) and allowed less ambitious timing of targets. Non-Article 5 countries are primarily developed countries and under KA divided into two groups with 45 countries in a first group (non-Article 5, earlier start) and five in a second group allowed to start somewhat later (non-Article 5, later start). We estimate HFC emissions (CO 2 eq using global warming potential GWP 100 from IPCC/AR5 (ref. 13 )) for all analysed scenarios (Fig. 2a ). In a pre-KA baseline, HFC emissions increase to about 4.2 GtCO 2 eq by 2050, which is within the range of previous estimates (4.0–5.3 GtCO 2 eq) by Velders et al. 14 . With full KA compliance, global HFC emissions drop to 0.32 GtCO 2 eq by 2050, achieving 56% reduction compared to 2010 levels. Technology exists that, if deployed globally to a maximum extent, could achieve near-complete mitigation of HFC emissions one-decade sooner than the KA phase-down schedule, resulting in a cumulative reduction of ~77 GtCO 2 eq HFC emissions until 2050 (Supplementary Table 2 ). Such a rapid reduction is, however, infeasible on practical grounds and also inconsistent with the MP’s history of a phased step-wise approach to refrigerant management. Instead, we have analysed a set of more realistic reduction scenarios. Fig. 2: HFC emissions. a , b , The baseline and alternative scenarios ( a ); the year 2050 under KA, Maximum Technically Feasible Reduction (MTFR) and additional alternative scenarios ( b ). The dashed purple line SSP1–1.9 scenario in a describes a world where global CO 2 emissions are cut to net zero around 2050 and meets the Paris Agreement’s goal of keeping global warming to well below 2 o C, preferably to 1.5 o C, compared to pre-industrial levels. The SSP1–1.9 scenario has a different baseline and modelling assumptions as compared to the IIASA study 8 . The orange box in b indicates the estimated range of HFC emissions consistent with the 1.5 o C target 7 . Source data Full size image First, we analyse whether aligning Article 5 Group 2 countries with the higher ambition level of the Article 5 Group 1 (Article 5 Groups 1 and 2 alignment scenario) would result in achieving the Paris Agreement targets. The results show that this would not be the case (Fig. 2b , where the orange box indicates the 75–80% threshold set by 1.5 °C-consistent scenarios) 7 . Next, we increase the ambitions of both Article 5 and non-Article 5 parties, resulting in achieving the Paris Agreement targets globally by 2050, however with different cumulative emissions until 2050 due to variations in the timing of adapted KA targets. If Article 5 and non-Article 5 party groups follow the KA phase-down schedules but step-up ambitions in their final phase-down step (starting in 2036 for non-Article 5 with Article 5 following only in year 2050) to 95% below baseline in the year 2050 (95% scenario), then the resulting cumulative reduction is 61 GtCO 2 eq (Supplementary Table 2 ). If the Article 5 Groups 1 and 2 alignment scenario and the 95% scenario are combined (combined scenario), the resulting cumulative reduction is 63 GtCO 2 eq. Finally, if the combined scenario is accelerated with a more ambitious target timeline (accelerated combined scenario), with non-Article 5 countries achieving 95% reduction already in 2036 and Article 5 countries starting earlier and achieving 95% in 2045, then a cumulative HFC-reduction potential of 69 GtCO 2 eq can be achieved. The last scenario follows the example of the accelerated phase-out of HCFCs under the MP from 2007. In particular, for the period 2021–2030, the cumulative emissions are lower in the accelerated combined scenario compared to other scenarios. The developing countries are less than 3 years away from the first HFC consumption compliance obligation of the KA applicable to Article 5 Group 1 countries. Much still needs to be done to ensure that all these countries are ready to comply with the 2024 cap on HFC consumption and production. In countries where HFC consumption is projected to exceed their baselines by 2024, there is an urgent need to implement actions towards a rapid transition to low-GWP refrigerants. In countries where HFC consumption is projected lower than their baselines by the agreed freeze year, there are opportunities for faster implementation of the KA to achieve HFC emission reductions earlier than strictly required under KA. KA provides an important opportunity and framework to control the production and consumption of HFCs resulting in reductions of both direct and indirect emissions from the cooling sector. Combining benefits from energy efficiency and climate-friendly cooling is vital to developing markets with rising cooling demand. Harnessing such opportunities by ensuring the transition to low-GWP refrigerants, combined with the adoption of energy-efficient cooling equipment, can potentially double the climate benefits of the HFC phase-down 3 and save as much as 20% of the expected future global electricity consumption 8 . Lower electricity consumption also offsets the need to build new power plants and increases energy access across emerging economies. For example, transitioning to low-GWP refrigerants with enhanced energy efficiency in room air-conditioners in China could avoid the construction of ~300 coal-fired power plants (500 MW each) by 2050 15 . Therefore, an early HFC phase-down will foster sustainable growth with energy-efficient, innovative technologies that provide jobs, increase energy access and reduce air pollution while reducing consumer energy bills. KA is a work in progress, but one that needs to be embraced and expanded upon in the global interests of mitigating climate change, just as the original MP has been instrumental in the recovery of the stratospheric ozone layer. An example of progressive legislation could be the HFC-reduction steps under the European Union F-gas regulation that is more ambitious than what is included in the KA 16 . Finally, if parties to the MP do not align early HFC phase-down policies with their economic transformation plans in the post-COVID era, they might not only become more vulnerable to climate shocks but also miss out on new technologies, investment and market access in a rapidly shifting global economy. Methods In developing the baseline HFC emissions under the IIASA study 8 , the historical consumption of HFCs for major sources, that is, stationary and mobile air-conditioning and domestic refrigerators, have been derived in a consistent manner across countries, starting from a compilation of data on underlying drivers; for example, the number of vehicles by type, commercial floor space area, cooling degree days, per capita income, average household sizes, current equipment penetration rates and so on. Estimating HFC consumption in commercial, industrial and transport refrigeration, foams and other smaller HFC sources (for example, aerosols, fire extinguishers and solvents) is more challenging since it varies greatly between countries due to differences in industrial structures and consumption patterns 17 . For these sectors, historical HFC consumption, as reported by industrialized countries to the UNFCCC, has been adopted when available. For developing countries, information on HFC consumption in these sectors has been compiled from various published sources, alternatively, derived consistently from underlying activity data using default factors from literature 8 , 18 . For the development of the baseline scenarios until 2040 we use the existing model setup in GAINS 9 , which for global scenarios uses drivers consistent with macroeconomic and energy sector projections from the International Energy Agency 19 . The extension in demand for cooling services between 2040 and 2100, expressed in tonnes of HFC consumed 8 , is consistent with the growth in population and macroeconomic indicators of the third Shared Socioeconomic Pathway (SSP3) 20 and the expected future increase in regional cooling degree days. In addition to the KA and Maximum Technically Feasible Reduction (MTFR) scenarios as explained in the IIASA study 8 , we analyse four scenarios for HFC phase-down to achieve the Paris Agreement targets by 2050. In the Article 5 Groups 1 and 2 alignment scenario, we assume that Article 5 Group 2 countries join the Article 5 Group 1 phase-down schedule immediately as shown in Fig. 1b . Note that Article 5 Group 2 countries have a later freeze date (Supplementary Table 1 ) and delayed phase-down steps compared with Article 5 Group 1 under the KA. In the 95% scenario, we assume that all party groups will follow the KA phase-down schedules but, in addition, the final phase-down step by 2050 will be 95% of baseline, valid for all party groups, as shown in Fig. 1c . In the combined scenario, we assume that Article 5 Group 2 countries join the Article 5 Group 1 phase-down schedule immediately as in the case of the Article 5 Groups 1 and 2 alignment scenario; in addition, the final phase-down step by 2050 will be 95% of baseline and, just like in the case of the 95% scenario, be valid for all party groups, as shown in Fig. 1d . Finally, the accelerated combined scenario is designed following the example of the accelerated phase-out of HCFC in 2007 (Extended Data Fig. 1 ) as shown in Fig. 1e . In the case of the HCFC example, the accelerated phase-out was agreed upon 11 years after the freeze date set for non-Article 5 countries and 3 years after the first phase-out step but before the freeze date set for Article 5 countries. Data availability Global HFC emissions data used in the baseline and alternative scenarios by MP party groups are available at: . Source data are provided with this paper.

To have a better chance of holding global warming to 1.5 degrees Celsius, we need to accelerate the phase-down of HFC refrigerants under the Montreal Protocol. This could also reduce pollution and improve energy access. An air conditioner may freshen the atmosphere in your home, but in doing so, it is probably degrading the atmosphere of Earth. Along with other cooling technologies such as refrigerators and heat pumps, today's aircon commonly relies on chemicals called HFCs (hydrofluorocarbons), which are very powerful greenhouse gasses. HFCs have been used to replace ozone-depleting substances, and their emissions have increased rapidly in the past two decades. To meet the Paris climate goals, the world now needs to wean itself off HFCs quickly, according to a new study led by IIASA researchers published in the journal Nature Climate Change. As a bonus, this process could reduce global power consumption substantially, bringing many benefits such as lower pollution. HFCs can be replaced with various gasses that have a far lower climate impact per kilogram, including ammonia, CO2, and hydrocarbons such as propane. Indeed, a phase-down of HFCs is already required by international law. In 2016, these chemicals were brought into the Montreal Protocol, a treaty originally set up to curb ozone-depleting substances. The protocol's 2016 Kigali Amendment lays out HFC cuts for four groups of countries up to 2047, requiring consumption to fall by 80 to 85% relative to their respective baselines. The problem is that HFC emissions lag years behind consumption. They can leak out of cooling devices during manufacture and use, and when equipment is scrapped. The new study considers this lag and examines how various HFC consumption scenarios would affect future emissions, using the IIASA Greenhouse Gas and Air Pollution Interactions and Synergies (GAINS) model. The study projects that, if left uncontrolled, HFC emissions from 2019 to 2050 would have amounted to more than 92 billion tons CO2 equivalent. (Note that cumulative emissions until 2050 determines the effectiveness of HFC reduction for climate mitigation). Controlled by the Kigali Amendment, the total should be about 32 billion tons. That is however still far above the roughly 16 billion tons in SSP1‐1.9 consistent climate scenarios, in which global warming is limited to around 1.5 degrees Celsius above pre‐industrial temperatures. "Current ambitions for HFC emissions reductions are not sufficient to meet the Paris Agreement's 1.5 degrees Celsius goal. A more ambitious target under the Kigali Amendment could still help achieve the Paris goal if countries act early," says Pallav Purohit, lead author on the study and a senior researcher in the Pollution Management Research Group of the IIASA Energy, Climate, and Environment Program. The Montreal protocol has a history of ratcheting up ambition, so the authors looked at various options for stronger HFC cuts. For example, in the existing Kigali Amendment, one group of developing nations is allowed to delay cuts by a few years because they have especially high ambient temperatures—so what if they were required to keep the same pace as other developing nations? This turns out to make little difference to total emissions. Or, what if all nations had to reach 95% emissions cuts by 2050, instead of 80 to 85% in 2047? Again, this barely reduces cumulative emissions to 2050, but it leaves emissions at a lower level for the rest of the century, which is more in line with 1.5 degrees Celsius scenarios. The most effective option involves all countries not only hitting 95% by 2050, but making accelerated deep cuts before that (for example, developed countries reaching 55% cuts in 2025, instead of the 35 to 40% required in the Kigali Amendment, and developing countries reaching 35% cuts in 2030, compared with 0 to 10% in the Amendment). This leads to 2050 cumulative emissions of less than 24 billion tons CO2 equivalent—much closer to the 1.5 degrees Celsius climate scenario. Better still, this early move would be an opportunity to replace old cooling equipment with more efficient hardware. This could save up to 20% of expected future global electricity consumption, which would double the climate benefits of the HFC phase-down, reduce air pollution, improve energy access and cut consumer energy bills. "Drawing on the Montreal Protocol's start-and-strengthen approach, accelerated HFC phase-down would increase the chances of staying below 1.5 degrees Celsius," Purohit concludes.

10.1038/s41558-022-01310-y

Chemistry

Novel potent antimicrobial from thermophilic bacterium

Arnoldas Kaunietis et al, Heterologous biosynthesis and characterization of a glycocin from a thermophilic bacterium, Nature Communications (2019). DOI: 10.1038/s41467-019-09065-5 Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-019-09065-5

https://phys.org/news/2019-03-potent-antimicrobial-thermophilic-bacterium.html

Abstract The genome of the thermophilic bacterium, Aeribacillus pallidus 8, encodes the bacteriocin pallidocin. It belongs to the small class of glycocins and is posttranslationally modified, containing an S -linked glucose on a specific Cys residue. In this study, the pallidocin biosynthetic machinery is cloned and expressed in Escherichia coli to achieve its full biosynthesis and modification. It targets other thermophilic bacteria with potent activity, demonstrated by a low minimum inhibitory concentration (MIC) value. Moreover, the characterized biosynthetic machinery is employed to produce two other glycopeptides Hyp1 and Hyp2. Pallidocin and Hyp1 exhibit antibacterial activity against closely related thermophilic bacteria and some Bacillus sp. strains. Thus, heterologous expression of a glycocin biosynthetic gene cluster including an S -glycosyltransferase provides a good tool for production of hypothetical glycocins encoded by various bacterial genomes and allows rapid in vivo screening. Introduction Ribosomally synthesized and posttranslationally modified peptides (RiPPs) are produced in all three domains of life. Part of the RiPPs overlap with a group of antibacterial peptides produced by bacteria, and this group historically is designated as bacteriocins 1 , 2 , 3 . They are active against other bacteria that are mostly closely related to the producer. These peptides exhibit considerable diversity with respect to their size, structure, mechanism of action, inhibitory spectrum, immunity mechanisms, and targeted receptors 4 . In the era of emergence of antibiotic-resistant bacteria 5 , bacteriocins have been suggested as a potential alternative to antibiotics in clinics and veterinary settings, but also as food preservatives against spoilage and pathogenic microorganisms 2 , 6 , 7 . Thermophilic bacteria have shown a great potential in biofuel production because of their higher metabolic rate and enzyme stability at elevated temperatures. Moreover, growth at high temperature facilitates recovery of volatile products, like ethanol 8 , and reduces requirement for cooling. Thermophilic fermentations are less prone to contaminations by mesophiles, although there are still risks that bioreactors will be contaminated by other thermophiles 9 , 10 . In addition, contamination by thermophiles is also a problem in production of dairy products 11 . This shows the need of discovery of new natural compounds that have activity against thermophilic bacteria. Glycocins are posttranslationally glycosylated bacteriocins. The sugar moieties are linked to the side chains of either Cys, Ser, or Thr residues. A glycocin can be regarded as being “glycoactive” when sugar moieties are essential for the antimicrobial activity 12 . Only a few glycocins have been identified and reported to date: sublancin 168, produced by Bacillus subtilis 13 ; glycocin F, produced by Lactobacillus plantarum 14 ; ASM1 (homologous to glycocin F), produced by Lb. plantarum 12 , 15 ; enterocin F4-9, produced by Enterococcus faecalis 16 and thurandacin, encoded by Bacillus thuringiensis and identified by genomic data mining and chemo-enzymatical synthesis in vitro 16 . The understanding of their mechanism of growth inhibition of target bacteria is far from complete. It is known that a specific phosphoenolpyruvate:sugar phosphotransferase system (PTS) is a factor affecting glycocin F and sublancin antibacterial activities 12 , 17 and that sublancin does not affect the integrity of the cell membrane and acts bacteriocidal 12 , 17 . In contrast, glycocin F and enterocin F4-9 have been reported to be bacteriostatic 14 , 16 . The synthetic machinery of the best-studied glycocin, sublancin, is encoded by a gene cluster containing sunA, sunS, sunT, bdbA , and bdbB genes. The precursor peptide, SunA, is modified by the S -glycosyltransferase SunS, which forms a very unusual S -linkage between the Cys residue and glucose. SunA glycosylation in vitro by SunS has been confirmed by chemo-enzymatical synthesis of mature sublancin 13 . Based on the SunT sequence similarity to bacteriocin ABC- transporters/peptidases, it is assumed that SunT cleaves the leader sequence and transports the core peptide to the outside of the cell. Two thiol-disulfide oxidoreductases, BdbA and BdbB, might be responsible for disulfide bond formation in the peptide 18 , 19 . In addition, it has been confirmed that the same gene cluster encodes the immunity protein SunI 20 . A similar genetic organization was found in gene clusters encoding the putative synthetic machineries of glycocin F 14 , thurandacin 21 and enterocin F4-9 16 . To date only two bacteriocins, i.e. geobacillin I and geobacillin II, produced by the thermophilic bacteria Geobacillus thermodenitrificans NG80-2, are well characterized 22 , 23 , 24 . Other bacteriocin-like antibacterial compounds from thermophilic microorganisms have been described in much less detail 25 , 26 , 27 , 28 , 29 , 30 , 31 . These reasons prompted us to find and to study new bacteriocins of this group. Thus, we have chosen the thermophilic Aeribacillus pallidus 8 strain that was previously isolated from soil above oil wells in Lithuania 32 . Previous studies have shown that this strain secretes an antibacterial compound that is active against other thermophilic bacteria. Unfortunately purification of this compound and identification of its amino acid sequence were not successful 32 , 33 . In this study we have identified genes in the genome of A. pallidus 8 that encode a biosynthetic machinery for a hypothetical bacteriocin (i.e. pallidocin), which belongs to the small class of glycocins. We also demonstrate the functional expression of the whole biosynthetic gene cluster of a glycocin in Gram-negative Escherichia coli , which facilitates further engineering and mechanistic studies. Following characterization of pallidocin demonstrated that it exhibits extremely strong activity against specific thermophilic bacteria, such as (Para)Geobacillus sp. and Caldibacillus sp. In addition, we identified and synthesized a variety of hypothetical glycopeptides and determined their properties by employing heterologous expression system of pallidocin. The characterized pallidocin S -glycosyltransferase PalS could be used not only for biosynthesis of hypothetical glycocins, but also for introduction of unique posttranslational modifications into other peptides with the aim to improve their bioactivities. Pallidocin is a good candidate to prevent bacterial contaminations in industrial fermentations operating under elevated temperatures. Previously, full maturation of recombinant glycocins was only reported in vitro for thurandacin and sublancin. Glycosylation and leader cleavage were performed enzymatically, followed by chemical oxidative folding 13 , 21 . The in vitro experiments limit the yield of the final product, are time consuming and expensive. Recently, a system was developed for the heterologous expression of sublancin in E. coli SHuffle T7 Express cells that in vivo installs the glycosylation and oxidative folding following a single in vitro step of proteolytic leader cleavage 34 . The SHuffle T7 Express strain expresses the disulfide bond isomerase DsbC, aiding oxidative folding of proteins in the cytoplasm 35 . Here, we demonstrate a different in vivo heterologous expression system to produce completely mature glycocins in E. coli BL21(DE3), evading the in vitro chemical and enzymatic steps. Results Identification and heterologous biosynthesis of pallidocin The pallidocin producer strain was identified as A . pallidus 8 (previously referred to as Geobacillus sp. 8 32 ) by a Genome-to-Genome Distance Calculator (GGDC) using a digital DNA−DNA hybridization (dDDH) analysis tool 36 . The A. pallidus 8 genome 37 was processed with the BAGEL4 bioinformatics tool 38 for the identification of gene clusters for bacteriocins biosynthesis. An operon coding for a putative glycocin was identified (Fig. 1a ). It contains five genes which were named palA , palS , palT , paldbA , and paldbB . Proteins encoded by these genes display between 38 and 53% sequence similarity with proteins of known functions, encoded in other bacteria. Based on this analysis we presumed that palA encodes the 61 amino acid precursor peptide (Fig. 1b ), palS encodes the SunS-like family peptide S -glycosyltransferase, palT encodes the SunT-like superfamily peptidase/ABC transporter protein, paldbA encodes the thioredoxin-like enzyme, and paldbB encodes the DsbB-like disulfide bond formation protein B. Fig. 1 Gene cluster and precursor peptides of glycocins. a Pallidocin biosynthetic gene cluster (4805 bp) pal identified by BAGEL4 in A. pallidus 8 genome. b Alignment of glycocin precursor peptides. Conserved regions are highlighted in green color, Cys forming disulfide bonds are underlined in orange, glycosylated amino acids are underlined in blue. Red dots between amino acids indicate predicted or experimentally determined leader cleavage sites. Numbers in the end of the sequences indicate the length of peptides in amino acids. PalA pallidocin precursor, SunA sublancin precursor, Hyp putative glycocin precursor Hyp1, Hyp2 putative glycocin precursor Hyp2, EnfA4-9 enterocin F4-9 precursor, GccF glycocin F precursor Full size image We decided to try to express this gene cluster in the easy-to-handle Gram-negative host E. coli . The whole biosynthetic gene cluster ( pal ) of the hypothetical glycocin (Fig. 1a ), which was named pallidocin, was amplified by PCR and cloned into the expression vector (Supplementary Fig. 1 ). The pal operon, starting with the start codon for the PalA, was cloned into the MCS behind the arabinose promoter and RBS of the vector. The expression of the pal genes in the heterologous host E. coli BL21(DE3) was induced with arabinose. The produced active antibacterial peptide was purified from 2L of supernatant (Supplementary Fig. 2 ) using liquid chromatography methods. The yield of the peptide was enough for mass spectrometric (MS) analysis and initial antibacterial activity screening, but not for the quantification by measuring the absorbance at 280 nm wavelength. Structural and functional characterization of pallidocin The predicted monoisotopic mass [M + H] + of the unmodified pallidocin core peptide is 4061.76. The monoisotopic mass [M + H] + of purified native pallidocin observed by LC-ESI-MS was 4219.79 (Supplementary Fig. 3a ). The monoisotopic mass [M + H] + of the peptide, after treatment with tris(2-carboxyethyl)phosphine-hydrochloride (TCEP), was 4223.82 (Supplementary Fig. 3b ). This suggests that the peptide has a posttranslational modification with a mass of +162 and two disulfide bonds (−4). To identify the modified residue, pallidocin was fragmented with chymotrypsin and further analyzed by LC-ESI-Q-MS/MS mass spectrometry (Supplementary Figs. 4 - 9 ). This showed that Cys25 of the core peptide has a modification with a mass of +162.05, which might be a hexose moiety. To identify the sugar, pallidocin was analyzed by GC-MS (Supplementary Fig. 10 ). The results demonstrated that the moiety attached to the Cys25 residue is glucose. Based on the analysis by the secondary structure prediction tool PSIPRED 39 , pallidocin has two α-helices (Fig. 2 ). Far-UV CD spectra analysis of mature pallidocin (Supplementary Fig. 11 ) revealed that the peptide contains substantial amounts of helical structure, judging from the pattern of the spectra from 193 to 240 nm. Similar spectra patterns were observed for sublancin 34 , 40 , 41 and glycocin F 14 , also. The estimate of the secondary structure content, made by the method of Raussens et al. 42 , predicted predominantly helical structure, with an estimate of 47% helix. The peptide was also estimated to contain 13% β-turn and 11% β-sheet structure. Fig. 2 Proposed structures of pallidocin, Hyp1 and Hyp2 glycocins based on PSIPRED predictions from the sequence and the known tertiary structure of sublancin 168 and GccF. α-helical structure highlighted in blue color. Coil structure highlighted in purple. Black dots indicate hydrophobic amino acids. Orange indicates acidic amino acids, whereas green indicates basic amino acids Full size image A wide range of temperatures and pHs were applied to evaluate pallidocin’s stability (Supplementary Table 1 ). The peptide was stable at room temperature for 10 days, while after 30 days it retained 12% of the activity. Fifty percent antibacterial activity was retained after autoclaving pallidocin at 121 °C for 15 min. Incubation at the pH range 2−10 did not affect pallidocin’s antibacterial activity, demonstrating its exceptionally high stability. Functional assessment of the genes palS and palT The genes palS and palT were cloned (Supplementary Fig. 12 ) and coexpressed with palA-his in E. coli BL21(DE3) to determine their functions. Gene palA was fused with a His7-tag in the C-terminus (Supplementary Figs. 12 and 13 ) to facilitate the purification of its product. The highest yield of synthesized peptides was observed in the insoluble fraction of the cell lysate. Peptides were purified from the insoluble fraction, tested for activity and analyzed by MALDI-TOF-MS to evaluate the presence of modifications (Table 1 ). Table 1 Results of glycocin precursor genes coexpression with palS and palT Full size table As expected, expression of only palA-his resulted in the synthesis of pre-PalA-His, the precursor peptide with the leader still attached (Supplementary Fig. 14 ), and with no antibacterial activity (Fig. 3 ). Coexpression of palA - his and palS genes resulted in the biosynthesis of pre-PalA-His-Glc (the precursor peptide with the leader attached and a mass increment of 162), which portrays glucosylation (Supplementary Fig. 15 ). Notably, this compound was active against the sensitive thermophilic strain Parageobacillus genomospecies 1 NUB36187 (Fig. 3 ). Coexpression of palA - his with palT resulted in the biosynthesis of the PalA-His core peptide with a mass corresponding to the leaderless peptide and this compound lacked antibacterial activity (Fig. 3 ). Coexpression of the genes palA - his with palS and palT resulted in the biosynthesis of a PalA-His-Glc core peptide (mature pallidocin) with the mass corresponding to the leaderless peptide with a mass increment of 162, corresponding to glucosylation (Supplementary Fig. 16 ). The glucosylated core peptide had antibacterial activity too (Fig. 3 ). Fig. 3 Antibacterial activities of synthesized and purified peptides. The peptides were derived after coexpression of glycocin precursors with PalS/PalT/PalST proteins in E. coli , subsequent purification by Ni 2+ affinity chromatography (IMAC) and RP-HPLC. Antibacterial activity was indicated as a clearly visible inhibition zone of sensitive strain P. genomospecies 1 NUB36187 and assessed by a spot on a lawn assay Full size image The role of disulfide bonds on the antibacterial activity To confirm the presence and importance of disulfide bonds in pallidocin, purified glycosylated precursor peptide with leader pre-PalA-His-Glc and leaderless PalA-His-Glc core peptide (mature pallidocin) were treated with TCEP and iodoacetamide (IAA) (Supplementary Table 2 ). These peptides were derived after coexpression of palA-his with palS or palST . After the treatment, the masses of pre-PalA-His-Glc and PalA-His-Glc core peptide measured by MALDI-TOF-MS increased by +234 and +228, respectively. The expected mass increment for one alkylated Cys is 57, for four alkylated Cys is 228. After all, despite the observed mass difference (6), which is tolerated as the machine error, it is clear that the disulfide bonds were reduced and all free Cys residues were alkylated. Notably, the disruption of disulfide bonds resulted in the loss of antibacterial activity against the indicator strain P . genomospecies 1 NUB36187 (Supplementary Table 2 ). After treatment with only IAA, no loss of antibacterial activity and mass increment was observed (Supplementary Table 2 ), confirming that all disulfide bonds are present in the pre-PalA-His-Glc and the PalA-His-Glc core peptide. In vitro leader peptide cleavage of pallidocin precursor To develop a more efficient way for pallidocin production, the genes his-Xa-palA and palS were coexpressed in E. coli BL21(DE3). PalA was engineered by adding a His-tag at the N-terminus and a Factor Xa peptidase cleavage site (IEGR) in front of the core peptide (Supplementary Fig. 13 ) for convenient leader removal. After coexpression, the glycosylated precursor peptide pre-His-Xa-PalA-Glc was purified and the leader was cleaved off in vitro using Factor Xa. The generated leaderless PalA-Glc core peptide (mature pallidocin) was further purified by RP-HPLC followed by mass spectrometry analysis for mass confirmation (Supplementary Fig. 17 ). The yield of synthesized active mature pallidocin was ~15 μg per 100 mL of bacterial culture. Activities of the glycosylated precursor peptide with leader and mature pallidocin were compared by the agar well diffusion assay using P. genomospecies 1 NUB36187 as a sensitive strain (Supplementary Fig. 18 ). The assay shows the highest serial twofold dilution of bacteriocin sample, which still displays antibacterial activity. Results indicate that pre-His-Xa-PalA-Glc had approximately 500 times lower activity than the mature pallidocin. Pallidocin alignment with other glycocins BLAST analysis of PalA identified two hypothetical peptides (Hyp1 and Hyp2) which have low sequence similarity to PalA (Fig. 1b ). They were encoded in Bacillus megaterium BHG1.1 (Hyp1) and Bacillus sp. JCM19047 (Hyp2) genomes. Hyp1 consists of a 23-residue leader sequence, a 41-residue core peptide, a 20-residue leader sequence Hyp2, and a 51-residue core peptide. Each of the peptides Hyp1 and Hyp2 has five Cys residues in the core sequence. Genome analyses of B. megaterium BHG1.1 and Bacillus sp. JCM19047 by the BAGEL4 tool did not find any gene clusters related to bacteriocin biosynthesis. However, BLASTp analysis of the genomic context of hyp1 revealed a gene cluster coding for putative glycocin biosynthetic machinery (Fig. 4 ). Genes in the cluster alongside the Hyp1 precursor gene encode for: Hyp1S protein with 50% sequence similarity to SunS-like family peptide S -glycosyltransferases; Hyp1T protein with 68% sequence similarity to SunT-like superfamily peptidase domain-containing ABC transporters; Trx protein with 69% sequence similarity to thioredoxin-like superfamily proteins and DsbB protein with 74% sequence similarity to DsbB-like superfamily disulfide bond formation proteins B. Moreover, BLASTp analysis of the genomic context of hyp2 revealed that a gene cluster may encode for putative glycocin biosynthetic machinery (Fig. 4 ). Genes in the cluster alongside the Hyp2 precursor gene encode the Hyp2S protein with 42% sequence similarity to SunS-like family peptide S -glycosyltransferases; Hyp2T protein with 40% sequence similarity to SunT-like superfamily peptidase domain-containing ABC transporters and the Trx protein with 43% sequence similarity to thioredoxin-like superfamily proteins. Two putative glycocin precursors, i.e. Hyp1 and Hyp2, were investigated and examined further for possible posttranslational modifications by the biosynthetic machinery of pallidocin. Fig. 4 The predicted biosynthetic gene clusters of glycocins Hyp1 and Hyp2. Hyp1 biosynthetic gene cluster encoded in the genome of B. megaterium BHG1.1 is 5013 bp in length and encodes for proteins: Hyp1, Hyp1S, Hyp1T, Trx, DsbB, and Hyp1U. The Hyp2 biosynthetic gene cluster encoded in the genome of Bacillus sp. JCM19047 is 3932 bp in length and encodes for proteins: Hyp2, Hyp2T, Trx, and Hyp2S Full size image Alignment of all glycocins (Fig. 1b ) showed that the leader sequences of SunA, ThuA, PalA, and Hyp1 share motifs of conserved residues. In contrast, the leader sequences of GccF and EnfA49 display no significant similarities. All glycocins have a double-glycine type motif (a proteolytic cleavage site at the end of the leader sequence). GccF and EnfA49 precursors have leader cleavage sites following the Gly-Gly motif. SunA, ThuA, and PalA precursors have a Gly-Ser-Gly motif, where the leader sequence is cleaved between the Ser and second Gly residue. In case of peptide Hyp2, it has putative leader cleavage site with motif Gly-Ser-Gly, whereas peptide Hyp1 has Gly-Lys-Gly. All glycocins characterized to date have five Cys residues in the core peptide and form two disulfide bonds between them. Hyp1 and Hyp2 core peptides have also five Cys residues and it is very probable that they form these bonds too. Early studies and our current one show that the Cys residue in the conserved region Gly-Cys-Gly-Gly/Ser of SunA, ThuA and PalA is glucosylated. This motif is present in Hyp1 and Hyp2 peptides as well and it suggests that Cys in this region is the target for glycosylation. The secondary structure prediction tool PSIPRED 39 proposes that two α-helices are present in the Hyp1 and Hyp2 core peptides as well as in pallidocin. Four out of five Cys residues reside in these helical structures of the peptides. Structures of sublancin 168 and glycocin F elucidated by NMR method have also two α-helices, which are nested by two disulfide bonds 43 , 44 and giving rise to the predicted structures for Hyp1 and Hyp2 (Fig. 2 ). Heterologous biosynthesis of hypothetical glycocins To determine whether PalS and PalT are able to modify and process different heterologous glycocin precursors in a heterologous Gram-negative host, they were coexpressed with genes coding for glycocin precursor peptides with leaders and His6-tags (Supplementary Figs. 13 and 19 ), i.e. pre-SunA-His, pre-GccF-His, pre-EnfA49-His, pre-Hyp1-His and pre-Hyp2-His. As before, the highest yield of peptides was observed in the insoluble fraction of cell lysate. After coexpression, the peptides were purified from the insoluble fraction and analyzed by mass spectrometry (Table 1 ). The analysis could not confirm that PalT cleaved off the leaders. However, it demonstrated that in addition to pre-PalA-His, the PalS glycosyltransferase does modify pre-SunA-His (Supplementary Fig. 20 and 21 ), pre-Hyp1-His (Supplementary Figs. 22 and 23 ), and pre-Hyp2-His (Supplementary Fig. 24 and 25 ), with a mass of +162 consistently pointing to monoglycosylation. The antibacterial activity analysis of the peptides showed that only pre-SunA-His-Glc and pre-Hyp1-His-Glc (glycosylated precursors with leaders) but not the pre-Hyp2-His-Glc were active against indicator strain P. genomospecies 1 NUB36187 (Table 1 and Fig. 3 ). We designed the genes core_hyp1-his and core_hyp2-his encoding the leaderless Hyp1 and Hyp2 core peptides with a His6-tag sequence at the C-terminus (Supplementary Figs. 13 and 19 ). As previously, palS was coexpressed with the core_hyp1-his and core_hyp2-his . The peptides were produced, purified and analyzed by mass spectrometry. LC-ESI-MS analysis confirmed that leaderless Hyp1-His and Hyp2-His core peptides (Supplementary Figs. 26 and 27 ) were glycosylated. However, only the Hyp1-His-Glc core peptide showed antibacterial activity against P. genomospecies 1 NUB36187 (Fig. 3 ). To confirm the importance of glycosylation, core_hyp1-his gene was expressed in E. coli without palS coexpression. The recombinant Hyp1-His core peptide did not show any antibacterial activity (Fig. 3 ), indicating that glycosylation indeed plays a crucial role in the antimicrobial activity of Hyp1, as well as for pallidocin, pre-SunA-His-Glc (this work) or mature sublancin 13 , 34 (glucosylated and oxidatively folded SunA core peptide). In addition, reduced and alkylated pre-Hyp1-His-Glc (Supplementary Table 3 ) and Hyp1-His-Glc core peptides (Supplementary Table 4 ) lost their antibacterial activities, confirming that disulfide bonds are also important for glycocin activity. Antibacterial activity of identified glycocins The antibacterial activity spectrum of purified pallidocin was tested against a number of Gram-positive and Gram-negative bacteria using a spot on lawn assay. Purified pallidocin exhibited antibacterial activity against Bacillus cereus ATCC14579, B. megaterium DSM319 and some thermophilic bacteria: Geobacillus stearothermophilus B4109, B4111, B4112, B4114 strains, P. genomospecies 1 NUB36187, Parageobacillus toebii B4110, B4162 strains, Parageobacillus caldoxylosilyticus B4119 and Caldibacillus debilis B4165. The activity spectrum of the synthesized glycocin Hyp1-His-Glc (glycosylated leaderless core peptide) was partially similar to the pallidocin spectrum (Supplementary Fig. 28 ). Glycocin was active against Geobacillus sp. B4113 and G. stearothermophilus B4163 strains, but exhibited no activity against G. stearothermophilus B4112, P. toebii B4110, P. caldoxylosilyticus B4119 and B. cereus ATCC14579 strains, in contrast to pallidocin. Minimum inhibitory concentrations (MICs) were determined for some pallidocin susceptible strains (Supplementary Tables 5 , 6 and 7 ). In liquid NB medium, the MIC of pallidocin against B. megaterium DSM319 was 37 nM, P. genomospecies 1 NUB36187 was 2.4 pM (i.e ±10 ng/L), G. stearothermophilus B4114 was 246 pM, Geobacillus toebii B4162 was 493 pM, and P. caldoxylosilyticus B4119 was 985 pM. MIC tests showed that Geobacillus spp. strains used in this assay rapidly acquired resistance for pallidocin as observed by growing colonies in the halo area or the growth in some 96-plate wells with pallidocin concentrations higher than MIC. Discussion We identified and characterized a posttranslationally modified bacteriocin (pallidocin). The gene cluster of the pallidocin biosynthetic machinery is encoded on the chromosome of the thermophilic bacterium A. pallidus 8, which is unprecedented. The characterized bacteriocin belongs to the small class of glycocins and shares genetic and structural similarities with sublancin 168 13 , glycocin F 14 , thurandacin 21 and enterocin F4-9 16 . Here, we show that the whole glycocin biosynthetic gene cluster, derived from a thermophilic bacterium, can be cloned and functionally expressed in a heterologous host E. coli BL21(DE3). Surprisingly, mature bacteriocin (glycosylated, oxidatively folded and leaderless) from Gram-positive bacteria could be synthesized and secreted by this Gram-negative host. Structural characterization of the purified recombinant pallidocin revealed posttranslational modifications: glycosylation and two disulfide bonds within two predicted α-helices. These two features appear to be specific for the class of glycocins 12 . Pallidocin has an S -linked glucose moiety to the Cys25, which is very uncommon among bacteria. Only a few cases of S -linked glycopeptides have been described and confirmed to date, i.e. for glycocin F, sublancin 168 and thurandacin 13 , 14 , 21 . Pallidocin exhibits high stability after exposure to high temperatures and a wide range of pH values. To our knowledge, only sublancin and enterocin F4-9 have been properly characterized for their stability. The stability of sublancin is decreased by 50% after 30 min incubation at 70 °C temperature. Sublancin is not very stable at acidic conditions; after incubation at pH 2 and 3 for 30 min, it retains only 20 and 40% of its activity, respectively 40 , while enterocin F4-9 after incubation at 80 °C for 15 min retains its full activity only at pH values from 2 to 8. After incubation at 100 °C for 15 min, enterocin F4-9 retains its full activity only at pH 4. Its activity is completely lost after incubation at 121 °C as well as at pH 10 16 . Compared to sublancin and enterocin F4-9, pallidocin is much more stable at high temperatures. Its activity decreases 50% only after 15 min incubation at 121 °C and is completely stable at 90 °C for 3 h. In contrast to sublancin and enterocin F4-9, pallidocin retains its full activity at acidic and basic conditions (pH 2−10). In vitro studies on the glycosylation of the sublancin precursor showed that S -glycosyltransferase has relaxed substrate specificity. It is able to attach other sugars: xylose, mannose, N -acetylglucosamine or galactose, as well. The native glycopeptide sublancin purified from B. subtilis contains glucose 13 . We do not know which sugar would be present in native pallidocin if the peptide was derived from A. pallidus 8. We can assume that native pallidocin has an S -linked glucose as well, as this sugar was found in recombinant pallidocin produced by E. coli . Experiments in which only two or three genes were coexpressed revealed that palS codes for an S -glycosyltransferase, which introduces glucose to the Cys25 residue in pallidocin. When palA-his was coexpressed with palS and palT , the mature pallidocin (PalA-His-Glc core peptide) was purified also. Additionally, the PalT protein shares sequence similarity to other bacteriocin ABC transporters, which leads to the conclusion that PalT has a dual, i.e. peptidase and transport, function. Oxidative folding of the peptide in the cytoplasm is unlikely. The formation of structural disulfide bonds in E. coli appears to be strictly segregated according to subcellular compartmentalization 45 . Because a reducing environment is necessary for enzymatic activity of glycosyltransferases 13 most probably PalS glycosylates peptides in the cytoplasm. When the whole gene cluster pal is expressed, the synthesized glycosylated precursor peptide should be transported to the periplasm by PalT, where the oxidative folding could take place. Following the precursor peptide coexpression with PalS, these bonds could be formed spontaneously by air oxidation 46 , 47 during peptide extraction and/or the purification process. In case of precursor peptide coexpression with PalST, disulfide bonds in the glycosylated core peptide could be formed in the periplasm or spontaneously by air oxidation 46 , 47 during the peptide extraction and purification process. We applied PalS for possible modifications of other glycocins and produced two glycosylated peptides: Hyp1 and Hyp2. PalS specifically monoglycosylated a certain group of glycocin precursors (pre-PalA-His, pre-SunA-His, pre-Hyp1-His, and pre-Hyp2-His), but it was not able to modify pre-GccF-His and pre-EnfA4-9-His. Interestingly, the 42%, 50% and 53% sequence similarities of Hyp1S, Hyp2S and SunS, respectively, to the PalS S -glycosyltransferase show quite surprisingly that this similarity is enough to allow modification of heterologous substrates. The distinctive feature of the glucosylated peptides is a Cys residue flanked by glycines (Gly-Cys-Gly-Gly/Ser) in the interhelical loop. PalS, ThuS, and SunS form a sugar S -linkage with a Cys residue in Gly-Cys-Gly-Gly/Ser motif of interhelical loop, and this motif is present only in SunA, ThuA, PalA, Hyp1, and Hyp2. Therefore, we can assign PalA, Hyp1, and Hyp2 to the sublancin- type glycocins, which is composed of SunA and ThrA 12 . Previous research on the sublancin S -glycosyltransferase, SunS, suggested that this enzyme recognizes an α-helical segment of the substrate and glycosylates only a Cys residue in the flexible loop following this helix 48 . Studies on thurandacin glycosyltransferase, ThuS, showed that it glycosylates ThuA at Cys28 or both Ser19 and Cys28. ThuS represents glycosyltransferase that catalyzes both O - and S -glycosylation of proteins. Earlier studies also demonstrated that SunS is not able to modify ThuA, although ThuS is able to modify SunA generating its bisglucosylated product. Moreover, SunA and ThuA with changed short sequences in their interhelical loops were also glucosylated by ThuS in these regions. On basis of this knowledge, it was suggested that the peptide sequence selectivity of ThuS is much more relaxed than that of SunS 21 . Here, we demonstrate that PalS has quite flexible substrate selectivity too and that it may monoglycosylate various precursor peptides: pre-PalA-His, pre-SunA-His, pre-Hyp1-His, and pre-Hyp2-His. In addition, we show that leaderless Hyp1-His and Hyp2-His core peptides can be modified by PalS, resulting in highly active antibacterial peptide Hyp1-His-Glc but not Hyp2-His-Glc. All sublancin-type glycocins, including the pallidocin, Hyp1 and glycosylated core peptide Hyp2, have a relatively rich content of hydrophobic residues at the N-terminus, and charged residues at the C-terminus. Comparing the core peptides of glycocins, Hyp2 has a relatively long C-terminus “tail”, not characteristic to other sublancin-type glycocins, and is relatively rich in charged residues (Glu20, Arg21, Arg22) in the interhelical loop (Fig. 2 ). These two features or one of them might be the reason why glycosylated Hyp2 core peptide and precursor did not have antibacterial activity against the strains tested. This could be the subject for future research on glycocin variability. In contrast to previous work on sublancin 48 , a newly published study showed that nonglycosylated and oxidatively folded core peptide of sublancin has the same topology of disulfide bonds as the native sublancin 34 . In fact, the previous assumptions that the free thiol of unmodified Cys disrupts the formation of the correct disulfide bridges by thiol-disulfide exchange and the blocked Cys residue can aid to form correct disulfide bonds between four free Cys residues 48 were proven wrong. We show that the disruption of disulfide bonds in the glycosylated precursor peptides pre-PalA-His-Glc, pre-Hyp1-His-Glc or the PalA-His-Glc and Hyp1-His-Glc core peptides leads to the loss of antibacterial activity. These data support observations of earlier investigations on sublancin. It confirms that disulfide bonds are crucial for antibacterial activity of glycocins as well as glycosylation 13 , 34 . Earlier studies on sublancin and thurandacin showed that the leader must be cleaved off to gain antibacterial activity 13 , 21 . In contrast, our synthesized glycosylated precursors with leaders (pre-PalA-His-Glc, pre-SunA-His-Glc and pre-Hyp1-His-Glc) still showed activity against a sensitive strain, suggesting that the leader removal from the glycosylated glycocin precursor is not essential for activity. In fact, glycocins with a leader attached exhibit substantial antibacterial activity. However, the absence of the leader increases the activity. It should be noted that in contrast to previous studies on glycocins, we have used a thermophilic indicator strain, which, as we show, exhibits extreme susceptibility, even to a leader-containing glycocin. Previously, only full maturation of recombinant glycocins was reported in vitro for thurandacin and sublancin. Glycosylation and leader cleavage was performed enzymatically, followed by chemical oxidative folding 13 , 21 . The in vitro experiments limit the yield of the end product, are time consuming and expensive. Recently, a system was developed for the heterologous expression of sublancin in E. coli SHuffle T7 Express cells that in vivo installs the glycosylation and oxidative folding following a single in vitro step of proteolytic leader cleavage 34 . SHuffle T7 Express strain expresses the disulfide bond isomerase DsbC, aiding oxidative folding of proteins in the cytoplasm 35 . Here, we demonstrate a different in vivo heterologous expression system for completely mature glycocins in E. coli BL21(DE3), evading the in vitro chemical and enzymatic steps. With the aid of PalS and PalT we can synthesize completely mature and active pallidocin, which is glycosylated, oxidatively folded and leaderless. Because pallidocin glycosyltransferase has flexible substrate selectivity, we propose that PalS could be a good tool for in vivo biosynthesis and screening of hypothetical glycocins, as we showed with Hyp1 and Hyp2. This approach demonstrates that after in vivo peptide glycosylation the disulfide bonds most probably are formed spontaneously during the purification process. It means that the in vitro chemical oxidative folding is not absolutely necessary. Moreover, the in vivo glycosylation of core peptides evades the enzymatic leader cleavage. Pallidocin exhibits antimicrobial activity against specific Gram-positive bacteria. Most of the tested thermophilic bacteria were susceptible to pallidocin. Among Bacillus sp. only B. megaterium DSM319 and B. cereus ATCC14579 strains were susceptible. It indicates a rather narrow activity spectrum restricted to closely related bacteria. Enterocin F4-9 exhibits antimicrobial activity against Enterococcus faecalis and E. coli JM109 16 . Glycocin F has narrow activity spectrum also that is restricted to the Lactobacillus genus 14 . Notably, sublancin is active against several Gram-positive bacteria, like Staphylococcus and, especially Bacillus species 20 . Pallidocin MIC values for thermophilic bacteria are extremely low when compared to the values for B. megaterium DSM319 and B. cereus ATCC14579. Notably, glycosylated leaderless Hyp1 core peptide demonstrated a different activity spectrum against thermophilic bacteria as compared to pallidocin. As in case of sublancin, the susceptible strains relatively easily develop resistance to pallidocin. The mechanism of resistance development will be the subject of our future study. ******Studies on sublancin and thurandacin have generally been focused on in vitro analysis and on the glycosyltransferase as potential tool for antibody generation and other purposes. S -linked glycopeptides have been highlighted to be more chemically and biologically stable compounds than O -linked glycopeptides 13 , 21 . Here, we identified and characterized antibacterial peptide pallidocin and expanded the currently small glycocin family. In addition, we demonstrated that glycosyltransferase PalS can be used as a tool for the biosynthesis of glycosylated antibacterial peptides in vivo. We have developed a system for heterologous expression and screening for hypothetical sublancin-type precursors with a minimal number of genes required. Pallidocin or Hyp1 could be applied in industrial processes facing thermophilic bacterial contaminations. Methods Mass spectrometry analysis Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF-MS) analysis was carried out using a Voyager-DE-Pro (Applied Biosystems) at the Interfaculty Mass Spectrometry Center (IMSC) of the University Medical Center Groningen (UMCG). One microliter of analyte was spotted onto an MALDI target and dried under ambient conditions. One microliter of matrix (saturated α-cyano-4-hydroxy-cinnamic acid matrix in 50% ACN/50% water with 0.1% TFA) was spotted onto the dried sample on the MALDI target and dried under ambient conditions prior to analysis. The Voyager-De-Pro was set in linear positive mode. The spectrum range was 3000–10,000. Voltage settings were 25,000 V acceleration, 93% grid, 0.1% guide wire. A data Explorer 4.9 (Applied Biosystems) was used to process and analyze the acquired data. High-resolution mass spectrometry (LC-ESI-Q-MS and MSMS) was carried out on a Q Exactive Hybrid Quadrupole-Orbitrap MS system (Thermo Scientific) combined with an UltiMate 3000 RSLC system (Thermo Scientific) at the IMSC of the UMCG. Each sample was injected into the UltiMate 3000 UHPLC system consisting of a quaternary pump, an autosampler and a column oven, which was coupled by a HESI-II electrospray source to the Q Exactive Orbitrap mass spectrometer (all Thermo Scientific). A Kinetex EVO-C18 (2.6 µm particles, 100 × 2.1 mm) column (Phenomenex) was used. The eluents for the LC separation were (A) water and (B) acetonitrile (ACN) both containing 0.1% formic acid. The following gradient was used: 5% B until 0.5 min, then linear gradient to 90% B in 4.5 min. This composition was held for 2 min, after which a switch back to 5% B was performed within 0.1 min. After 2.9 min of equilibration, the next injection was performed. The LC flow rate was 500 µL/min, the LC column was kept at 60 °C and the injection volume was 10 µL. The HESI-II electrospray source was operated with the parameters recommended by the MS software for the LC flow rate used (spray voltage 3.5 kV (positive mode)); other parameters were sheath gas 50 AU, auxiliary gas 10 AU, cone gas 2 AU, capillary temperature 275 °C, heater temperature 400 °C. The samples were measured in positive mode from m/z 500 to 2000 at a resolution of 70,000 @ m/z 200. The instrument was calibrated in positive mode using the Pierce LTQ Velos ESI Positive Ion Calibration Solution (Thermo Scientific). The system was controlled using the software packages Xcalibur 4.1, SII for Xcalibur 1.3 and Q-Exactive Tune 2.9 (all Thermo Scientific). The Xtract-algorithm within Xcalibur was used for deconvolution of the isotopically resolved data to a monoisotopic spectrum represented in Supplementary Figures 3 - 9 , 16 , 17 , 21 , 26 and 27 . For spectra represented in Supplementary Figures 14 , 15 , 20 , and 22 - 25 , manual deconvolution was calculated by formula 1. Theoretical peptide masses were calculated using a web tool ProteinProspector (MS-Product/MS-Isotope) at . $$\begin{array}{l}{\mathrm{m/z \times charge - }}\left( {{\mathrm{mass}}\,{\mathrm{of}}\,{\mathrm{the}}\,{\mathrm{attached}}\,{\mathrm{H}}^ + } \right) + \left( {{\mathrm{1 \times mass}}\,{\mathrm{of}}\,{\mathrm{H}}^ + } \right) \\ \quad = {\mathrm{peptide}}\,{\mathrm{molecular}}\,{\mathrm{weight}}\,\left[ {{\mathrm{M}} + {\mathrm{H}}} \right]^ + .\hfill \end{array}$$ (1) Genomic DNA analysis The pallidocin producer strain was identified as Aeribacillus pallidus 8 (previously referred to as Geobacillus sp . 8 32 ). A Genome-to-Genome Distance Calculator (GGDC) 36 , 49 for digital DNA−DNA hybridization (dDDH) analysis at was applied for the analysis of the genome with accession number LVHY00000000.1 ), which has been sequenced previously 37 . Genomic DNA was submitted for bacteriocin mining to the BAGEL4 server 38 at . An operon coding for a putative glycocin was identified (Fig. 1a ) in the genome. The information on the biosynthetic gene cluster pal , responsible for in vivo production of pallidocin and containing the genes palA , palS , palT , paldbA , and paldbB , was accessed via the National Center for Biotechnology Information (NCBI) at . The gene cluster is of chromosomal origin with the length of 4805 nucleotides and is found between base pairs 2768 and 7572 in the contig with accession number NZ_LVHY01000134.1 . Peptide and protein sequences were submitted to the NCBI database for BLASTp analysis at . BLASTp analysis revealed that palA encodes for the 61 amino acid precursor peptide (Fig. 1b ) with 39% sequence similarity to the sublancin 168 precursor SunA of B . subtilis (accession number NP_390031.1 ). palS encodes for a protein with 53% sequence similarity to the SunS-like family peptide glycosyltransferase of B . pseudomycoides (accession number WP_097849814.1 ). palT encodes for a protein with 38% sequence similarity to the SunT-like superfamily leader cleavage/ABC-type transporter of Paeniclostridium sordellii (accession number WP_057576215.1 ). paldbA encodes for a protein with 50% sequence similarity to the thioredoxin-like enzyme of B . megaterium (accession number WP_061859981.1 ). paldbB encodes a protein with 47% sequence similarity to the DsbB-like disulfide bond formation protein B of B . cereus (accession number WP_098593227.1 ). BLASTp analysis revealed that Hyp1 (accession number WP_061859978.1 ) is encoded in the genome of Bacillus megaterium BHG1.1 (accession number LUCO00000000 ). The genome encodes the Hyp1 biosynthetic gene cluster, which is 5013 bp in length and encodes for proteins Hyp1S (accession number WP_081113339.1 ), Hyp1T (accession number WP_061859980.1 ), Trx (accession number WP_061859981.1 ), DsbB (accession number WP_061859982.1 ), and Hyp1U (accession number WP_061859983.1 ). BLASTp analysis revealed that Hyp2 (accession number WP_035395705.1 ) is encoded in the genome of Bacillus sp. JCM 19047 (accession number BAWC00000000 ). The genome encodes the Hyp2 biosynthetic gene cluster, which is 3932 bp in length and encodes for proteins: Hyp2T (accession number WP_035395706.1 ), Trx (accession number WP_035395707.1 ), and Hyp2S (accession number GAF22634.1 ). Bacteriocin activity assays A colony of indicator strain P. genomospecies 1 NUB36187 (BGSC 9A11) was spread on a Nutrient broth (NB) medium agar plate containing: 1% tryptone (BD Bacto), 0.5% beef extract (BD BBL), 0.5% NaCl (Merck), 1.5% agar (BD Bacto), and incubated overnight at 60 °C. Grown biomass was collected with sterile inoculum loop and spread on a fresh NB agar plate and incubated for 4 h at 60 °C. After the incubation (before cells starts to form spores), all biomass from the plate was washed with NB medium and the cell suspension was adjusted to OD (600 nm) of 1. The suspension was mixed with liquid NB agar medium (55 °C) at the ratio 1:100 and mixed thoroughly. Fifteen milliliters of the resulting cell suspension was dispersed in a Petri plate and left to solidify. For a well diffusion assay, wells were cut out in the solid medium in the Petri plate and filled with 50 µL of serial twofold dilutions of samples. The titer was defined as the reciprocal of the highest dilution that resulted in inhibition of the indicator strain. For a spot on lawn assay, samples of 10 µL were spotted on the solid medium in the Petri dish and incubated at 60 °C overnight. P. genomospecies 1 NUB36187 (BGSC 9A11) was used as sensitive strain for all experiments. Antibacterial activities of hypothetical glycocins were tested against other thermophilic and mesophilic strains by the same approach. When the mesophilic strains were tested the incubations were performed at 37 °C. Analyzed bacteriocins were screened for their antibacterial activities against B. cereus ATCC14579, B. megaterium DSM319, Geobacillus sp. B4113, G. stearothermophilus B4109, B4111, B4112, B4114, B4161, B4163, P. toebii B4110, B4162, P. caldoxylosilyticus B4119 and C. debilis B4165. DNA amplification by PCR A single PCR mix included Phusion HF Buffer (Thermo Scientific), 2.5 mM dNTPs mix (Thermo Scientific), 1.5 mM MgCl 2 (Thermo Scientific), PfuX7 DNA polymerase (homemade), primers (0.5 μM each), and 1 ng/µL DNA template. Target DNA was PCR-amplified by 30 cycles of denaturing (94 °C for 30 s), annealing (5 °C or more lower then T m for 30 s), and extending (68 °C for 1 min per 1 kbp). Amplifications were confirmed by 1 or 2% agarose gel electrophoretic analyses. The list of primers and their sequences is provided in Supplementary Table 8 . DNA cloning DNA digestion was performed with restriction endonucleases purchased from Thermo Scientific and according to the manufacturer’s recommendations. Amplified or digested DNA was always cleaned with a NucleoSpin Gel and PCR Clean-up Extraction Kit (Macherey−Nagel), unless stated otherwise. A T4 DNA Ligase (Thermo Scientific) was used for DNA ligations, according to the manufacturer’s recommendations, unless stated otherwise. The ligation products were transformed to E. coli TOP10 cells by electroporation. Cells were plated on Lysogeny broth (LB) agar plates with appropriate antibiotics and grown at 37 °C overnight. Several colonies were picked and tested by colony PCR, to confirm whether the insert is present in the vector. Colonies with correct inserts were inoculated into LB medium with the appropriate antibiotic. The cultures were grown at 37 °C overnight, and plasmids were isolated using a NucleoSpin Plasmid Extraction Kit (Macherey−Nagel). DNA sequences of inserts in isolated plasmids were always confirmed by DNA sequencing. For gene expression, plasmid DNA was transformed to E. coli BL21(DE3) by electroporation. Cells were plated on LB agar plates with appropriate antibiotics and grown at 37 °C overnight. Several colonies were picked and inoculated into LB medium with the appropriate antibiotic, grown overnight at 37 °C, mixed with glycerol (glycerol end concentration 20%) and stored at −80 °C for further use. Cloning of pallidocin biosynthetic gene cluster pal A. pallidus 8 was grown in a Bacto Brain Heart Infusion medium (BD Diagnostics) at 55 °C in a shaking incubator. Genomic DNAs were extracted with a GenElute Kit (Sigma-Aldrich) according to the manufacturer’s recommendations. The whole gene cluster pal encoding the pallidocin biosynthetic machinery (Fig. 1a ) was amplified by PCR as a single unit (4805 bp) using F-PalA-USER and R-PalA-USER primers and A. pallidus 8 genomic DNA as template. The pBAD24 vector was PCR-amplified using F-pBAD-USER and R-pBAD-USER primers and pBAD24 plasmid as the template. Obtained PCR products were ligated by USER Enzyme (NEB), according to the manufacturer’s protocol, and transformed to E. coli TOP10 cells by electroporation. Colony PCR method using F-pBAD24 and R-pBAD24 primers was used for selection of positive transformants. Obtained construct pBAD24-pal was propagated in E. coli TOP10 cells and isolated. The presence of the cloned insert in the construct was confirmed by PCR (Supplementary Fig. 1 ). To amplify the insert the following pairs of primers were used together with pBAD24-pal as the template: the whole gene cluster pal (F-PalA-BspHI and R-BdbB); palA (F-PalA-BspHI and R-PalA-HindIII); palS (F-PalS-In-Fusion and R-PalS-In-Fusion); palT (F-PalT-In-Fusion and R-PalT-In-Fusion); bdbA (F-BdbA and R-BdbA); bdbB (F-BdbB and R-BdbB). The sequence of the insert ( pal gene cluster) was confirmed by DNA sequencing. DNA sequencing was performed with primers: F-pBAD24, R-pBAD24, Pal0, Pal1, Pal2, Pal3, Pal4, Pal5, and Pal6. The insert pal , starting with the start codon for the PalA, was cloned into the MCS behind the arabinose promoter and RBS of the pBAD24 vector. Sequences of primers are provided in Supplementary Table 8 . For the protein expression pBAD24-pal was transformed to E. coli BL21(DE3). Overexpression of pal and purification of pallidocin E. coli BL21(DE3) cells transformed with pBAD24-pal was grown overnight at 37 °C in LB medium containing ampicillin (50 µg/mL) and inoculated to 1 L of M9-ampicillin minimal medium (12.8 g/L Na 2 HPO 4 × 7H 2 O, 3 g/L KH 2 PO 4 , 0.5 g/L NaCl, 1 g/L NH 4 Cl, 0.24 g/L MgSO 4 , 0.11 g/L CaCl 2 , and 4 mL glycerol in MiliQ water) at the ratio 1:100. The cells were grown at 37 °C to the OD (600 nm) of 0.6–0.7, then arabinose was added to the final concentration of 2 mM and the culture was incubated at 37 °C for additional 16 h. Cells were harvested by centrifugation at 10,000 × g for 15 min at 4 °C. Supernatant was collected, immediately filtered through a 0.45 µm filter and uploaded on an Econo-Column chromatography column, 2.5 × 30 cm (Bio-Rad) filled with a 50 g of Amberlite XAD16N hydrophobic polyaromatic resin (Sigma-Aldrich), which was previously equilibrated with deionized water. After sample loading, the column was washed with 500 mL deionized water. Elution was performed with 250 mL of 100% methanol. The eluate was collected, diluted with deionized water (ratio 1:3) and lyophilized in a freeze-dryer. Pellets were dissolved in 100 mL of 50 mM lactic acid buffer (pH 4.5) and filtered through a 0.45 µm filter. Then, it was loaded on an NGC system (Bio-Rad) equipped with a HiTrap SP HP 5 mL cation exchange column (GE Healthcare Life Sciences), which was previously equilibrated with 50 mM lactic acid buffer (pH 4.5). The column was then washed with 50 mM lactic acid buffer (pH 4.5) and elution performed with 50 mM lactic acid buffer containing 300 mM NaCl (pH 4.5). The eluate was mixed with trifluoroacetic acid (TFA) to the end concentration of 0.1% and loaded on an RP-HPLC system (Agilent) equipped with a Jupiter Proteo, C-12, 250 × 10 mm column (Phenomenex). The column was equilibrated in 5% of solvent B (solvent A = MiliQ water with 0.1% TFA, solvent B = ACN with 0.1% TFA). Bacteriocin was eluted by an increase of solvent B up to 60% over 80 min with a flow rate of 2 mL/min. Elution fractions were tested for antibacterial activity against P. genomospecies 1 NUB36187 using a drop on a lawn assay. Active fractions were lyophilized, pellets dissolved in a solution containing 6 M guanidine-HCl and 0.1% TFA, and applied on a Jupiter Proteo, C-12, 250 × 4.6 mm column (Phenomenex) which was equilibrated in 5% of solvent B. Bacteriocin was eluted by an increase of solvent B up to 60% over 80 min with a flow rate of 1 mL/min. Elution fractions were tested for antibacterial activity against P. genomospecies 1 NUB36187 using a spot on lawn assay. Samples of elution fractions with antibacterial activity were analyzed by MALDI-TOF-MS, and the rest were lyophilized and stored at −80 °C until further use. Cloning for gene coexpression The gene palA was amplified by PCR using F-PalA-BspHI and R-PalA-HindIII primers, the gene palS was amplified by PCR using F-PalS-In-Fusion and R-PalS-In-Fusion primers. The gene palT was amplified by PCR using F-PalT-In-Fusion and R-PalT-In-Fusion primers. The DNA region containing two genes palS and palT was amplified by PCR using F-PalS-In-Fusion and R-PalT-In-Fusion primers. A. pallidus 8 genomic DNA was used as the template to amplify the genes. Amplified palA was double digested with BspHI and HindIII. pRSFDuet-1 vector was double digested with NcoI and HindIII. Both digestion products were cleaned and ligated by a conventional cloning. The ligation mixtures were transformed to E. coli TOP10 cells by electroporation. Positive colonies were selected by colony PCR using F-pRSFDuet-1 and R-pRSFDuet-1 primers. Positive clones were propagated for the plasmid isolation. The sequences of the inserts in the constructs were confirmed by DNA sequencing using F-pRSFDuet-1 and R-pRSFDuet-1 primers. A site-directed mutagenesis approach was used to introduce the His7-tag sequence (GGHHHHHHH) in the C-terminus of the PalA peptide. A new construct pRSFDuet-1 encoding palA-his was generated by PCR amplification using 5′-phosphorylated primers F-PalA-His and R-PalA-His. Construct pRSFDuet-1- palA was used as template. The PCR product was cleaned, ligated and transformed to E. coli TOP10. Positive clones were confirmed by colony PCR and propagated for plasmid isolation. The His7-tag encoding sequence in the resulting construct (pRSFDuet-1- palA-his ) was confirmed by DNA sequencing using F-pRSFDuet-1 and R-pRSFDuet-1 primers. To introduce a Factor Xa cleavage site and a His-tag (N-terminus) in the PalA peptide, a his-Xa-palA gene was engineered. First, the gene palA was amplified by PCR using F-PalA-BamHI and R-PalA-HindIII 2 primers and A. pallidus 8 genomic DNA as the template. PCR product palA and vector pRSFDuet-1 were double digested with BamHI and HindIII according to the manufacturer’s recommendations. Resulting products were cleaned, ligated by conventional cloning and transformed to E. coli TOP10 by electroporation. Positive clones were selected by colony PCR using F-pRSFDuet-1 and R-pRSFDuet-1 primers and propagated for isolation of plasmids. The sequence of the insert was confirmed by DNA sequencing using F-pRSFDuet-1 and R-pRSFDuet-1 primers. The insert, palA gene, in the resulting construct pRSFDuet-1-his-palA was introduced behind the His6-tag encoding sequence (MGSSHHHHHHSQDP). Next, a site-directed mutagenesis approach was used to incorporate a Factor Xa proteolytic cleavage site to the N-terminal part of the PalA peptide (Supplementary Fig. 13 ). Four wild-type peptide residues LQGS were changed to IEGR. pRSFDuet-1 coding for his-Xa-palA was amplified by PCR using 5′-phosphorylated primers F-PalA-Xa and R-PalA-Xa and template pRSFDuet-1-his-palA, then ligated and transformed to E. coli TOP10 by electroporation. Positive clones were selected by colony PCR followed by plasmid isolations. Correct sequence of constructed pRSFDuet-1-his-Xa-palA vector was confirmed by DNA sequencing using F-pRSFDuet-1 and R-pRSFDuet-1 primers. Sequences of the primers are provided in Supplementary Table 8 . Synthetic genes of sublancin 168 ( sunA-his ), glycocin F ( gccF-his ), enterocin F4-9 ( enfA4-9-his ), hypothetical peptide 1 ( hyp1-his ), and hypothetical peptide 2 ( hyp2-his ) were synthesized by GenScript with codon optimization for E. coli and delivered in a pUC57 vector. The genes encode glycocin precursors with leaders and His6-tag sequences (HHHHHH) in the C-terminuses of the peptides. All synthesized genes were cloned into the pRSFDuet-1 vector and transformed to E. coli TOP10. The presences of the inserts in the vector pRSFDuet-1 were confirmed by PCR (Supplementary Fig. 19 ). To amplify the inserts, the following primer pairs were used: sunA-his (primers F-SunA-BamHI and R-SunA-HindIII); hyp1-his (primers F-Hyp1 and R-Hyp1); hyp2-his (primers F-Hyp2 and R-Hyp2); gccF-his (primers F-GccF and R-GccF); enfA4-9-his (primers F-EnfA4-9 and R-EnfA4-9). Constructed pRSFDuet-1 vectors coding for the cloned genes were used as the templates, respectively. Sequences of the inserts in the plasmids were also confirmed by DNA sequencing using F-pRSFDuet-1 and R-pRSFDuet-1 primers. Sequences of the primers are provided in Supplementary Table 8 . A site-directed mutagenesis approach was used to engineer core_hyp1-his and core_hyp2-his genes coding for Hyp1-His and Hyp2-His core peptides without leader sequences and with a His6-tag (HHHHHH) in the C-terminuses (Supplementary Fig. 13 ). pRSFDuet-1-core_hyp1-his was amplified by PCR using 5′-phosphorylated primers F-Hyp1-Leaderless and R-Leaderless. pRSFDuet-1-core_hyp2-his was amplified by 5′-phosphorylated primers F-Hyp2-Leaderless and R-Leaderless. pRSFDuet-1-hyp1-his and pRSFDuet-1-hyp2-his vectors were used as templates, respectively. Obtained PCR products were ligated and transformed to E. coli TOP10 by electroporation. Positive clones were selected by colony PCR using F-pRSFDuet-1 and R-pRSFDuet-1 primers and propagated for isolation of plasmids. The presences of the genes coding for leaderless peptides in the vectors were confirmed by PCR (Supplementary Fig. 19 ). We amplified the insert core_hyp1-his by primers F-Hyp1-Leaderless and R-Hyp1-HindIII, and pRSFDuet-1-core_hyp1-his vector as template. The insert core _ hyp2-his was amplified by primers F-Hyp2-Leaderless and R-Hyp2, and pRSFDuet-1-core_hyp2-his vector as template. In addition, the inserts encoding Hyp1-His and Hyp2-His core peptides were confirmed by DNA sequencing using F-pRSFDuet-1 and R-pRSFDuet-1 primers. Sequences of the primers are provided in Supplementary Table 8 . To generate pBAD24 vectors coding for genes palS , palT and palST , the pBAD24 vector was double digested with NcoI and PstI, separated by DNA electrophoresis and extracted from agarose gel using a NucleoSpin Gel and PCR Clean-up Extraction Kit (Macherey−Nagel). Next, PCR-amplified palS , palT and palST were ligated with the double digested pBAD24 vector using a Quick-Fusion Cloning Kit (Bimake) according to the manufacturer’s recommendations. After the ligation, the mixtures were diluted with MiliQ in ratio 1:5 and transformed to E. coli TOP10 by electroporation. Positive clones were selected by colony PCR using F-pBAD24 and R-pBAD24 primers and propagated for isolation of plasmids. The presences of the cloned inserts in the constructs were confirmed by PCR (Supplementary Fig. 12 ). We amplified the insert palA-his by primers F-PalA-BspHI and R-PalA-HindIII, and pRSFDuet-1-palA-his vector as template. The insert palS was amplified by primers F-PalS-In-Fusion and R-PalS-In-Fusion, and pBAD24-palS vector as template. The insert palT was amplified by primers F-PalT-In-Fusion and R-PalT-In-Fusion, and pBAD24-palT vector as template. The insert palST was amplified by primers F-PalS-In-Fusion and R-PalT-In-Fusion, and pBAD24-palST vector as template. Sequences of the inserts in the resulting constructs (pBAD24- palS , pBAD24- palT and pBAD24- palST ) were also confirmed by DNA sequencing. We used F-pBAD24, R-pBAD24, and Pal1 primers for pBAD24- palS sequencing; F-pBAD24, R-pBAD24, Pal3, Pal4, and Pal5 primers for pBAD24- palT sequencing; F-pBAD24, R-pBAD24, Pal1, Pal2, Pal3, Pal4, and Pal5 primers for pBAD24- palST sequencing. Genes of precursors were cloned in MCS1 behind the IPTG promoter and RBS of the pRSFDuet-1 vector. Genes palS/palT/palST were cloned into the MCS behind the arabinose promoter and RBS of the pBAD24 vector. Sequences of primers are provided in Supplementary Table 8 . The vectors pRSFDuet-1 and pBAD24 coding for palS/palT/palST were transformed to E. coli BL21(DE3) for protein expression experiments. Coexpression of glycocin precursors with PalS and PalT E. coli BL21(DE3)-containing vectors pRSFDuet-1- palA-his and pBAD24 coding for either palS, palT or palST were grown at 37 °C overnight in LB medium containing ampicillin (50 µg/mL) and kanamycin (30 µg/mL). The next day it was inoculated into 100 mL of fresh LB−ampicillin−kanamycin medium in the ratio 1:100 and grown at 37 °C to the OD (600 nm) of 0.6−0.7. Arabinose and IPTG were added to a final concentration of 1 mM of each and the culture was grown for additional 4 h at 37 °C. After the induction, cells were harvested by centrifugation at 7000 × g for 15 min at 4 °C and resuspended in 5 mL of binding buffer (20 mM NaH 2 PO 4 , 500 mM NaCl, pH 7.4). pRSFDuet-1 vectors with cloned glycocin precursor genes: sunA-his/gccF-his/enfA4-9-his/hyp1-his/hyp2-his , were coexpressed in E. coli BL21(DE3) together with the pBAD24 vector coding for palS/palT/palST , respectively. The pRSFDuet-1 vector coding for his-Xa-palA was coexpressed in E. coli BL21(DE3) together with the pBAD24 vector coding for palS . The coexpression was performed in the same approach as the coexpression of pRSFDuet-1-palA-his with pBAD24-pasS. pRSFDuet-1 vector encoding core_hyp1-his/core_hyp2-his was coexpressed with pBAD24 vector encoding palS in E. coli BL21(DE3). The coexpression was induced for 7 h by the same approach as was coexpression of pRSFDuet-1-palA-his with pBAD24-pasS. Purification of the peptides The obtained cell suspension after coexpression was sonicated on ice for 20 min using a VCX 130 Sonicator with cycle 10 s ON and 10 s OFF. The highest concentration of the produced peptide was found in the insoluble fraction of lysed cells. Cell debris was removed by centrifugation at 15,000 × g for 20 min at 4 °C. The supernatant was discarded and the pellet of insoluble fraction was resuspended and sonicated at the same conditions in 5 mL of binding buffer containing 6 M guanidine-HCl. The sample was filtered through a 0.45 μm filter and applied to an NGC system (Bio-Rad) equipped with a HisTrap FF 1 mL (GE Healthcare Life Sciences) immobilized metal affinity chromatography (IMAC) column pre-equilibrated with binding buffer containing 4 M guanidine-HCl. After sample application, the column was washed with binding buffer containing 4 M guanidine-HCl. The peptide was eluted with elution buffer (20 mM NaH 2 PO 4 , 500 mM NaCl, 500 mM imidazole, pH 7.4) containing 4 M guanidine-HCl. The fractions containing eluted peptides were further purified by an RP-HPLC system (Agilent) equipped with Jupiter Proteo, C-12, 250 × 10 mm column (Phenomenex). The eluate was mixed with TFA to reduce the pH to 2–3 and loaded on the column equilibrated in 5% of solvent B. The peptides were eluted by an increase of solvent B up to 60% over 60 min with a flow rate of 2 mL/min. All fractions were tested for antibacterial activity against P. genomospecies 1 NUB36187 using a drop on lawn assay. Additionally, elution fractions were analyzed by MALDI-TOF-MS and LC-ESI-MS. Leader cleavage of the pre-His-Xa-PalA-Glc peptide Lyophilized pellets of purified glycosylated pallidocin precursor pre-His-Xa-PalA-Glc were dissolved in 1 mL of 6 M guanidine-HCl and loaded on gel filtration column PD-10 (GE Healthcare Life Sciences). The buffer exchange of the sample was performed according to the manufacturer’s recommendations. The column was pre-equilibrated with 50 mM tris-HCl, 100 mM NaCl, pH 7.5 buffer. The eluate was collected and mixed with CaCl 2 to the end concentration of 2 mM. 10–20 µL of Factor Xa peptidase (enzyme concentration 1 mg/mL, NEB) was added to 0.5 mL of previously gel filtrated peptide solution (peptide concentration 0.5 mg/mL). The mixture was stored at room temperature for 3–6 h. After sample treatment with the peptidase, 4 mL of 6 M guanidine-HC and TFA, to quench the pH to 2–3, were added. Then, the sample was loaded on an RP-HPLC system (Agilent) equipped with a Jupiter Proteo, C-12, 250 × 4.6 mm column (Phenomenex) equilibrated in 5% of solvent B. The peptide was eluted by an increase of solvent B from 20% up to 60% over 80 min with a flow rate of 1 mL/min. Elution fractions were tested for antibacterial activity against P. genomospecies 1 NUB36187 using a drop on lawn assay. The elution fractions were analyzed by MALDI-TOF-MS. Elution fractions containing active and glycosylated pallidocin core peptide PalA-Glc were lyophilized by freeze-dryer. Pellets were stored at −20 °C or dissolved in MiliQ containing 50% ACN and 0.1% TFA solution for further use. The quantity of purified peptide was measured by a NanoPhotometer N60 (Implen). The molar absorptivity (extinction coefficient) was calculated (at 280 nm and 13,200/M/cm or at 205 nm and 171,330/M/cm) based on the peptide sequence. Calculations were performed by a web tool, provided at 48 . Determined peptide quantities at 280 and 205 nm were the same. The yield of synthesized pallidocin from 200 mL of bacterial culture was ~30 µg. Iodoacetamide assays for detection of free cysteines To detect the presence of free cysteine thiols in peptides, an iodoacetamide (IAA) assay was used. For the detection of free Cys residues in the native peptide, reactions contained 100 mM tris-HCl (pH 8.3), 40 mM IAA and the peptide. For the detection of free Cys residues in a reduced peptide, reactions contained 100 mM tris-HCl (pH 8.3), 10 mM TCEP, 40 mM IAA and the peptide. All reactions were in total volume of 1 mL, incubation conditions were 2 h at room temperature in the dark. The reaction mixtures were quenched with TFA to pH < 4. Samples were loaded on an RP-HPLC system (Agilent) equipped with a Jupiter Proteo, C-12, 250 × 4.6 mm column (Phenomenex) equilibrated in 5% of solvent B. Peptides were eluted by an increase of solvent B from 20% up to 60% over 40 min with a flow rate of 1 mL/min. Elution fractions containing peptides were tested for antibacterial activity against P. genomospecies 1 NUB36187 using a drop on lawn assay and further analyzed by MALDI-TOF-MS. The detection of free cysteines was determined by the presence or absence of thiol modifications, carboxyamidomethyl (CAM). Effects of pH on bacteriocin stability 50 mM buffer solutions of KCl-HCl pH 2, citric acid-sodium citrate pH 4, phosphate pH 6, tris-HCl pH 8 and sodium carbonate-sodium bicarbonate pH 10 were prepared for the following assay. Pallidocin was dissolved in MiliQ water containing 50% ACN and 0.1% TFA to the end concentration of 1 ng/µL. 13.5 µL of each buffer solution was mixed with 1.5 µL of pallidocin solution and the resulting mixtures were stored for 3 h at room temperature. After the incubation, 15 µL of 500 mM tris-HCl pH 7.5 buffer solution and 120 µL of NB medium were added to the mixtures. Next, serial twofold dilutions with NB medium were made for each mixture. Bacteriocin activity was tested for each sample with agar well diffusion assay. Effect of temperature on bacteriocin stability Pallidocin was dissolved in MiliQ water containing 50% ACN and 0.1% TFA to the end concentration of 1 ng/µL. One hundred and fifty microliters of NB medium was mixed with 1.5 µL pallidocin solution and stored at room temperature for 24 h, 10 days and 30 days. In addition, samples were stored at different temperatures: 60 °C, 90 °C for 3 h and autoclaved for 15 min at 121 °C. After the incubation, serial twofold dilutions in NB medium were made for each mixture. Bacteriocin activity was tested for each sample with agar well diffusion assay. Determination of minimum inhibitory concentrations MICs were determined as described by Wiegand et al. 50 with some modifications. One colony of a sensitive strain was picked from an NB agar plate, inoculated to liquid NB medium and grown at 55 °C, in a shaking incubator until OD (600 nm) of 0.1 was reached. Then, the culture was diluted with NB medium till concentration of 10×10 5 CFU/mL. One hundred and fifty microliters of fresh NB medium was mixed with 5 µL of pallidocin solution (1 ng/µL in 50% ACN and 0.1% TFA) and serial twofold dilutions with NB medium were made. Seventy-five microliters of resulting diluted pallidocin mixtures were transferred to a 96-well plate and mixed with 75 µL previously prepared cell suspension of sensitive strain. The final volume of the mixture in the well was 150 µL and the end concentration of the sensitive strain was 5×10 5 CFU/mL. Positive controls, 150 µL mixture of NB medium with the sensitive strain (5×10 5 CFU/mL), and negative controls, 150 µL mixture of NB medium with 5 µL of pallidocin solution (1 ng/µL in 50% ACN and 0.1% TFA), were prepared and dispersed in the same 96-well plates. The plate with a lid was placed in a plastic box (12 cm × 20 cm × 6 cm) with a wet paper towel to keep high humidity and prevent medium evaporation at high temperature. The plate was incubated for 18 h at 55 °C in a shaking incubator. After incubation the growth of bacteria was evaluated visually and by a plate reader. The analyses were performed in triplicate. It should be noted that because of the high mutation rate and emergence of resistant mutants in some wells in the plate, the calculation of average MIC from three replicates are prone to variation. The final MIC value was determined by the lowest amount of pallidocin required to inhibit cell growth in a well. Because bacteria were grown at 55 °C, it was not possible to use a plate reader at this condition, as the instrument is not suited for measurements at high temperatures. Determination of the sugar modification of pallidocin The presence of a sugar moiety on Cys25 of pallidocin core peptide was confirmed via acid-catalyzed methanolysis and derivatization of the sugar, analysis by gas chromatography-mass spectrometry (GC-MS), and comparison to derivatized sugar standards. Purified pallidocin was lyophilized in a glass tube, internal standard-mannitol and 0.5 mL of methanolic 1 M HCl (Sigma-Aldrich) were added. The mixture was heated at 85 °C overnight. The reaction was cooled at room temperature and neutralized by adding solid Silver Carbonate (Sigma-Aldrich) till pH 7. For re-N-acetylation, two drops of Acetic Anhydride (Sigma-Aldrich) was added, mixed and stored overnight at room temperature in the dark. Next day, the reaction mixture was centrifuged at 1000 × g for 2 min. Supernatant was transferred to a new glass tube. 0.5 mL of methanol was added to the silver salt pellets, mixed and centrifuged again. Supernatant was collected and pooled with the first one. The procedure was repeated twice. Collected supernatant was evaporated in a vacuum evaporator. Tri-methylation of the sugar was performed by adding 0.3 mL of silylation reagents (pyridine:hexamethyldisilazane:trimethyl-chlorosilane = 5:1:1). The mixture was incubated at room temperature for 30 min. Sugar standards mix : d -mannose, d -galactose, d -glucose, N -acetylgalactosamine and N -acetylglucosamine were also treated and derivatized using the conditions described above. The derivatized sugar of pallidocin and sugar standards mix were analyzed individually by a GCMS-QP2010 Plus system (Shimadzu) equipped with a Zebron ZB-1HT column, L = 30 m × I.D. = 0.25 mm × df = 0.25 μm (Phenomenex). The temperature gradient was from 140 to 240 °C at 4 °C per min. The carrier gas was helium and the flow rate set at 2.0 mL/min. Circular dichroism spectroscopy Circular dichroism (CD) spectra were recorded using a J-815 CD Spectrometer (JASCO) with a cuvette of 0.1 cm path length in the range from 190 to 260 nm at 0.5 nm intervals. For CD spectrum measurement lyophilized pallidocin was dissolved in a solvent (40% ACN, 0.1% TFA) to a final concentration of 9 µM. The spectrum of scan was obtained with a 2 nm optical bandwidth. The baseline scans were collected with the solvent alone and then subtracted from the sample scan. The estimate of secondary structure content was made using spectra from 193 nm to 260 nm by the method of Raussens et al. 42 at . Code availability The BAGEL4 38 server ( ) was used for bacteriocin mining in the genome sequences. The strain was identified by submitting its genomic DNA sequence to a Genome-to-Genome Digital Calculator (GGDC) 36 , 49 for digital DNA−DNA hybridization (dDDH) analysis at . BLAST analyses of peptide and protein sequences were performed using the NCBI database at . Calculations of extinction coefficients for peptide’s absorbance at 280 and 205 nm were performed by a web tool Protein Calculator 51 at . Theoretical peptide mass calculations were performed using a web tool ProteinProspector (MS-Product/MS-Isotope) at . The estimate of secondary structure content based on far-UV CD spectroscopy was made by the method of Raussens et al. 42 at . The estimate and modeling of secondary structures based on amino acid sequences were made by PSIPRED 22 web tool at . Reporting summary Further information on experimental design is available in the Nature Research Reporting Summary linked to this article. Data availability Data supporting the findings of this work are available within the paper and its Supplementary Information files . A reporting summary for this article is available as a Supplementary Information file . The source data underlying Supplementary Figures 1 , 11 , 12 , 19 and Supplementary Tables 5 − 7 are provided as a Source Data file. All other data are available from the corresponding author on request.

University of Groningen microbiologists and their colleagues from Lithuania have discovered a new glycocin, a small antimicrobial peptide with a sugar group attached, which is produced by a thermophilic bacterium and is stable at relatively high temperatures. They also succeeded in transferring the genes required to produce this glycocin to an E. coli bacterium. This makes it easier to produce and investigate this compound, which could potentially be used in biofuel production. These findings were published in Nature Communications on 7 March. The rise of antibiotic resistance has spurred the search for new antimicrobials. Bacteriocins—peptide toxins produced by bacteria to inhibit growth in similar or related bacterial strains—are a possible alternative to the more traditional antibiotics. Bacteriocins would also be useful to protect high-temperature fermentations mediated by thermophilic bacteria. But this would require the use of bacteriocins that are stable at higher temperatures. Mystery "That is why we were interested to find that the thermophilic bacterium Aeribacillus palladius, isolated from the soil above an oil well in Lithuania, appeared to produce an antibacterial peptide," says University of Groningen Professor of Molecular Biology Oscar Kuipers. Thus far, purification and identification of the compound had not been successful. Therefore, Ph.D. student Arnoldas Kaunietis from Vilnius University spent almost two years in Kuipers' lab to solve the mystery. He is the first author on the new paper. By analyzing genomic information from the Lithuanian bacteria using BAGEL4 software developed by Anne de Jong and Auke van Heel in Kuipers' group, the researchers discovered the genes that are responsible for the production of the bacteriocin, and the final gene product was named pallidocin. The BAGEL4 software searches for gene clusters with the potential ability to produce novel antimicrobials. Sugar The antimicrobial turned out to be a glycocin belonging to a class of post-translationally modified peptides. This means that after its production, one or more functional groups are added to the peptide. In the case of glycocins, this functional group is a sugar. "Only five other glycocins were known thus far," says Kuipers. In order to facilitate further research and engineering of this peptide, the genes responsible for the production of pallidocin were transferred to E. coli BL21 (DE3) bacteria. "The expression of the genes worked well, which is a real breakthrough, as it is difficult to express a whole antimicrobial gene cluster from a gram-positive bacterial strain directly in a gram-negative bacterium and to get the product secreted." Biofuel After isolating pallidocin, the scientists were able to confirm that it is highly thermostable and exhibits extremely strong activity against specific thermophilic bacteria. Furthermore, by using the sequence of pallidocin biosynthesis genes in BAGEL4, two similar peptides were discovered in two different strains of Bacillus bacteria. These peptides, named Hyp1 and Hyp2, were also successfully expressed in the E. coli strain. "This shows that the expression system works well for various glycocins; it is able to produce them in vivo," says Kuipers. Pallidocin might be useful in high-temperature fermentations, which are used to produce biofuels or chemical building blocks. The higher temperature makes it easier to recover volatile products such as ethanol but also reduces the risk of contamination with common bacteria. However, contamination with thermophilic bacteria is possible. "Both pallidocin and Hyp1 appear to be active against thermophilic bacteria and some Bacillus species," says Kuipers. And there could be more applications: "Contamination by thermophiles is also a problem in the food industry."

10.1038/s41467-019-09065-5

Biology

Great tits living in cities are genetically different from great tits in the countryside

Pablo Salmón et al, Continent-wide genomic signatures of adaptation to urbanisation in a songbird across Europe, Nature Communications (2021). DOI: 10.1038/s41467-021-23027-w Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-021-23027-w

https://phys.org/news/2021-05-great-tits-cities-genetically-countryside.html

Abstract Urbanisation is increasing worldwide, and there is now ample evidence of phenotypic changes in wild organisms in response to this novel environment. Yet, the genetic changes and genomic architecture underlying these adaptations are poorly understood. Here, we genotype 192 great tits ( Parus major ) from nine European cities, each paired with an adjacent rural site, to address this major knowledge gap in our understanding of wildlife urban adaptation. We find that a combination of polygenic allele frequency shifts and recurrent selective sweeps are associated with the adaptation of great tits to urban environments. While haplotypes under selection are rarely shared across urban populations, selective sweeps occur within the same genes, mostly linked to neural function and development. Collectively, we show that urban adaptation in a widespread songbird occurs through unique and shared selective sweeps in a core-set of behaviour-linked genes. Introduction Urban development is rapidly expanding across the globe, and although urbanisation is regarded a major threat for wildlife 1 , its potential role as an evolutionary driver of adaptation has not been explored until recently 2 , 3 , 4 , 5 . Many species show phenotypic adaptations to the multiple urban challenges, such as higher levels of noise, artificial light at night, air pollution, altered food sources or habitat fragmentation 6 . These adaptations include changes in behaviour, e.g., refs. 7 , 8 , 9 , morphology and locomotion, e.g., refs. 10 , 11 , 12 , 13 or toxin tolerance 14 . Indeed, there is now evidence that the phenotypic divergence between urban and rural populations may have a genetic basis in some species 14 , 15 , in line with the finding that micro-evolutionary adaptations in natural populations can occur within short timescales, particularly in response to human activities 16 , 17 , 18 . The study of the genetic signals of adaptation could provide important insights into the magnitude of the evolutionary change induced by urbanisation on wildlife. However, the short evolutionary timescale, the dependence of evolution on local factors, and the polygenic nature of many phenotypic traits, make detecting the genomic signals of adaptation difficult 17 , 19 . The majority of studies on the genomic basis of urban adaptation have either focused on a limited number of markers and genes 3 , 20 or on a narrow geographical scale, e.g., refs. 4 , 15 , 21 , 22 , thereby limiting the inferences that can be made on the consistency of genomic responses to urbanisation 5 . Conversely, the study of multiple populations on a broad geographical scale provides a powerful framework to identify the evolutionary forces shaping the genomic responses and to test the repeatability of urban adaptation 15 , 23 , 24 . In particular, comparing across populations enables the distinction between random demographic non-adaptive processes, such as genetic drift, and signatures of natural selection, as selection might be expected to affect the same genomic regions (nucleotides or genes) and/or functional pathways across cities 23 , 25 . The current evidence on the genomic basis of urban adaptations ranges from high parallelism, involving a single or few genes 15 , 23 , to polygenic, and to putative genetic redundancy, e.g., ref. 22 . This discrepancy might be due to the variable and sometimes local urban selection pressures together with the polygenic nature of many adaptive phenotypic traits, which could lessen the likelihood for shared genetic changes at the nucleotide level 26 . In this study, we present a multi-location analysis of the evolutionary response to urbanisation, using the great tit ( Parus major ), to identify the genetic basis of urban adaptation. The great tit is a widely distributed songbird and a model species in urban, evolutionary and ecological research, e.g., refs. 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , with demonstrated phenotypic changes in response to urban environments in several populations 29 , 34 , 35 , 37 . In addition, genomic resources are well developed for this species 38 and it is known that across its European range, the species presents low genetic differentiation 21 , 39 , 40 , 41 . In order to examine genomic responses to urbanisation on a broad geographical scale, we analyse pairs of urban and rural great tit populations from nine localities across Europe (Fig. 1 and Supplementary Table 1 ). We combine several complementary approaches to: (1) detect allele frequency shifts across many loci, which will facilitate the exploration of the polygenic aspect in the adaptation to urbanisation; (2) search for selective sweeps in urban populations, which will provide information on population-specific signatures of selection; and (3) identify enriched functional pathways associated with genes putatively under selection in urban populations, which will help us to infer which particular phenotypes, known and unknown, underlie urban adaptation in great tits. Overall, the present study deepens our understanding of the evolutionary drivers and forces shaping the genomic landscape of urban adaptation on a continental scale. Fig. 1: Study sampling locations and urbanisation intensity. a Centre: map of Europe, showing the targeted cities where the sampling of great tits ( Parus major ) was carried out. Red areas indicate the main dense urban areas. The nine inset zooms flanking the central map exemplify the landscape and degree of urbanisation for each of the urban–rural sampling locations. Green shading represents canopy or green areas (i.e., vegetation cover) and roads and buildings are represented in grey. See “Methods”, Supplementary Fig. 1 and Supplementary Table 1 for more details. Map data sources: Stamen Design under CC BY 3.0 ( ); data licensed by OpenStreetMap Foundation (OSMF) under ODbL ( ). Shapefiles: Natural Earth. b Urbanisation scores (principal component, PC urb ) for all nine urban–rural pairs. In both figures, the coloured circles represent each city (in b , urban: open circles; rural: closed circles). BCN Barcelona, GLA Glasgow, GOT Gothenburg, LIS Lisbon, MAD Madrid, MAL Malmö, MIL Milan, MUC Munich, PAR Paris. Great tit illustration made by Pablo Salmón. Source data for b is provided as a Source data file. Full size image Results and discussion Genetic diversity and population structure across European urban and rural populations A total of 192 great tits from the nine paired urban–rural populations were genotyped at 517,603 filtered SNPs, with 10–16 individuals per sampling site (Supplementary Table 1 ). We quantified the relative degree of urbanisation for each site (urbanisation score: PC urb , from principal component analysis, PCA; see “Methods”, Fig. 1b , Supplementary Fig. 1 and Supplementary Table 1 ) to inform our genetic downstream analyses. Population structuring based on 314,351 LD (linkage-disequilibrium)-pruned SNPs (excluding small linkage groups and the Z-chromosome) was overall low across the 18 studied sites (Supplementary Fig. 2 ), with each of the first two principal components explaining <3% of the overall variation across populations (Fig. 2a and Supplementary Fig. 3a ). Thus, we used a UMAP (Uniform Manifold Approximation and Projection) approach to summarise the genetic variation along the first 20 PC axes. The UMAP analysis revealed the presence of distinct genetic clusters for some of the localities, although there was still a strong clustering of individuals from Gothenburg, Munich, Milan and Paris (Fig. 2a and Supplementary Fig. 3b ; results were comparable when including the Z-chromosome, Supplementary Fig. 3c ). This analysis also suggested that the levels of divergence differ strongly between each urban and adjacent rural population (Fig. 2a and Supplementary Fig. 3b ). Furthermore, the population pairs from Glasgow and Lisbon showed the highest levels of divergence, with Lisbon and Glasgow separating along PC1 and PC2, respectively (Supplementary Fig. 3a ), and also separating strongly in the UMAP plot (Fig. 2a ). This increased divergence of Glasgow and Lisbon, both located in the range edge of the great tit distribution, could be explained by their slightly reduced heterozygosity, particularly in the urban populations (Supplementary Table 1 ). Indeed, population tree analyses using TreeMix supported the presence of increased drift in both urban populations (Fig. 2b ). However, overall, we did not find lower heterozygosity levels in any of the nine-urban compared to the rural populations, which suggests that urban colonisation was in general not associated with significant bottlenecks (see Supplementary Table 1 for details; Wilcoxon test: W = 30, P = 0.377). Fig. 2: Great tit population structure. a UMAP reduction of the first 20 principal component axes for an LD-pruned SNP dataset, excluding the Z-chromosome and small linkage groups. The insert shows the percent of variance explained (PVE) by each of the first 50 PC axes. The corresponding principal component plots can be found in Supplementary Fig. 2 . b Population tree generated with TreeMix showing the relationship of urban (open circles) and rural (closed circles) populations from all sampling locations. c Comparison of F ST value distributions (mean ± SD) across population and habitats. The effect size plots (mean ± 95% bootstrap CI) show that overall rural populations (rural vs rural) display lower differentiation than the studied population pairs (urban vs rural), but not lower than urban populations (urban vs urban). d Isolation-by-distance analyses (Mantel’s test) for urban (open circles, light shaded, n = 72) and rural (closed circles, dark shaded, n = 72) populations separately. Shaded area denotes the 95% CI. e Effective migration surfaces for great tits across Europe. Negative log migration rates [log( m )] depict areas with less gene flow than expected under an isolation-by-distance model, whereas positive migration rates indicate stronger gene flow. Note that migration rates outside the central part of the sampling distribution are generally low, also between closely related cities. BCN Barcelona, GLA Glasgow, GOT Gothenburg, LIS Lisbon, MAD Madrid, MAL Malmö, MIL Milan, MUC Munich, PAR Paris. Colour code is given in panel a . Source data are provided as a Source data file. Full size image Interestingly, the population structure analysis indicated that multiple urban populations, namely Glasgow, Lisbon, Madrid, Malmö, Milan, and to weaker extent Barcelona and Gothenburg, formed distinct genetic clusters and independent drift units together, in some cases, with their adjacent rural counterparts (Fig. 2a, b ). Nonetheless, the TreeMix analysis also suggests the presence of migration events across some populations, including long distance migration events, e.g., from Lisbon to Glasgow (Supplementary Fig. 4b ). Although, there was no clear pattern in relation to predominant gene flow between urban populations (Supplementary Fig. 4 ). In contrast to the mentioned population clusters, we did not detect any particular signs of genetic divergence between populations from Munich and Paris (neither PC or UMAP axis see, e.g., Fig. 2a and Supplementary Fig. 3 ). Likewise, we detected significant genetic differentiation between all pairwise comparisons ( P < 0.05; F ST permutation test; see “Methods”) except for the urban populations in Munich and Paris (Fig. 2e and Supplementary Table 2 ). Moreover, the genetic population differentiation ( F ST ) within population pairs (urban vs adjacent rural population), as well as between urban populations (urban vs urban), was on average higher than the differentiation between rural populations (rural vs rural) across Europe (Fig. 2c ). These results support the idea that gene flow is generally stronger between rural compared to urban habitats in the nine population pairs, and indicates reduced gene flow between urban and adjacent rural populations. This is further supported by the observed isolation-by-distance patterns, which showed slightly stronger IBD for urban (Mantel’s r = 0.46, P = 0.008) compared to rural populations ( r = 0.43, P = 0.007; Fig. 2d ). Lastly, analysis of the effective migration surfaces confirmed reductions in gene flow in many parts of Europe, compared to neutral expectations, including those in close proximity, e.g., Gothenburg and Malmö, and simultaneously highlighted strong gene flow in Central/Western Europe, i.e., Munich and Paris (Fig. 2e ). Overall, our results reveal weak population structuring across Europe, with slightly increased genetic differentiation between urban populations compared to the more admixed rural populations. Previous results in another European songbird, the blackbird ( Turdus merula ) showed that urban birds likely independently colonised multiple times European cities from forest sources 42 . In our study, the finding of independent clusters for some urban populations also point towards a similar scenario, i.e., an independent and repeated colonisation of urban habitats from largely admixed rural populations. However, the overall weak genetic divergence and the detection of significant migration events across distant populations suggests a possible role of gene flow in the facilitation of urban adaptation, via the spread of adaptive alleles. Still, in our subsequent analyses we treated all our urban–rural population pairs as independent, including Munich and Paris, as selection pressures might differ across cities and result in idiosyncratic genomic responses to selection despite any ongoing gene flow. Identification of urbanisation-associated allele frequency shifts To identify genomic regions with consistent allele frequency shifts associated with urbanisation, we used two complementary genotype–environment association (GEA) approaches, LFMM (latent-factor mixed models) and an additional Bayesian approach (using BayPass). Testing for GEAs with LFMM (using PC urb as a continuous habitat descriptor, Fig. 1b ) revealed 2758 SNPs associated with urbanisation (0.52% of the full SNP dataset, false-discovery rate (FDR) < 1%; Fig. 3a ). Urbanisation-associated SNPs were widely distributed across the genome and did not cluster in specific regions. Larger chromosomes ( R 2 = 0.97) and those with more genes ( R 2 = 0.90; Fig. 3b ) contained more urbanisation-associated SNPs, highlighting the polygenic nature of urban adaptation. A PCA (PC GEA ), based on all urban-associated SNPs detected by LFMM, clearly separated urban and rural populations along PC1 GEA (proportion of variance explained—PVE—by PC1 = 1.98%; Supplementary Fig. 5c ), suggesting consistent allele frequencies in those loci across European cities. Fig. 3: Genome-wide association with urbanisation. a Manhattan plot showing the genotype–urbanisation (PC urb ) association estimated with LFMM across the genome. The red and blue dotted lines show the 0.1 and 1% FDR significance thresholds, respectively. Red dots highlight “core urbanisation SNPs” that were also identified as urbanisation associated using BayPass (Supplementary Fig. 6 ). The heatmap below the Manhattan plot highlights the SNP density across the genome. b Correlation between the number of urbanisation-associated SNPs (from LFMM) and the number of genes per chromosome (linear model: F 1,30 = 284.60, P < 2.2 × 10 −16 ). Shaded area denotes the 95% CI. c Main axis of variation in a PCA GEA based on “core urbanisation SNPs”. Grey circles show individuals, and yellow dots and lines show the mean ± SD for urban and rural populations, n urban = 96 and n rural = 96 independent birds. d Effect sizes (partial η 2 ) for the effects of “habitat” (urban vs rural) and “habitat × locality” interaction on urban–rural allele frequency, using linear model are shown for all significant LFMM SNPs on the x -axis and y -axis, respectively. “Core urbanisation SNPs” are highlighted in yellow circles. SNPs that lie above the dashed line show strong consistent shifts in allele frequencies across localities, whereas SNPs below the line show more variable allele frequency shifts across localities. Source data are provided as a Source data file. Full size image In contrast, the results obtained by BayPass only identified 70 SNPs strongly associated with urbanisation (Bayes factor ≥ 20; Supplementary Fig. 6a ). The lower number of identified SNPs might respond to the stronger population structure correction applied by BayPass, which correlates to some extent with the direction of local adaptation 43 . Of these significant SNPs, 34 were shared with the LFMM analysis (more than expected by chance: χ 2 = 31.20, P = 2.35 × 10 −08 ; Fig. 3a and Supplementary Fig. 6b ). These shared SNPs, which we term “core urbanisation SNPs ” (Supplementary Table 3 ), are likely involved in the local adaptation of great tits to urban habitats, and indeed, they more strongly discriminated urban and rural individuals across Europe (PVE by PC1 GEA-shared = 11.7%; Fig. 3c and Supplementary Fig. 7a ). In order to gain a better understanding of the importance of habitat (i.e., urban vs rural) on allele frequency shifts in the detected SNPs, we assessed using univariate linear models, the explanatory power of habitat, locality (city), and their interaction on the direction and strength of allele frequency shifts per-SNP across all paired populations. A significant “habitat” term indicates consistent allele frequency shifts across cities, particularly when the effect size (partial η 2 ) is higher compared to the effect size of a significant “habitat × locality” interaction term, which describes differences in the direction and magnitude of the allele frequency change between cities 44 . Applying these models to each urbanisation-associated SNP from the LFMM model (2758 SNPs), we detected large variation in allele frequency shifts across localities (Fig. 3d and Supplementary Fig. 7b ), with a slightly larger proportion of SNPs showing differences in allele frequency shifts across localities (“habitat × locality” η 2 > “habitat” η 2 ). In contrast to this, most of the “core urbanisation SNPs” showed similar allele frequency differences across localities, with 76% of them showing a main effect of the urban habitat (Fig. 3d ), and with the same allele increased in frequency in seven or more urban populations (Supplementary Fig. 7c ; see Supplementary Fig. 8 for the minor allele frequency trajectory between habitats). We refined this analysis by accounting for the effect of allele frequency correlations between SNPs through the use of the principal component axis from all the LFMM urbanisation-associated SNPs (PC1 GEA ) and the “core urbanisation SNPs” (PC1 GEA-shared , Supplementary Fig. 7d ). In both cases, the “habitat” term explained the majority of the total variation in allele frequency divergence ( η 2 PC1 GEA, habitat = 0. 87, P < 0.001; η 2 PC1 GEA-shared, habitat = 0. 73, P < 0.001). In comparison, both the effect of “locality”, which corresponds to the distinct evolutionary history of the local populations ( η 2 PC1 GEA, locality = 0.59, P < 0.001; η 2 PC1 GEA-shared, locality = 0.20, P < 0.001), and the interaction of “habitat × locality” ( η 2 PC1 GEA, habitat × locality = 0.13, P = 0.002; η 2 PC1 GEA-shared, habitat × locality = 0.13, P = 0.001), explained much smaller proportions of the total variation (smaller effect size) than the habitat term by itself. In contrast, genetic variation along PC2 and PC3 was mostly inconsistent across localities and likely specific for each city (Supplementary Fig. 7d ). Overall, the PCA-based results are in line with those obtained by the per-SNP results, reinforcing the idea of consistent continent-wide responses to urbanisation, in particular regarding the “core urbanisation SNPs”. Together, we show that there is local adaptation to urban habitats in great tits, and that this occurred across Europe through shifts in allele frequency of the same loci. This might have occurred via the standing genetic variation observed for the species, as putatively adaptive alleles were shared across large parts of the species’ European distribution 41 , or the exchange of adaptive alleles through gene flow across urban populations. Nonetheless, in line with a polygenic basis of adaptation via subtle allele frequency shifts in several loci 26 , the detected urban–rural differences were generally small (Supplementary Fig. 9 ) with only a few SNPs showing relatively strong allele frequency shifts with |∆AF| (allele frequencies difference between urban and rural) >0.5. Signatures of selection in urban populations While GEA approaches are powerful tools for identifying even subtle allele frequency shifts associated with local adaptation 43 , they have difficulty detecting selective sweeps unique to one or few populations or sweeps of different haplotypes associated with the same gene. Selective sweeps have been associated with urban adaptation in other species, for example, in the case of New York City brown rats ( Rattus norvegicus ) and white-footed mice ( Peromyscus leucopus ) 4 , 45 . Therefore, we further performed genome-wide scans of differentiation and selective sweep analyses in each of our studied populations to test if these contributed to the urban adaptation in great tits and, importantly, to explore if selective sweeps were population specific or widely shared across Europe. Since the exchange of genetic material between urban and rural populations, and the selection pressures are likely recent and ongoing in great tits, we opted to use a cross-population statistic that can identify ongoing and recently completed hard and soft selective sweeps ( XP-nSL ) 46 . Using this approach, we found that between 436 and 700 genomic windows (200 kb sliding windows with 50 kb steps, see “Methods”), which clustered into 127–173 wider genomic regions, showed signatures of selection in urban populations (Fig. 4a ). Genomic outlier regions had an average size of 355.8 kb ± 251.6 SD, with the largest region (3.05 Mb) detected on chromosome 1 in Glasgow and were widely distributed across the genome with a few distinct population-specific peaks (Fig. 4a ). The large number of genomic outlier regions are in line with the high number of significant GEAs across the genome and confirms the inferred polygenic nature of urban adaptation in great tits (Fig. 3a ). However, in contrast to the highly concordant allele frequency shifts of GEA loci, XP-nSL values were in general weakly correlated across populations (Fig. 4b ), and outlier windows were shared across a maximum of five populations (Supplementary Fig. 10 ). The top three shared windows clustered in a 300 kb region on chromosome 11 (16.05–16.35 Mb) and were detected in Gothenburg, Munich and the Iberian Peninsula (Barcelona, Madrid and Lisbon), with one partially overlapping outlier window detected in Glasgow. In general, the most pronounced sharing of outlier windows did not seem to occur between the geographically closest populations, as expected for shared genetic variation or adaptive introgression (see Supplementary Fig. 10 ). For example, Gothenburg and Paris, and Lisbon and Glasgow, showed the highest number of shared outlier windows (53 and 47, respectively). Indeed, the number of shared outlier windows between urban populations was not correlated with their genetic (Mantel’s r = 0.28, P = 0.150) or geographic distance (Mantel’s r = 0.05, P = 0.400), suggesting that shared adaptive genetic variation through introgression and gene flow between neighbouring urban populations likely did not facilitate urban adaptation. However, we cannot fully exclude the role of gene flow as migration was generally strong, even across large distances (Supplementary Fig. 4 ). Fig. 4: Signatures of selection in urban populations. a Manhattan plots showing the distribution of cross-population nSL scores ( XP-nSL ) for each urban–rural population pair in 200 kb sliding windows with a 50 kb step size. Positive scores depict selection in urban populations. Significant autosomal outlier windows are highlighted in red and outlier windows on the Z-chromosome in blue (see “Methods”). Autosomal and sex-chromosome outlier windows were detected separately. b Violin plots showing the distribution of correlation coefficients (Pearson’s r ) for the different inter-population summary statistics for all pairwise comparisons. Note that all distributions overlap zero and are not consistent across pairs. The shaded grey areas of the violin plot denote the kernel density estimation of the data distribution. The bottom and top edges of the white boxplot denote the upper and lower interquartile range (25th and 75th) with the vertical black line showing the median, and the thin horizontal line representing the rest of the distribution. Black dots represent values outside the distribution. c Violin plots for the distribution of correlation coefficients between inter-population summary statistics with proxies for recombination rate (within population LD, GC-content). The boxplot shows the interquartile range (white area), median (vertical black line) and the remaining distribution (horizontal black line). d Example signatures of divergence and proxies for recombination along chromosome 1 and 6, to highlight the heterogeneity of divergence across populations. Window-based measures were loess-smoothed across each chromosome. Populations with selective sweeps on the respective chromosomes are highlighted, with Glasgow (blue) on chromosome 1 and Lisbon (orange) on chromosome 6. All other populations are in grey in the background. Note that the highest sweep window, marked by the dotted line, corresponds to a likely low-recombination region on chromosome 6, but a normal recombination region on chromosome 1. e Non-significant correlation of the geographic distance between urban and rural populations from the same locality and their genetic differentiation. (linear model: F 1,7 = 0.23, P = 0.645). Shaded area denotes the 95% CI. BCN Barcelona, GLA Glasgow, GOT Gothenburg, LIS Lisbon, MAD Madrid, MAL Malmö, MIL Milan, MUC Munich, PAR Paris. Source data are provided as a Source data file. Full size image The observed landscape of variation in XP-nSL is in agreement with the genomic landscape of genetic differentiation (standardised Z -transformed F ST [ ZF ST ]) between each adjacent urban and rural pair across the same windows (Fig. 4b , Supplementary Figs. 11 and 12 ). ZF ST outlier windows ( ZF ST > 4, equivalent to four SD) were largely population specific (Supplementary Fig. 12 ), suggesting that population-specific selective and demographic processes have partly shaped the genomic landscape 47 . To strengthen our inference of the selective sweeps landscape associated with urban adaptation in our species, we additionally searched for older and completed selective sweeps using the haplotype-based Rsb statistic 48 . Similarly to the results obtained using the XP-nSL statistic and ZF ST , window-based Rsb scores were weakly and inconsistently correlated across populations (Fig. 4b and Supplementary Fig. 13 ). We detected fewer Rsb outlier windows (209–538) and clustered genomic outlier regions (39–98; average size: 396.9 kb ± 329.2 SD; max = 3.45 Mb on chromosome 3 in Gothenburg) compared to XP-nSL . This result suggests that ongoing/recent sweeps outnumber older and completed selective sweeps in urban great tits. Similar to XP-nSL , the Rsb outlier windows were only shared across a maximum of five populations and in general widely distributed across the genome (Supplementary Fig. 14 ). Overall, the selection analyses showed that selective sweeps, and in particular those that are recent and ongoing ( XP-nSL ), were pervasive across the genomes of urban great tits, supporting our inference of a highly redundant polygenic adaptation to urban environments. The underpinning genomic changes associated with urban adaptation were largely population specific, likely the result of differential selective pressures, adaptation through complex multivariate phenotypic changes and/or selection on standing genetic variation 49 , 50 . Nonetheless, we still detected shared selective sweeps across urban populations, suggesting the presence of common genomic responses to the urban selective pressures, either through repeated selective sweeps or sharing of adaptive genetic variants. Evolutionary drivers of genetic differentiation and signatures of selection The largely population-specific landscapes of differentiation are in line with observations in other young divergences, and have been previously explained by the weaker effect of linked selection at early stages of the divergence process 47 . However, to exclude a driving role of non-adaptive processes, we corroborated that the detected signatures of selection were indeed caused by divergent selection and not by genetic drift or linked selection in low-recombination regions 47 . To achieve this, we evaluated the correlation between signatures of selection ( XP-nSL and Rsb ) and differentiation ( F ST ) with two proxies of genomic recombination rate, i.e., LD and intronic GC-content (both in 200 kb non-sliding windows). LD is correlated with population-specific recombination rates and the strength of selection, while intronic GC-content is a surrogate for the long-term recombination rate variation across the great tit genome, i.e., a higher GC-content will indicate a higher recombination rate 51 , 52 . To confirm that both proxies provide similar pictures of the recombination landscape, we tested their correlation across all population pairs. Indeed, LD and GC-content, were (negatively) correlated in all populations (Fig. 4c ), suggesting that the general effect of recombination on the genomic diversity landscape is broadly consistent. Under a scenario of linked selection in low-recombination regions, we would expect strong and consistent positive correlations between GC-content or LD with our estimates of genetic differentiation and selection, i.e., F ST , XP-nSL and Rsb 53 , 54 , 55 . Yet, this was not the case, (Fig. 4c ), which indicates that processes other than linked selection are the main driver for population differentiation in our study. Indeed, the overall patterns are more in line with divergent selection, as e.g., the population-specific selection peak on chromosome 1 in Glasgow is not associated with low-recombination regions (low GC-content) or a pronounced peak in LD (Fig. 4d ). However, it is important to note that these correlations are positive in some populations, see e.g., XP-nSL and GC-content (Fig. 4c ), and that some population-specific peaks are located on low-recombination regions, and likely caused by linked selection or other non-adaptive processes (Fig. 4d ). While the contribution of non-adaptive processes likely differed across populations, we did not detect a general signature of linked selection as a major driver of divergence between urban and rural populations. Divergent selection has been suggested to be the main driver in the early stages of differentiation in other avian systems 47 , 56 , and it is not surprising that a similar incipient pattern is observed between urban–rural populations. Although other non-adaptive processes, such as stochasticity via population-specific drift, could also generate variable genomic landscapes in an urban context 57 , that scenario would be characterised by a decreased gene flow and increased genetic differentiation with geographical distance between adjacent urban and rural populations. However, genetic differentiation between paired urban and rural populations did not significantly increase with geographic distance ( R 2 = −0.11, P = 0.645; Fig. 4e ), further excluding drift as the main process underneath the variability in the genomic and the genetic differentiation landscape between urban and rural populations. Nonetheless, genetic drift potentially led to increased genetic differentiation in some urban populations, i.e., Glasgow and Lisbon, which both showed reduced heterozygosity. Hence, we conclude that the mosaic of genomic signatures of differentiation between urban and rural populations and selective sweeps in the studied populations are most likely driven by genomic responses to selection on standing genetic variation, with potentially minor contributions of linked selection in low-recombination regions or other idiosyncratic non-adaptive processes. Genes and pathways associated with adaptation to urbanisation While responses to selection at the haplotype level were highly variable across urban populations, selective sweeps might affect the same genes or different genes with similar functional impacts (e.g., same pathway). Such consistent genomic responses on the gene or functional level could explain the similarity in phenotypic responses to environmental variation despite a large variation on the haplotype level 14 , 15 , 58 . Following this rationale, we identified genes associated with signatures of selective sweeps in each urban population, but without requiring outlier windows to overlap with each other. Between 976 and 1366 genes were associated with signatures of ongoing or recent selective sweeps ( XP-nSL ), and 362–923 with completed selective sweeps ( Rsb ). Of these genes, 64–207, and 5–76, were detected in at least two urban populations (based on XP-nSL or Rsb , respectively; Supplementary Figs. 15 and 16 ). The higher number of genes associated with recent and ongoing selective sweeps compared to complete selective sweeps is in line with the colonisation pattern of urban habitats, under ongoing or recent gene flow. Although it is statistically unlikely to detect selective sweeps in the same gene in more than three urban populations by chance alone (permutation test; χ 2 1 = 77.95, P < 2.2 × 10 −16 ), we conservatively focused our functional interpretation only on those genes that were associated with signatures of selection ( XP-nSL or Rsb ) in more than half of all urban populations (≥5 urban populations), as these likely play more important roles in urban adaptation. In total, we detected 42 genes associated with recent or ongoing selective sweeps in at least five populations ( XP-nSL : max. six populations), and 15 genes associated with older completed selective sweeps ( Rsb : max. seven populations; Fig. 5a and Supplementary Table 4 ), with none of the genes detected by both selection statistics. It is noteworthy that we did not detect an increase in the number of shared genes under selection in relation to geographical (Mantel’s r = 0.27, P = 0.080) or genetic distance (Mantel’s r = 0.05, P = 0.480), which is in accordance with the results regarding shared outlier windows and further supports that many selective sweeps likely occurred independently in multiple urban populations. In addition, 12 of the 57 shared genes under selection were also associated with urbanisation in the LFMM analysis (Supplementary Table 4 ). Interestingly, two of these, GMDS and SLC6A15 , were not only associated with complete selective sweeps ( Rsb scores) in seven and six populations, respectively, but also previously shown to be differentially expressed in blood and/or liver between urban and rural birds from Malmö 59 . Moreover, following Malmö as an example, one gene found to be differentially expressed in Watson et al. 59 , VPS13A , was highly differentiated ( ZF ST ) and associated with selective sweeps in that particular population, but importantly, also under selection in Madrid ( Rsb ) and Lisbon ( XP-nSL ) and in general, associated with urbanisation (LFMM analysis). This exemplifies that some genes show signatures of urban adaptation on a continental scale, making them strong candidates for adaptation in these urban centres. Fig. 5: Candidate genes and pathways underlying urban adaptation. a Map showing the spatial distribution of the top shared candidate genes putatively under selection in urban populations across Europe, detected in at least six urban populations. CDH18 , marked with an asterisk, was also detected in the GEA analysis. Genes in bold font were associated with ongoing selection ( XP-nSL ), and those with grey font with older completed sweeps ( Rsb ). The map was plotted in R using ggplot with the maps package. b Haplotype strips showing haplotype structure around two representative candidate genes ( GMDS and HRT7 ) on chromosome 2 and 7, respectively. Individuals (rows) are ordered by haplotype similarity and colour-coded by habitat of origin. Note that individuals from the same habitat do not share the same haplotype. c Distribution of SNP-based Rsb scores for each of the six populations, in which GMDS is associated with significant outlier windows, and XP-nSL scores around HRT7 . Red shades depict SNPs with the strongest positive selection scores. Gene boundaries are shown by red dotted lines and grey background box. Note that not always the same regions (upstream, downstream or genic) are under selection in urban populations (positive scores) around the candidate genes. d Enrichment of associated genes in significantly enriched gene ontology (GO) groups by GO category in the LFMM and BayPass analysis (see Supplementary Table 4 for a detailed list). The colour of the bars represents the false-discovery rate (FDR) from the gene ontology overrepresentation analyses. BCN Barcelona, GLA Glasgow, GOT Gothenburg, LIS Lisbon, MAD Madrid, MAL Malmö, MIL Milan, MUC Munich, PAR Paris. Full size image It is noteworthy that despite finding signatures of selection overlapping the same gene in five or more urban populations (e.g., GMDS, HTR7 or CDH18 ), the exact locations of these signatures were not consistent (Fig. 5b, c ). Under a shared selective sweep or adaptive introgression scenario, we would have expected the sharing of urban haplotypes between populations. However, haplotype plots do not show any noticeable clustering across urban individuals (e.g., GMDS or HTR7 , Fig. 5b ), suggesting that different adaptive haplotypes swept to high frequency in urban populations across Europe, which is in line with soft selective sweeps from standing genetic variation 60 . Furthermore, we find that the strongest selective sweep signatures are, in some cases, not within the genes but up- or downstream and without a consistent pattern among populations. For example, signatures of selection around HTR7 were downstream in Glasgow and Munich, but upstream in Malmö, Paris, Madrid and Lisbon (Fig. 5c ). This might suggest that regulatory changes rather than protein-coding ones are associated with urban adaptation and the nature of regulatory variation likely differs across populations. However, high-resolution genomic studies based on whole-genome resequencing coupled with multi-tissue functional genomic analyses are needed to resolve this question, which is beyond the scope of this study. Many of the genes associated with urbanisation have previously been linked to behavioural divergence, and cognitive and learning functions, suggesting adaptive phenotypic shifts related to behaviour. In particular, HTR7 (Chr 7) codes for a receptor of the neurotransmitter serotonin, a pathway consistently identified as target of changes linked to urbanisation. Indeed, two previous studies in birds, one on European blackbirds using a candidate gene approach and another on burrowing owls ( Athene cunicularia ), using full-genome resequencing, also described changes in this pathway as the main difference between urban and rural populations 3 , 22 . Both species also show consistent urban–rural differences in two behaviours associated with serotonin: down-regulation of the stress axis and altered nocturnal behaviour under artificial light at night, respectively 9 , 61 , 62 . The same gene, HRT7 , has also been directly linked to behavioural differences between migratory and resident rainbow trout ( Onchorynchus mykiss ) 63 , supporting its potential role in behavioural changes in urban great tits. The CDH18 gene (Chr 2), is part of a superfamily of membrane proteins involved in synaptic adhesion and revealed as a candidate gene in phonological alterations in humans 64 . The PTPRD and VPS13A genes (both in Chr Z), are suggested to be involved in hippocampal neural development 65 and were previously linked to bird navigation, flight performance and migratory behaviour 66 , 67 , 68 . Moreover, two other genes under selection and associated with urbanisation are involved in the regulation of cognitive processes and behavioural disorders in vertebrates ( DLG2 (ref. 69 ) and NRXN3 (ref. 70 ), Chr 1 and 5, respectively), and were also previously linked to urban divergence in burrowing owls 22 . Thus, our selection analyses suggest that natural selection repeatedly acts on behavioural traits and sensory and cognitive performance, all previously shown to be among the most widespread differences between urban and rural wildlife populations 71 , 72 . These findings were furthermore supported by gene ontology (GO) analysis of the 2758 urbanisation-associated SNPs (LFMM analysis), linked to 984 genes (1501 SNPs in genic regions). Accordingly, most of the GO terms were related to neural functioning and development (e.g., GO:0016358, FDR = 4.30 × 10 −6 ), cell adhesion (e.g., GO:0098742, FDR = 3.43 × 10 −6 ) and sensory perception (e.g., GO:0050954, FDR = 3.42 × 10 −3 ; Fig. 5d and Supplementary Table 5 ). These GO terms were mainly clustered into two interacting networks, one related to sensory recognition and the other to neural development and cell adhesion (Supplementary Fig. 17 ). These findings reinforce the previous idea on cognitive and behavioural changes as key responses to urbanisation, as song structure and escape or distress behaviour have been previously shown to differ between urban and rural great tit populations across Europe 7 , 35 , 73 . Nonetheless, whether this is the result of a genetic response to selection or phenotypic plasticity was to a large extent still unknown. Our present study suggests a strong genetic component for these putatively urban adaptative phenotypes in great tits across Europe. Detailed functional genomic and phenotypic analyses are now needed to understand the role of these genes and pathways in the adaptive divergence of urban and rural great tits and other songbirds. Our study demonstrates clear genomic signals of local adaptation to urban habitats in a common songbird on a continent-wide scale across Europe. We found that a combination of polygenic allele frequency shifts and spatially varying selective sweeps are associated with adaptation to urban environments. Our results strongly suggest that a few genes, which have known neural developmental and behavioural functions, experienced selective sweeps in urban populations. This suggests a strong consistency in the functional processes associated with urbanisation, despite the fact that the underlying haplotypes are often not shared. Thus, our study exemplifies evolutionary adaptation to urban environments on a European scale, and highlights behavioural and neurosensory adjustments as important phenotypic adaptations in urban habitats. Methods Sample collection and DNA extraction During the years 2013–2015, 20 or more individual great tits were sampled at paired urban–rural sites from nine European cities (Fig. 1a and Supplementary Table 1 ). We sampled a total of 192 individuals (aged >1 year old) with 10–16 individuals per site (Supplementary Table 1 ). Sexes were balanced between pairs (urban–rural) in the dataset (GLMM; sex: χ 2 1 = 1.33, P = 0.280). Each of the paired sampling sites (urban or rural, hereinafter populations) was sampled within the same season. Barcelona and Munich were sampled during winter, however, in both cases only known birds (recaptures) were included in the study, thus, all birds can be considered resident. All urban populations were located within the city boundaries, the areas are characterised with significant proportion of human-built structures, such as houses and roads with managed parks as the only green space (Supplementary Fig. 1 ). Rural populations were chosen to contrast the urban locations regarding degree of urbanisation and were always natural/semi-natural forests, and contained only a few isolated houses. All urban and rural populations were separated by a distance above the mean adult and natal dispersal distance of this species (i.e., see Supplementary Table 1 ) 74 . Blood samples (~25 µl) were obtained either from the jugular or brachial vein and stored at 4 °C in ethanol or SET buffer and subsequently frozen at −20 °C. In each case, procedures were identical for the paired urban and rural populations. All procedures employed during field work were approved by national Ethical Committees and authorisations to collect samples were delivered by the Environment Department of the Generalitat de Catalunya (permit no. AECC/SF/0438), the Scottish Natural Heritage (permit no. 52463) and UK Home Office (license no. 70/7899), the Malmö-Lund animal Ethical Committee (permit no. M454 12:1), the CEMPA and Portuguese Ministry of Environment (permit nos. 40/2014 and 164/2014), the Ministry of the Environment, Housing and Territorial Planning of Madrid (permit nos. 10/103329.9/14, 10/169940.9/13, 10/045383.9/14, 10/127641.9/14 and 10/055393.9/14), the Institute for Environmental Protection and Research (ISPRA, license nos.15510 and 15944) and the Lombardy Region (permit no. 3462), the Tierschutzgesetz (TierSchG, German animal protection law) and the Regierung von Oberbayern (permit nos. 55.2-1-54-2532.2-7-07 and 55.2-1-54-2532-140-11) and the Minister of Higher Education, Research and Innovation (Ethics Committee for Animal Experimentation license no. 005 and permit no. APAFIS#19941-2019032516275025), the Prefect of Paris and the Prefect of Seine et Marne (permit nos. DRIEE-2012-31 and DRIEE-2012-32) and the CRBPO (National Museum of Natural History, permit no. 537 and licence no. 1454). DNA was extracted from ~5 µl samples of red blood cells in 195 µl of phosphate-buffered saline, using Macherey-Nagel NucleoSpin Blood Kits (Bethlehem, PA, USA), and following the manufacturer’s instructions or manual salt extraction (ammonium acetate). The quantity and purity of the extracted genomic DNA was high as measured using a Nanodrop 2000 Spectrophotometer (Thermo Fisher Scientific) and Qubit 2.0 Fluorometer (Thermo Fisher Scientific). Urbanisation score To quantify the degree of urbanisation at each site, we used the UrbanizationScore image analysis software 75 , based on aerial images from Google Maps (Google Maps 2017), and following the methods described in different studies assessing the effect of urbanisation on wild bird populations 76 . Briefly, each sampling site was represented by a 1 × 1 km rectangular area around the capture locations. We opted to use this spatial resolution around the sampling sites to approximately cover individuals’ territory 32 , but also to ensure that we considered the heterogeneity within each selected site. The content in each rectangle was evaluated dividing the image in 100 × 100 m cells and considering three land-cover characteristics in each: proportion of buildings, vegetation (including cultivated fields) and paved surfaces (Supplementary Fig. 1 ). The different land-cover measures obtained per site were used in a PCA to estimate an urbanisation score variable (PC urb ) for each of the urban or rural populations per locality, see Table S1 . The PC urb values were transformed to obtain negative values in the less urbanised and positive values in the more urbanised sites. We used the average of the urbanisation estimates if birds were captured in more than one location within each site (>2 km apart, mean ± SD: 931.22 ± 1005.26 m). All quantifications were done in triplicates by the same person (P.S.) and the estimates were highly repeatable (intra class correlation coefficient, ICC = 0.993, 95% confidence interval (CI) = 0.997–0.987, P < 0.001). SNP genotyping All 192 individuals were successfully genotyped using a custom made Affymetrix© great tit 650 K SNP chip at Edinburgh Genomics (Edinburgh, United Kingdom). The Affymetrix SNP chip was developed based on whole-genome sequencing data from great tits sampled across multiple European populations, largely corresponding to our sample sites 38 . Thus, this SNP chip allowed us to genotype a large number of individuals on a genome-wide basis at high SNP density without a strong ascertainment bias. SNP calling was done following the “Best Practices Workflow” in the software Axiom Analysis Suite 1.1.0.616 (Affymetrix©) and all the individuals passed the default quality control steps provided by the manufacturer (dish quality control values >0.95) and previous studies using the same SNP chip 38 , 77 . A total of 544,610 SNPs were then exported to a variant-calling format (VCF) and Plink v. 1.9 (ref. 78 ), and further filtered and assigned to chromosomes using the P. major reference genome build 1.1 (annotation release ID 101; “ GCA_001522545.2 ”). A total of 155 SNPs were not found in the used assembly and 26,852 SNPs were not in chromosomic regions; thus, these SNPs were removed from further analysis leaving a total of 517,603 SNPs. Genetic diversity and population structure We calculated the genome-wide genetic diversity as expected heterozygosity ( H e ) for each population using Plink v. 1.9 (ref. 78 ), and tested if genetic diversity significantly differed between urban–rural populations from the same location using t tests in R, and overall across urban–rural pairings using a Wilcoxon rank-sum test in R v.3.6.1 (ref. 79 ). Furthermore, we estimated pairwise F ST between all population pairs (urban–rural per locality), using VCFtools v. 0.1.15 (ref. 80 ). Mean average F ST was computed across all comparisons after setting negative values to zero. For analyses of population structure, we pruned the SNP dataset based on LD in Plink v.1.9 (ref. 78 ), using a variance inflation factor threshold of 2 (“-indep 50 5 2”), retaining 358,149 SNPs. Using this LD-pruned dataset, we performed a PCA using Plink. We performed two different PCAs, (i) only with autosomes and excluding all small linkage groups and the Z-chromosome, and (ii) including the Z-chromosome but excluding all small linkage groups. Subsequently, we performed additional dimensionality reduction for each PCA dataset using the first 20 principal components, using the UMAP v.0.2.7.0 R-package with default settings 81 . We also compared the results to the UMAP based on the first ten PCs. Genetic ancestry analysis was performed using the software package fastStructure v.1.0 (ref. 82 ), with K ranging from 2 to 9 and cross-validation. In addition, we inferred a population tree based on allele frequency co-variances using Treemix v.1.3 (ref. 83 ), with blocks of 100 SNPs. We fitted up to five migration edges and determined the optimal number of migration edges that lead to the best fitting population tree by (i) comparing the likelihoods of trees with different number of migration edges, (ii) estimating the total variance explained by each tree and (iii) by comparing the significance of the fitted migration edges. To infer migration rates across our entire sampling area, we used EEMS 84 to estimate effective migration surfaces across all our samples based on the complete LD-pruned SNP dataset, with the following settings: nDemes = 1000, numMCMCIter = 2,000,000, numBurnIter = 1,000,000 and numThinIter = 9999. The results were plotted using the R-scripts provided with the EEMS software packages. Environment-associated SNPs We used two different approaches to identify SNPs associated with the degree of urbanisation, based on the “urbanisation score”. First, we used a univariate latent-factor linear mixed model implemented in LFMM v.1.5 for examining allele frequency–environment associations 85 . Based on the number of ancestry clusters (K) inferred with fastStructure v.1.0, we ran LFMM with two and four latent factors, respectively. Each model was run five times for 10,000 iterations with a 5000-iteration burn-in. We calculated the median z -score for each locus across all ten runs, and selected SNPs with a FDR < 1% to be associated with urbanisation. The results with two or four latent factors were highly concordant and the same candidate loci were recovered; thus, we only used the results obtained with four latent factors for further analyses. We also assessed the distribution of P values to check for the impact of confounding factors (Supplementary Fig. 5a ). In addition, to assess if associations are putatively false positives, we performed a permutation analysis. We randomised habitat-assignments 20 times and performed LFMM association analyses on each randomised dataset. Following Fuller et al. 86 we determined a significance threshold as the 95th percentile of the Z -score distribution and identified SNPs from the initial LFMM analyses above this threshold as significant. Using the significance cut-off based on randomisation (>99th percentile), we detected a far higher number of associated SNPs (4358 SNPs; Supplementary Fig. 5b ) than using a FDR < 1% (2758 SNPs; Fig. 3a ); therefore, and to avoid a larger number of false positives, we opted for a more conservative approach and focus on the FDR results in the analysis (see “Results and discussion”). Second, we analysed associations with urbanisation using the auxiliary covariate model implemented in BayPass v.2.1 (ref. 87 ). We estimated the allele frequency–environment association for each SNP with the urbanisation score for each population accounting for population structure, using a covariance matrix. We estimated the covariance matrix using the LD-pruned SNP dataset in the core model using default parameters: 20 pilot runs of 1000 iterations, a run length of 50,000 iterations, sampling every 25th iteration, and a burn-in of 5000 iterations. The resulting covariance matrix was used as input for five replicated runs of the auxiliary covariate model using the above settings. The strength of association is given in the test by estimated Bayes factor (measured in deciban; dB). We calculated the median Bayes factor across all five replicated runs and considered all SNPs with a deciban unit (dB) > 20 as urbanisation associated. This is the strictest criterion and is considered as “decisive evidence” for the association 87 . Using a resampling-without-replacement-based permutation approach and a chi-square test, we determined if the overlap between the LFMM and BayPass candidate SNPs is higher than expected by chance. Patterns of genetic differentiation ( F ST ) To identify genomic regions distinguishing adjacent urban and rural great tits, we estimated the genetic differentiation (Weir and Cockerham’s F ST 88 ) for each urban and rural pair for each SNP, using VCFtools v. 0.1.15 (ref. 80 ). We subsequently summarised and plotted F ST values in 200 kb sliding windows with 50 kb steps, using the WindowScanR v.0.1 R-package ( ), and standardised window-based F ST estimates using a z -transformation ( ZF ST ). Standardisation was performed separately for autosomes and the Z-chromosome. To determine if the extent of genetic differentiation between all possible population pairs was significantly different from zero, we estimated P values using a permutation analysis in Genodive v.3 (ref. 89 ). For computational reasons, the analysis was performed on a thinned SNP dataset (21,062 SNPs), for which we randomly selected one SNP every ~50 kb. Haplotype-based selection analyses To identify genomic regions showing signs of selective sweeps in urban–rural population pairs, we used two haplotype-based selection scans. First, we scanned the genome for regions showing differences in extended haplotype homozygosity (EHHS) between urban and rural populations (see Rsb score below) to identify genomic regions showing signs of older and completed selective sweeps. We used fastPHASE v.1.4 (ref. 90 ) to reconstruct haplotypes and impute missing data independently for each chromosome using the default parameters, except that each individual was classified by its population (“-u” option). We used ten random starts of the EM algorithm (“-T” option) and 100 haplotypes (“-H” option). The fastPHASE output files were analysed using rehh v.2.0 (ref. 91 ) to calculate Rsb statistics per focal SNP. The Rsb score is the standardised ratio of integrated EHHS (iES, which is a site-specific EHHS) between two populations 48 , 91 . This statistic measures the extent of haplotype homozygosity between two populations and follows the rationale that if a SNP is under selection in one population compared to the other, the region around this locus will show an unusually high level of haplotype homozygosity compared to the neutral distribution. In accordance to Gautier and colleagues 91 , we considered significant SNPs putatively under selection in urban populations based on a threshold of Rsb ≥ 4. To identify genomic regions consistently showing strong signals of selection, we summarised Rsb scores in 200 kb sliding windows (50 kb steps) and selected windows under selection as those with (i) an average Rsb scores above the 95th percentile of the genome-wide distribution (separately for autosomes and the Z-chromosome) and (ii) a proportion of urban outlier SNPs ( Rsb > 4) above the 95th percentile of the genome-wide distribution (separately for autosomes and the Z-chromosome). The proportion of outlier SNPs per window was the number of SNPs with Rsb values >4 compared to the total number of SNPs per window 92 . Because recombination rates can be assumed to be conserved between closely related urban and rural populations, the cross-population comparative nature of the Rsb statistic provides an internal control that cancels out the effect of heterogeneous recombination across the genome 48 . Second, we used a cross-population nSL 93 statistic ( XP-nSL ) that tests for signatures of ongoing or recently completed selective sweeps and is implemented in selscan v.1.3.0 (ref. 46 ). XP-nSL contrasts nSL , the number of segregating sites by length, between two populations and thus tests for differential local adaptation. Analogous to Rsb , we conservatively detected genomic regions by a 200 kb sliding windows as those with (i) average XP-nSL scores above the 95th percentile of genome-wide distribution (separately for autosomes and the Z-chromosome) in each population and (ii) proportions of outlier SNPs ( XP-nSL > 2) above the 95th percentile. We also plotted haplotype patterns around candidate genes using the haplostrips software v.1.3 (ref. 94 ). To determine larger genomic regions associated with signals of selection, we merged adjacent outlier windows for each selection statistic ( XP-nSL and Rsb ) into outlier regions, if windows were not >200 kb apart, using csaw v.3.12 (ref. 95 ). To determine the consistency of selection across urban centres, we implemented a resampling approach to assess the likelihood of genes showing signs of selection in two, three, four or more populations. We resampled with replacement n genes ( n = number of genes with signatures of selection in each urban population) for each population from the list of all SNP-linked genes using the “resample ” function in R, assessed the amount of overlap between populations (from two to eight populations) and repeated the sampling 100,000 times for each comparison. We then calculated the mean and 95% CI for each comparison and compared the number of observed shared candidate genes to the expected number of candidate genes. The expected number of genes showing signs of selection in three or more populations was zero, thus we focused on genes showing signs of selection ( Rsb and XP-nSL ) in three or more populations. Decomposing habitat and locality effects for urbanisation-associated SNPs To determine the explanatory power of urbanisation-associated SNPs and consistency in allele frequency changes across populations, we used linear models (i.e. “ Y ∼ habitat × locality × habitat × locality”) to quantify the effects of (i) “habitat” (i.e., consistent change in allele frequency across urban–rural pairs), (ii) “locality” (i.e., effect of city-of-origin on absolute allele frequency) and (iii) “habitat × locality” interaction (i.e., inconsistent change in allele frequency across urban–rural pairs) 44 , 96 . We performed this analysis for all individual SNPs (coded as 0, 1 or 2) and based on PC scores from the first three PC axes for all LFMM candidate SNPs (PC LFMM ), and those overlapping between the LFMM and BayPass analyses (“core urbanisation SNPs”, PC GEA , see above). The PCA was estimated for each SNP dataset using SNPrelate v.3.12 (ref. 97 ). We used the “EtaSq” function implemented in BaylorEdPsych v.0.5 (ref. 98 ) to extract the effect sizes (partial η 2 ) for the model terms in each linear model. We further estimated the directionality of allele frequency changes across populations by counting in how many urban populations the same allele was the minor allele for all significant LFMM candidate SNPs and the “core urbanisation SNPs”. We then estimated for each SNP dataset the proportion of SNPs that showed concordant minor alleles in five, six, seven, eight or nine urban populations. Patterns of linkage disequilibrium and intronic GC-content across the genome To estimate the impact of variation in recombination rate and linked selection in low-recombination regions on patterns of divergence (i.e., regions showing signs of selective sweeps), we estimated the correlations of LD and intronic GC-content in 200 kb windows with measures of selection/differentiation ( Rsb, XP-nSL and ZF ST ). First, we estimated patterns of LD ( r 2 ) across the genome using Plink 1.9 (ref. 78 ) for pairs of SNPs located up to 200 kb apart for each urban population (–r2–ld-window-r2 0–ld-window-kb 200,000). LD estimates were averaged in 200 non-sliding windows using the WindowScanR v.0.1 R-package. Second, we estimated the GC-content for all introns across the great tit ( P. major ) reference genome build 1.1 (annotation release ID 101; “ GCA_001522545.2 ”), as low intronic GC-content is a good proxy for long-term reduced recombination in that region 52 . Intron coordinates were extracted from the great tit reference genome using the plyranges v.3.12 (ref. 99 ) and GenomicRanges v.3.12 (ref. 100 ) R-packages, and the GC-content for each intron inferred using the “nuc ” function in BEDtools v.2.28 (ref. 101 ). Intronic GC-content values were further averaged across 200 kb non-sliding windows using WindowScanR. Lastly, we estimated the genome-wide Pearson correlations between LD, GC-content and all estimators of selection ( Rsb , XP-nSL and ZF ST ) across all 200 kb windows for each population, using the “cor.test ” function in R. Functional characterisation of candidate SNP We obtained the gene annotations for all candidate SNPs from the great tit ( P. major ) reference genome build 1.1 (annotation release ID 101; “ GCA_001522545.2 ”). We used all genes containing SNPs associated with urbanisation (LFMM and BayPass, n = 1501 SNPs within genes). To analyse the enrichment of functional classes, we identified overrepresented GOs (biological processes, molecular functions and cellular components), using the WebGestalt software tool 102 . The gene background was set using annotated great tit genes (annotation release ID 101) containing SNPs from the SNP chip and with Homo sapiens orthologues. H. sapiens genes were used as a reference set as human genes are better annotated with GO terms than those of any avian system (e.g., Gallus gallus ) 16 , 39 . We focused on non-redundant GO terms to account for correlations across the GO graph topology and GO terms as implemented in WebGestalt 102 . An FDR < 0.05 was used as a threshold for significantly enriched GO terms. Furthermore, we searched the public record for functions of individual candidate genes. We also used GOrilla 103 to visualise the connections of GO terms associated with LFMM candidate genes and Cytoscape v.3.6.1 (ref. 104 ) to visualise the GO network and identify all enriched GO terms (biological processes, P < 0.001), including redundant terms. Candidate genes associated with signatures of selection were those that overlapped with significant XP-nSL or Rsb outlier windows. Overlaps between outlier windows and annotated genes were assessed using the plyranges R-package v.3.12. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The genotyping data that support the findings of this study is available in variant call format (VCF) via the European Variation Archive (EVA) with the accession number PRJEB44069. Source data are provided with this paper. Code availability No custom code or mathematical algorithm was developed for this project or is considered crucial to the conclusions. All relevant software and R-functions that were used are referred to in the “Methods” section.

Great tits living in cities are genetically different from great tits in the countryside. Researchers made the discovery after a unique study where they examined populations of great tits in nine large European cities, including Glasgow. The researchers compared the city bird genes with the genes of their relatives in the countryside. The findings, which are published today in Nature Communications, showed that it didn't matter if the great tits lived in Barcelona, Malmö or Glasgow: in order to handle an environment created by humans, the city birds all evolved in a similar way. The European research, which was led by Lund University in collaboration with researchers at the University of Glasgow, included a study of birds from Kelvingrove Park together with birds from around the forest in the University of Glasgow's SCENE (Scottish Center for Ecology and the Natural Environment) facility, located near Loch Lomond. The researchers found that different genes linked to important biological functions regulated by serotonin, including aggression and circadian rhythms, were found to have been selected and passed on from generation to generation in the city birds. In rural populations, these behaviors are also important, however, the genes that control them do not confer the same advantage as they do in an urban environment. Caroline Isaksson, senior lecturer at Lund University, led the study together with Dr. Pablo Salmón, now a research fellow at the University of Glasgow. She said: "This indicates that these behaviors, and cognition, are very important in order to live in urban environments with a lot of stress in the form of noise pollution, light at night, air pollution and constant proximity to people." Fig. 1: Study sampling locations and urbanisation intensity. From: Continent-wide genomic signatures of adaptation to urbanisation in a songbird across Europe The study is the largest carried out on how urban environments affect the genome, and thus the genetic material of the animals that live there. In total, 192 great tits were examined among populations in Malmö, Gothenburg, Madrid, Munich, Paris, Barcelona, Glasgow, Lisbon and Milan. For each urban population, the researchers had a control group of great tits living nearby, but in a rural environment. Blood samples have been taken from the birds and analyzed genetically. Dr. Pablo Salmón, from the University of Glasgow, said: "It is surprising that cities, which from an evolutionary perspective are a recent phenomenon, are already leaving their footprint in the genome of birds." The researchers analyzed more than half a million single-nucleotide polymorphisms (SNPs) spread over the entire genome, and found that a handful of genes that had clearly changed in response to the urban environment. Great tits are common throughout Europe, and it has long been known that they are quite similar genetically. Despite this, researchers have now identified clear genetic differences between great tits in the city and great tits in the countryside. The study, "Continent-wide genomic signatures of adaptation to urbanization in a songbird across Europe," has been published in Nature Communications.

10.1038/s41467-021-23027-w

Medicine

Manipulating mitochondrial shape may limit metastatic cancer, study finds

Pravat Kumar Parida et al, Limiting mitochondrial plasticity by targeting DRP1 induces metabolic reprogramming and reduces breast cancer brain metastases, Nature Cancer (2023). DOI: 10.1038/s43018-023-00563-6 Journal information: Nature Cancer

https://dx.doi.org/10.1038/s43018-023-00563-6

https://medicalxpress.com/news/2023-06-mitochondrial-limit-metastatic-cancer.html

Abstract Disseminated tumor cells with metabolic flexibility to utilize available nutrients in distal organs persist, but the precise mechanisms that facilitate metabolic adaptations remain unclear. Here we show fragmented mitochondrial puncta in latent brain metastatic (Lat) cells enable fatty acid oxidation (FAO) to sustain cellular bioenergetics and maintain redox homeostasis. Depleting the enriched dynamin-related protein 1 (DRP1) and limiting mitochondrial plasticity in Lat cells results in increased lipid droplet accumulation, impaired FAO and attenuated metastasis. Likewise, pharmacological inhibition of DRP1 using a small-molecule brain-permeable inhibitor attenuated metastatic burden in preclinical models. In agreement with these findings, increased phospho-DRP1 expression was observed in metachronous brain metastasis compared with patient-matched primary tumors. Overall, our findings reveal the pivotal role of mitochondrial plasticity in supporting the survival of Lat cells and highlight the therapeutic potential of targeting cellular plasticity programs in combination with tumor-specific alterations to prevent metastatic recurrences. Main Metastatic relapses are common in cancer patients considered disease-free after primary diagnosis and treatment 1 . Stem-like disseminated tumor cells with specialized metastatic traits adapt and survive as latent metastases in distal organs by overcoming oxidative stress, nutrient limitation, microenvironmental and immune defenses. Latent metastatic cells that persist as subclinical disease are responsible for late recurrences. Understanding the traits and vulnerabilities of these cells is critical for developing strategies to prevent metastasis 2 . Metabolic reprogramming is a hallmark of cancer that evolves during metastatic progression 3 . Nutrient availability and dependencies are likely to dictate metabolic reprogramming of disseminated cancer cells. Given that tumor cells in the primary and metastatic sites are metabolically distinct, targeting extracellular nutrient and microenvironmental driven dependencies of disseminated tumor cells is a rational approach to limit progression of metastasis 4 , 5 , 6 , 7 , 8 , 9 . Tumor cells that undergo metabolic reprogramming overcome stress imposed by the microenvironmental and persist or proliferate in distal organs 6 , 10 , 11 . However, the precise molecular processes that facilitate metabolic reprogramming in latent metastatic cells to overcome stress and persist after dissemination remain unknown. Mitochondria, the cellular energy powerhouses, are integral to stress response, as they not only support cellular bioenergetic needs but also generate metabolites that facilitate molecular and epigenomic responses. Mitochondria are dynamic organelles that undergo coordinated cycles of fusion and fission to facilitate adaptations to cellular and extracellular cues 12 . Cancer cells exploit adaptive mitochondrial dynamics to meet energy needs, regulate reactive oxygen species, reprogram cellular metabolism and survive environmental or nutrient stress 13 , 14 . In this Article, using mouse models and patient samples, we investigated how disseminated latent metastatic cells in the lipid-rich brain microenvironment meet their cellular energetic demands and persist to initiate overt metastasis. We report mitochondrial plasticity enables fatty acid oxidation (FAO) and survival of latent metastatic cells as subclinical disease. Furthermore, we demonstrate targeting mitochondrial plasticity by attenuating expression of mitochondrial fission regulator dynamin-related protein 1 (DRP1) and through pharmacologic inhibition of DRP1 using a small brain-permeable molecule suppressed growth of latent metastatic cells and reduced brain metastatic burden in preclinical models. Results Latent cells uptake FAs secreted by reactive astrocytes Cancer metastases to the central nervous system are lethal. Brain metastatic incidence in breast cancer patients varies with disease subtype and approximately 25–50% of patients with advanced stage breast cancer present with brain metastases 15 , 16 , 17 . Moreover, metachronous brain metastases are common in HER2 + (human epidermal growth factor receptor 2) and hormone receptor-positive breast cancer patients that have undergone primary therapy and considered disease-free 17 , 18 , 19 , 20 , 21 , 22 . Current treatment options for brain metastasis are limited and not curative. To understand how disseminated tumor cells persist as subclinical disease and initiate metachronous metastasis, we performed in vivo phenotypic selection in mice and isolated latent brain metastatic (Lat) cells from HCC1954 and SKBR3 HER2 + breast cancer cells 6 , 23 , 24 , 25 . Lat cells are stem cell like, immune evasive and metabolically distinct. Compared to HCC1954 HER2 + breast adenocarcinoma parental cells, brain tropic latent metastatic cells were enriched in metabolites and genes associated with fatty acid (FA) metabolism (Extended Data Fig. 1a,b Supplementary Table 1 ). Correspondingly, significant enrichment in content, namely, carnitine-conjugated FA, lauroyl- l -carnitine (carnitine 12:0) and palmitoyl- l -carnitine (carnitine 16:0), was observed in Lat cells (Extended Data Fig. 1a ). Lipidomic profiles also indicated a higher neutral FA content in Lat cells ( Fig. 1a and Supplementary Table 2 ). Cancer cells can synthesize FAs de novo or uptake them from the surrounding microenvironment 26 . However, the expression of FASN, ACC1 and phospho-ACC1 (Ser-79), enzymes associated with FA synthesis, was notably lower in Lat cells (Extended Data Fig. 1c ). Moreover, 13 C 6 -glucose tracing showed significantly low enrichment of palmitate isotopologues (Extended Data Fig. 1d and Supplementary Fig. 1a ), suggesting that Lat cells are better equipped to uptake exogenous lipids. Fig. 1: Latent cells uptake FAs secreted by reactive astrocytes. a , Heatmap showing total FA profiles of neutral lipid content generated by GC–MS in HCC1954 Pa and Lat cells ( n = 4, each group). b , IF image showing reactive astrocytes (GFAP + , magenta) surrounding Lat cells (GFP + , green) in mouse brain 5 weeks after intracardiac injection. c , IF image showing accumulation of LDs (red) in HCC1954 Pa and Lat cells (green) cultured with BODIPY-558/568-C12 labeled GFAP + reactive astrocytes (gray). Scale bar, 100 µm. d , Quantification showing average LDs in Pa and Lat cells (HCC1954 and SKBR3) cultured with BODIPY-558/568-C12 labeled astrocytes for 24 h, n = 5 each group. e , Quantification showing time-dependent increase in lipid transfer from astrocytes to HCC1954 Pa and Lat cells, n = 6, each group. Isolated astrocytes from mice pups were labeled with C12-BODIPY(2 μM) for overnight and then washed three times. Next, HCC1954 Pa and Lat cells cultured in R3F were transferred to the astrocytes cultured plates and live cell imaging was performed using confocal microscope. d , e , Box and whiskers plot showing minima to maxima. with all points. f , TEM images showing LDs and mitochondria in HCC1954 Pa and Lat cells. g , Quantification of TEM image showing mitochondrial length in HCC1954 (Pa, n = 6; Lat, n = 6) and SKBR3 (Pa, n = 6; Lat, n = 5) model. h , Anti-TOMM20 antibody IF images showing mitochondrial morphology of HCC1954 Pa and Lat cells. Scale bar, 50 µm. i , Classification of cells based on percentage distribution of tubular, intermediate and punctate mitochondrial morphology in HCC1954 Lat and Pa cells. n = 8, each group. j , Western blot showing expression of DRP1 along with phosphorylated p-DRP1 S616 (activated form) and p-DRP1 S637 (inactivated form) in HCC1954 Pa and Lat cells. In a , d , e , g and i , ‘ n ’ represents biologically independent samples. Data are presented in d , e , g and i as mean ± s.e.m., and P values in d , e and g were calculated by two-tailed unpaired t -test. The experiments shown in b , c , f , h and j were repeated independently at least two or three times with similar results. Source data Full size image The brain, a lipid-rich organ composed of neurons and supporting glial cells, is a unique microenvironment encountered by disseminated breast cancer cells. Astrocytes, brain resident glial cells, become reactive and undergo molecular and functional remodeling in response to pathologic conditions such as injury, infection or cancer 27 . Indeed, immunofluorescence (IF) staining of mouse brain 5 weeks after intracardiac injection revealed that Lat cells in the brain are surrounded by reactive astrocytes (Fig. 1b ). Tumor-associated reactive astrocytes are known to secrete anti-inflammatory cytokines, growth factors, and FAs that promote immune evasion and tumor cell growth 28 , 29 , 30 . Therefore, we established astrocyte cancer cell co-culture assay protocol to assess FA transfer from astrocytes to cancer cells using BODIPY-558/568-C12 (4,4-difluoro-5-(2-thienyl)-4-bora-3a,4a-diaza-s-indacene-3-dodecanoic acid). We first incubated astrocytes with BODIPY-C12 overnight, washed three times with warm culture media and cultured them with GFP + cancer cells. Live cell time-lapse imaging was performed on these co-cultures to assess transfer of fluorescent labeled FAs from astrocytes to cancer cells. Significantly higher number of BODIPY-C12 + lipid droplets (LDs) were observed in HCC1954 and SKBR3 Lat cells compared to the Pa cells (Fig. 1c–e , Extended Data Fig. 1e,f , Supplementary Videos 1 and 2 and Supplementary Fig. 1b ). Likewise, HCC1954 and SKBR3 exposed to lauric acid (12:0) and palmitic acid (16:0) had increased number of LDs (Extended Data Fig. 1h,i ), whereas no such differential effect was observed in the presence of oleic acid (18:1) (Extended Data Fig. 1j ). Ability to store excess free FAs from the cytoplasm in the form of LDs attenuates lipo-toxicity and promotes survival 31 , 32 . Indeed, Lat cells had higher viability compared to parental cells in astrocyte co-culture experiments or upon supplementation of astrocytic media (Extended Data Fig. 1g ). Likewise, administration of BODIPY-C12 increased cell death in parental cells compared to Lat cells (Extended Data Figs. 1k and 2a,b and Supplementary Videos 3 and 4 ). Thus, we conclude Lat cells uptake FAs from astrocytes, store excess as LDs and resist lipo-toxicity induced cell death. Fragmented mitochondria puncta enriched in Latent cells To visualize and quantify LDs in Lat and Pa cells we performed transmission electron microscopy. In support of earlier observations, the number of lipid droplets was significantly elevated in Lat cells compared to Pa cells from both cell line models (Fig. 1f and Extended Data Fig. 2c ). Noticeably, smaller fragmented mitochondria were observed in Lat cells compared to parental cells that have large tubular mitochondria (Fig. 1f,g and Extended Data Fig. 2c ). TOMM20 staining confirmed an increased number of intermediate and punctate mitochondria in Lat cells (Fig. 1h,i ). Similarly, increased number of LDs and punctate mitochondria was also observed in HCC1954 Lat cells compared to Pa cells when cultured with lauric acid (12:0) and palmitic acid (16:0) (Extended Data Fig. 2d–i ). In agreement with these observations, expression of mitochondrial fission protein, DRP1 and its phosphorylation at serine 616 (p-DRP1 S616 ) that promotes DRP1 translocation to mitochondria was elevated in HCC1954 and SKBR3 latent cells (Fig. 1j and Extended Data Fig. 2j ) 33 , 34 . Increased DRP1 acetylation has also been reported to promote DRP1 phosphorylation and mitochondrial fission 35 . Mass spectrometry analysis of DRP1 further confirmed high abundance of p-DRP1 S616 and identified several acetylated lysine residues (K10, K92, K160, K238 and K597) in Lat cells (Extended Data Fig. 3a ). Lat cells oxidize internalized FAs and maintain redox homeostasis Internalized FAs could be directed to mitochondria for oxidation or directly stored in LDs. Increased β-oxidation of FAs in mitochondria could generate acetyl-CoA, which feeds into the tricarboxylic acid (TCA) cycle and promotes cell survival during stress or nutrient depleted conditions 36 , 37 . Indeed, 14 C-labeled palmitate treatment induced higher 14 CO 2 production in Lat cells compared to Pa (Fig. 2a ) suggesting increased FAO. Further supporting this observation, carnitine palmitoyl-transferase IA (CPT1A), which converts long-chain fatty acyl-CoA to fatty acylcarnitine and helps translocate FAs into the mitochondria, was enriched in HCC1954 and SKBR3 Lat cells (Extended Data Fig. 3b ). Fig. 2: Lat cells oxidize internalized FAs and maintain redox homeostasis. a , Bar graph showing oxidation of 14 C-palmitic acid (1μCi ml −1 and production of 14 CO 2 in presence of palmitic acid (100 μM), L-carnitine (1 mM) and glucose (10 mM) in Pa and Lat cells. n = 5, each group. b , Schematic illustration showing distribution of palmitate-derived carbon in TCA cycle intermediates and glutathione. Empty circles represent 12 C and violet circles represent 13 C (illustration was made using BioRender.com). c–g , 13 C 16 palmitic acid tracing showing enrichment of isotopologues of citrate ( c ), glutamate ( d ) and malate ( e ), GSSG ( f ) and GSH ( g ) in HCC1954 Pa and Lat cells, respectively. Briefly, cells were grown in R3F for 24 h and then treated with bovine serum albumin (BSA)-conjugated 13 C 16 palmitic acid (100 µM) for 24 h. Samples were collected in 80% methanol in water, and enrichment of metabolites was analyzed by LC–MS. n = 4, each group. h , Flow cytometry analysis of cellular ROS (CellROX Deep Red staining) in Pa and Lat cells ( n = 8, each group). i , MTT cell viability assay showing survival of HCC1954 Pa and Lat cells ( n = 6, each group) grown in palmitic acid (100 μM and 400 μM) for 48 h. j , Showing APC Annexin V/propidium iodide (PI) + cells in HCC1954 Pa and Lat cells ( n = 4, each group) treated with palmitic acid for 48 h. In a and c – j , ‘ n ’ represents biologically independent samples. Data are presented in a and c – j , as mean ± s.e.m., and P values in a and c – j were calculated by two-tailed unpaired t -test. Source data Full size image To further assess FA utilization patterns and their contribution to the observed differences in metabolite pools, we performed 13 C 16 palmitate isotope tracing analysis (Fig. 2b ). Carnitine-conjugated FAs were significantly high in Lat cells (Extended Data Fig. 3c,d ). Furthermore, the M+2 isotopologue of citrate, glutamate, and malate were significantly enriched in HCC1954 and SKBR3 Lat cells compared to Pa cells (Fig. 2c–e and Extended Data Fig. 3e–g ). Also, palmitate-derived carbons were enriched in the GSH (reduced glutathione) and GSSG (oxidized glutathione) forms of glutathione in Lat cells (Fig. 2f,g and Extended Data Fig. 4a,b ). Steady-state glutathione quantification indicated an elevated GSH/GSSG ratio in Lat cells with reduced cellular ROS (Fig. 2h and Extended Data Fig. 4c ). As noted earlier, Lat cells are better equipped to store excess FAs as LDs and survive than parental cells (Fig. 2i,j and Extended Data Fig. 4d ). Thus, latent cells oxidize internalized FAs, which then undergo oxidation through the TCA cycle and maintain cellular redox homeostasis that is critical for metastasis initiating cells. DRP1-driven mitochondrial dynamics enable FAO and redox homeostasis As noted earlier, HCC1954 and SKBR3 latent cells had augmented levels of mitochondrial fission protein DRP1 and the activated phosphorylated form p-DRP1 S616 (Fig. 1j and Extended Data Fig. 2j ) that promotes DRP1 translocation to mitochondria and mitochondrial fission (Extended Data Fig. 4e ). Moreover, punctate mitochondria are associated with increased FAO 35 . Therefore, we checked to see if altered mitochondrial dynamics facilitated increased FAO and metabolic reprogramming in Lat cells. To test this possibility, we generated DRP1-depleted HCC1954 and SKBR3 Lat cells using doxycycline (dox)-inducible small hairpin RNA (shRNA) (Extended Data Fig. 4f and Supplementary Table 3 ) targeting DNM1L (dynamin 1 like) gene that encodes DRP1 protein. As expected, transmission electron microscope (TEM) and IF analyses show DRP1 depletion in Lat cells results in increased mitochondrial length, tubular mitochondria and reduced number of punctate mitochondria compared to controls (Fig. 3a–c and Extended Data Fig. 4g ). LD accumulation was augmented in DRP1-depleted latent cells (Extended Data Fig. 4h,i ). In agreement, lipidomic profiles indicated increased FA content in DRP1-depleted latent cells (Extended Data Fig. 4j and Supplementary Table 4 ). Fig. 3: DRP1-driven mitochondrial dynamics enable FAO and redox homeostasis. a , TEM images showing LD and mitochondria in Ctrl and DRP1-depleted HCC1954 Lat cells. b , IF images highlighting altered mitochondrial dynamics in Ctrl and DRP1-KD Lat cells. c , Classification of cells based on percentage distribution of tubular, intermediate and punctate mitochondrial morphology in Ctrl ( n = 5) and DRP1-KD ( n = 4) HCC1954 Lat cells. d , LC–MS data showing enrichment of M+2 isotopologue of citrate, glutamate, malate, GSSG and GSH from 13 C 16 -palmitic acid in Ctrl and DRP1-KD HCC1954 Lat cells. n = 4, each group. e , Flow cytometry analysis of cellular ROS (CellROX Deep Red staining) in Ctrl and DRP1-depleted Lat cells ( n = 4, each group). f . Flow cytometry analysis showing APC Annexin V/PI + cells in HCC1954 Ctrl and DRP1-depleted Lat cells ( n = 4, each group) treated with 100 μM palmitic acid for 48 h. g , Heatmap of neutral lipids in Lat cells showing differential FA content upon CPT1A depletion in HCC1954 Lat cells ( n = 3). Morpheus, Broad Institute software was used for generating heatmap. h , IF images showing mitochondrial morphology in Ctrl and CPT1A knockdown HCC1954 Lat cells. i , Quantification of mitochondrial morphology (percentage distribution of tubular, intermediate and punctate mitochondria) in Ctrl and CPT1A-depleted Lat cells, n = 5, each group. j , Labeling of M+2 isotopologue of citrate, glutamate, malate, GSSG and GSH from 13 C 16 -palmitate in Ctrl and CPT1A-depleted HCC1954 Lat cells ( 13 C 16 -palmitate tracing in DRP1 and CPT1A-depleted cells was performed in a single experimental set up with same control). n = 4, each group. k , CellROX Deep Red staining showing ROS in Ctrl and CPT1A-depleted HCC1954 Lat cells, n = 4, each group. l , MTT cell viability assay showing survival of Ctrl and CPT1A-depleted HCC954 Lat cells grown in palmitic acid (100 μM) for 48 h. n = 8, each group. In c – g and i – l , ‘ n ’ represents biologically independent samples, and data are presented as mean ± s.e.m. P values were calculated by two-tailed unpaired t -test ( d – f ) or ordinary one-way ANOVA ( j – l ). The experiments shown in a , b and h were repeated independently at least two or three times with similar results. Source data Full size image Next, we asked whether altered mitochondrial dynamics impacts FA uptake and oxidation in Lat cells. 13 C 16 palmitate tracing in control and DRP1-depleted Lat cells indicated no notable differences in palmitate uptake. However, DRP1 depletion led to reduced synthesis of carnitine-conjugated FAs (Extended Data Fig. 5a ) and attenuated fractional enrichment in M +2 isotopologue of citrate, glutamate and malate (Fig. 3d and Extended Data Fig. 5b-g ). Moreover, DRP1 depletion drastically reduced the levels of glutathione GSH and GSSG (Fig. 3d and Extended Data Fig. 5h–k ). In agreement with these observations, DRP1-depleted Lat cells had a significant drop in steady-state glutathione and increased cellular ROS (Fig. 3e and Extended Data Fig. 5l ). Moreover, DRP1 depletion results in decreased cell viability due to increased cell death upon palmitate treatment and attenuated formation of oncospheres (Fig. 3f and Extended Data Fig. 5m–q ) and administration of antioxidant N-acetyl- l -cysteine (NAC) was able to rescue these observed differences (Extended Data Fig. 5r ). These results suggests that DRP1-driven mitochondrial dynamics enable FAO and redox homeostasis in Lat cells. As we observed, DRP1-driven mitochondrial plasticity promotes FAO and helps survival of Lat cells. We checked the effect of enriched CPT1A expression in Lat cells by knocking down CPT1A with dox-inducible shRNA (Extended Data Fig. 6a and Supplementary Table 3 ). Consequences of limiting CPT1 function are increased accumulation of LDs and lipotoxicity 38 , 39 . Indeed, oil red staining showed an increase in number of LDs (Extended Data Fig. 6b,c ). Moreover, FA profiles also indicated an increase in FA content upon CPT1A depletion (Fig. 3g and Supplementary Table 4 ). Noticeably distinct changes in mitochondrial dynamics, a notable reduction in punctate mitochondria and increased mitochondrial length or tubular mitochondria, were observed upon CPT1A depletion (Fig. 3h,i ). Moreover, p-DRP1 S616 levels were attenuated upon CPT1A depletion (Extended Data Fig. 6d ). 13 C 16 palmitate tracing showed no substantial differences in palmitate uptake between Ctrl and CPT1A-depleted cells. However, synthesis of carnitine-conjugated FAs and labeling of M+2 isotopologue of citrate, glutamate, and malate was significantly attenuated upon CPT1A depletion in Lat cells (Fig. 3j and Extended Data Fig. 6e–k ). CPT1-depleted HCC1954 and SKBR3 Lat cells had a drastic reduction in glutathione (GSH and GSSG) levels (Fig. 3j and Extended Data Fig. 6l–o ). Further phenocopying DRP1 depletion, reduced ability to form oncospheres, increased cellular ROS levels, and increased cell death upon palmitate administration were observed in CPT1-depleted Lat cells (Fig. 3k,l and Extended Data Fig. 6p,q ). DRP1-driven mitochondrial plasticity promotes metastatic latency To assess the effect of DRP1 depletion on metastatic latency, we intracardially injected Ctrl and DRP1 knockdown Lat cells into mice. One week after injection, mice were supplemented with a dox diet to deplete DRP1. Five weeks after injection, we euthanized mice and collected brains and performed immunohistochemical (IHC) analysis to assess the number of surviving latent cells. Compared to controls, mice bearing DRP1-depleted HCC1954 Lat cells had a significant reduction in the number of GFP + metastatic lesions (Fig. 4a ). In contrast, no significant difference in primary tumor (PT) burden (tumor weight and tumor volume) was observed upon DRP1 depletion (Extended Data Fig. 6r ). Previously we reported natural killer (NK) cells can limit the proliferation of Lat cells and enforce metastatic latency. Depletion of NK cells with anti-asialo-GM1 polyclonal antibody after dissemination in athymic mice bearing Lat cells results increased metastatic outbreaks 23 . No overt metastasis was observed upon NK cell depletion in mice injected with DRP1-depleted Lat cells (Fig. 4b,c ), indicating elimination of residual disease. Fig. 4: DRP1-driven mitochondrial plasticity promotes metastatic latency. a , Brain serial section and quantification of GFP + brain tropic latent metastatic cells/lesions in Ctrl and DRP1-depleted HCC1954 Lat cells, n = 4, each group. Illustration was made using BioRender.com. BLI, bioluminescence imaging. b , c , Ex vivo brain images and bioluminescence signal showing brain-only photon flux in Ctrl ( n = 6), Ctrl+ anti-asialo-GM1( n = 6), KD DRP1 (n = 8), and KD DRP1 + anti asialo-GM1 ( n = 8). Here, mice were injected with Ctrl and DRP1-depleted Lat cells, followed by vehicle or anti asialo-GM1 treatment. d , Brain serial section and quantification of GFP + brain tropic latent metastatic cells/lesions in Ctrl and CPT1A knockdown HCC1954 Lat cells, n = 4, each group. e , f , Showing oncosphere image and quantification in Ctrl, DRP1-depleted and DRP1-rescued HCC1954 Lat cells. n = 6, each group. g , IF images showing mitochondrial morphology in Ctrl, DRP1-depleted and DRP1-rescued HCC1954 Lat cells. h , Brain serial section and quantification of GFP + brain tropic latent metastatic cells/lesions in HCC1954 Ctrl, DRP1-depleted and DRP1-rescued Lat cells, n = 4, each group. In a , c , d , and h , ‘ n ’ represents the number of mice, and in f represents biologically independent samples, with data presented as mean ± s.e.m. P values were calculated by two-tailed unpaired t -test ( a ), two-tailed Mann–Whitney U -test ( c ) or ordinary one-way ANOVA ( d , f , h ). The experiment shown in g was repeated independently two times with similar results. Source data Full size image As CPT1A loss phenocopies DRP1 depletion, we assessed the effect of CPT1A depletion on metastatic latency by injecting mice intracardially with either Ctrl or CPT1A-depleted HCC1954 and SKBR3 Lat cells. CPT1 depletion was induced by administrating dox diet a week after injection. Analogous to DRP1 depletion, CPT1A depletion resulted in a significant reduction in the number of GFP + metastatic events in mice brains (Fig. 4d and Extended Data Fig. 6s ). Next, we performed DRP1 rescue experiments to show DRP1 is essential for survival of Lat cells. For these studies, we generated inducible HA-tagged DNM1L full-length constructs and transduced their expression in DRP1-depleted Lat cells (Extended Data Fig. 6t and Supplementary Table 5 ). Inducible DRP1 expression resulted in reversing the mitochondrial morphology and augmented the oncosphere-forming ability in DRP1-depleted cells (Fig. 4e–g ). As expected, increased number of surviving Lat cells were observed in the brain upon DRP1 rescue (Fig. 4h ). Together, our data demonstrate that DRP1-dependent mitochondrial plasticity facilitates FAO and promotes survival of latent metastatic cell in mice. Phospho-DRP1 is elevated in human metachronous brain metastases High DRP1 expression correlated with poor distant metastasis free survival in HER2 + breast cancer patients (Extended Data Fig. 7a ). Obtaining biopsies from patients who show no detectable disease after their primary diagnosis is impractical. Therefore, we investigated p-DRP1 status in metachronous brain metastatic lesions from HER2 + breast cancer patients 6 , who had undergone a considerable latent metastatic phase (Supplementary Table 6 ). IHC analysis of matched PT and metachronous brain metastatic lesions from seven HER2 + breast cancer patients revealed high expression of p-DRP1 S616 in brain metastatic lesions compared to corresponding matched PTs (Fig. 5a,b ). Likewise, increased DRP1 and p-DRP1 S616 expression and altered mitochondrial dynamics were observed in metachronous brain metastatic (M-BM) cells compared to Pa and Lat cells (Extended Data Fig. 7b–d ) 6 . Akin to Lat cells, depletion of DRP1 in M-BM cells resulted in reduced oncosphere formation which was rescued by administration of antioxidant NAC (Extended Data Fig. 7e,f ). M-BM cells predominantly metastasize to the brain with occasional metastasis to the spine/bone. Whole body, spine and brain photon flux analysis indicated a significant decrease in brain and spine/bone metastatic burden upon DRP1 depletion, resulting from increased apoptosis (Fig. 5c–e and Extended Data Fig. 7g–i ). Similar to Lat cells, no significant difference in tumor burden (tumor volume and tumor weight) was observed between orthotopically injected Ctrl and DRP1-depleted M-BM cells (Extended Data Fig. 7j ). Furthermore, ectopic expression of DRP1 in DRP1-depleted M-BM cells was able to rescue mitochondrial dynamics, oncosphere-forming ability and metastatic incidence in mice (Fig. 6a–c and Extended Data Fig. 8a,b ). Fig. 5: Phospho-DRP1 is elevated in human metachronous brain metastases. a , IHC staining for p-DRP1 S616 (1:200, DAB, 10×) in matched human HER2 + PT and metachronous brain metastases (Br. Met). b , Representative graph showing histoscore of p-DRP1 S616 in HER2 + PT and matched brain metastatic samples ( n = 7, each group). c , Bar graph showing spine and brain-only photon flux in mice bearing Ctrl ( n = 9) and DRP1-depleted ( n = 9) M-BM cells. d , e , Ex vivo brain images ( d ) and brain-only photon flux ( e ) showing metastatic burden in mice bearing Ctrl ( n = 9) and DRP1-depleted ( n = 8) M-BM cells. In b , ‘ n ’ represents number of human patients, and in c and d , ‘ n ’ represents number of mice. Data are presented as mean ± s.e.m. P value in b was calculated by two-tailed paired t -test, and in c and d , P values were calculated by two-tailed Mann–Whitney U -test. Source data Full size image Fig. 6: Genetic depletion or pharmacologic inhibition of DRP1 attenuates brain metastasis. a , IF images showing mitochondrial morphology of HCC1954 Ctrl, DRP1-depleted and DRP1-rescued M-BM cells. b , c , Ex vivo brain images ( b ) and brain-only photon flux ( c ) showing metastatic burden in mice bearing Ctrl ( n = 10), DRP1-depleted ( n = 9) and DRP1-rescued ( n = 9) M-BM cells. d , Quantification of GFP + brain tropic latent metastatic cells/lesions in mice bearing HCC1954 Lat cells treated with vehicle (10% DMSO in corn oil) and Mdivi-1 (40 mg kg −1 ) for 4 weeks (once daily). n = 4, each group. e , f , Ex vivo brain images ( e ) and brain-only photon flux ( f ) showing metastatic burden in mice bearing M-BM cells. After injection of M-BM cells, mice were treated with either vehicle (10% DMSO in corn oil, n = 8), or Mdivi-1 (40 mg kg −1 , n = 7) for 4 weeks (once daily). In c , d and f , ‘ n ’ represents number of mice, and data are presented as mean ± s.e.m. P values in c , d and f were calculated by Kruskal–Wallis test, two-tailed unpaired t -test and two-tailed Mann–Whitney U -test, respectively. The experiment shown in a was repeated independently two times with similar results. Source data Full size image Pharmacologic inhibition of DRP1 attenuates brain metastasis Mitochondrial division inhibitor 1 (Mdivi-1) is reported to inhibit DRP1-dependent mitochondrial fission and attenuate neuronal apoptosis in several models of brain ischemia and neurodegeneration 40 , 41 , 42 . As Mdivi-1 is brain-permeable and not neurotoxic, we assessed its effect on Lat and M-BM cells. Similar to DRP1 depletion, Mdivi-1 treatment resulted in increased LDs and reduced viability of HCC1954 and SKBR3 Lat and M-BM cells (Extended Data Fig. 9a–f ). Treatment with Dynasore, which interferes with the GTPase activity of the mitochondrial dynamin DRP1 and dynamin 1 and dynamin 2, but not of other small GTPases, resulted in increased LD and reduced cell viability in Lat and M-BM cells (Extended Data Fig. 9g–i ) 43 . Next, we assessed the effect of Mdivi-1 on brain metastasis. To perform this analysis, we injected HCC1954 Lat and M-BM cells intracardially into athymic mice. One week after injection, we administrated Mdivi-1 orally every day for 4 weeks. No significant differences in mice body weights were observed during the course of treatment (Extended Data Fig. 10a ). IHC and bioluminescence imaging analyses showed that DRP1 inhibition results in a significant reduction of the number of surviving latent cells and attenuated brain metastatic incidences in mice bearing M-BM cells (Fig. 6d–f and Extended Data Fig. 10b,c ). Thus, limiting DRP1 function and consequently impeding mitochondrial plasticity attenuates survival of latent and metachronous brain metastatic cells. As depletion of DRP1 in Lat and M-BM cells resulted in reduced oncospheres formation and reduced brain metastatic potential of Lat and M-BM cells, we also assessed the impact of the DRP1 inhibitor alone and in combination with the current standard of care 44 , 45 . HCC1954 and SKBR3 Lat and M-BM cells were resistant to the HER2 tyrosine kinase inhibitors (TKIs) lapatinib or tucatinib as single agents compared to Pa cells. In contrast, Mdivi-1 was effective in limiting oncosphere formation in HCC1954 and SKBR3 Lat and M-BM cells alone and in combination with HER2 TKIs (Extended Data Fig. 10d,e ). In summary, our study highlights the vital role of mitochondrial plasticity in facilitating the oxidation of FAs in latent metastatic cells. Depleting or pharmacologically inhibiting DRP1 in preclinical breast cancer metastasis models disrupts mitochondrial dynamics, cellular bioenergetics and redox homeostasis, resulting in attenuated brain metastasis (Fig. 7 ). These findings suggest that targeting mitochondrial plasticity is a promising therapeutic approach to prevent metastatic recurrences. Fig. 7: Schematic presentation highlighting the role of DRP1-driven mitochondrial plasticity and metabolic reprogramming in HER2 + breast cancer brain metastasis. Illustration was made using BioRender.com. Full size image Discussion Brain metastases are lethal, and their treatment remains an unmet clinical need. Current systemic therapies are primarily directed toward inhibiting oncogenomic alterations or reactivating immune surveillance to limit disease progression. Resistance to targeted therapies are widely observed, and durable responses to immunotherapies are not universal. Moreover, although PTs and metastases share common ancestral driver gene mutations, they evolve independently 46 , 47 , 48 . Reversible adaptations that promote survival and organ colonization are critical for establishing metastasis. We report metabolically flexible, latent disseminated tumor cells rely on mitochondria, the cellular powerhouse and a dynamic organelle, to utilize astrocyte secreted FAs. Fragmented mitochondrial puncta facilitate FAO that supports cellular bioenergetic energy needs and redox homeostasis that is critical for survival of disseminated latent metastatic cells. In agreement with these findings, brain tumor-initiating cells have also been reported to be enriched for fragmented mitochondria, and targeting DRP1 results in increased cell death and reduced tumorigenesis. Moreover, DRP1 activation correlates with poor prognosis in glioblastoma and breast cancers 7 , 8 , 49 . These observations highlight targeting mitochondrial dynamics is a viable therapeutic opportunity to limit both brain tumors and metastasis. Slow-cycling or quiescent latent metastatic cells are also functionally plastic, as they retain ability to proliferate, initiate metastasis and colonize distal organs. Therefore, latent metastasis-promoting alterations are also likely to be reversible and not necessarily driven by fixed oncogenic alterations or signaling responses. Thus, mitochondrial plasticity enables FAO and metabolic programming in response to cellular stress, nutrient state and energy needs and thereby facilitating cellular adaptation is critical for survival of latent disseminated tumor cells. In addition, a consequence of increased FAO is increased pools cellular acetyl-CoA that could reversibly alter signaling responses and epigenomic landscape in metastatic cells 50 . Overcoming technical limitations to accurately characterize the metabolic state of disseminated latent cells in both mice and humans remain a significant challenge. In this study, we used in vitro metabolic tracing studies to gain a better understanding of the metabolic state of latent cells. However, to further advance our understanding, longitudinal experimental assays to assess the metabolic state of latent and overt metastatic cells are needed. Such assays would provide valuable insights into the metabolic flexibility of disseminated cancer cells and help in developing targeted therapeutic strategies. Likewise, increased DRP1 acetylation is also implicated in promoting mitochondria fission by promoting DRP1 phosphorylation and translocation 33 . Here, we identified several acetylation sites on DRP1 in brain tropic cells. How these modifications affect DRP1 function and the role of mitochondrial dynamics in mediating epigenomic alterations and metastatic latency needs further investigation. Abnormal mitochondrial dynamics are associated with neurodegenerative diseases, aging and cancer 51 , 52 , 53 . Increased mitochondrial fission in neurons results in mitochondrial dysfunction, and limiting mitochondrial fission results in improved synaptic function and reduced cognitive decline in preclinical models 54 . Moreover, inhibition of mitochondrial fission can also protect myocardial ischemia 33 , 55 , 56 and acute kidney ischemia 57 , 58 . On the other hand, disseminated Lat breast cancer cells and glioma-initiating cells display fragmented mitochondrial morphology, and DRP1 inhibition results in increased tumor cell death 49 . Thus, targeting DRP1 and mitochondrial fission may have broader therapeutic applicability to cancers with a propensity to disseminate and colonize brain. We show pharmacological inhibition of mitochondrial plasticity using DRP1 inhibitor Mdivi-1 limits residual disease and delays metastatic relapses in preclinical models. Likewise, high-potency small-molecule DRP1 GTPase inhibitors (Drpitor1, Drpitor1a) could also be effective in limiting brain metastasis 59 . Finally, these findings may also have broader therapeutic application in other cancer types with a tendency to metastasize to the brain. Methods Ethics statement The University of Texas Southwestern Medical Center Institutional Animal Care and Use Committee approved this study (Animal Protocol no. 2017-102099). All animal studies were conducted in compliance with the ethical guidelines. The maximal tumor size of animal experiments was 2,000 mm 3 , and all experiments did not exceed this limit. Patients consented to use of any deidentified tumor tissues for research purposes under Vanderbilt Ingram Cancer Center institutional review board-approved protocol (Vanderbilt Ingram Cancer Center no. 160606). Validation studies were performed on deidentified tumor tissues (Supplementary Table 6 ). Patients were not recruited for this specific study. Ages of the patients were undisclosed. Cell lines HCC1954 (Pa; ATCC, Cat. CRL-2338, Lat, M-BM) and SKBR3 (Pa; ATCC, Cat. HBT-30, Lat, M-BM) cells were cultured in RPMI-1640 (Cat. R8758, Sigma-Aldrich) media supplemented with 10% FBS, 2 mM glutamine, 100 units/l –1 penicillin, 100 mg l −1 streptomycin and 1 μg ml −1 amphotericin B. Cells were maintained in 37 °C incubator with 5% CO 2 and split every 3 days at 1:4 dilution 6 . Lenti-X-HEK 293 T (ATCC, Cat. CRL-1573) cells were grown in DMEM, high-glucose media supplemented with 1 mM sodium pyruvate, 100 U l −1 penicillin, 100 mg l −1 streptomycin and 10% FBS. Astrocytes were cultured in poly- l -lysine-coated plates with DMEM high-glucose media, (Cat. D6429, Sigma-Aldrich) supplemented with 10% heat-inactivated FBS (Cat. 10437028, Thermo Fisher Scientific), 100 U l −1 penicillin, 100 mg l −1 streptomycin (Cat. P0781, Sigma-Aldrich) and astrocyte growth supplement (Cat. 1852, ScienCell Research Laboratories). Animals Female athymic mice (Hsd: athymic nude mice-Foxn1nu) aged 4–5 weeks were purchased from Envigo (Cat. 069) and allowed to acclimate to the animal facilities before experimental procedures. The mice were housed in a UT Southwestern animal facility room under a 12/12-h light/dark cycle, at a temperature of 20–26 °C with 30–50% humidity and provided with ad libitum access to a standard Teklad diet (Envigo, Cat. 2916) and water. A dox diet (Envigo, Cat. TD.08541) was given to induce shRNA-mediated specific knockdown or overexpression of the gene of interest in tumor cells. We monitored animal health once daily throughout the experiment timeline. To isolate astrocytes, we obtained 3- to 4-day-old NSG mice (NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ; Cat. 005557, Jackson Laboratory) pups from the UTSW animal breeding facility. Isolation and culture of mouse cortical astrocytes NSG mice pup brains aged 3 or 4 days were isolated and kept in chilled 1× HBSS buffer. Using a stereomicroscope, meninges were peeled off from the cortex, and the cortex hemispheres were extracted, cut into small pieces and dissociated using 1:1 HBSS:trypsin (0.125% final) for 6–8 min at 37 °C with occasional shaking. The cells were centrifuged at 400 g for 5 min and then grown in poly- l -lysine-coated T75 flasks using astrocyte culture media. The media was changed daily for the first 3 days to remove cellular debris. BODIPY-558/568-C12 transfer assays Astrocytes were cultured for 48 h in six-well plates with complete media, followed by treatment with 2 μM BODIPY-558/568 (Cat. D3835, Thermo Fisher Scientific) for 24 h. After washing the plates three times in warm media, the cells were incubated for 1 h in R3F (RPMI 3% FBS, 10 mM glucose and 0.2 mM glutamine) media. Cancer cells, cultured in R3F for 48 h, were detached by trypsinization and labeled with green cell tracker dye (Cat. C7025, Thermo Fisher Scientific) and co-cultured with astrocytes in R3F media. Live cell time-lapse imaging was performed using a Nikon CSU-W1 Spinning Disk Confocal at 37 °C and 5% CO 2 . BODIPY-558/568-C12 pulse-chase assays Cancer cells were cultured in R3F (six-well plates) media for 24 h, media was discarded and replenished with fresh R3F media (phenol red-free) and then treated with 1 μM BODIPY-558/568 (Cat. D3835, Thermo Fisher Scientific). Time-lapse imaging was performed using an IncuCyte3 imager for 24 or 48 h. Transmission electron microscopy Briefly, 2 × 10 5 HCC1954 Pa and Lat cells were cultured in MatTek dishes for 24 h with R3F media followed by treatment of sodium palmitate (100 µM) for 24 h. Cells were fixed on MatTek dishes with 2.5% (v/v) glutaraldehyde in 100 mM sodium cacodylate buffer. Samples were processed following the method described previously 60 . To acquire the images, we used a Tecnai G2 Spirit TEM (FEI) with a LaB6 source operating at 120 kV, using a Gatan Ultrascan CCD camera. IF imaging for mitochondria visualization To visualize mitochondria, 2 × 10 5 cells were cultured on coverslips and immune-stained with anti-TOMM20 antibody. Cells were fixed in 4% paraformaldehyde, permeabilized with 0.1% Triton-X in PBS, blocked with 3% BSA (Cat. A9647, Sigma-Aldrich), and incubated with the primary antibody. DRP1 mitochondrial localization was observed in live cells stained with MitoTracker Deep Red FM (Cat. M22426, Thermo Scientific) before fixation. after primary antibody staining coverslips were washed and incubated with Alexa Fluor conjugated secondary antibody (antibody information is included in Reporting Summary). Coverslips were mounted and visualized with an EVOS fluorescence microscope. Steady-state carnitine-conjugated FA analysis Next, 3 × 10 6 cells were plated in R3F media (10 cm dish) and grown for 24 h at 37 °C and 5% CO 2 . After aspirating out the media, cells were washed with 1 ml ice-cold 0.9% NaCl solution and then scraped in 1 ml chilled 80% methanol. The cell lysate/methanol mixture was subjected to three freeze/thaw cycles using liquid nitrogen and a 37 °C water bath, vortexed for 1 min and centrifuged at 13,000 g for 10 min in a refrigerated centrifuge. The supernatant was collected, dried using a SpeedVac and subjected to liquid chromatography–mass spectrometry (LC–MS) analysis. Cell viability assays To determine cell proliferation in various conditions, MTT (Cat. M5655, Sigma-Aldrich) assay was performed 61 . CellTiter-Glo luminescent cell viability assay (Cat. G9241, Promega) and Trypan blue exclusion assays were performed as per manufacturer’s protocol. ROS measurement Next, 5 × 10 5 cells were plated in 60-mm plates for 24 h followed by treatment of 100 μM palmitic acid conjugated with BSA for 24 h. Cells were collected and washed with 3% FBS in PBS and incubated with 5 μM CellROX Deep Red reagent (Cat. C10422, Thermo Fisher Scientific) for 30 min. Postincubation cells were washed twice and fixed with 4% paraformaldehyde for 15 min. Next, cells were washed twice and resuspended in PBS with 3% FBS for flow cytometric analysis (FACSCalibur-Becton Dickinson). Analysis was performed using FlowJo v10 software (Supplementary Fig. 2 ) APC Annexin V/PI staining Briefly, after treatment, 0.5 × 10 6 cells were washed in ice-cold 1× PBS and resuspended in 100 µl staining buffer and incubated with 5 µl APC Annexin V (Cat. 640932, BioLegend) and 10 µl PI for 15 min at room temperature in a dark place. Flow cytometric analysis was immediately performed using FACSCalibur (Becton Dickinson). Analysis was performed using BD FACSDiva v9.0.1 software (Supplementary Fig. 3 ). Colorimetric caspase-3 activity assay Colorimetric caspase-3 activity assay was performed using a kit (Cat. ab39401, Abcam). Briefly, cells were grown without palmitate, and cell lysates were prepared using RIPA buffer. An aliquot of 300 µg protein was used for each condition. DTT and DEVD-p-NA substrates were added according to the manufacturer’s protocol. Samples were mixed well and incubated at 37 °C for 90 min, and finally, OD was measured at 405 nm. Total FA profiling in tumor cells by GC–MS Total FA profiles of neutral lipid content was generated by a modified gas chromatography mass spectrometry (GC–MS) method previously described 62 . Samples were analyzed on an Agilent 7890/5975 C by electron capture negative ionization equipped with a DB-5MS column (40 m × 0.180 mm with 0.18 µm film thickness) from Agilent. Hydrogen was used as the carrier gas. FAs were analyzed in selected ion monitoring mode and normalized to internal standards. Data were processed using MassHunter v.B08 software (Agilent). To obtain the neutral lipid content, a three-phase liquid extraction (3PLE) protocol was used 63 . BSA-palmitic acid conjugation For BSA-palmitic acid conjugate preparation, prewarmed RPMI-1640 media (Biological Industries, Cat. SKU: 01-101-1 A) was stirred at 37–40 °C on a heated magnetic stirrer. Next, 10% BSA (Cat. A6003, Sigma-Aldrich) was added to achieve a final BSA concentration of 0.8%. A stock of 10 mM sodium palmitate (Cat. P9767, Sigma-Aldrich) was prepared by dissolving it in isopropanol: water (1:1). Palmitate was added to the RPMI-1640-BSA media while maintaining the final concentration of isopropanol to 0.5% and palmitate to 100 μM. The solution was stirred for 1 h at 37–40 °C, followed filtration. The same protocol was applied for the preparation of BSA and 13 C 16 palmitate (Cat. CLM-6059, Cambridge Isotope Laboratories) conjugate. In vitro 13 C 16 palmitic acid tracing For in vitro isotope tracing experiments, 1 × 10 6 cells were grown in 6-cm dishes with 4 ml R3F media for 24 h. The next day, media was discarded and replenished with R3F palmitic acid media (R3F, 100 μM 13 C 16 palmitic acid, 0.8% BSA, 0.5% isopropanol) for 24 h. After incubation, media supernatant was collected, and cells were rapidly washed using 400 μl ice-cold 0.9% NaCl and fixed with 300 μl of 80% chilled LC–MS grade methanol in water. Cells were scraped, collected and preserved at −80 °C. Collected samples were subjected to three rounds of freezing/thawing using liquid nitrogen and 37 °C water bath. Supernatant was collected by centrifuging at 13,000 g for 10 min and lyophilized using a SpeedVac (Thermo). Lyophilized samples were resuspended in 100 μl 80:20 acetonitrile/water solution (HPLC grade) for LC–MS analysis. FAO assay To assess FAO, SKBR3 Pa and Lat cells were seeded into 96-well plates. Radiolabeled 1- 14 C palmitic acid (1µCi ml −1 ) was given in PBS supplemented with 10 mM HEPES and 1 mM L-carnitine. Nonradiolabeled substrates were added to obtain final concentrations of palmitic acid (100 µM) and glucose (10 mM). CCCP (3 µM) was added for uncoupled substrate oxidation. CO 2 was captured in an activated UniFilter-96w GF/B microplate with 1 M NaOH (25 μl per well). MicroBeta2 Microplate Counter (PerkinElmer) was used to measure the radioactivity after addition of scintillation liquid (30 µl; MicroScint PS PerkinElmer). Nile red and Oil Red O staining for LDs For Nile red staining, 4% paraformaldehyde fixed cells were washed and stained with 2 µg ml −1 Nile red for 10 min. 5 mg ml −1 Oil Red O stock solution was prepared by dissolving it in isopropanol. To make working solution, Oil Red O stock solution was added to distilled water in a ratio of 3:2, kept for 10 min and filtered using Whatman filter paper and 0.45-µm syringe filter. Next, Oil Red O working solution was added to the fixed cells for 10 min and rinsed with distilled water. Cells were washed three times, and DAPI staining was done to visualize nucleus. Oncosphere assays To generate oncospheres, 50–100 cells were resuspended in HuMEC serum-free media supplemented with bFGF (10 ng ml −1 ), EGF (20 ng ml −1 ), insulin (5 μg ml −1 ) and 1xB27 supplement. The suspension was incubated at 37 °C and 5% CO 2 for 7–8 days. The oncospheres were imaged using EVOS fluorescence microscope. Percentage oncosphere formation was determined either by quantification or CellTiter-Glo assay. Western blotting Cells cultured in R3F for 48 h were lysed with RIPA lysis and extraction buffer added with protease and phosphatase inhibitor. For immunoblotting, 20–35 µg protein was resolved in 7.5–12% SDS-PAGE, transferred onto nitrocellulose membrane (Millipore), using a semi-dry transfer apparatus and blocked in 5% BSA or TBST-milk for 1–2 h. Membranes were washed incubated with primary antibody for overnight at 4 °C on a shaker. Membranes were washed with TBST (three or four times for 10 min), and secondary antibody was added. Next, washing was performed and membranes were developed using West Femto Super Signal (Thermo Fisher Scientific) and a Bio-Rad imager. Images were exported with Image Lab v6.1 software. Detail information about antibodies used for western blotting is included in Reporting Summary. Steady-state glutathione measurement The level of total glutathione reduced (GSH) and oxidized glutathione (GSSG) was measured using a kit (Cat. K264, BioVision) according to the manufacturer’s protocol. All data presented were normalized to 10 6 cells. Paraffinized brain tissue section staining Paraffinized brain tissue section staining was performed as described before 6 . To prepare the samples for IHC analysis, 5- to 10-µm sections were cut and mounted on slides. The slides were then incubated at 65 °C for 30 min, followed by deparaffinization and rehydration through sequential incubations in xylene (two times for 3 min), 100% ethanol (two times for 3 min), 95% ethanol (two times for 3 min), 70% ethanol (two times for 3 min), 50% ethanol (two times for 3 min) and water (two times for 3 min).Antigen retrieval was carried out by placing the slides in 700 ml antigen retrieval buffer (10 mM Tris HCl, 1 mM EDTA, 10% glycerol, pH 9.0) at 95 °C for 25 min. The slides were then cooled to room temperature for 30 min and rinsed with PBS (two times for 2 min). To block nonspecific binding, the slides were treated with 10% horse serum in 0.1% PBS-T for 1 h at room temperature. The slides were incubated with Anti-GFAP at a dilution of 1:500 in 2% serum PBS-T overnight at 4 °C. The next day, the slides were washed in PBS (three times for 10 min) and incubated with secondary antibody for 1 h at room temperature. Slides were washed with PBS (two times for 10 min) and mounted following Hoechst 33342(Cat. 62249, Thermo Fisher Scientific) staining. IHC analysis of patient samples was performed using a Dako Autostainer Link 48 system. The slides were baked at 60 °C for 20 min, followed by deparaffinization and hydration. Antigen retrieval was performed at pH 9.0 for 20 min using the Dako PT Link. The tissue was then incubated with a peroxidase block and anti-phospho-DRP1 S616 antibody (Cat. 4494, Cell Signaling Technology) at a dilution of 1:200 for 35 min. Finally, the staining was visualized using the EVOS microscope. Mass spectrometry for DRP1 posttranslational modifications Experiment was performed as described previously 64 . Briefly, DRP1 pull down protein samples from Lat cells were digested overnight with trypsin (Pierce) after reduction and alkylation with DTT and iodoacetamide (Sigma-Aldrich). The resulting samples were subjected to solid-phase extraction cleanup with an Oasis HLB plate (Waters) and then injected onto an Orbitrap Fusion Lumos mass spectrometer coupled to an Ultimate 3000 RSLC-Nano liquid chromatography system. The samples were injected onto a 75 µm inner-diameter, 75-cm-long EasySpray column (Thermo Fisher Scientific) and eluted with a gradient from 1% to 28% buffer (80% (v/v) ACN, 10% (v/v) trifluoroethanol and 0.1% formic acid in water) over 90 min. The positive ion mode was used in operating the mass spectrometer, using a source voltage of 1.8 kV and maintaining an ion transfer tube temperature of 275 °C. The Orbitrap was used to acquire MS scans at a resolution of 120,000. Peptide identification was performed using Proteome Discoverer v2.4 SP1 (Thermo Fisher Scientific), with Sequest HT searching against the human protein database from UniProt (downloaded April 8, 2022; 20,361 entries) with a false-discovery rate cutoff of 1% for all peptides. Phosphorylation of serine, threonine, tyrosine and acetylation of lysine were set as variable modifications. Virus production and infection for knockdown and rescue To generate shRNA knockdown cells, Lenti-X 293 T cells were prepared and co-transfected with either pTRIPZ or pTRIPZ-shRNA of target gene CPT1A (Clone Id: V3THS_359757, V3THS_359760, Horizon Discovery), DRP1 (Clone Id: V2THS_29084, only commercially available and validated construct in Horizon Discovery with high efficiency) with packaging plasmid PAX (7 µg) and envelop plasmid MD2.G (2.7 µg) by lipofection method using lipofectamine 3000 (Invitrogen). Eighteen hours after lipofection, media was replenished with fresh media. Virus particles were harvested after 48-h incubation and precipitated using PEG-it virus precipitation solution (Cat. LV810A-1, System Biosciences). Viral precipitates were resuspended with 100 µl Iscove’s modified Dulbecco’s medium (312440053, Thermo Fisher Scientific). Cancer cells cultured overnight were transduced with 20 µl virus solution in 2 ml growth media with 5 µg ml −1 polybrene. After 6-h incubation, cells were replenished with 2–3 ml fresh growth media and incubated for 24 h. Knockdown was induced by 1 µg ml −1 dox hydrochloride (Cat. 3072, Sigma-Aldrich) supplementation. For DRP1 rescue in DRP1-depleted cells, synthetic Homo sapiens dynamin 1-like ( DNM1L ), transcript variant 5 (the shRNA targeting site was codon optimized without changing the amino acids to minimize the shRNA binding; Supplementary Table 5 ) was procured from Integrated DNA Technology and cloned into pINDUCER21 (plasmid 46948) where DRP1 was in frame with HA tag. Virus was produced and infected as described above. Intracardiac injections and in vivo mice studies This experiment was performed as described before 6 , 23 , 24 . Briefly, 5 × 10 4 cells were resuspended in 100 μl 1× PBS were intracardially injected into the left ventricle of mice with the help of 26 G tuberculin syringe. Weekly monitoring of tumor growth and the occurrence of metastasis was measured through bioluminescence imaging. For dox-inducible gene depletion experiments, female athymic mice (Hsd: athymic nude mice-Foxn1nu; Envigo) aged 5–6 weeks were injected intracardially either with TRIPZ control or shRNA or rescue cells. Following a week of injection, mice were fed with dox diet and monitored metastatic incidence using noninvasive small animal imager AMI-HTX. Data were analyzed by aura spectral instrument imaging software (v. 4.0). NK depletion experiment Five- to six-week-old female athymic mice (Hsd: athymic nude mice-Foxn1nu; Envigo) were injected with 5 × 10 4 Ctrl or DRP1-depleted Lat cells intracardially. Postinjection mice were randomized into groups based on bioluminescence imaging signal. NK depletion was performed 1 day after injection by administrating 100 μl anti-asialo GM1 (FUJIFILM Wako Shibayagi, Cat. 986-10001, RRID:AB_516844) intraperitoneally. Orthotopic tumor xenograft model Briefly, 2×10 7 cells Ctrl and DRP1-depleted Lat and M-BM cells were resuspended in PBS and Matrigel (1:1 ratio) in 1 ml. Female athymic mice (Hsd: Athymic nude mice-Foxn1nu; Envigo) aged 5–6 weeks were anaesthetized by controlled isoflurane administration through a nose cone in a sterile hood. An incision was made between the fourth and fifth nipple of the mouse to expose the mammary fat pad, and 100 μl cell suspension was injected using a 28 G insulin syringe. Four weeks after injection, tumors were collected and tumor volume and weight was measured. Mdivi-1 oral gavage A total of 5 × 10 4 cancer cells were intracardially injected into 5- to 6-week-old female athymic mice (Hsd: Athymic nude mice-Foxn1nu; Envigo). Postinjection mice were randomized into two groups based on bioluminescence imaging signal. Mice were then administrated with vehicle control (0.2 ml of 10% DMSO in corn oil) and Mdivi-1 (Cat. S7162, Selleckchem, 40 mg kg −1 ) through oral gavage once daily for 4 weeks. Statistics and reproducibility No statistical method was used to pre-determine sample size, but our sample sizes are similar to those reported in previous publications. Data collection and analysis were not performed blind to the conditions of the experiments and outcome assessment. For in vivo experiments in necessary conditions, mice were randomly grouped. In vitro experiments were not randomized. Individual data points were represented as dots in graphs. All statistical analysis tests used in this study are indicated in the figure legends. Statistical significance between two comparative groups was determined using unpaired t -test or Mann–Whitney U -test. For statistical comparisons between multiple study groups, ordinary one-way analysis of variance (ANOVA) or Kruskal–Wallis test was used. GraphPad prism 9 software was used for all statistical analysis. No data were excluded from the analysis. Data distribution was assumed to be normal, but this was not formally tested. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Data availability The RNA-seq data that support the findings of this study have been deposited in the Gene Expression Omnibus under the accession number GSE180098 .The mass spectrometry data have been deposited in Mass Spectrometry Interactive Virtual Environment (MassIVE) with accession number MSV000091574 . Source data are provided with this paper. All other data supporting findings of this study can be available from the corresponding author on reasonable request. Code availability All of the analyses in our study were conducted using standard workflows and open source software as detailed in Methods , without any custom code being used or developed.

Mitochondria that power cellular activity fragment and change shape in breast cancer cells that migrate to the brain, an adaptation that appears necessary for the cells to survive, UT Southwestern Medical Center researchers report in a new study. The findings, published in Nature Cancer, could lead to new ways to prevent brain metastases, or the spread of cells from primary tumors to the brain. "Through mitochondrial plasticity, these cancer cells undergo metabolic reprogramming that aids their survival in the brain niche that otherwise would not be available to them. Exploiting this vulnerability could offer a way to prevent brain metastases," said study leader Srinivas Malladi, Ph.D., Assistant Professor of Pathology at UT Southwestern and a member of the Harold C. Simmons Comprehensive Cancer Center. Metastatic cancer, which is treated as stage IV cancer, is responsible for the majority of cancer deaths. The Malladi lab focuses on understanding how cells that escape from a primary tumor can live in different locations in the body, often for years, before emerging as metastatic cancer. Using breast cancer, a disease that commonly metastasizes to the brain, as a model, Dr. Malladi and his colleagues discovered that cancer cells that migrate to the brain reprogram their metabolism to depend on fatty acids rather than carbohydrates as a main energy source. This switch is necessary to survive in the brain, which is a completely different environment, Dr. Malladi explained. But how the cells accomplish this metabolic switch was unclear. To answer this question, Dr. Malladi and his team isolated latent metastatic (Lat) cells—cancer cells that had migrated from the primary tumor but had not begun actively forming new tumors—from the brains of mouse models. They observed that these Lat cells have distinctly shaped "punctate," or dot-like, mitochondria compared to the primary tumor cells with elongated tubular mitochondria. Moreover, the Lat cells readily used fatty acids. This suggested that the mitochondrial shape change or plasticity was necessary for fatty acid metabolism. Further experiments showed that the fragmentation was driven by an increase in a protein known to be involved in mitochondrial fission called dynamin-related protein 1 (DRP1). When the researchers used a genetic technique to decrease the amount of DRP1, the Lat cells' mitochondria regained their tubular shape and lost the ability to metabolize fatty acids. Similarly, when they used a chemical that inhibited DRP1, Lat cells residing in mouse brains formed fewer metastases and were significantly less likely to survive. A separate examination of metastatic tumors that formed in breast cancer patients showed that a phosphorylated form of DRP1 was elevated, suggesting that this phenomenon occurs in humans as well, Dr. Malladi said. He and his colleagues plan to test DRP1 inhibitors to determine whether they might prevent, slow, or reverse metastatic disease, an important next step toward developing a treatment.

10.1038/s43018-023-00563-6

Nano

Newly developed nanoparticles help fight lung cancer in animal model

Yang Liu et al, An inhalable nanoparticulate STING agonist synergizes with radiotherapy to confer long-term control of lung metastases, Nature Communications (2019). DOI: 10.1038/s41467-019-13094-5 Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-019-13094-5

https://phys.org/news/2019-11-newly-nanoparticles-lung-cancer-animal.html

Abstract Mounting evidence suggests that the tumor microenvironment is profoundly immunosuppressive. Thus, mitigating tumor immunosuppression is crucial for inducing sustained antitumor immunity. Whereas previous studies involved intratumoral injection, we report here an inhalable nanoparticle-immunotherapy system targeting pulmonary antigen presenting cells (APCs) to enhance anticancer immunity against lung metastases. Inhalation of phosphatidylserine coated liposome loaded with STING agonist cyclic guanosine monophosphate–adenosine monophosphate (NP-cGAMP) in mouse models of lung metastases enables rapid distribution of NP-cGAMP to both lungs and subsequent uptake by APCs without causing immunopathology. NP-cGAMP designed for enhanced cytosolic release of cGAMP stimulates STING signaling and type I interferons production in APCs, resulting in the pro-inflammatory tumor microenvironment in multifocal lung metastases. Furthermore, fractionated radiation delivered to one tumor-bearing lung synergizes with inhaled NP-cGAMP, eliciting systemic anticancer immunity, controlling metastases in both lungs, and conferring long-term survival in mice with lung metastases and with repeated tumor challenge. Introduction Immunotherapy is providing tremendous promise in the new era of cancer treatment. Checkpoint inhibitors and adoptive T cell transfer therapy have shown improved survival in melanoma, non-small-cell lung cancer, and renal cell cancer patients 1 , 2 . However, only a fraction of patients benefit from immunotherapy, and lack of specific tumor targeting is frequently associated with immune-related toxicity. Thus, many attempts are being made to improve anticancer immunity while reducing adverse effects. In contrast to systemic immunotherapy, intratumoral injection of immunomodulators is intended to focus the immune response locally on the malignancy and tumor draining lymph nodes (TDLNs) 3 . Moreover, given the heterogeneous nature of tumor antigens (TAs), intratumoral immunotherapy may have a potential for arousing a polyclonal antitumor immune response in situ against diverse cancer targets 3 , 4 . Among various types of immunomodulators, activators of the stimulator of interferon (IFN) genes (STING) pathway to elicit antitumor immunity have recently attracted much attention 5 . Being identified as a potent STING agonist, cyclic guanosine monophosphate–adenosine monophosphate (cGAMP) functions in the cytosol to ligate STING on endoplasmic reticulum (ER) membrane to activate STING pathway and type I IFN production 6 , 7 . Recent preclinical studies involving intratumoral injection of STING agonists alone or combined with irradiation (IR) have shown its aptitude for enhancing antitumor immunity 8 , 9 , 10 . Mechanistic studies indicate that activation of STING pathway within tumor-resident antigen-presenting cells (APCs) leading to type I IFNs production is indispensable for generation of adaptive immunity against tumors 11 , 12 , 13 . Although the in situ immunotherapy approach is attractive, there are several disadvantages associated with the intratumoral injection of immunostimulants. This approach is generally limited to those accessible tumors, and becomes even more challenging if repeated injections are needed. Structurally, cGAMP contains the phosphodiester bond that is susceptible to degradation by extracellular phosphodiesterase, also the two phosphodiester bonds in cGAMP restrict its penetration through the plasma membrane. Thus, in order to achieve adequate biological activity, cGAMP is commonly used at relatively high concentrations. However, excessive intratumoral cGAMP may induce programmed death-ligand 1 (PD-L1) overexpression on tumor cells and increase tumor-infiltrating regulatory T cells (Tregs), resulting in a negative impact on antitumor immunity 10 , 14 , 15 , 16 . Importantly, several studies have shown that intratumoral immunostimulants generally induce local immune response at the injected site, but have limited effect on distant, uninjected tumor sites, implying that the local approach may be inadequate to elicit systemic immunity, or that the systemic response even if induced, may be rendered inactive when exposed to the immunosuppressive tumor microenvironment (TME) at distant naive tumor sites 3 , 10 , 17 . To overcome these limitations, we set out to develop a nanotechnological strategy that enables targeted delivery of immunostimulants to intratumoral APCs. Here, we assemble phosphatidylserine (PS) on the surface of nanoparticle-cGAMP (NP-cGAMP) because membrane-exposed PS can be recognized and engulfed by macrophages and dendritic cells through their PS receptors 18 , 19 . Because of STING located on ER membrane, it is critical to ensure intracellular delivery and subsequent release of cGAMP to cytosol while bypassing lysosome. We thus utilize calcium phosphate (CaP) to precipitate cGAMP in the liposomal core. After ingestion by APCs, in response to low endosomal pH, cGAMP is released to the cytosol and binds to STING, initiating the STING pathway. We then demonstrate that inhalation of aerosolized NP-cGAMP enables rapid distribution of NP-cGAMP to individual lesions in both lungs bearing lung metastases and subsequent uptake by APCs to stimulate STING signaling and type 1 IFNs production (Fig. 1a ). Last, we tested if inhaled NP-cGAMP can enhance radiotherapy in lung metastasis mouse models of B16-OVA melanoma and 4T1 breast carcinoma. Indeed, fractionated IR (8 Gy × 3) delivered to a lung bearing metastases combined with NP-cGAMP inhalation synergizes to control the metastases not only in the IR but also in the non-IR-treated lung. In addition to enhancing APC sensing of immunogenic IR at irradiated tumor sites and cross-presentation of TAs to prime effector T cells at TDLNs, inhaled NP-cGAMP promotes proinflammatory TME in non-irradiated tumors and facilitates recruitment of cytotoxic CD8 + T cells, which contribute to robust anticancer immunity observed in this study (Fig. 1a ). These data demonstrate that inhalation of NP-cGAMP may represent a pharmacological approach to enhance APC-mediated adaptive immune response against lung metastases. Fig. 1 Characterization of PS-coated NP-cGAMP. a Schematic of the mode of action of the inhalable NP-cGAMP for enhancing antitumor immunity against lung metastases. Inhalation of PS-coated NP-cGAMP enables targeted delivery of the STING agonist, cGAMP to APCs in both irradiated and non-irradiated lung metastases. In addition to enhancing APC maturation, innate immune sensing of immunogenic IR, and cross-presentation of tumor antigen (TA) to prime effector T cells at the irradiated tumor sites, the inhaled NP-cGAMP promotes proinflammatory response in the non-irradiated tumors and facilitates recruitment of TA-specific effector T cells to lung metastases in both lungs, which contribute to therapeutic effects on both the IR and non-IR treated tumors. b Size distribution, PDI, and surface charge of NP-cGAMP, NP labeled with DIR (NP-DIR), and NP labeled with Rhob (NP-Rhob) by dynamic light scattering (DLS). c TEM images of NP-cGAMP. d Diameter change and PDI change of NP-cGAMP at pH 7.4 in 10% FBS (37 °C) over time. e Cumulative release of cGAMP from PS-coated NPs under pH 7.4, 6.5, or 5.0 at 0.5, 1, 2, 4, 8, 12, 24, 36, 48, and 72 h. Data are shown as mean ± SD of n = 3 biologically independent experiments. Source data are provided as a Source Data file Full size image Results Preparation and characterization of PS-coated NP-cGAMP Liposomal NP-cGAMP was prepared in two steps using the water-in-oil reverse microemulsion method 20 , 21 . Both layers of liposome membrane are composed of anionic PS, of which PS exposed on the outer layer serves as the “eat me” signal for APCs, while the inner PS interacts with excessive cationic Ca 2+ of CaP to complex with cGAMP in the core. NP-cGAMP had an average diameter of 118.8 nm and a negative surface charge of −40.7 mV (Fig. 1b, c ). High-performance liquid chromatography (HPLC) analysis indicated high encapsulation efficiency of cGAMP (71.9%) and high payload of cGAMP (31.6 μg/mg lipid). To study its stability, NP-cGAMP was incubated in 10% fetal bovine serum (FBS) at 37 °C. Dynamic light scattering showed that the size remained relatively unchanged for up to 120 h (Fig. 1d ). Furthermore, to mimic physiological extracellular and acidic endosomal environments, drug release profiles of NP-cGAMP in phosphate-buffered saline (PBS) at pH 7.4, 6.5, 5.0 were determined (Fig. 1e ). NP-cGAMP exhibited little cGAMP release at pH 7.4, but an acidic pH-responsive release (~40% at pH 6.5 and ~80% at pH 5.0 after 12 h). These data demonstrate that NP-cGAMP is stable at pH 7.4, but releases cGAMP in a pH-dependent manner. Preferential uptake of NP-cGAMP by APCs and STING activation To evaluate uptake of PS-coated NPs (PS-NPs) by APCs, DiR-labeled NP-DiR was incubated with various APCs including alveolar macrophage (AM), bone marrow-derived dendritic cells (BMDCs), and bone marrow-derived macrophage (BMDM). After 30 min incubation, robust uptake of NP-DiR by all three types of cells was observed (Fig. 2a ). By contrast, there was minimal DiR signal in 4T1-luc breast cancer cells, B16F10 melanoma cells, or mouse vascular endothelial bEnd.3 cells (Supplementary Fig. 1a ). To exclude the possibility that specific uptake by APCs was simply due to the anionic lipid coating, we reconstructed the liposome by replacing PS with anionic phosphatidic acid (PA) on the outer layer. Incubation with PA-NP (surface charge −42 mV) showed much less DiR in APCs compared to PS-NP (Supplementary Fig. 1b ). We further investigated if the specific uptake by APCs is indeed PS-dependent. Our data clearly showed that the uptake of PS-NP by APCs was largely abolished if PS-NP was pretreated with anti-PS antibody (Ab) to block the surface PS (Supplementary Fig. 1c, d ). We also demonstrated the uptake of NP-DiR by APC cells, but not cancer cells by flow cytometry (Supplementary Fig. 1e–h ). These data indicate that PS-coated NPs are recognized and ingested by APCs in a PS-mediated process. Fig. 2 NP-cGAMP uptake by APCs activates STING pathway and CD8 T cell cross-priming. a BMDM, BMDC or AM cells were cultured with DiR-labeled PS-coated NPs (Red) for 30 min and fixed and co-stained with phalloidin (green) and DAPI (blue). Intracellular NP signals were clearly observed. Scale bar = 20 μm. b Real-time PCR of changes in messenger RNA (mRNA) levels of inflammatory cytokine genes in BMDM, BMDC, and AM after treated with free cGAMP (100 nM), NP-cGAMP (100 nM cGAMP), or NP-CTR (2′5′-GpAp as a control of cGAMP) for 4 h. c Western blot detection of STING pathway activation in BMDM, BMDC, and AM after treatment, as indicated, for 8 h. d FACS analysis of expression of the co-stimulatory molecule CD86 and MHC-II in BMDM, BMDC, and AMs after indicated treatment for 8 h. e B16-OVA cells treated with/without a single dose of 20 Gy IR were continued to culture for 72 h and then co-cultured with BMDCs under indicated treatment for 18 h. Expression of the OVA peptide SIINFEKL–MHC-I molecule Kb complex on the surface of BMDCs was analyzed by FACS. f ELISA assay of IFN-γ production from OT-1 CD8 + T cells in vitro. CD8 + T cells isolated from the OT-1 mouse spleen were added into the above mixture of BMDCs with IR- or non-IR-treated B16-OVA cells (BMDCs:T cells = 1:5) and further incubated for 18 h. IFN-γ concentrations in supernatant were determined by ELISA assay. Data were shown as mean ± SD of n = 3 biologically independent experiments. * P < 0.001 by Student’s T test. Source data are provided as a Source Data file Full size image To assess whether NP-cGAMP can enhance cytosolic delivery of cGAMP to activate STING pathway and type I IFN production in APCs, BMDMs, BMDCs, and AMs were incubated with 100 nM free cGAMP or NP-cGAMP for 4 h. Relative expression of type I IFN and other inflammatory response genes were evaluated by real-time PCR. As shown in Fig. 2b , NP-cGAMP induced a drastic increase in expression of Ifnb1 and Ifna1 , as well as other proinflammatory genes, including Tnf , Il1b , Il6 , Il12b , and Cxcl9 , 10 , while only a modest increase was observed with free cGAMP (Fig. 2b ). Consistent with the PCR results, enzyme-linked immunosorbent assay (ELISA) measurements detected a significant increase in the corresponding cytokines in the culture medium of APCs treated with NP-cGAMP (Supplementary Fig. 2 ). Western blot analysis revealed higher levels of phosphorylated TBK1 and IRF-3 in NP-cGAMP than free cGAMP-treated APCs (Fig. 2c ), indicating activation of the STING pathway. These data demonstrate that NP-cGAMP enables efficient cytosolic delivery of cGAMP to activate the STING pathway and production of type I IFNs and other inflammatory cytokines in APCs. NP-cGAMP stimulates APC activation and cross-presentation The APCs that were treated with NP-cGAMP for 8 h were also analyzed by flow cytometry (fluorescence-activated cell sorting (FACS)) to study the expression of major histocompatibility complex class II (MHC-II) and CD86. Marked increase (right shift) in both MHC-II and CD86 was observed in the NP-cGAMP-treated BMDM, BMDC, and AM (Fig. 2d ). These data, together with previous observations of upregulated proinflammatory cytokines (Fig. 2b and Supplementary Fig. 2 ), suggest that NP-cGAMP stimulates APC maturation. To investigate if NP-cGAMP can enhance APC sensing and cross-presenting of TA to prime T cells, we then chose melanoma B16-OVA cells that stably express chicken ovalbumin (OVA) as a model antigen, and treated the cells with a single 20 Gy IR. Using anti-mouse SIINFEKL-2Kb Ab, we detected a significant increase in antigen presentation on the B16-OVA cells (Supplementary Fig. 3a, b ). The B16-OVA cells with/without IR were then incubated with BMDCs in the presence of free cGAMP or NP-cGAMP for 18 h. FACS analysis showed that NP-cGAMP led to a significant increase in the OVA peptide–MHC-I complex on BMDCs ( p < 0.05, Student’s T test), and even higher expression when BMDCs were incubated with the irradiated B16-OVA cells ( p < 0.001; Fig. 2e and Supplementary Fig. 3c, d ). Last, we isolated CD8 + T cells from transgenic mice expressing T cell receptor specific for the OVA peptide SIINFEKL (OT-1) and placed the OT-1 CD8 + cells into the above mixture of BMDCs and B16-OVA cells. ELISA results showed significantly higher levels of IFN-γ in the culture medium where the BMDCs were preincubated with the irradiated B16-OVA in the presence of NP-cGAMP (Fig. 2f ), indicating activation of TA-specific CD8 + T cells. These data demonstrate that NP-cGAMP efficiently promotes APC activation and cross-presentation of TA to prime effector CD8 + T cells. Targeted delivery of NP-cGAMP to lung APCs by inhalation To enable delivery of NP-cGAMP to deep lungs via inhalation, aerosolized NP-cGAMP was generated with a nebulizer system (Supplementary Fig. 4a ). These aerosols had a mass mean aerodynamic diameter (MMAD) of 1.38 μm and geometric standard deviation (GSD) of 1.25 (Supplementary Fig. 4b ), which falls into the size range optimal for deep lung deposition 22 , 23 . After inhaling NP-DiR, the mice with established 4T1-luc lung metastases were sacrificed at different times, and major organs were dissected and imaged ex vivo with IVIS. As shown in Fig. 3a, b , fluorescence signals were observed exclusively in lungs during the time course of 48 h. Concurring with the IVIS data, quantitative HPLC revealed predominant accumulation of the Rhod-b-labeled NPs in lungs with negligible amount in blood and other tissues (Fig. 3c ). By applying various concentrations or durations of inhaled NP-Rhod-b, NP-Rhod-b concentrations in metastases-bearing lungs were measured (Supplementary Fig. 4c ), which was used to estimate the NP-cGAMP dose deposited to lungs. Under our experimental protocol with a 28 min inhalation of NP-cGAMP (37 µM cGAMP) in 5 mL PBS, we estimated 0.1 μg cGAMP deposited in an animal’s lungs, less than one-hundredth of the dose of free cGAMP used for intratumoral injection in several studies 8 , 10 . Based on a previously published mathematical model 22 (Supplementary Fig. 4d ), we can control the cGAMP dose delivered to lungs by changing inhalation duration or initial concentration of NP-cGAMP. Accuracy of the inhalation dose calculation was validated by direct measurements of cGAMP concentrations in tissues by HPLC using NP-cGAMP-FITC (Supplementary Fig. 4e ). Fig. 3 PS-NP via inhalation targets pulmonary antigen-presenting cells in lung metastases. a Representative ex vivo fluorescence imaging of major organs dissected from 4T1-luc lung metastases-bearing mice at 1, 24, and 48 h post inhalation of DiR-labeled PS-coated NPs. Light signals were exclusively from both lungs and b the lung signals were quantified ( n = 3 biologically independent mice/time; ** p < 0.01, Student’s t test). c HPLC measurements of concentrations of PS-coated NPs labeled with RhoB in various tissues of the 4T1-luc lung metastasis mice post inhalation ( n = 3/time) were consistent with the IVIS imaging data. With the lung concentrations decreasing over time, NP increased in TDLNs. d Metastases-bearing lung tissues obtained 24 and 48 h post inhalation were subjected to immunofluorescent staining. The merged images clearly showed that DiR-labeled NPs (red) co-localized predominantly with CD11c + APCs (green). Co-staining of 4T1 tumor cells with anti-luciferase (purple) indicated that the PS-NPs were distributed well into individual metastases and captured by intratumoral APCs. DAPI (blue), scale bar 20 μm. e Representative FACS characterization of pulmonary APC subsets in 4T1 lung metastasis-bearing lungs: alveolar macrophages (AMs; CD11c + F4/80 + ), interstitial macrophages (IMs; CD11c − F4/80 + ) and DCs (CD11c + F4/80 − ). f The percentage of DiR+ AMs, IMs, and DCs within their respective subpopulations was determined at 1, 24, 48, and 96 h after inhalation of NP-DiR. g The percentage of DiR+ migratory mAM, mDC103+, mCD11b + DCs in the total of APCs migrated to TDLNs was determined at different times post inhalation. Data are shown as mean ± SD of n = 3 biologically independent mice. Source data are provided as a Source Data file Full size image It has been documented that the NPs become exposed to lung environment after they are delivered in aerosols to bronchioles and alveoli, at this point their physicochemical properties, for example, size and surface charge, most likely determine their fate 23 . We formulated NP-cGAMP with a mean diameter of ~120 nm because it is generally accepted within an optimal size range for a liposome drug delivery system after taking into account multiple factors such as payload and intratumoral diffusibility 24 , 25 . Anionic surface charge is also considered preferable for intratumoral distribution of liposome, while minimizing non-specific uptake by cell types other than APCs 26 , 27 . Indeed, immunohistochemical studies of lung tissues post inhalation of NP-DiR clearly showed that DiR signals were located primarily within CD11c + APCs, and co-staining with anti-luciferase to label tumor cells further revealed that the NP-DiRs were distributed well intratumorally in individual metastases and engulfed by intratumoral CD11c + APCs (Fig. 3d ; Supplementary Fig. 5 ). Quantitative analysis showed that only a small fraction of NP-DiR was found outside APCs (<13%, Supplementary Fig. 5e ). We further compared APC populations and DiR signals in metastases versus non-tumoral lung tissues. Our data revealed that there were significantly more APCs in tumors than non-tumoral lung tissue (Supplementary Fig. 5c ), despite a slightly higher percentage of APCs found to contain DiR-NP in the non-tumoral lung tissues (67% vs. 60% at 48 h; Supplementary Fig. 5f ). Taken together, these data demonstrate the ability of PS-NPs to penetrate tumor tissues and target intratumoral APCs. We also conducted FACS to quantify the uptake of NP-cGAMP by each subset of lung APCs. As shown in Fig. 3e , lung APCs were classified into AMs (CD11c + F4/80 + ), interstitial macrophages (IMs; CD11c − F4/80 + ), and DCs (CD11c + F4/80 − ). One hour post NP-DiR inhalation, 41.7 ± 7.0% (SD) of AMs, 13.2 ± 3.6% of IMs, and 7.6 ± 3.2% of DCs were DiR+, which increased over 48 h, and then decreased (Fig. 3f ). Meanwhile, the lung-derived DiR+ mAMs and mCD103+ or mCD11b + DCs were found to increase over time up to 96 h in TDLNs (Supplementary Fig. 6 ; Fig. 3g ), indicating the ability of these NP-captured lung APCs to migrate to TDLNs. CD103 + DCs were sparsely populated in tumor-bearing lungs (Supplementary Fig. 7 ), but have been shown to play a potent role in cross-priming cytotoxic CD8 + T cells in TDLNs 28 , 29 . Together, these data demonstrate that inhalation enables effective delivery of NP-cGAMP to intratumoral APCs in lung metastases. Inhaled NP-cGAMP synergizes with IR against lung metastases Radiotherapy is commonly used in clinic to treat lung cancer. In this study, we investigated if inhalation of NP-cGAMP can enhance radiotherapy against lung metastases. We established a melanoma lung metastasis model in immunocompetent mice by injecting B16-OVA cells intravenously. Five days later, after confirming multifocal lung metastases in both lungs, the mice were treated with IR alone, NP-cGAMP inhalation, or both. Fractionated IR (8 Gy × 3) was delivered to a specified area of the right lung while avoiding mediastinum and other normal tissues (Supplementary Fig. 8a–d ). The 8 Gy × 3 dose schedule has previously been shown to induce immunogenic cell death and synergize with immunotherapy against breast cancer in mouse models 30 , 31 . Moreover, a recent study also reports that focal lung radiation with 8.5 Gy/fraction is safe without causing adverse effects 32 . For the combination treatment, NP-cGAMP was inhaled 24 h after each of three IR fractions. On day 18, the mice were sacrificed and treatment efficacy was analyzed by counting the number of metastases on the surface of lungs. As shown in Fig. 4a , for the mice treated with IR alone, there was a clear distinction between the irradiated right lung and the non-irradiated left lung, indicating that IR only affected the irradiated tumors. Despite decreased tumor volume, there was no significant change in the total number of lesions on the irradiated right lung, compared to the control treatment (Fig. 4b ). NP-cGAMP inhalation alone led to a decrease in the number of metastatic foci in both lungs ( p < 0.05, Student’s T test). NP-cGAMP plus IR achieved the highest therapeutic efficacy, inhibiting metastases in both the IR- and non-IR-treated lungs, and causing complete regression of lung metastases in some mice ( p < 0.001, Fig. 4a, b ). These data demonstrate that NP-cGAMP inhalation plus IR induces strong antitumor immunity that leads to regression of lung metastases in both the irradiated and non-irradiated lungs. Fig. 4 NP-cGAMP inhalation synergizes with radiotherapy by eliciting APC-mediated adaptive immunity. The B16-OVA melanoma lung metastasis model was established by intravenously (i.v. injecting 2 × 10 5 B16-OVA cells into C57BL/B6 mice. On day 5, after confirming development of multifocal metastases in both lungs, the mice were treated with fractionated radiation to the right lung (IR, 8 Gy × 3), inhalation of NP-cGAMP (24 h after each IR for three doses) or both. NP-CTR (2′5′-GpAp) served as a control of NP-cGAMP. To deplete pulmonary APCs, NP-clodronate was administered via inhalation 6 h before each of the three NP-cGAMP inhalations. To deplete CD4 + or CD8 + T cells, anti-CD4 Ab or anti-CD8α Ab was injected i.p. (400 µg), respectively, one day before IR and repeated 7 days later. a The mice ( n = 6/group) were sacrificed on day 18 and representative lungs ( n = 3) from the treatment groups were displayed. b Both lungs were examined under a dissecting microscope and a total of metastatic lung foci on each lung were counted. c , d Expressions of the complex of OVA peptide SIINFEKL–MHC-I molecule Kb on CD103 + (CD11c + CD103 + CD11b − ) DCs in the left, right lung and TDLNs ( n = 6) were analyzed by FACS on day 9 (24 h after the last inhalation). e – g The number of CD8 + and CD4 + T cells in metastases-bearing left and right lung ( n = 6) was quantified based on FACS. h , i FACS analysis indicated percentages of the SIINFEKL tetramer + CD8 + T cells in the total of CD8 + T cells in both lungs obtained on day 9. j Percentages of intracellular IFN-γ + SIINFEKL tetramer + CD8 + T cells in the total of CD8 + T cells were also quantified for the left and right lung in treatment groups as indicated. Data are shown as mean ± SD of n = 6 biologically independent samples. * P < 0.05, ** p < 0.01, and *** p < 0.001 by Student’s t test. Source data are provided as a Source Data file Full size image To study plausible mechanisms underlying the enhanced immunity, we first assessed if NP-cGAMP inhalation improved cross-presentation of TA in vivo. A subset of the mice from the above treatment groups was sacrificed 24 h after the last inhalation (48 h after the last IR). Both tumor-bearing lungs and TDLNs were dissected. Because CD103 + /CD8α + DCs have been implicated as the most competent APCs for cross-priming CD8 + T cells in mice 33 , 34 , 35 , we applied FACS gating strategies to differentiate CD103 + DCs (CD103 + CD11b − CD11c + ) from CD11b + DCs (CD11b + CD103 − CD11c + ), and further analyzed the expression of the OVA peptide SIINFEKL–MHC-I complex on these two types of DCs (Supplementary Fig. 7 ). As opposed to IR alone, which led to an increase only in the irradiated lung, IR plus inhalation induced significantly upregulated antigen presentation on CD103 + DCs in both lungs (Fig. 4c ), and the CD103 + DCs with high antigen presentation were also detected in TDLNs (Fig. 4d ), implicating migration of these APCs from tumor sites to TDLNs where they cross-prime T cells. Similarly, expression of SIINFEKL–MHC-I complex was detected on CD11b + CD103 − DCs (Supplementary Fig. 9g ), which also increased after treatment with inhalation with/without IR. These data are consistent with previous reports that both types of DCs are capable of ingesting and processing TA and cross-presenting TA within the MHC-I complex 28 , 29 . However, CD103 + DCs have been found to be more potent on cross-priming CD8 + T cells, whereas CD11b + DCs may be involved in priming CD4 + T cells through their MHC-II–peptide complex 28 , 29 , 35 . Consistent with our previous in vitro observations (Fig. 2d ), NP-cGAMP inhalation activated the expression of co-stimulatory molecule, CD86, and MHC-II on APCs in both lungs (Supplementary Fig. 9 ). We next investigated whether NP-cGAMP inhalation with/without IR drove expansion of TA-specific T cells. As shown in Fig. 4e–g , significantly increased numbers of CD4 + and CD8 + T cells were observed in both lungs post IR, NP-cGAMP inhalation, or both. However, FACS analysis after SIINFEKL–MHC tetramer staining showed that the combination treatment led to a ~10-fold and >5-fold increase in the number of OVA-specific CD8 + T cells in both lungs compared to the control and IR alone, respectively (Fig. 4h, i ). Moreover, inhalation alone or combined with IR activated these tumor-specific CD8 + T cells, evidenced by their higher levels of intracellular IFN-γ (Fig. 4j ). Examinations of TDLNs from the combination treatment also revealed significant expansion of tumor-specific CD8 + T cells (Fig. 5a, b ). To interrogate if the combination treatment elicited systemic tumor-specific immunity, we examined spleens of the treated mice and found that there was indeed a significant increase in tetramer-positive CD8 + T cells ( p < 0.001, Student’s T test; Fig. 5a, c ). We further conducted in vivo VITAL assay by injecting the carboxyfluorescein succinimidyl ester (CFSE) fluorescence-labeled OVA splenocytes into the previously treated mice (Supplementary Fig. 10 ). Compared with the relatively constant level of the non-OVA-labeled splenocytes, significantly more killing (>60%) of the OVA splenocytes was observed in the mice treated with IR plus inhalation ( p < 0.001, Student’s T test; Fig. 5d, e ), confirming induction of systemic tumor-specific immunity. Fig. 5 NP-cGAMP inhalation in combination with IR induces systemic tumor-specific immunity. a The lung TDLNs and spleen obtained on day 9 (24 h after the last inhalation) from the mice with indicated treatment were analyzed by FACS for SIINFEKL tetramer + OVA-specific CD8 T cells. Frequency of SIINFEKL tetramer + CD8 + T cells in TDLNs ( b ) and spleen ( c ). d In vivo VITAL assay. Spleen cells from naive C57BL/6 mice were isolated and half of the cells were pulsed with OVA 257–264 for 2 h in complete medium. The non-pulsed and OVA-pulsed cells labeled with high (0.5) or low (0.05) CFSE, respectively, were equally mixed and injected intravenously (i.v.) into the mice with indicated treatment (day 18 post tumor implant), and 16 h later blood was drawn for FACS analysis. Representative percentages of the low CFSE were indicated. e The combination treatment achieved maximal killings of the OVA splenocytes. Data are shown as mean ± SD of n = 6 biologically independent experiments. ** P < 0.01 and *** p < 0.001 by Student’s t test. Source data are provided as a Source Data file Full size image To determine whether NP-cGAMP induced anticancer immune response is APC-dependent and what subset of effector T cells is required to execute the response, we conducted depletion studies in the mice receiving IR plus NP-cGAMP. To deplete pulmonary APCs, we used the same nanoconstruct of PS-NP to encapsulate clodronate (NP-Clod) and delivered NP-Clod via inhalation 6 h before each of the three NP-cGAMP inhalations. To deplete CD4 + or CD8 + T cells, anti-mouse CD4 or CD8 antibodies were injected intraperitoneal (i.p.) 1 day before IR and repeated 7 days later. We found that depletion of lung APCs or CD8 + T cells (Fig. 4b ) significantly abrogated the antitumor function of the combination treatment, indicating that both APCs and CD8 + T cells are required for the induced antitumor immunity. Depletion of CD4 + T cells had no significant impact on the therapeutic response. These observations are in good agreement with previous reports, in which radiation alone or combined with STING agonists or other immunomodulators significantly enhance MHC-I expression and priming of cytotoxic CD8 + T cell via activated APCs 8 , 10 , 36 , 37 . However, our current study cannot disentangle possible contribution of individual subsets of CD4 + T cells, for example, the possibility of positive impact by eliminating Treg cells. Collectively, these data demonstrate that NP-cGAMP inhalation promotes APC immune sensing and cross-priming CD8 + T cells, and synergizes with radiotherapy to elicit robust anticancer immunity that results in inhibition of both the irradiated and non-irradiated B16-OVA lung metastases. Inhalation of NP-cGAMP promotes proinflammatory TME It is well recognized that TME is extremely immunosuppressive, which may largely counteract the effect of antitumor immunity 3 , 38 , 39 . To investigate if NP-cGAMP inhalation improves the immunosuppressive TME, we analyzed the B16-OVA metastases-bearing lung tissues from the previous treatment groups. ELISA results showed that NP-cGAMP inhalation led to significantly elevated levels of IFN-1β as well as tumor necrosis factor-α, IFNγ, interleukin-6 (IL-6), IL-12p40 in both lungs (Fig. 6a ). By contrast, IR alone caused no significant change in the non-irradiated lung despite a moderate increase in the irradiated lung. Of interest, CXCL10, a ligand of CXCR3, significantly increased in both lungs after NP-cGAMP inhalation with/without IR (Fig. 6a ), which could contribute to the marked increase in tumor-infiltrating T cells observed in both lungs 34 (Fig. 4e–g ). Fig. 6 NP-cGAMP inhalation stimulates proinflammatory cytokines and improves the ratio of CD8 + T/Tregs. a The metastases-bearing lungs obtained on day 9 (24 h after the last inhalation) from the mice with indicated treatment were homogenized separately for the left and right (IR) lung. Various cytokines in the supernatant were measured by ELISA assay. Data shown as mean ± SD ( n = 5). b FACS analysis of FoxP3 + CD4 + regulatory T cells in the above metastases-bearing lung tissues and c quantitative data showed a significant increase in the ratio of CD8 + T cells/Tregs in both lungs in the inhalation alone or the combination treatment group, while a significant decrease in the ratio was detected in the irradiated right lung. Data are shown as mean ± SD of n = 6 biologically independent samples. ** P < 0.01; *** p < 0.001 by Student’s t test. Source data are provided as a Source Data file Full size image There is increasing evidence that the level of antitumor immunity is controlled by the balance of tumor-specific effector T cells and Tregs 14 , 40 . As presented previously in Figs. 4 and 5 , IR alone was able to promote a moderate but significant increase in the number of tumor-infiltrating T cells in both lungs. However, IR also led to an increased number of tumor-infiltrating FoxP3 + CD4 + Tregs, and consequently a significantly decreased ratio of CD8 + T/Tregs in the irradiated tumors ( p < 0.05, Student’s T test; Fig. 6b, c ). Similar observations have been reported by others 10 , 41 . However, this negative effect was abrogated by NP-cGAMP inhalation, which significantly increased the ratio of CD8 + T/Treg in both lungs (Fig. 6b, c ). These findings are in good agreement with several recent studies reporting activation of STING pathway or type I IFNs to directly inhibit co-stimulation-dependent Treg activation and proliferation 37 , 42 . Thus, unlike the direct intratumoral injection, our inhalation strategy enables delivery of NP-cGAMP to the lesions in both the irradiated and non-irradiated lungs to stimulate proinflammatory cytokines and suppress Tregs. The data further reiterate the necessity to overcome the immunosuppressive TME not only in irradiated tumors but also non-irradiated tumors to elicit robust anticancer immunity. Efficacy is confirmed in 4T1 breast cancer lung metastases To determine whether potent antitumor immunity generated by IR plus NP-cGAMP inhalation was confined to B16-OVA lung metastases, we extended our study to 4T1 breast cancer lung metastases. To mimic clinical development of breast cancer lung metastases, we implanted 4T1-luc cells orthotopically into a mammary fat pad. When the tumor reached ~500 mm 3 on day 14, the primary tumor was surgically removed. On day 18, after confirming establishment of lung metastases by bioluminescence imaging (BLI; Fig. 7a ), the mice were randomly grouped and treated with IR alone (right lung), NP-cGAMP inhalation, or both, with the same dose and schedule as in the previous B16-OVA studies. Both BLI and magnetic resonance imaging (MRI) were applied to monitor growth of lung metastases post treatment. Compared to the control group, mice treated with IR or NP-cGAMP alone exhibited significantly lower BLI signals on days 32 and 39, indicating delayed tumor growth (Fig. 7a, b ). IR plus NP-cGAMP inhalation led to further reduced BLI signals, even complete signal loss observed in some animals (Fig. 7a ). MRI quantitatively evaluated tumor volume change by measuring individual lung metastases and summing to obtain the total tumor volume for each animal. The total tumor volume was significantly smaller in the combination group than IR or inhalation alone (Fig. 7c, d ). The imaging data were consistent with ex vivo examination of lung metastases (Fig. 7e ). For the long-term survival study, as presented in the Kaplan–Meier survival curves, the mice treated with IR plus inhalation survived significantly longer ( p < 0.001, log-rank test; Fig. 7f ), 50% of them ( n = 4) were completely cured, showing no sign of disease for at least 150 days (Fig. 7f ). Similar to the B16-OVA study, we found that the survival benefit was abrogated after depletion of pulmonary APCs with NP-Clod (Fig. 7f ), reiterating the indispensable role of APCs in the observed antitumor immunity. Moreover, the long-term surviving mice resisted secondary tumor challenge, indicating that the combination treatment triggers antitumor memory in this model (Fig. 7g ). Fig. 7 NP-cGAMP inhalation plus IR significantly is efficacious against 4T1 breast cancer lung metastases. Female BALB/c mice were implanted orthotopically with 1 × 10 6 4T1-Luc cells into the fourth mammary fat pad. On day 14 when the tumors size reached around 500 mm 3 , the primary tumor was removed. On day 18, after confirming establishment of lung metastases by visualizing bioluminescence imaging (BLI) signals in the chest, the mice were treated with fractionated radiation to right lung (IR, 8 Gy ×3), inhalation of NP-cGAMP, or both. NP-cGAMP was inhaled 24 h after each IR for a total of three inhalations. Longitudinal BLI and MRI were conducted to monitor growth of lung metastases on days 18, 32, and 39. a IVIS images of three representative animals from each treatment group and b quantitative BLI light intensity of the chest. Data were shown as mean ± s.e.m. of n = 6 biologically independent mice. c Longitudinal MRI T2-weighted imaging follow-up of a representative mouse chest from each treatment group and d MRI quantitative tumor volume. Data were shown as mean ± s.e.m. of 4–5 mice. * P < 0.05; ** p < 0.01; *** p < 0.001 by Student’s T test. e Ex vivo lung image represented from each treatment group on day 39. Inhalation of NP-CTR (2′5′-GpAp as a control of cGAMP) for three doses; an additional group with NP-cGAMP plus IR, in which NP-clodronate was administered via inhalation 6 h before each of the three NP-cGAMP inhalations to deplete pulmonary APCs. f Kaplan–Meier survival curves o f the treatment groups up to 140 days after tumor ( n = 8/group) were plotted and statistically analyzed by log-rank test, * p < 0.05; ** p < 0.01; *** p < 0.001. g Mice cured with NP-cGAMP + IR were re-challenged 56 days later with 4T1-Luc cells. Naive mice challenged at the same time served as positive controls. Data showed the percent of mice rejecting 4T1-Luc tumor re-challenge in each group. Source data are provided as a Source Data file Full size image Consistent with previous observations in the B16-OVA model, mechanistic studies revealed that inhalation of NP-cGAMP activated APCs in both 4T1-luc metastases-bearing lungs, as evidenced by significantly increased expression of CD86 and MHC-II on CD103 + DCs, CD11b + DCs, and AMs (Supplementary Fig. 11 ). NP-cGAMP inhalation also led to drastic increase in IFN-1β and other proinflammatory cytokines in both lungs (Supplementary Fig. 12 ). Moreover, NP-cGAMP inhalation with/without IR increased the number of tumor-infiltrating T cells (Supplementary Fig. 13 ) and activated CD8 + T cells in both lungs (Fig. 8a, b ). Like the B16-OVA, IR alone induced a significant increase in tumor-infiltrating FoxP3 + CD4 + Tregs and a decrease in the ratio of CD8 + T/Tregs in the irradiated lung (Fig. 8c, d ). However, NP-cGAMP inhalation offset the negative impact of IR by significantly increasing the CD8 + T/Treg ratio ( p < 0.01, Student’s T test; Fig. 8c, d ). Immunohistochemical staining of CD8 + T cells and FoxP3 + Tregs in lung metastases coincided with the FACS data (Fig. 8e, f ). Together with previous studies in the B16-OVA model, NP-cGAMP inhalation plus IR effectively controls lung metastases. Fig. 8 NP-cGAMP inhalation activates CD8 T cells and improves the ratio of CD8 + T/Treg in 4T1 lung metastases. Twenty four hours after the last inhalation, the mice under indicated treatment ( n = 6 biologically independent mice/group) were sacrificed and both metastases-bearing lungs were dissected for FACS and immunohistochemical studies. a After ex vivo re-stimulation of T cells with phorbol 12-myristate 13-acetate/ionomycin, representative FACS dot plots ( n = 5000) of activated CD8 + T cells by staining intracellular IFNγ from the left and right (IR) lung of each indicated treatment. b Quantitative analysis of frequency of IFN-γ + CD8 + T cells in the left and right lung. c Representative FACS analysis and d quantitative frequency of FoxP3 + CD4 + Treg cells and ratio of CD8/Treg. e Immunohistochemical staining of CD8 + T cells and FoxP3 + Tregs in metastases-bearing lung tissues of n = 3 biologically independent mice under indicated treatment, and quantitative data showing the mean ± SD ratio of CD8/Treg in lung metastases. f NP-CTR: NP-2′5′-GpAp as a negative control of NP-cGAMP; IR (L) and IR (R): left (non-irradiated) and right (irradiated) lung of IR alone; the same applies to NP-cGAMP + IR(L) or IR(R). Scale bar = 20 μm. * P < 0.05; *** p < 0.001 by Student’s T test. Source data are provided as a Source Data file Full size image To further investigate therapeutic efficacy of the combination treatment, we modified the above 4T1 lung metastasis protocol by retaining the primary tumor without surgical removal. In addition to the same combination lung treatment, the primary tumor was treated with/without intratumoral injection of NP-cGAMP. Our data showed that the combination lung treatment alone had modest effects on controlling the primary tumor (Supplementary Fig. 14a–d ). However, in combination with intratumoral injection of NP-cGAMP (5 µg × 2), IR plus NP-cGAMP inhalation led to significant growth delay of the primary tumor and improved survival (Supplementary Fig. 14a–d ). FACS analysis of the primary tumors clearly showed significantly increased APC maturation, tumor-infiltrating activated CD8 + T cells, and the CD8 + /Treg ratio with the combination lung treatment plus intratumoral NP-cGAMP (Supplementary Fig. 14e–i ). By contrast, lung treatment without intratumoral NP-cGAMP induced essentially no change in the TME of primary tumor even though increased activation of effector CD8 + T cells was evident in spleens of these mice (Supplementary Fig. 14e–j ). These results are in line with a number of recent publications 43 , 44 . Chao et al. 43 reported that treatment of radiation plus intratumoral CpG on primary 4T1 tumor had modest effects on lung metastases, whereas an addition of systemic anti-CTLA4 Ab regressed lung metastases and prolonged survival 43 . Immune checkpoint inhibitors are known for their ability to activate tumor-infiltrating effector T cells, which are otherwise exhausted in the TME of the distant non-treated tumor. Taken together, these data support our hypothesis that the immunosuppressive TME negatively impacts anticancer immunity, and it is indispensable to overcome it to elicit durable antitumor immunity. Inhalation of NP-cGAMP is safe Liposome is biocompatible and safe as most clinically approved NP drugs are formulated in liposome. Inhalation of NP-cGAMP was well tolerated by unanesthetized mice. To investigate whether NP-cGAMP inhalation potentially induces systemic immune toxicity, we conducted systematic studies by monitoring body weight, measuring cytokine levels and liver enzymes in blood and examining major organs histopathologically in both healthy mice and lung metastases-bearing mice. In the healthy cohort, there was no significant difference in body weight, liver enzymes aspartate aminotransferase and alanine aminotransferase or cytokines in blood between control treatment and inhalation with/without IR at day 1 or 22 post treatment (Supplementary Fig. 15 ). Similarly, no significant change was detected in lung metastases-bearing mice (Supplementary Fig. 16 ). Hematoxylin and eosin (H&E) staining of major organs at day 1 or 22 showed no visible morphological change in healthy or lung metastases-bearing mice treated with inhalation with/without IR (Supplementary Fig. 17 ). To investigate whether the treatment caused any short- or long-term lung toxicity, we first examined the healthy mice on days 1 and 22 post treatment. H&E staining showed no significant morphological change, for example, hemorrhage, despite some increased infiltration of cells, likely leukocytes, observed in the lung treated with IR plus inhalation on day 1. On day 22, there was no sign of inflammation in the lung (Supplementary Fig. 18a, b ). We then studied the four mice that were cured with the combination treatment. After sacrifice on day 150, visual examinations of major organs, including lungs, liver, and kidney, revealed no obvious lesions; histology found no microscopic lesions in lungs, confirming no residual metastases (Supplementary Fig. 18c ). The shape of alveoli, alveolar wall thickness, and microvasculature in the non-irradiated left lung all looked similar to their healthy counterparts. While a few small regions in the irradiated lung were seen to contain clustering macrophage-like cells and some slightly thickened alveolar walls (Supplementary Fig. 18c ), there was no obvious pathological change, for example, fibrosis. The lack of local and systemic toxicity may be attributed to the APC-targeting ability of NP-cGAMP and the low dose used for inhalation, which led to negligible amounts of cGAMP taken up by other lung cells (Fig. 3d ; Supplementary Fig. 5 ) or entering blood (Fig. 3a–c ). Nevertheless, NP-cGAMP inhalation in combination with IR is safe and effective against lung metastases. Discussion Radiation is known for its potential to kill cancer cells to release TA 36 , 43 , 45 , 46 , 47 , 48 , which may trigger systemic immunity against cancer metastases 31 , 48 , 49 . However, radiation effect on non-irradiated tumors is rarely seen in clinic. In this study, we demonstrate an inhalable NP-immunotherapy system for APC-targeted delivery of STING agonists to activate intratumoral APCs and enhance immune sensing of immunogenic radiotherapy in lung metastases mouse models. NP-cGAMP inhalation synergizes with fractionated IR to generate potent antitumor immunity against both the irradiated and non-irradiated lung metastases. The ability of NP-cGAMP inhalation to remodel the TME by converting “cold” into “hot” tumors in both the irradiated and non-irradiated tumors contributes to the therapeutic efficacy. Intratumoral injection of immunomodulators to elicit anticancer immunity has shown promising results in preclinical studies and is being investigated extensively in clinical trials. Although we also demonstrated the utility of intratumoral injection of NP-cGAMP, we were not intending to compare it with the inhalation approach in this study, because the individuals who would receive this therapy come from different patient populations. In the context of clinical cancer development and treatment strategy, primary cancer, for example, breast cancer or melanoma, is commonly resected by surgery or locally treated once detected. However, lung metastases may develop years later in these patients, and in many cases when the primary cancer is well controlled. We are specifically considering the potential of our inhalation approach for treating this population of cancer patients. From a clinical perspective, these patients have the highest mortality, thus development of effective treatment, for example, immunotherapy, is the most needed. Given that lung is an extremely common site for breast cancer and melanoma to metastasize 50 , 51 and in many cases multiple lesions develop at peripheral lung 52 , 53 , the inhalation approach, which has advantages including its non-invasiveness, feasibility for repeated procedures, and accessibility to multiple lung lesions/lobes at the same time, may have particular translational relevance to deliver immunomodulators to these lung metastases. With a similar strategy of using hypofractionated stereotactic body radiation therapy to treat a single lesion or few lesions in a lung 54 , and combined with NP-cGAMP inhalation to activate antitumor immunity in both irradiated and non-irradiated tumors, we believe that this nano-immunotherapy system may have the potential for treating lung metastases arising from a variety of primary cancer types. Methods Preparation of PS-coated NP-cGAMP The liposomal NP-cGAMP was prepared in two steps using the water-in-oil reverse microemulsion method 20 , 21 . DSPC (1,2-distearoyl- sn -glycero-3-phosphocholine,18:0 PC), cholesterol, brain PS ( l -α-phosphatidylserine), DSPA (1,2-distearoyl- sn -glycero-3-phosphate,18:0 PA), and Rhod-b (18:1 Liss Rhod PE, 1,2-dioleoyl- sn -glycero-3-phosphoethanolamine- N -(lissamine rhodamine B sulfonyl)) were purchased from Avanti Polar Lipids. cGAMP (2′3′-cGAMP, cyclic [G(2′,5′)pA(3′,5′)p]) and cGAMP control (2′5′-GpAp) were purchased from InvivoGen. Both layers of the liposome membrane are composed of anionic PS. Briefly, 150 μL CaCl 2 (2.5 M in ddH 2 O, pH = 7.0) was added to 5 mL mixed oil phase 1 (cyclohexane:igepal co-520 = 80:20) in 50 mL flask and stirred at 600 r.p.m. for 20 min to form a well-dispersed microemulsion phase. Oil phase 2 was prepared by adding 50 μL lipid mixture (20 mM, PS:DSPC:cholesterol = 5:4:1) to 5 ml mixed oil phase (cyclohexane:igepal co-520 = 80:20). A similar microemulsion containing sodium phosphate was prepared by adding 150 μL Na 2 HPO 4 (25 mM in ddH 2 O, pH = 9.0) to 5 ml mixed oil phase 2 with the calcium (Ca):phosphate (P) ratio of 100:1. The CaP core with single layer anionic lipid coating was formed by mixing oil phase 1 and oil phase 2 and stirring at 600 r.p.m. for 20 min. To collect the lipid-coated CaP cores, 10 ml ethanol was added to the mixture and centrifuged at 16,000 × g for 15 min, followed by washing with ethanol three times. The collected CaP cores were dispersed in 1 ml chloroform and centrifuged at 2000 × g for 5 min to remove CaP precipitates without lipid coatings. The supernatants containing a single layer of lipid-coated CaP were further mixed with 70 μL lipid mixture (20 mM, brain PS:DSPC:cholesterol = 5:4:1, molar ratio, in CHCl 3 ) into 50 ml flask, followed by chloroform evaporation under reduced pressure to form a lipid film. NP-CaP NPs with bilayer lipid coating were formed by adding 1 ml PBS (0.1 M, pH = 7.4) and rehydrating under water bath sonication for 5 min and sonic probe at 20 W for another 2 min at 70 °C. The resulting NPs were further filtered with 0.45 μM membrane to remove the free lipid aggregates and stored at 4 °C. For DSPA-exposed CaP NPs, DSPA at 20 mM was used instead of brain PS during the preparation. To load cGAMP, half of the desired content of cGAMP was mixed with CaCl 2 solution and Na 2 HPO 4 . For fluorescently labeling NPs, 18:1 Liss Rhod PE (1,2-dioleoyl- sn -glycero-3-phosphoethanolamine- N -(lissamine rhodamine B sulfonyl); Rhod-b) was added to the second lipid mixture with a molar ratio of 1%. DiR was further used for labeling the NPs by adding DiR directly to the second lipid mixture at a molar ratio of 5%. Characterization of NP-cGAMP The size, size distribution, and zeta potential of NP-cGAMP in aqueous solution were measured by a Malvern Zetasizer Nano ZS90. Transmission electron microscopy (TEM) measurements were performed on an FEI Tecnai Bio Twin TEM. To determine cGAMP loading efficiency, 0.2 ml of NPs were incubated with 0.2 ml 0.8 M HCl for 24 h to dissolve CaP NPs and release cGAMP. The mixture was further centrifuged at 14,000 × g for 15 min, and the supernatant was gathered to quantify the cGAMP concentrations by HPLC analyses using an Agilent 1100 HPLC system. The release of cGAMP from NPs was assessed by the dialysis of NP-cGAMP solution against release medium at different pH (pH 7.4, 6.5 and 5.0). The release medium was removed for analysis at 0.5, 1, 2, 4, 8, 12, 24, 36, 48, and 72 h. cGAMP content in the release medium was determined by HPLC. The NPs were incubated in PBS (pH 7.4, 0.01 M) with 10% FBS (v/v) at 37 °C for 5 days to study particle stability. The change in particle size was monitored at specific time intervals by Zetasizer. Cell lines The OVA-transfected mouse melanoma B16F10 cells (B16-OVA) and parental B16F10 cells (ATCC), and the 4T1 breast cancer cells stably transfected with firefly luciferase (4T1-luc; provided by Dr. David Soto Pantoja, Wake Forest), mouse vascular endothelial cells, bEnd.3 (ATCC) were cultured in Dulbecco’s modified Eagle’s medium (DMEM), supplemented with 10% FBS, 100 U/mL penicillin, and 100 μg/mL streptomycin and maintained in a humidified atmosphere containing 5% CO 2 at 37 °C. Induction of BMDM and BMDC and isolation of AM cells BMDMs were generated by culturing bone marrow cells flushed from C57BL/6 mouse femurs. Briefly, the gathered bone marrow cells were incubated in completed α-MEM (minimum essential medium Eagle-alpha modification) containing 10% heat-inactivated FBS and penicillin/streptomycin. After 18 h, the floating cells were collected and cultured in complete BMDM medium (complete α-MEM supplemented with 50 ng/mL M-CSF). Three days later, the adherent cells were used as macrophages and phenotyped by determining the expression of CD11b and F4/80 (typically 70–85% CD11b + F4/80 + ). BMDCs were also generated by culturing bone marrow cells flushed from the femurs of C57BL/6 mice in BMDC medium: RPMI-1640 containing 10% heat-inactivated FBS, penicillin/streptomycin, 20 ng/mL granulocyte–macrophage colony-stimulating factor, 5 ng/mL IL-4, and 1× 2-mercaptoethanol. The culture medium was half-replaced every 2 days, and the non-adherent and loosely adherent immature DCs were collected on day 8 and phenotyped by determining the expression of CD11c (typically 60–80% CD11b + ). To obtain AMs, C57BL/6 mice were sacrificed by CO 2 . The trachea was then cannulated with a blunt 22-gauge needle and tied in place. Bronchoalveolar lavage was flushed with PBS containing 0.5 mM EDTA. Aliquots of 1 mL were instilled into the lungs and aspirated back into the syringe. This procedure was repeated three times per mouse. The gathered cells were cultured in DMEM medium with 10% heat-inactivated FBS, 2-mercaptoethanol (1×), and penicillin/streptomycin. Cellular uptake of PS-coated NPs in vitro Briefly, 5000 BMDM, BMDC, and AM cells were seeded in 24-well plates with poly-lysine-coated coverslips for 24 h. Then, the media were replaced with the complete medium containing PS-exposed or DSPA-exposed NPs labeled with DiR and cultured for another 0.5 h. For the blocking study, PS-NPs were pretreated with/without anti-PS antibody PGN635 (25 µg/mL; Peregrine Pharmaceuticals, Tustin, CA) for 2 h before incubating with AM or BMDC cells for 30 mins 55 . Cells were then washed twice with cold PBS and fixed with 4% formaldehyde for 15 min. After 1% Triton X-100 treatment for 15 min, the cells were stained with Alexa Fluor 488 phalloidin (Thermo Fisher) for 30 min and incubated with 1 μg/mL DAPI (4′,6-diamidino-2-phenylindole) for 5 min. The cellular uptake of NPs was observed by fluorescence microscope. To determine whether the PS-coated NPs are preferentially recognized and ingested by APCs, we repeated the above experiment with 4T1-luc or B16F10 cancer cells or normal mouse vascular endothelial bEnd.3 cells. Real-time quantitative PCR of type I IFN and other inflammatory genes BMDM, BMDC, or AM cells (10 5 ) were seeded into 6-well plates and cultured for 24 h. The cells were then incubated at 37 °C with free cGAMP or NP-cGAMP (100 nM cGAMP) suspended in complete culture medium for 4 h. All cellular RNA was collected using TRIzol reagent (Thermo Fisher). One microgram of total RNA was transcribed into complementary DNA (cDNA) using High-Capacity cDNA Reverse Transcription Kit (Thermo Fisher), and real-time quantitative PCR was performed using PowerUp SYBR Green Master Mix (Thermo Fisher) with primers listed in Table S1 . Messenger RNA levels were normalized against the housekeeping gene GAPDH (glyceraldehyde 3-phosphate dehydrogenase). Western blot of STING pathway activation by NP-cGAMP For western blot analysis, the above APC cells incubated with free cGAMP or NP-cGAMP (100 nM cGAMP) for 8 h were washed and lysed in 50 µL of lysis buffer containing a protease inhibitor cocktail (Roche). The cell lysates with total protein (40 µg) were electrophoresed. Anti-phospho-IRF-3 (1:1000), anti-IRF-3 (1:1000), anti-phospho-TBK1 (1:1000), and anti-TBK1 (1:1000) were from Cell Signaling Technology. Anti-β-actin (1:1000) (Santa Cruz Biotechnology) was used as housekeeping protein. Goat anti-mouse or goat anti-rabbit horseradish peroxidase (HRP)-conjugated antibodies (1:10,000) (Santa Cruz Biotechnology) were used as the secondary antibody. The membranes were visualized using the ECL system (Bio-Rad) and the expression levels of protein were normalized to actin protein expression levels. Western blots source data provided in the Source Data file. Flow cytometry Flow cytometry was performed on a BD Canto II flow cytometer and analyzed using the FlowJo software (BD Biosciences). A list of antibodies used here was summarized in Table S2 . For intracellular staining of IFN-γ, fresh isolated cells are treated with phorbol 12-myristate 13-acetate/ionomycin cocktail according to the manufacturer’s specification (BioLegend). The cells were then washed, stained with antibodies against CD3, CD4, and CD8α, fixed with fixation buffer and subsequently stained intracellularly with antibodies against IFN-γ in Intracellular Staining Permeabilization Wash Buffer (BioLegend). Doublets and debris of dead cells were excluded before various gating strategies were applied. Gates and quadrants were set based on isotype control staining, and the mean fluorescence intensity (MFI) values are calculated by minus the MFI of isotype control antibodies. ELISA assay Cytokines in cell culture medium, blood samples, or lung tissues were analyzed with ELISA Max Deluxe Sets from BioLegend by following the manufacturer’s instructions. After adding the HRP substrates, optical densities were determined at a wavelength of 450 nm in an ELISA plate reader (Bio-Rad). Liver enzyme assay Serum aspartate transaminase activity (Enzychrom Aspartate Transaminase Assay Kit, BioAssay Systems, Haymard, CA) and alanine transaminase activity (Enzychrom Alanine Transaminase Assay Kit, BioAssay Systems) were performed following the manufacturer’s instructions. In vitro DC cross-presentation assay BMDCs were generated and phenotyped by determining the expression of CD11c. Semi-confluent B16-OVA cells were irradiated with 0 or 20 Gy of a single dose and then immediately seeded into 24-well plate at 1 × 10 5 cells per well and cultured for 72 h. Wells were washed twice and 2 × 10 4 BMDCs were added and cultured in the presence of NP-cGAMP, free cGAMP, or controls at 37 °C for 18 h. OVA 257–264 presented with MHC-I on the cell surface was detected by anti-H-2Kb bound to SIINFEKL-PE-Cy7, an antibody that specifically recognizes OVA peptide SIINFEKL bound to H-2Kb of MHC-I by FACS. In vitro CD8 T cell priming assay CD8 + T cells were isolated from spleens of C57BL/6-Tg (TcraTcrb)1100Mjb/J (OT-I) mice (Jackson Laboratory) by magnetic separation (STEMCELL Technologies) according to the manufacturer’s instructions. The purity of CD8 + T cells was >95%. A total of 1 × 10 5 cells CD8 + T cells were added into the mixture of BMDCs with the B16-OVA cells pretreated with/without IR in the presence of NP-cGAMP or controls, as described above. After 18 h, cell culture supernatants were collected and measured for IFN-γ content as a surrogate of activation of tumor-specific CD8 + T cells, by ELISA assay. In vivo lung metastasis models All animal experiments complied with all relevant ethical regulations for animal testing and research and were performed with approved of the Institutional Animal Care and Use Committee at the Wake Forest University School of Medicine. For the B16-OVA lung metastasis model, C57BL/6 mice (6–8 weeks, female:male at 1:1; Charles River Laboratories, Wilmington, MA) were injected intravenously with 2 × 10 5 B16-OVA cells. Five days later, the mice developed multifocal metastases on both lungs. For the 4T1-luc lung metastasis model, 4T1-luc cells were injected orthotopically into the right fourth mammary fat pad of female BALB/c mice (8–10 weeks; Charles River). When the tumor volume reached ~500 mm 3 , the primary tumor was surgically removed in a subset of animals. Development of lung metastases was confirmed before treatment by BLI or MRI. In vivo mechanistic studies and treatment were subsequently conducted on the mice, as described below in details. Inhalation of aerosolized PS-coated NPs A clear plastic box with a wire-netting floor was used for inhalation treatment. Aerosol was generated via a medical-grade nebulizer attached by medical tubing to the animal chamber (Supplementary Fig. 4 ). Three unanesthetized animals were placed in the sealed chamber at each time and exposed to aerosol at an air flow rate of 7 L/min for 28 min, during which ~5 mL of solution loaded in the nebulizer were aerosolized. The MMAD and GSD of aerosolized particles and aerosol concentration were measured using a TSI 3321 Aerodynamic Particle Sizer Spectrometer. For in vivo biodistribution study, DiR- or Rhod-b-labeled PS-coated NP-CaP (DiR 660 μg; Rhod-b 80 μg) in 5 mL PBS was loaded into the nebulizer. For treatment, NP-cGAMP (37 μM cGAMP) in 5 mL PBS was loaded into the nebulizer. In vivo biodistribution and quantification of NP-cGAMP BALB/c healthy mice and mice with 4T1-luc lung metastasis were placed in the chamber to inhale DiR-labeled PS-coated NP for 28 min. The animals were sacrificed at 1, 24, and 48 h ( n = 3/time) and major organs were dissected and ex vivo imaging was conducted using an IVIS ® Lumina system (Caliper Life Sciences). The signal intensity was analyzed using the Living Image ® 3.1 software. Immediately after ex vivo imaging, the lung tissues were preserved and sectioned for immunofluorescence microscopy. Cryosections (6 μm thick) were co-stained with anti-mouse CD11c-FITC (1:200; BioLegend, N418) and anti-luciferase (1:500; Sigma-Aldrich, L0159), followed by cy3-anti-rabbit secondary antibody (1:800; Jackson Immuno) and observed using a fluorescence microscope. DiR signals were recorded and merged with the CD11c image and the luciferase-stained image of the same field. DiR + cells in lung tissues and TDLNs were also assessed by FACS to quantify tissue concentrations of PS-coated NPs, the 4T1-Luc lung metastasis mice were sacrificed at 1, 24, and 48 h after inhaling Rhod-b-labeled PS-coated NP ( n = 3/time). Major organs and blood were collected for HPLC analyses 56 . For Rhod-b detection, HPLC grade acetonitrile and water (90:10, v/v) with 0.1% trifluoroacetic acid (TFA) was used as mobile phase A and HPLC grade tetrahydrofuran with 0.1% TFA was used as mobile phase B. The mobile phase was delivered at 1.0 mL/min with A:B = 7:3 at 30 °C. The fluorescence detector was set at 540 nm for excitation and 590 nm for emission, and linked to ChemStation LC 3D system for data analysis. Linear calibration curves for concentrations in the range of 0.004–100 ng/μL were plotted using the peak areas by linear regression analysis and concentration of Rhod-b in each sample was determined. Based on HPLC measurements of Rhod-b-NPs in lungs, the concentration of cGAMP deposited in lungs 1 h after inhalation was determined. The fluorescent dye-labeled cGAMP (Biolog Life Science Institute GmbH & Co. KG) was also used to quantify cGAMP in tumor-bearing lungs after inhalation. For the cGAMP detection, HPLC grade acetonitrile with 0.1% TFA was used as mobile phase A and HPLC grade water with 0.1% TFA was used as mobile phase B. The mobile phase was delivered at 1.0 mL/min with A:B = 5:5 at 30 °C. The fluorescence detector was set at 494 nm for excitation and 517 nm for emission, and linked to ChemStation LC 3D system for data analysis. In vivo VITAL cell killing assay Spleen cells from naive C57BL/6 mice were isolated and pulsed with/without OVA 257–264 for 2 h in complete medium. The non-pulsed and OVA-pulsed cells were then labeled with high (0.5) or low (0.05) CFSE, respectively, in serum-free medium for 15 min. Equal numbers (1 × 10 7 ) of CFSE high (non-pulsed) and CFSE low (OVA-pulsed) cells were mixed and injected intravenously into the B16-OVA mice of different treatment groups (on day 18 post tumor implant). After 16 h, blood was collected and subjected to flow cytometry analysis. The number of CFSE high and CFSE low cells were determined and used to calculate the percentage of OVA peptide-specific cell killing based on the following equation: $${\mathrm{Percentage}}\;{\mathrm{of}}\;{\mathrm{specific}}\;{\mathrm{killing}} = \left( {1 - {\mathrm{non}{\mbox{-}\mathrm{transferred}}}\;{\mathrm{control}}\;{\mathrm{ratio/experimental}}\;{\mathrm{ratio}}} \right)100.$$ (1) Irradiation Animals were anesthetized with inhaled 3% isoflurane and positioned with the lung 50 cm below the aperture of an X-RAD 320 orthovoltage irradiator (Precision X-ray). Fractionated IR (8 Gy × 3 daily fractions) was delivered at a rate of 176 cGy/min (300 kVp voltage and 10 mA current) through a custom-fabricated lipowitz alloy shield. The home-made half circle-shaped lipowitz alloy shield with a radius of 5 mm was placed 3.5 cm above the animal. This shield yields a semi-circular field of ~1.08 cm in diameter, allowing a specified area of the right lung to be irradiated while minimizing the radiation dose to important mediastinum and tissues outside the right lung. To ensure good reproducibility, animal landmarks such as the sternum and ribs were used for initial radiation, the light field of which was marked on the animal, where the radiation light field was aligned with the 10 mm line on the animal for the second and third IR (Supplementary Fig. 8 ). In vivo experimental treatment For the B16-OVA lung metastases, after confirming visible multifocal lung lesions on the surface of both lungs by sacrificing randomly selected animals on day 5, the mice ( n = 6/group) were randomly grouped and treated as follows: (i) PBS (inhalation), (ii) NP-CTR (inhalation of NP-2′5′-GpAp), (iii) IR (8 Gy × 3 to the right lung), (iv) inhalation of NP-cGAMP, (v) NP-cGAMP + IR, (vi) NP-cGAMP + IR + Clod (inhalation of NP-Clod), (vii) NP-cGAMP + IR + anti-CD4, and (viii) NP-cGAMP + IR + anti-CD8α. Inhalation of NP-cGAMP or controls occurred 24 h after each IR for a total of 3 doses. To deplete pulmonary APCs, the mice inhaled NP-Clod (200 µg) 6 h before each NP-cGAMP for 3 doses. For depleting CD4 + or CD8 + T cells, 400 µg anti-mouse CD4 + antibody (BioXCell, cloneGK1) or anti-mouse CD8α antibody (BioXCell, clone2.43) was injected i.p. one day before treatment and repeated 7 days later. The mice were sacrificed for enumeration of metastatic lung foci on day 18. Both the left and right lungs were evaluated under a dissecting microscope and the blinded quantification conducted. For the 4T1-luc model, a subset of animals was subjected to surgical removal of the primary tumor, while the primary tumor in the other subset of mice was kept without resection, as described above. Lung metastases were visualized by BLI and MRI and confirmed by ex vivo examination before treatment in both studies. The mice ( n = 8/group) were then randomly grouped and treated as follows: (1) without primary tumor: (i) PBS, (ii) NP-CTR, (iii) IR (8 Gy × 3 to the right lung), (iv) inhalation of NP-cGAMP, (v) NP-cGAMP + IR, (vi) NP-cGAMP + IR + Clod; (2) with primary tumor: (i) PBS, (ii) inhaled NP-cGAMP + IR to the right lung + intratumoral PBS, (iii) inhaled NP-cGAMP + IR to the right lung + intratumoral NP-cGAMP (1 µg × 2; first dose, post IR, but 24 h before the last inhalation; second dose, 24 h post the last inhalation), (iv) inhaled NP-cGAMP + IR to the right lung + intratumoral NP-cGAMP (5 µg × 2). Lung metastases burden was monitored longitudinally by BLI and MRI, while a caliper was used for measuring the primary tumor volume. Survival of the mice was followed for up to 150 days. For tumor re-challenge study, long-term surviving mice from the combination treatment group (56 days) were implanted with 4T1-luc cells (10 6 ) into the contralateral mammary fat pad. Additional control mice were implanted to confirm tumor growth. BLI of initiation and development of 4T1-luc lung metastases Longitudinal BLI was conducted after orthotopic implantation of 4T1-luc breast cancer. The mice were anesthetized with 2% isoflurane inhalation and injected i.p. with 150 mg/kg d -luciferin. BLI was acquired 10 min later using the IVIS Lumina Imaging System (Caliper Life Sciences). Data were quantified with the Living Imaging software by using absolute photon counts (photons/s/cm 2 /Sr) in a region of interest, manually drawn to outline the BLI signal of the chest. MRI follow-up of lung metastases volume MRI was conducted on a 7 T Bruker BioSpec small animal scanner (Bruker Biospin, Rheinstetten, Germany). The imager of MRI was blinded to the group allocation. MRI was initiated on day 18 and followed on days 32 and 39. The mice were anesthetized with isoflurane (3% induction, 1.5% maintenance). Respiration was monitored with a respiratory bulb under the chest and a SHARPII animal monitoring system was used for respiratory gating. Anatomical T2-weighted imaging was conducted using a RARE sequence with TR/TE = 1600/23 ms; ETL: 8; NSA: 8; matrix size: 128 × 128. Individual tumor volumes were measured on T2-W images by manually outlining the enhancing portion of the mass on each image and the total of lung metastases volume for each animal was the sum of individual volumes. Histology and immunohistochemistry H&E staining was performed on cryosections (10 µm) of different tissues, including normal heart, lung, liver, and spleen, as well as lung metastases-bearing lungs. For immunohistochemical staining of tumor-infiltrating lymphocytes, cryosections (10 µm) of 4T1-luc lung metastases-bearing lung tissues obtained from the above treatment group on day 22 (24 h after the last inhalation) were immunostained with anti-mouse CD8α (1:500; BioLegend) or anti-mouse FoxP3 antibody (1:500; BioLegend), followed by HRP-conjugated goat anti-rat secondary antibody (1:500; Jackson Immuno). The sections were then developed with DAB Kits (3,3′-diaminobenzidine; Vector Laboratories) and counterstained by hematoxylin. Statistical analysis Statistical analysis was performed using Microsoft Excel and Prism 7.0 (GraphPad). Data were presented as mean ± SD. Statistical significance was determined by Student’s t test. All t tests were one-tailed and unpaired and were considered statistically significant if p < 0.05. The survival assay was analyzed using a log-rank test and considered statistically significant if p < 0.05. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The authors declare that data supporting the findings of this study are available within this article and its Supplementary Information, and all additional data are available from the corresponding author on reasonable request. A Source Data file, including the source data for Figs. 1 – 8 in the main text and Supplementary Figs. 1 – 5 , 9 , and 11 – 16 , has been provided.

Scientists have reported a new approach to treating lung cancer with inhaled nanoparticles developed at Wake Forest School of Medicine, part of Wake Forest Baptist Health. In this proof of concept study, Dawen Zhao, M.D., Ph.D., associate professor of biomedical engineering at Wake Forest School of Medicine, used a mouse model to determine if metastatic lung tumors responded to an inhalable nanoparticle-immunotherapy system combined with the radiation therapy that is commonly used to treat lung cancer. The study is published in the current issue of Nature Communications. Lung cancer is the second most common cancer and the leading cause of cancer death among both men and women. More people die of lung cancer than of colon, breast and prostate cancers combined. Although immunotherapy is promising, it currently works in less than 20% of patients with lung cancer. Significant clinical evidence suggests that at the time of diagnosis most patients' tumors are poorly infiltrated by immune cells. This "cold" immune environment in tumors prevents the body's immune system from recognizing and eliminating the tumor cells. Overcoming this immunosuppressive tumor environment to attack the cancer efficiently is currently an area of great interest in the scientific community, Zhao said. Previous approaches have involved direct injection of immunomodulators into tumors to boost their immune response. However, this approach is generally limited to surface and easily-accessed tumors, and can become less effective if repeated injections are needed to sustain immune response. "The goal of our research was to develop a novel means to convert cold tumors to hot, immune-responsive tumors," Zhao said. "We wanted it to be non-invasive without needle injection, able to access multiple lung tumors at a time, and be safe for repeated use. We were hoping that this new approach would boost the body's immune system to more effectively fight lung cancer." The nanoparticle-immunotherapy system that Zhao and his team developed delivered immunostimulants via inhalation to a mouse model of metastatic lung cancer. The immunostimulant-loaded nanoparticle which, when deposited in the lung air sacs, were taken up by one specific type of immune cells, called antigen-presenting cells (APC). The immunostimulant, cGAMP, in the nanoparticle was then released inside the cell, where stimulation of a particular immune pathway (STING) activated the APC cell, which is a critical step to induce systemic immune response. The team also showed that combining the nanoparticle inhalation with radiation applied to a portion of one lung led to regression of tumors in both lungs and prolonged survival of the mice. In addition, the team reported that it completely eliminated lung tumors in some of the mice. Through mechanistic studies the team then confirmed that the inhalation system converted those initially cold tumors in both lungs to hot tumors favorable for robust anti-cancer immunity. Zhao's inhalable immunotherapy presents several key advantages to previous methods, especially the ability to access deep-seated lung tumors because the nanoparticulate-carrying aerosol was designed to reach all parts of the lung, and the feasibility of repeated treatment by using a non-irritating aerosol formulation. The treatment was shown to be well tolerated and safe without causing adverse immune-related distress in the mice. The Wake Forest School of Medicine researchers have filed a provisional patent application for the inhalable nanoparticle-immunotherapy system.

10.1038/s41467-019-13094-5

Biology

Researchers clarify how cells defend themselves from viruses

Hong Cao et al. The anti-viral dynamin family member MxB participates in mitochondrial integrity, Nature Communications (2020). DOI: 10.1038/s41467-020-14727-w Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-020-14727-w

https://phys.org/news/2020-03-cells-defend-viruses.html

Abstract The membrane deforming dynamin family members MxA and MxB are large GTPases that convey resistance to a variety of infectious viruses. During viral infection, Mx proteins are known to show markedly increased expression via an interferon-responsive promoter to associate with nuclear pores. In this study we report that MxB is an inner mitochondrial membrane GTPase that plays an important role in the morphology and function of this organelle. Expression of mutant MxB or siRNA knockdown of MxB leads to fragmented mitochondria with disrupted inner membranes that are unable to maintain a proton gradient, while expelling their nucleoid-based genome into the cytoplasm. These findings implicate a dynamin family member in mitochondrial-based changes frequently observed during an interferon-based, anti-viral response. Introduction Mitochondria are dynamic organelles that undergo fission and fusion events in response to changes in metabolic needs and important cellular processes such as cell division and apoptosis 1 , 2 , 3 . Central to these alterations in mitochondrial form are the dynamin family of large GTPases, which are known to bind and remodel membranes 4 , 5 . Distinct from the conventional dynamins (Dyn1, 2, 3) that are known to support membrane scission at distinct cellular sites such as the plasma membrane 6 , the Golgi apparatus 7 , as well as endosomes 8 and autolysosomes 9 , 10 , 11 , the dynamin-related proteins have been implicated in the dynamics of both mitochondria 12 and peroxisomes 13 . Although highly conserved within the N-terminal GTPase domains, these dynamin-related proteins are only modestly conserved in the middle domain and generally lack pleckstrin binding and proline-rich domains found in the conventional family members. The mitofusins (MFN1/MFN2) are believed to support the outer mitochondrial membrane fusion and optic atrophy 1 (OPA1) has been implicated in the fusion of the inner membrane 14 , 15 , 16 , 17 . Reciprocally, dynamin-related protein 1 (DRP1) has been shown to constrict the outer mitochondrial membrane 4 with the final scission event perhaps driven by conventional Dyn2 18 . In addition to the mitochondria-associated dynamins are the myxovirus resistance GTPases (MX1/2 human and MxA/B mouse), which are <40% similar to the conventional dynamins but over 60% identical in sequence to each other. These related proteins are expressed in cells upon exposure to type 1 interferon (IFN) and convey a cellular innate immune response to a variety of different pathogenic viruses 19 , 20 , 21 , 22 , 23 , 24 including influenza, hepatitis B (MxA), and HIV-1 (MxB). We and others have found that MxA can assemble into polymers that act to deform membranes 25 , 26 and appears to reside on components of the smooth endoplasmic reticulum (ER) 25 , 27 . MxB also assembles into polymeric structures 28 and was originally found to associate with nuclear pores where it was predicted to restrict the access of a viral genome to the host transcriptional machinery 29 . MxB has been described to bind to the HIV-1 genome, while impairing its chromosomal integration 24 , 30 , 31 , and recently has been identified as a key factor behind IFN-mediated suppression of hepatitis C virus infection 32 . In this study we report that although markedly upregulated in cells treated with IFN, MxB is expressed constitutively in a variety of cell types, particularly in primary hepatocytes and hepatoma cell lines. In addition to reported localization at nuclear pores, we find that MxB is intimately associated with the inner mitochondria, where it plays an essential role in the maintenance of mitochondrial cristae and DNA stability of this organelle. Results MxB associates with mitochondria in cultured cells As the related family member MxA has been shown to convey a partial resistance to the hepatocyte-centric HBV 33 , while associating with membranous structures in hepatocytes, our goal is to compare the distribution and expression of the highly related family member MxB in hepatocyte cell lines and in the whole liver. Using several different antibodies and tagged constructs to MxB, we observe the punctate nuclear staining previously observed by others 34 , 35 , 36 . In addition to this localization, we observe a fusiform and linear staining pattern in the hepatoma cell line Hep3B and HeLa cells, using a previously published MxB antibody and exogenously expressed green fluorescent protein (GFP)/Myc-tagged constructs (Fig. 1a–d and Supplementary Fig. 1a–c ). Co-staining these cells with a variety of different organelle marker antibodies reveals that the MxB labeling represents mitochondria as confirmed by antibodies to cytochrome c oxidase 4 (CoxIV) (Fig. 1a–d and Supplementary Fig. 1a–c ). This localization is consistent when staining with three distinct MxB antibodies or expressing either GFP or Myc-tagged MxB, whereas the characteristics of labeling using these different probes varies from fine to large puncta coating the mitochondria. Fig. 1: MxB is localized to mitochondria in cultured cells. Fluorescence images of Hep3B cells or HeLa cells, stained with a polyclonal antibody (IJ-GP) to MxB ( a , c ), or expressing exogenous human MxB-GFP ( b , d ). All cells were stained for the inner mitochondrial membrane marker CoxIV. MxB is visualized as small or large puncta that align closely with mitochondrial membranes. Boxed regions provide a higher magnification for color overlays showing the alignment of MxB (green) with mitochondria (red) in both cell types. Scale bars, 10 μm. e Western blot analysis of different human tissues (Br = brain, Lng = lung, Liv = liver, LN = lymph node, SP = spleen, Tst = testis, Tns = tonsil) for endogenous MxB that appears enriched in the liver, lymph node, testis, tonsil, and pig liver. f Western blotting comparing MxB expression in HeLa cells, three different hepatocyte cell lines, and a human monocyte cell line (THP-1), as well as primary human and pig hepatocytes. Expression levels are exceptionally high in primary hepatocytes, the HepG2 and Huh7 cell lines (H3B = Hep3B, HG2 = HepG2, Hu7 = Huh7, THP = THP-1, Hu = human hepatocytes). g Western blotting of four distinct cell lines comparing endogenous MxB expression under control conditions vs. treatment with 1,000 IU/ ml -1 of IFN-α−2Α. HeLa, Hep3B and HepG2 cells induced MxB expression by IFN treatment, but not Huh7 cells. Full size image Western blot (WB) analysis of various human tissues for endogenous MxB shows that MxB appears enriched in the liver, lymph node, testis, tonsil, and pig liver. Of note is the fact that human lymph node and tonsil tissues possess two bands, whereas the liver, testis, and pig liver indicate a single band (Fig. 1e ). Accordingly, appreciable levels of endogenous MxB are observed in several human hepatoma cell lines (Hep3B, HepG2, Huh7) and primary hepatocytes isolated from human and pig livers (Fig. 1f ). HeLa cells appear to have modest MxB levels that become more apparent upon longer exposures (Supplementary Fig. 4 ). As reported previously 34 , MxB expression is markedly increased in isolated cells treated with 1000 IU ml −1 of IFNα (Fig. 1g ). MxB has been shown to be expressed in two forms, one of which contains an N-terminal sequence required for nuclear targeting. The distribution of the long and short MxB forms have been examined in significant detail by others 33 , 34 . In our hands we find that the short form has modest affinity for the nucleus or mitochondria regardless of the tag used in comparison with the long form that associates with both organelles (Supplementary Fig. 1d-j ). As our initial morphological observations (Fig. 1 and Supplementary Fig. 1 ) show MxB associating with mitochondria as either small or large puncta, or more diffuse in nature, we think it is important to analyze live cells expressing MxB-mCherry to test for changes in MxB localization over time. One example of this dynamic distribution is depicted in Fig. 2a , showing a single Hep3B cell expressing MxB-mCherry viewed over a 16 h time period. Dramatic and unexpected changes in MxB localization are observed, as the tagged protein appears to cycle between diffuse and punctate forms during the extended viewing period. Subsequently, cells co-transfected with Mito-GFP 37 as a mitochondrial marker test whether the coalescing MxB puncta might associate with this organelle. Indeed, as shown in the images from movies of two distinct live cells (Fig. 2b, c ), MxB-mCherry appears diffuse and barely detectable at early viewing times but subsequently coalesces into numerous large puncta that associate with the tips of undulating mitochondria (inserts). As observed in Fig. 2a , these puncta appear to dissolve over time, leaving the MxB in a diffuse state. Fig. 2: MxB distribution in living cells is dynamic. a Time-lapse imaging of Hep3B cells expressing mCherry-MxB and viewed over a 16 h time period. Initially, MxB is diffuse in distribution, although condenses into several fine puncta by 5 h that subsequently dissolve but reform into much larger structures by 9 h. Surprisingly, these large aggregates also dissolve or disassemble several hours later (see Supplementary Movie 1 ). b , c Time-lapse imaging of Hep3B cells co-expressing mCherry-MxB (red) and Mito-GFP (white), and viewed over a 16 h time period to test for dynamic associations of MxB with mitochondria. Low-magnification images of cells show the tagged MxB changing from a diffuse to punctate, to diffuse distribution in close proximity of mitochondria. Higher magnification viewing (inserts) reveals that many of these formed puncta associate along or at the tips of the dynamic mitochondria (arrows) (see Supplementary Movies 2 and 3 ). Scale bars, 10 μm. Full size image MxB manipulations disrupt mitochondrial structure and function Based on the morphological observations described above, we next test whether manipulation of MxB levels might alter mitochondrial morphology. Hep3B cells transfected with either GFP-tagged or untagged wild-type (WT) or GTPase-defective mutant K131A MxB for 24 h are followed by staining with the CoxIV marker antibody. Representative images displayed in Fig. 3a, b show that mutant MxB-transfected cells possess fragmented mitochondria and significantly reduced CoxIV staining compared with adjacent untransfected cells. Importantly, in many transfected cells, mitochondrial distribution, size, and shape appear abnormal. These altered phenotypes include fragmented, twisted, or clustered morphologies (Fig. 3c and Supplementary Fig. 3a ). It is particularly surprising that 25–30% of the cells examined have markedly attenuated CoxIV staining (Fig. 3d ). To define these changes at a higher resolution, Hep3B cells transfected to express WT or K131A mutant MxB for 2 days are then fixed and embedded for electron microscopy (EM). Although untransfected cells display normal mitochondrial morphologies, many Hep3B cells expressing either of the constructs possess mitochondria with a greatly reduced number of cristae and/or numerous vacuoles or blisters (Fig. 3e–g ). Similar results are seen in transfected HeLa cells, which lose almost all of their mitochondrial cristae, leaving hollow, elongated, organelle shells (Supplementary Fig. 2a–c ). To better understand the effects of the mutant K131A form, we compare the distribution of this mutant with the WT form in Hep3B cells that are transfected with GFP-tagged constructs then fixed and viewed by fluorescence microscopy. Quantification of mitochondrial and nuclear fluorescence using ImageJ software shows a near-equal preference of the WT form for both organelles, while the K131A mutant exhibits increased association with mitochondria (Supplementary Fig. 2d ). As a biochemical comparison, Hep3B cells expressing either the WT or K131A form subjected to subcellular fractionation to obtain nuclear, mitochondrial enriched, and cytosol fractions, are blotted for corresponding organelle markers and MxB (Supplementary Fig. 2e ). By this approach, a near-equal distribution of both forms is observed. Fig. 3: Manipulation of MxB expression induces marked changes in mitochondrial morphology. a , b Fluorescence micrographs of transfected Hep3B cells (outlined in white border) expressing either an untagged ( a ) or GFP-tagged ( b ) GTPase-defective K131A mutant of MxB. Transfected cells possess mitochondria that are fragmented or distorted, or exhibit attenuated staining with antibodies to CoxIV compared with adjacent untransfected cells. c Bar graph representing morphological changes of mitochondria in control, WT, or mutant (K131A) MxB-expressing cells. Morphologies of clustered, fragmented, deformed, or normal mitochondria were counted in images of 636 Hep3B cells from three distinct experiments. d Graph representing the mean percentage of cells exhibiting a reduced CoxIV staining from 636 control or MxB-transfected cells from three distinct experiments. e – g Electron micrographs of mitochondria from Hep3B cells transfected to express MxB WT or K131A. Control cells ( e ) possess mitochondria with normal cristae, while cells expressing WT ( f ) or mutant (K131A) MxB ( g ) display mitochondria with reduced cristae and dark membrane vesiculations in the matrix (arrowheads). Scale bars, 10 μm ( a , b ), 0.5 μm ( e – g ). h – o Reducing MxB levels by siRNA or shRNA treatment promotes mitochondrial fragmentation . h Western blot of Hep3B cells treated for 5 days with siRNA or shRNA to MxB shows reduced protein levels. i – k Immunofluorescence of control Hep3B cells ( i ) or cells treated with siRNA ( j ) or shRNA ( k ) to reduce endogenous MxB levels and stained for CoxIV. Both treatments induced a marked change in mitochondrial morphology. l Bar graph representing morphological changes of mitochondria from 568 Hep3B cells from three distinct experiments. m – o Electron microscopy of control NTsi ( m ) or siRNA MxB-treated cells showing fragmented mitochondria with vesiculated cristae (arrowheads) and large vacuoles (arrows). ( c , d , l Data were analyzed using a two-tailed t -test, error bars indicate SD. Scale bars, 10 μm ( i – k ), 0.5 μm ( m – o ). Full size image To further define the role of MxB in maintaining mitochondrial morphology, Hep3B cells are subjected to MxB knockdown using both small interfering RNA (siRNA) and short hairpin RNA (shRNA) approaches. Following 5 days of treatment, cells are fixed and stained for CoxIV (Fig. 3h–k ) or prepared for EM (Fig. 3m–o ). As observed in cells expressing exogenous WT or mutant (K131A) MxB protein, these knockdown cells display altered mitochondrial morphologies including fragmented, twisted, or clustered shapes. A comprehensive table showing the effects of MxB knockdown, or overexpression, on mitochondrial shape is provided in Supplementary Fig. 3a . In contrast to the cells overexpressing tagged MxB proteins, the knockdown cells did not exhibit reduced staining for CoxIV or Tom20. EM of siRNA-treated Hep3B cells, however, did reveal highly altered mitochondria with deformed shapes, large vacuoles, and vesiculated cristae (Fig. 3m–o and Supplementary Fig. 2f–h ). Based on our findings from the expression of either WT or mutant MxB proteins, and two distinct knockdown strategies, we conclude that this dynamin family member plays an important role in the maintenance of mitochondrial shape and integrity. A central property of the Mx proteins is a substantially increased expression induced by IFN. Although MxB, but not MxA, is constitutively expressed in some cell and tissue types examined (Fig. 1e–g ), the expression levels of both Mx proteins is markedly increased in cells treated with IFN (Fig. 1g, Supplementary Fig. 3j ). We found that Hep3B cells treated with IFN for 48 h possess fragmented mitochondria in comparison with control cells (Supplementary Fig. 3b–d, h ) and slightly increased levels of MxB on the nuclear envelope vs. mitochondria by immunostaining (Supplementary Fig. 3e-g,i ). As a biochemical comparison, Hep3B cells treated as above with IFN for 48 h prior to lysis and differential centrifugation provides a rudimentary separation of nuclei from the mitochondria and cytosol. These fractions are run on SDS-polyacrylamide gel electrophoresis (PAGE) and probed by WB with antibodies to MxB and MxA (Supplementary Fig. 3j ). This blotting is consistent with the cell staining and suggests that although there is a large increase in MxB associated with both organelles, a slight preference is observed with the nuclear fraction. As a control comparison the IFN -induced MxA protein is observed exclusively in the cytosolic fraction. As light microscopy imaging of MxB in cultured cells (Fig. 1 ) suggests an intra-mitochondrial localization consistent with the structural defects we observe by EM (Fig. 3 ), we utilize biochemical approaches to further define MxB localization. To this end, standard subcellular fractionation methods using pig liver as a tissue source 38 , 39 are used. As displayed in Fig. 4a , a substantial enrichment of CoxIV is observed through sequential gradient centrifugation steps when comparing liver homogenate with a highly enriched mitochondrial fraction (Pure-Mito). Importantly, MxB protein levels are markedly increased in the “pure” mitochondrial fraction compared with all other isolated fractions, including the nuclear pellet fraction in which MxB is expected to reside. We next subject this “pure” mitochondrial preparation to a protease protection assay that is used often to define the topology of proteins associated with membranous organelles 40 . A freshly isolated, enriched, mitochondria fraction is treated with different concentrations of trypsin for 20 min at 4 °C then solubilized in sample buffer prior to SDS-PAGE and WB analysis for MxB and the mitochondrial proteins OPA1, Tom20, or CoxIV (Fig. 4b ). Each of these four proteins display distinct sensitivities to the protease with the outer mitochondrial transmembrane protein Tom20 showing the most degradation followed by OPA1. MxB appears to be largely protected from the protease treatment while the inner membrane protein CoxIV remains intact, perhaps due to a tight association with the 12 other members of the respiratory Complex IV. Fig. 4: MxB resides within the mitochondrial matrix and is essential for normal mitochondrial function. a Western blot analysis of MxB distribution observed by subsequent steps of biochemical mitochondrial enrichment from pig liver. Fractions include: homogenate (Hom), a nuclear pellet (Nucl), a crude mitochondria fraction (Crude), a pure mitochondria fraction (pure), a cytosolic fraction (Cyto), and the endoplasmic reticulum (ER); all fractions were probed with marker antibodies to the mitochondria (CoxIV), actin, endoplasmic reticulum (Calnexin), and Golgi (GM130). Equal protein loads were used for each lane. b Western blot analysis of a mitochondria protease protection assay. A freshly isolated, enriched, mitochondria fraction from pig livers was treated with 0, 0.2, or 1 mg ml −1 of trypsin for 20 min, then probed with antibodies to MxB, or the mitochondrial proteins OPA1, Tom20, and CoxIV. MxB and respiratory complex enzyme CoxIV appear resistant to the higher levels (1.0 mg ml −1 ) of trypsin, whereas Tom20 and OPA1 are digested. c – g APEX2-based EM of the mitochondria in Hep3B cells. EMs of Hep3B cells exposed to transfection reagent as a control ( c ) or transfected to express either WT MxB ( d , e ) or mutant (K131A) MxB ( f , g ) fused to APEX2. Mitochondria in control cells show normal staining and structure, whereas WT MxB-expressing cells ( d , e ) exhibit prominent labeling of nuclear pores (arrow heads) and mitochondrial membranes (arrows). Mitochondrial cristae staining also appear reduced in these organelles. In contrast, mitochondria of MxB mutant-expressing cells ( f , g ) are darkly labeled throughout with exceptionally dark staining regions observed along the cristae (arrows) and in the matrix. These organelles also display vacuoles, protrusions, and disordered cristae. h HeLa cells and Hep3B cells ( i ) were co-transfected to express the mutant MxB K131A protein and a Mito-GFP marker or co-stained for CoxIV. Cells were fixed and stained for MxB. Both cell types showed an increased colocalization of the mutant MxB with mitochondria compared with WT protein. Scale bars, 0.5 μm ( c – g ), 10 μm ( h , i ). Full size image To support the morphological and biochemical observations described above, we attempt to define the localization of MxB at the ultrastructural level using a recently developed APEX2 approach for EM, in which cultured cells express the protein of interest tagged with a 28 kDa ascorbate peroxidase (APEX2) 41 . The chimeric protein maintains its activity within cells despite aldehyde fixation and allows the focused deposition of an electron-dense reaction product where the protein is localized. This application has been used to provide accurate localization for a series of known cellular proteins including several to different parts of the mitochondria including the outer mitochondrial membrane, the intermembrane space, and the matrix 41 . Accordingly, Hep3B cells expressing an MxB-APEX2 construct are then processed and viewed by EM. As shown in Fig. 4c , mock-transfected cells that are subsequently treated the same as MxB-APEX2-transfected cells, possess normal mitochondria of low contrast due to the lack of post staining. In contrast, cells expressing MxB WT-APEX2 exhibit dense staining of nuclear pores (Fig. 4d , arrowheads), consistent with the findings of others 41 along with darkly stained outer and inner mitochondrial membranes (arrows). Many of these mitochondria also appear to a have reduced number of cristae. Cells expressing the mutant MxB K131A-APEX2 display mitochondria filled with considerably darker reaction product in the matrix (Fig. 4f, g ), which is also observed in cells viewed by fluorescence microscopy (Fig. 4h,i ). As the biochemical and morphological localizations of MxB are consistent with the observed changes in mitochondrial cristae in MxB-altered cells, it is important to test whether mitochondria within these cells had normal or impaired function. To this end, Hep3B cells are transfected to express vector, WT, or mutant MxB protein (Fig. 5a–c ) or, alternatively, are treated with either non-targeted siRNA or siRNA to MxB for 5 days (Fig. 5d, e ) prior to incubation with 50 nM Rho123 dye for 30 min as a probe for mitochondrial membrane potential. The capacity of the mitochondria to actively sequester this dye is widely used to measure the bioenergetics of this organelle 42 . Importantly, all cells in which MxB levels are manipulated show substantially less Rho123 fluorescence than did control cells (Fig. 5f ), suggesting that cells with morphologically altered mitochondria due to MxB manipulation have reduced cristae and are functionally compromised. Fig. 5: Expression of MxB disrupts mitochondrial function and genome stability. a – f Mitochondria in MxB-expressing cells exhibit loss of the proton gradient. Fluorescence micrographs of Hep3B cells that were transfected to express either pmCherry-N1 Vector, MxB WT-pmCherry, the MxB-pmCherry K131A mutant, or exposed to siRNA treatment for 5 days followed by incubation with 50 nM of the proton-gradient sensing dye Rho123. Transfected cells (white outlines) show modest mitochondrial fluorescence, indicating a metabolic impairment. f Graph quantification of the effects of MxB expression or siRNA knockdown on mitochondrial function ( n = 529 cells over 3 independent experiments). g – n Release of the mitochondrial genome (mtDNA) into the cytoplasm in MxB-manipulated cells. Fluorescence micrographs of Hep3B cells treated for 5 days with control vectors and either siRNA or shRNA to reduce endogenous levels of MxB, OPA1, or both MxB and OPA1. Cells were then fixed and stained for mitochondrial DNA nucleoids (green) and CoxIV (red). Each manipulation shows low magnifications of representative cells with corresponding high-magnification images of boxed regions. Control-treated cells ( g , j ) exhibited close alignment of DNA with mitochondria, whereas siRNA- or shRNA- treated cells resulted in fragmented mitochondria and a significant number of DNA puncta residing in the cytoplasm ( h , k ). In comparison, OPA1 KD cells displayed shorter mitochondria with DNA nucleoids at the tips ( i ). l Fluorescence quantification of mtDNA loss into the cytoplasm under different treatment conditions. The Y -axis reflects relative amount of mtDNA outside the mitochondria ( n = 244 microscopic fields over 4 independent experiments). m , n Graphs depicting quantitative PCR of either total cellular mtDNA ( m ) or cytosolic ( n ) mtDNA. Mean fold change in mtDNA copy number between NTsi, MxB, OPA1, and MxB/ OPA1 siRNA-treated cells was measured by qPCR using a mitochondrial D-loop primer set ( m , n ; n = 3 independent experiments). Data were analyzed using a two-tailed t -test, error bars represent SD. l NTsi vs. MxBsi, OPA1si, MxBsi /OPA1si, TRC2 vs. shRNA#44: all P < 0.0001, OPA1si vs. MxBsi/OPA1si: P = 0.0007; NS, not significant ( P = 0.1827). m NTsi vs. MxBsi: P = 0.00082, OPA1si: P = 0.00069, MxB/OPA1si: P = 0.00023. n NTsi vs. MxBsi: P = 0.0027, OPA1si: NS ( P = 0.2334), MxB/OPA1si P = 0.00305. Scale bars, 10 μm. Full size image Loss of MxB expells mitochondrial genome into the cytoplasm From the combined observations detailed above showing altered mitochondrial function, form, and disrupted cristae upon manipulations of MxB levels, we predict that these mitochondria might also exhibit alterations in their genome stability. Indeed, the mammalian dynamin family member OPA1 is also named MGM1 in yeast (for mitochondria genome maintenance), as the loss of this gene leads to disorganized cristae and loss of mitochondrial DNA 43 , 44 . The mitochondrial genome of 37 genes is packaged into spherical, punctate structures known as nucleoids, which can be observed using various DNA or protein antibodies 44 , 45 , 46 . To test the effects of MxB on the morphology and distribution of mitochondrial nucleoids, Hep3B cells are treated with siRNA or shRNA to MxB, as well as non-targeted siRNA or vector-alone shRNA (TRC2) as a control for 5 days prior to fixation and staining with antibodies to mitochondrial DNA (mtDNA) and CoxIV (Fig. 5g–k ). OPA1 levels also reduced via siRNA are used as a comparison (Fig. 5i ). In control cells, each mitochondria contains one to ten nucleoids (Fig. 5g, j ). In striking contrast, the mitochondria of MxB knockdown cells are markedly fragmented (Fig. 5h, k ) with fewer nucleoids and many of the nucleoids appear to reside within the cytoplasm. Morphological quantification of these cells shows a three to four fold increase in cytoplasmic distribution of mitochondrial nucleoids in MxB-knockdown cells vs. control cells (Fig. 5l ). This loss of mtDNA is substantially less pronounced in cells subjected to a knockdown of the inner mitochondrial dynamin OPA1 (Fig. 5i, l ). To support these morphological observations, we utilize quantitative PCR (qPCR) methods to measure total mtDNA as well as cytosolic mtDNA content from Hep3B cells subjected to siRNA knockdown of MxB and/or OPA1 (Fig. 5m, n ). Interestingly, total DNA levels are nearly doubled in the knockdown cells compared with control cells. Consistent with the morphological measurements, cytoplasmic mtDNA levels in OPA1-knockdown cells are modestly increased compared with control cells, whereas these levels are markedly increased (three to four fold) in MxB-knockdown cells. To test whether the MxB-induced loss of mtDNA is dependent upon OPA1 function, Hep3B cells treated with siRNAs to both OPA1 and MxB are subjected to morphometry and qPCR as described above. When both of these dynamin family members are reduced, the loss of mtDNA into the cytoplasm trends toward control levels, suggesting that OPA1 may play a role in the genomic instability induced by a reduction of the MxB protein. Discussion In this study, we report an unexpected subcellular distribution and function for the dynamin family member MxB at the mitochondria. Although MxB is well described to exhibit an IFN-inducible expression 20 , 47 , we find it is also expressed at basal levels in several untreated tissues including human liver, lymph node, testis, and tonsil, as well as pig liver (Fig. 1e ). Further, MxB is expressed in a variety of hepatocyte cells including primary hepatocytes (Fig. 1f ) and this expression is markedly increased upon IFN treatment (Fig. 1g ). This basal level of expression may suggest that MxB performs housekeeping functions in cells in addition to its anti-viral action. Although HeLa cells do not show an MxB band in the blot provided, this can be resolved upon higher protein loading and/or longer exposure of the blot. In addition, alterations of MxB function in these cells lead to disrupted mitochondria (Fig. 1 and Supplementary Fig. 2 ). Further, despite our continued efforts, we have been unable to isolate viable MxB-knockout clones using CRISPR/Cas9 in HeLa or Hep3B cells, suggesting a critical function of this protein. The higher molecular weight band of MxB in human liver (Fig. 1e ), Hep3B, HepG2, and human primary hepatocytes (Fig. 1f ) showed prominent expression. Interestingly, when treating cells with IFN, the short form of MxB is induced, indicating that it may exist constitutively with the long form of MxB in the liver/hepatocytes. However, this lower molecular weight form of MxB needs to be confirmed to indeed be a truncated form of MxB. The widely reported distribution of MxB at the cytoplasmic face of nuclear pores is consistent with this protein containing an N-terminal nuclear localization sequence (NLS) 34 , making it readily available to prevent the nuclear uptake and chromosome integration of lentiviruses such as HIV-1 23 into the host genome. In addition to the reported localization at nuclear pores, numerous other studies have described a cytoplasmic distribution for MxB, which we believe may have represented the mitochondria. From the data described above, the localization of MxB at the mitochondria is observed consistently in several different cell types, while using multiple MxB antibodies or tagged constructs (Fig. 1a–d and Supplementary Fig. 1a, b ). Although all of these reagents indicate a nuclear and mitochondrial distribution for MxB, distinctions in the types of mitochondrial staining were observed, which varied from evenly distributed to small or large puncta. Digital recording of live Hep3B cells expressing MxB-mCherry demonstrated that a single cell may display all of these distributions (Fig. 2 ). It was particularly surprising that such large aggregates of MxB can form and dissociate, as static images of these structures suggested the formation of insoluble or denatured aggregates of a tagged protein. The dissolution of these aggregates suggests otherwise and may provide insights into the function of this antiviral dynamin protein. Whether this cycling process leads to an association of MxB with the surface or interior of the mitochondria needs to be determined. We do observe that the mitochondrial distribution is significantly increased in cells expressing an MxB K131A mutant (Fig. 4h, i ), a finding that provided us with the initial motivation for pursuing the mitochondrial association. In addition to the observed fluorescence localization of MxB to the mitochondria, substantial amounts of MxB were found in isolated organelle fractions from pig liver that coincided with mitochondrial enrichment as assessed by CoxIV levels (Fig. 4a ). Equally supportive are the findings from the organelle protein protection assays, suggesting that, as observed for OPA1 (Fig. 4b ) 40 , 48 , MxB is protected from mild trypsin exposure of 0.2 mg ml −1 and it may reside within the inner mitochondrial membrane. In contrast to OPA1, MxB remains resistant to proteolysis at higher trypsin concentrations, as is the case for CoxIV, suggesting MxB may indeed reside on the inner membrane or in the matrix within a protective complex. The mitochondrial localization of WT MxB is less clear at the ultrastructural level using the APEX2-EM method, as both the inner and outer mitochondrial membranes appeared darker than observed in control untransfected cells (Fig. 4c–e and Supplementary Fig. 4 ). It is unclear, however, if this indicates that MxB resides on both the inner and outer membranes or the WT protein has a more modest, and perhaps transient, interaction with the inner membrane. A pronounced mitochondrial staining was observed in cells expressing an APEX2-tagged MxB K131A mutant protein. In these cells, the reaction product appeared to fill the matrix in its entirety with the most intense staining coating the cristae membranes while leaving, what we interpret to be, the intermembrane space unstained. This enhanced association of the GTPase mutant was also observed in transfected cells viewed by fluorescence microscopy (Fig. 4 h, i). It will be important to define how MxB is targeted to the mitochondria as opposed to the nuclear envelope and how these localizations are exclusive to MxB compared with MxA. Defining this targeting process will further strengthen the concept of MxB as a bona fide mitochondrial dynamin family member, while providing new insights into its function. GTPase defective mutants in the conventional dynamins are generally believed to bind GTP, at perhaps a reduced affinity, but remain in a bound state as hydrolysis is defective 49 , 50 , 51 . GTP-bound dynamin is thus more likely to favor polymer assembly while associating with membranes in cells 52 , a state also induced by the GTP analog GTPγS as observed using in vitro assays 53 , 54 . Thus, the analogous GTPase mutation in MxB could explain the marked increase in mitochondrial association and represent a key functional step during its GTP binding/hydrolysis cycle. In contrast to conventional dynamins, recent reports have suggested that GTP binding disassembles both MxB and MxA polymers 28 , 55 , 56 . As these dynamin-related family members are largely assumed to interact with virally derived proteins, the membrane affinity for the MxB K131A or MxA K83A GTPase-defective mutants appears to be understudied. The distribution of MxB reported in this study suggests that it likely has multiple functions in cells beyond that of restricting a single family of infectious virus. MxA has been implicated in attenuating the infection of several distinct virus families with either RNA or DNA genomes at multiple cellular sites that include the nucleus as well as the smooth ER 25 . This is likely to be the case for MxB, as an evolutionary analysis suggests that MxB plays a central and conserved role in the IFN response to a broader range of pathogens than is currently appreciated 57 . The very recent observation that MxB acts as a potent pan-herpes virus restriction factor 58 supports this premise and may represent the first of many other MxB sensitive viral pathways yet to be identified. As additional insights into the functions of the Mx proteins expand, it is likely that newly identified intracellular localizations will also increase. Conventional dynamins have been reported to participate in a wide variety of key cellular processes beyond the liberation of caveolae or clathrin-coated pits from the plasma membrane. Some of these functions include budding of nascent secretory vesicles from the trans-Golgi network, vesiculation of autophagosomes and lysosomes, as well as centriole separation and midbody scission during mitosis 5 , 6 , 50 . MxB has been reported to possess a NLS at the N-terminus and is likely to be expressed in at least two forms 34 . It is attractive to predict that the form lacking the N-terminal NLS is targeted to the mitochondria, although our expression studies have not supported this concept. It is also not clear whether the basally expressed form is more or less prone to associate with the nucleus or mitochondria compared with the IFN-induced form. The fact that alterations in MxB levels have a pronounced effect on mitochondrial shape and integrity is consistent with the fact that this protein appears to have a mitochondrial distribution. As displayed in Fig. 3 and Supplementary Fig. 2 , manipulation of MxB levels in cells by either siRNA or shRNA knockdown or overexpression of WT or mutant MxB protein leads to distortions in mitochondrial shape along with a profound disruption in the organization of mitochondrial cristae. These changes accompany a marked loss of mitochondrial function such as an inability to maintain a proton gradient (Fig. 5a–f ) and a substantial translocation of the mitochondrial nucleoids (mtDNA) into the cytoplasm (Fig. 5g–n ). We were surprised to find that in addition to the expulsion of the nucleoids from the mitochondria into the cytoplasm in MxB-knockdown cells, these cells also appeared to possess an increase in the total mtDNA content as assessed by qPCR. This was observed repeatedly in cells treated with siRNA to either MxB or OPA1. Consistent with this finding is a study of muscle fibers from patients with OPA1 mutations, showing a two- to fourfold increase in mtDNA copy number 59 . The authors suggest this increase represents a compensatory mitochondrial proliferative response to maintain a needed level of WT mtDNA. The stability of the mitochondrial genome appears to be directly related to the integrity of mitochondrial cristae, as has been suggested for the yeast protein MGM1, named after its involvement with mitochondrial genome maintenance 60 . A putative role for MxB in cristae fusion or genome tethering based on its localization and effects on mitochondrial structure is certainly less developed than what is now known about the yeast MGM1 and mammalian OPA1 proteins after many years of investigations 16 , 61 , 62 . Future detailed studies on the role of MxB on the inner mitochondrial membrane dynamics will be important, although challenging, as defining the functions of putative inner mitochondrial membrane proteins are particularly difficult. It is well documented that shape, as well as cristae number and structure, are often altered in response to disease, energy burden or metabolic state of the host cells 63 , 64 . Indeed, since the discovery of yeast DNM1 65 and the mammalian homolog DRP1/DLP1 66 almost two decades ago, its precise role in mitochondrial fission remains somewhat elusive, particularly with the observation that conventional Dyn2 is recruited to the constriction site and may provide the final scission event 18 . From our perspective, what makes the findings of this current study so exciting is the concept of a dynamin family member that is markedly induced in response to viral infection also participates in mitochondrial form, function, and integrity. Further, the two known viruses which MxB attenuates are known to utilize or modify mitochondria during infection. HIV-1 infection is well described to cause alterations in mtDNA and induce mitochondria-centric apoptosis 67 , 68 . In addition, herpes virus, a recently identified MxB target 58 , was originally observed to have an intimate association with the host’s cell mitochondria over 60 years ago 69 . It is now well known that infection by this virus results in a rapid elimination of mtDNA from the host cells 70 , 71 . Thus, it is especially attractive to postulate that MxB could be inducing genomic changes in the host’s cell mitochondria as part of the innate immune response. It is now clear that the inappropriate discharge of mtDNA into the cytoplasm is recognized by the host cell as a “prokaryotic infection event” that engages the DNA sensor cGAS to promote STING-dependent signaling 45 , 72 , 73 . It will be particularly interesting to test whether MxB is responsible for the observed mitochondrial release of mtDNA during infection by herpes virus and other pathogens. Methods Antibodies and reagents Rabbits were immunized with the keyhole limpet hemocyanin-conjugated human MxB N-terminal peptide, 5′-KPWPYRRRSQFSSRKYLKKEMNSFQQQP-3′, and the crude antisera were affinity purified using an agarose column conjugated with the appropriate high-performance liquid chromatography-purified synthetic peptide and a low pH elution buffer, according to the manufacturer’s directions (Pierce Chemical Co., Rockford, IL). The purified anti-MxB antibody (immunofluorescence staining (IF) 5 μg ml –1 ; WB 1:1000) was dialyzed (molecular weight cut-off 12,000–14,000 kDa) against Dulbecco’s phosphte-buffered saline (PBS) (D-PBS; 8.1 mM Na 2 HPO 4 , 1.2 mM KH 2 PO 4 pH 7.2, 138 mM NaCI, 2.7 mM KCI, 0.9 mM CaCI 2 , 0.5 mM MgCl 2 ) containing 0.04% NaN 3 , concentrated using polyethylene glycol, and in some cases, 2 μg ml –1 bovine serum albumin (BSA) was added 74 . The other anti-MxB antibodies [Guinea Pig (IF 1:200; WB 1:1000) and Rabbit (IF 5 μg ml –1 ; WB 1:500)] were a generous gift from Dr Ilkka Julkunen 34 and Dr Chen Liang 23 . The anti-MX2 rabbit polyclonal antibody (NBP1-81018) (IF 1:100; WB 1:500) was from Novus (Centennial, CO). The CoxIV (3E11) (IF 1:250; WB1:1000) rabbit mAb and (4D11-B3-E8) (IF 1:100; WB1:1000) mouse mAb; Tom20 (D8T4N); (WB1:1000) rabbit mAb; OPA1 (D6U6N) (WB1:1000) rabbit mAb; GM130 (D6B1) XP (WB1:1000) rabbit mAb; and GAPDH (D16H11) XP (WB1:1000) rabbit mAb were from Cell Signaling (Danvers, MA). The OPA1 mouse mAb (WB1:500) was from BD Biosciences (San Jose, CA). The mtDNA antibody (WB 1:100) was from EMD Millipore (Temecula, CA). The calnexin-ER marker (WB 1:1000) was from Abcam (Cambridge, MA). The anti-Actin (WB 1:2000) was from Sigma (St Louis, MO). Goat anti-rabbit and goat anti-mouse secondary antibodies conjugated to either Alexa-Fluor-488 or -594 used for IF staining (1:500) were all obtained from Thermo Fisher Scientific (Rockford, IL), and horseradish peroxidase-conjugated goat anti-rabbit and goat anti-mouse were from BioSource International, Inc. (Camarillo, CA), which were used for WB analysis. Miniprep Express TM Matrix was from MP Biomedicals (Solon, OH). Restriction enzymes were from New England Biolabs (Ipswich, MA). DNA ladders (1 kb and 1 kb plus) were from Invitrogen (Carlsbad, CA) and all other chemicals and reagents, unless otherwise stated, were from Sigma (St Louis, MO). Unprocessed blottings of MxB can be found in Supplementary Fig. 4 . Plasmid construction The following primers were used to amplify MxB WT-pCR3.1 (without tag) from the plasmid pBS-T7/MxB, a generous gift from Dr G Kochs: MxBWT5′-ATG TCTAAGGCCCACAAGCCTTGGCCCT-3′; MxBWT3′-TCAGTGGATCTCTTTGCT GGAG-3′. Full-length PCRs were performed using the XL PCR kit (Applied Biosystems, Branchburg, NJ) and the PCR fragments were ligated into the eukaryotic expression TA-vector pCR3.1 (Invitrogen, Carlsbad, CA). The MxB inserts from the pCR3.1 constructs were excised by digestion with the corresponding enzymes (XhoI and BamHI for GFP and mCherry tag; BamHI and XhoI for Myc tag) and sub-cloned into the expression vectors pEGFP-N1/pmCherry-N1 (Clontech, Palo Alto, CA) and pcDNA3.1 (Invitrogen, Carlsbad, CA). The constructs inserted into pEGFP-N1 /pmCherry-N1 and pcDNA3.1 have no intervening stop codons. MxB K131A was generated using PCR-based mutagenesis methods. To construct MxB WT/K131A-APEX2-pCR3.1, we purchased APEX2-OMM 41 plasmid from Addgene (Cambridge, MA) and reconstructed APEX2 into MxB WT/K131A-pCR3.1. The mito-GFP 37 was made using the MTS (mitochondrial-targeting sequence) from IVD (Isovaleryl CoA dehydrogenase) for the template and ligated into pEGFP N1 vector by EcoRI and BamHI cut sites. All plasmids were purified with the plasmid Maxi kit from Qiagen (Germantown, MD) and DNA constructs were verified by restriction enzyme analysis and sequencing (The Mayo Molecular Biology Core; GENEWIZ, South Plainfield, NJ). Sequences were analyzed using DNA* analysis software (DNA star, Madison, WI). siRNA/shRNA and transfections siRNAs targeting Human Mx2 (M-011736-01-0005) and OPA1 (M-005273-00-0005) were purchased from Dharmacon (Lafayette, CO) and transfected with Lipofectamine RNAiMAX (Invitrogen, Carlsbad, CA) following the standard protocol. Mx2 shRNA and TRC vector (shRNA control) were purchased from Sigma (St Louis, MO) and delivered into cells following the Lentiviral particles standard protocol (Santa Cruz). Tissue, cell culture, and transfection Human tissue was from the Mayo Clinic Biobank for Gastrointestinal Health Research (Institutional Review Board (IRB) 622-00) and was de-identified frozen tissue. Consent and approval were waived, because the IRB determined the frozen tissue did not constitute research involving human subjects as defined under 45 CFR 46.102. Primary pig hepatocytes isolated and cultured from female, Large Domestic Cross-bred White Pigs, 2–3 months old 74 , were a kind gift from Dr Scott Nyberg, Mayo Clinic. All animals received humane care and procedures were performed with approval of the Institutional Animal Care and Use Committee at Mayo Clinic and are in accordance with guidelines set forth by the National Institutes of Health. Human hepatocytes were from Bioivt (Westbury, NY). HeLa cells, an adenocarcinoma from human cervix (ATCC CCL-2), and Huh-7, a hepatocellular carcinoma from human liver 75 , were grown in Dulbecco’s modified Eagle’s medium, plus 10% fetal bovine serum, 100 U ml −1 penicillin, and 100 μg ml −1 streptomycin (Gibco, Waltham, MA). Hep3B and HepG2 cells, both hepatocellular carcinoma from human liver (ATCC HB-8064) and (ATCC HB-8065), respectively, were maintained in minimum Eagle’s medium with 10% fetal bovine serum, 100 U ml −1 penicillin, and 100 μg ml −1 streptomycin (Gibco, Waltham, MA). THP-1 cells are monocyte from acute leukemia (ATCC TIB-202). All cells are cultured in 5% CO 2 and 95% air at 37 °C incubator. Cells were cultured in T-75 flasks (Fisher Scientific, Pittsburgh, PA). Cells were transfected using Lipofectamine 2000 (Invitrogen, Carlsbad, CA) according to the manufacturer’s protocol. Immunofluorescence and electron microscopy For IF microscopy, cells on coverslips were rinsed with D-PBS and fixed for 20 min with 2.5% formaldehyde in PIPES buffer (0.1 M Pipes pH 6.95, 3 mM MgSO 4 , 1 mM EGTA). After rinsing with D-PBS, cells were permeabilized with 0.1% Triton X-100 in D-PBS for 2 min, rinsed in D-PBS, and incubated in blocking buffer (5% normal goat serum and 5% glycerol in D-PBS) for 1 h at 37 °C. Cells were incubated in primary antibodies diluted in blocking buffer and rinsed repeatedly in D-PBS before incubating in the appropriate fluorescently labeled secondary antibodies diluted in blocking buffer. Cells were then washed extensively with D-PBS, rinsed with distilled water, and mounted on a glass slide with mounting reagent Prolong gold (Invitrogen, Carlsbad, CA). Fluorescence micrographs were acquired using an AxioObserver D.1 epifluorescence microscope (Carl Zeiss, Thornwood, NY) equipped with a 100 W mercury lamp using a ×63, 1.4 NA objective lens, an Orca II digital camera (Hamamatsu Photonics, Hamamatsu, Japan) and ZEN software (Carl Zeiss Microscopy LLC, Thornwood, NY). Images were processed using Adobe Photoshop (Adobe Systems Incorporated, San Jose, CA). For standard transmission EM, cells on carbon-coated coverslips were rinsed in 37 °C Hank’s buffered salt solution and fixed with 37 °C primary fixative (100 mM cacodylate pH 7.4, 60 mM sucrose, 2.5% glutaraldehyde) for 1 h at room temperature, rinsed three times with washing buffer (100 mM cacodylate pH 7.4, 200 mM sucrose) then fixed in the secondary fixative (50 mM cacodylate pH 7.4, 100 mM sucrose, 1% OsO 4 ) for 1 h at room temperature, and rinsed three times in water then fixed in 1% uranyl acetate in water for 1 h at room temperature. Samples were then dehydrated in a graded ethanol series, embedded in Quetol 651 (Ted Pella, Redding, CA), and polymerized in a 65 °C oven overnight. After removal from the oven, the coverslip was removed from the bottom of the sample, the block trimmed down to a trapezoid 1 mm wide at the base, 100 nm thin sections were cut and viewed on a Joel 1200 transmission electron microscope (Jeol Ltd, Tokyo, Japan). For cells expressing the APEX2 tag, after primary fixation for 30 min, an electron-dense reaction product was generated by incubating the samples with 1 mg ml −1 DAB-HCl (3,3’-diaminobenzidine tetrahydrochloride; Electron Microscopy Sciences, Hatfield, PA) + 0.012% H 2 O 2 in washing buffer for 3 or 6 min at room temperature, washed four times in washing buffer, then processed the same as the other EM samples (secondary fix through embedding). Live-cell imaging Hep3B cells were transfected with MxB-mCherry and mito-GFP in a six-well plate. After transfection, cells were re-plated into 35 mm glass-bottomed imaging dishes (Cell E&G, San Diego, CA) in MEM media with 10% fetal bovine serum, 100 U ml −1 penicillin, and 100 μg ml −1 streptomycin (Gibco, Waltham, MA). The 35 mm imaging dish was placed in a stage top incubator at 37 °C and 5% CO 2 on a Zeiss AxioObserver microscope (Carl Zeiss, Thornwood, NY) equipped with a Colibri 7 LED light source, a Zeiss Axiocam 702 digital camera (Carl Zeiss, Thornwood, NY) and images were captured every 5 min for 16–24 h with Zen software (Carl Zeiss, Thornwood, NY). Quantitative methods and statistical analysis For optical measurements of mitochondrial DNA, fluorescence micrographs of Hep3B cells stained for CoxIV and mtDNA were analyzed with ImageJ software 76 and the degree of overlap was calculated using the Just Another Colocalization Plugin 77 . Briefly, nuclei were deleted in the mtDNA channel, then using the plugin, thresholds were set for the individual channels and the amount of overlap between the two channels was calculated using the Manders M2 coefficient (fraction of mtDNA overlapping with the mitochondria). For statistical analysis, graphs were generated and statistical analyses performed using the two-tailed t -test using GraphPad Prism software (San Diego, CA). Rhodamine 123 assay Fluorescence micrographs of Hep3B cells that were knocked down with siRNA or transfected to express either vector (pmCherry-N1), MxB WT–pmCherry-N1, or MxB K131A–pmCherry-N1 for 5 days. On day 5, cells were loaded with 50 nM Rho123 for 30 min, washed 3 times, and incubated at 37 °C for 90 min prior to any imaging. Images of cells were taken in both fluorescein and mCherry channels, and the mitochondrial fluorescence of the transfected cells was quantified and graphed. Vector-expressing cells or NTsi control cells were treated with carbonyl cyanide m-chlorophenylhydrazone (CCCP) to depolarize the mitochondria and their fluorescence intensity set to zero to calibrate the graphs. qPCR measurements of mitochondrial DNA The whole cell and cytosolic DNA was purified by QIAamp DNA Mini Kit (250) LN:154041343 (Qiagen, Valencia, CA). Quantitative real-time PCR was done using Lightcycler 480 system (Roche Life Science, Indianapolis, IN) and analysis was performed using the Roche analysis software. The thermo-cycler conditions were 95 °C for 5 min, 95 °C for 30 s, 60 °C for 30 s, 72 °C for 30 s, totalling 50 cycles. The primer sequences used to amplify human D-loop were: sense 5′-CATCTGGTTCCTACTTCAGGG-3′ and antisense 5′-CCGTGAGTGGTTAATAGG GTG-3′. Mitochondrial purification and protease protection assay Enriched mitochondria (mito) fractions were isolated from pig liver using a modified protocol at 4 °C 39 . Approximately 10 g of pig liver was resuspended in 50 ml of IB-1 buffer (225 mM mannitol, 75 mM sucrose, 0.5% BSA, 0.5 mM EGTA, and 30 mM Tris-HCl pH 7.4) and homogenized 10 or more strokes at 1500 r.p.m. Homogenate was then centrifuged twice at 740 × g for 5 min to separate unbroken cells and nuclei (pellet) from the rest of the fractions (supernatant). The latter was further centrifuged at 9000 × g for 10 min to separate crude mitochondrial pellet from cytosolic fraction containing microsomes/ER (supernatant). For isolation of the cytosol and ER, the supernatant was centrifuged at 95,000 × g for 30 min using a SW41 rotor (Beckman)—the resulting pellet contains ER with cytosol being present in the supernatant. The crude mito pellet was further gently resuspended (not to disrupt the outer mito membrane) in 20 ml of IB-2 (225 mM mannitol, 75 mM sucrose, 0.5% BSA, and 30 mM Tris-HCl pH 7.4) and centrifuged twice at 10,000 × g for 10 min. After the final centrifugation, crude mito pellet was resuspended in 2.25 ml of MRB (250 mM mannitol, 5 mM HEPES, and 0.5 mM EGTA). To isolate pure mito fraction, 1.8 ml of crude mito fraction was layered on Percoll gradient followed by centrifugation at 95,000 × g for 30 min. A dense band containing purified mito at the bottom of the tube was collected using a glass Pasteur pipette and diluted 10× with MRB followed by two centrifugations at 6300 × g for 10 min. The final pellet was resuspended in a small volume of MRB and used for mitochondria protease protection assay 40 . Briefly, 100 μg of freshly isolated pure mito fraction was treated with 0.2 or 1 mg ml −1 trypsin in 100 μl of mitochondria digestion buffer (10 mM sucrose, 0.1 mM EGTA, and 10 mM Tris/HCl pH 7.4) for 20 min on ice. The reaction was terminated by boiling the samples in Laemmli buffer + β-mercaptoethanol for 5 min. Rigor and reproducibility The type of statistical test, n -values, and P -values are all listed in the figure legends or in the figures. All microscopy data were quantified using ImageJ or Fiji, and figures were assembled using Adobe Photoshop and are representative of at least three independent experiments with similar results. All experiments were performed at least three times, except for Fig. 4a–b , which was fractionated one time and blotted three times, Supplementary Fig. 2a–c , which was done one time and represents 18 cells, and Supplementary Fig. 2d , i, which was each done one time. Supplementary Movie 1 is representative of 24 cells over 6 experiments and Supplementary Movies 2 and 3 are representative of 33 cells over 5 experiments. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The datasets generated and analyzed during this study are available from the corresponding author upon reasonable request. Figures 1 – 5 and Supplemental Figs. 1 – 5 have associated raw data. Data for the graphs in the manuscript are available in the associated source file.

A protein known to help cells defend against infection also regulates the form and function of mitochondria, according to a new paper in Nature Communications. The protein, one of a group called myxovirus-resistance (Mx) proteins, help cells fight infections without the use of systemic antibodies or white blood cells. The authors report that MxB, which is associated with immune response to HIV and herpes virus, is key to mitochondrial support. "Our work provides new insights into how this dynamin MxB protein assists in fighting viral infections, which could have substantial health implications in the future," says Mark McNiven, Ph.D., a Mayo Clinic cell biologist and senior author. Viral infection In response to infection, a cell releases interferon and neighboring cells ramp up Mx protein production. The authors replicated previous findings that MxB blocks nuclear pores and MxB increases markedly when cells are treated with interferon. But they also show that some MxB is present in most immune tissues, such as tonsil, prior to a "red alert" and that it has another role. "We were surprised to see MxB present on, and in, mitochondria," says Hong Cao, Ph.D., a Mayo Clinic research scientist and first author. "That it is both induced in response to infection and vital to mitochondrial integrity is exciting, considering that HIV and herpes alter mitochondria during infection." Protecting the generator The authors report that during infection, MxB dynamically condenses, dissolves and reforms over time, and traced MxB's travels to the nuclear pores, as well as to the tips and along mitochondria. They also show, via a cell line that can't make MxB in response to interferon, that mitochondrial cristae are affected by MxB, as well. "Without active MxB protein, mitochondria become nonfunctional, no longer produce energy, and kick out their DNA genome into the cytoplasm," says Dr. Cao. "These cells are not happy, but may have the capacity to survive a viral infection." History of mitochondrial investigation The work of Dr. Cao and team builds on the findings of mitochondrial investigators at Mayo. "Over two decades ago, our lab discovered a set of proteins that perform mechanical work to shape and pinch mitochondria," says Dr. McNiven. That discovery led to a variety of research initiatives across the international mitochondria field into not only basic research questions, but also into clinical areas. This work shows that mitochondrial dynamics, such as fission and fusion, are vital functions. They regulate cell death needed to retard cancer cell growth and the turnover of damaged mitochondria needed to prevent neurodegenerative disorders, and contribute to antiviral cell immunity, to name a few. The next steps, Dr. McNiven says, are to continue to investigate how MxB is targeted to and internalized by mitochondria, and how its association induces such drastic changes to biology of this organelle.

10.1038/s41467-020-14727-w

Biology

Tracing the rapid evolution of spermatogenesis across mammals

Henrik Kaessmann, The molecular evolution of spermatogenesis across mammals, Nature (2022). DOI: 10.1038/s41586-022-05547-7. www.nature.com/articles/s41586-022-05547-7 Journal information: Nature

https://dx.doi.org/10.1038/s41586-022-05547-7

https://phys.org/news/2022-12-rapid-evolution-spermatogenesis-mammals.html

Abstract The testis produces gametes through spermatogenesis and evolves rapidly at both the morphological and molecular level in mammals 1 , 2 , 3 , 4 , 5 , 6 , probably owing to the evolutionary pressure on males to be reproductively successful 7 . However, the molecular evolution of individual spermatogenic cell types across mammals remains largely uncharacterized. Here we report evolutionary analyses of single-nucleus transcriptome data for testes from 11 species that cover the three main mammalian lineages (eutherians, marsupials and monotremes) and birds (the evolutionary outgroup), and include seven primates. We find that the rapid evolution of the testis was driven by accelerated fixation rates of gene expression changes, amino acid substitutions and new genes in late spermatogenic stages, probably facilitated by reduced pleiotropic constraints, haploid selection and transcriptionally permissive chromatin. We identify temporal expression changes of individual genes across species and conserved expression programs controlling ancestral spermatogenic processes. Genes predominantly expressed in spermatogonia (germ cells fuelling spermatogenesis) and Sertoli (somatic support) cells accumulated on X chromosomes during evolution, presumably owing to male-beneficial selective forces. Further work identified transcriptomal differences between X- and Y-bearing spermatids and uncovered that meiotic sex-chromosome inactivation (MSCI) also occurs in monotremes and hence is common to mammalian sex-chromosome systems. Thus, the mechanism of meiotic silencing of unsynapsed chromatin, which underlies MSCI, is an ancestral mammalian feature. Our study illuminates the molecular evolution of spermatogenesis and associated selective forces, and provides a resource for investigating the biology of the testis across mammals. Main The rapid evolution of the testis across mammals is probably mainly explained by positive selection associated with sperm competition, which reflects the evolutionary pressure on males to achieve reproductive success 7 . Consequently, testis sizes, sperm production rates, sperm morphologies and other cellular traits substantially vary across mammals, even between closely related species such as the great apes, due to great mating system differences, especially regarding the extent of female promiscuity 7 . The rapid evolution of the testis is reflected at the molecular level. Previous gene expression comparisons for various organs across mammals revealed that rates of evolutionary expression change are highest in the testis, probably due to frequent adaptive changes but potentially also widespread relaxation of purifying selection 1 , 2 , 3 , 4 , 5 , 6 . Consistently, genes with testis-specific expression tend to be enriched with genes whose coding sequences have been shaped by positive selection 8 . In addition, new genes that emerge during evolution tend to be predominantly expressed in the testis and thus probably also contribute to its rapid phenotypic evolution 3 , 9 . The testis also shows several other unique molecular features. First, chromatin in spermatogenic cells is massively remodelled during spermatogenesis, a process that culminates in the tight packaging of DNA around protamines in the compact sperm head 10 . This remodelling leads to widespread leaky transcription in the genome 11 , which in turn probably facilitates the initial transcription and, hence, the frequent emergence of new testis-expressed genes and alternative exons during evolution 3 , 9 , 11 , 12 . Second, the differentiation of sex chromosomes from ancestral autosomes triggered the emergence of MSCI in eutherians and marsupials 13 (therians), which led to the establishment of backup gene copies that substitute for parental genes on the X during meiosis 3 , 14 . In spite of MSCI, the X chromosome has become enriched with testis-expressed genes during evolution 3 , 15 , 16 , 17 , 18 , 19 , 20 , presumably due to sexually antagonistic selective forces favouring the fixation of male-beneficial mutations on this chromosome 21 . Finally, translational regulation of transcriptomes is widespread across spermatogenesis 6 . Previous large-scale transcriptomic investigations of testis evolution were largely limited to bulk-organ samples 1 , 2 , 3 , 4 , 5 , 6 , 19 . Recent high-throughput single-cell (sc) or single-nucleus (sn) RNA-sequencing (RNA-seq) technologies enable detailed investigations of the cellular and molecular evolution of the testis, as exemplified by two scRNA-seq comparisons between human, macaque and mouse 22 , 23 , but a comprehensive investigation of the evolution of spermatogenesis across all main mammalian lineages is lacking. Here we provide an extensive snRNA-seq resource covering testes from ten representative mammals and a bird, enabling detailed comparisons of spermatogenic cells and underlying gene expression programs within and across mammals ( ). Our evolutionary analyses of these data unveiled ancestral as well as species- and lineage-specific cellular and molecular characteristics of mammalian spermatogenesis. Spermatogenesis across 11 species We generated snRNA-seq data for testes from ten species that cover the three main mammalian lineages and include key primate species, representing all simian (anthropoid) lineages (Fig. 1a ): eutherian mammals (representatives for five of the six extant ape lineages, including humans; rhesus macaque, an Old World monkey; common marmoset, a New World monkey; and mouse), marsupials (grey short-tailed opossum) and egg-laying monotremes (platypus). Corresponding data were generated for a bird (red jungle fowl, the progenitor of domestic chicken; hereafter referred to as ‘chicken’), to be used as an evolutionary outgroup. The dataset consists of 27 libraries, with one to three biological replicates per species and a median of roughly 275 million snRNA-seq reads per library (Supplementary Table 1 ). We refined and extended existing genome annotations across all species on the basis of bulk-testis RNA-seq data (seven libraries) (Supplementary Tables 1 and 2 and Methods ), to ensure optimal read-mapping and prevent biases in cross-species analyses. After quality controls and filtering steps ( Methods ), we obtained transcriptomes for a total of 97,521 high-quality nuclei for the 11 species, with a mean of 8,866 cells per species, a median of 1,856 RNA molecules (unique molecular identifiers (UMIs)) detected per cell and low percentages of mitochondrial UMIs (Supplementary Fig. 1 and Supplementary Table 1 ). Fig. 1: snRNA profiling across ten mammals and a bird. a , Species sampled and uniform manifold approximation and projection (UMAP) of snRNA-seq datasets. UMAP of the integrated primate dataset (dashed box), showing undifferentiated and differentiated SG (undif. SG and dif. SG, respectively), leptotene, zygotene, pachytene and diplotene SCs (lept. SC, zyg. SC, pach. SC and dipl. SC, respectively), spermatids (SD) and somatic cell types. Chimp., chimpanzee. b , Principal component (PC) analysis of cell-type pseudo-bulks. Species and lineages are encircled by a dashed line. Each symbol represents an individual. c , Gene expression phylogeny based on pseudo-bulk transcriptomes for whole testes. Bootstrap values (4,498 1:1 orthologous amniote genes were randomly sampled with replacement 1,000 times) are indicated by circles, ≥0.9 (white fill). Full size image We identified the main germ cell types 11 along the continuous cellular proliferation and differentiation path of spermatogenesis across all species: spermatogonia (SG), the mitotic cells fuelling spermatogenesis, including spermatogonial stem cells; spermatocytes (SC), where meiosis takes place; and the haploid round spermatids (rSD) and elongated spermatids (eSD), which together reflect spermiogenesis (Fig. 1a , Extended Data Figs. 1a–f , Supplementary Table 3 and Methods ). We also identified separate clusters corresponding to somatic testicular cells, in particular Sertoli cells (except for platypus, in which Sertoli cells could not be unambiguously distinguished), the main spermatogenesis support cells, but also other cell types, such as Leydig cells, peritubular cells, endothelial cells and macrophages (Extended Data Figs. 1a–e ). The close evolutionary relationship of the seven primates in our study enabled the direct integration of datasets across these species and thus the identification of sub-cell types that correspond to intermediate events during spermatogenesis (Fig. 1a ). For all species, we traced the dynamic gene expression programs underlying spermatogenic differentiation and key molecular events, thus also identifying a host of new marker genes (Extended Data Fig. 1g and Supplementary Table 4 ). To obtain an overview of cell-type relationships across species, we performed a principal component analysis (PCA) based on pseudo-bulk cell-type transcriptomes (Fig. 1b ). The first principal component (PC1) orders the spermatogenic cell types according to the progression of spermatogenesis for all species (Fig. 1b , from left to right). This observation suggests that our data capture ancestral aspects of spermatogenic gene expression programs that are shared across mammals or amniotes, despite the rapid evolution of the testis 1 , 2 , 3 , 4 , 5 , 6 and the long divergence times of 310 million years (Fig. 1a ). PC2 separates the data by species or lineages, reflecting diverged aspects of spermatogenesis, whereas PC1 and PC2 together separate somatic and spermatogenic cell types for each species. The close clustering of biological replicates is a further indicator of the high data quality. Rates of evolution along spermatogenesis To investigate rates of gene expression evolution across cell types, we reconstructed gene expression trees ( Methods ). A tree based on pseudo-bulk transcriptomes for the whole testis (Fig. 1c ) recapitulates the known mammalian phylogeny (Fig. 1a , except for the gibbon–macaque grouping), akin to previous trees based on bulk-tissue RNA-seq data across mammalian organs 1 , 2 . This observation is consistent with the view that regulatory changes steadily accumulated over evolutionary time 1 , with present-day RNA abundances reflecting the evolution of mammalian lineages and species. To trace the cellular source of the rapid evolution of the testis, we built expression trees for the different cell types, which also recapitulate the known species relationships (Extended Data Fig. 2a ). Notably, the total branch lengths of the trees, which reflect the amount of evolutionary expression change, vary substantially between cell types (Fig. 2a ). Whereas the rate of expression evolution is similar in Sertoli cells and diploid spermatogenic cells (and lower than that in other somatic cell types), it is substantially higher in the postmeiotic haploid cell types (rSD and eSD), consistent with a recent inference based on data for three eutherians 23 . The higher resolution afforded by a primate-specific analysis provides further details (Fig. 2a ). Starting in late meiosis (pachytene SC), evolutionary rates progressively increase until the end of spermiogenesis (late eSD). Thus, late spermatogenic stages drive the previously observed rapid evolution of the testis 1 , 2 , 3 , 4 , 5 , 6 . Fig. 2: Gene expression divergence and evolutionary forces. a , Total branch lengths of expression trees among testicular cell types for amniotes and primates. Box plots show the median (central value); upper and lower quartile (box limits) and 95% confidence intervals (whiskers) for 1,000 bootstrap replicates. b , Spearman’s correlations between humans and other species from 100 bootstrap replicates (dots). Lines correspond to linear regression trends (after log transformation of the time axis. Regression R 2 values range from 0.86 to 0.97. Ma, millions of years ago. c , Mean pLI value of expressed genes in human (the pLI score reflects the tolerance of a gene to a loss-of-function mutation; lower values mean less tolerance). d , Percentage of expressed genes leading to a lethal phenotype when knocked out in mouse (out of 4,742 knockouts 25 ). e , Mean normalized ratio of nonsynonymous (d N ) over synonymous (d S ) nucleotide substitutions of expressed genes in macaque. f , Percentage of expressed genes under positive selection (out of 11,170 genes tested for positive selection) in chimpanzee. g , Mean phylogenetic age of expressed genes in mouse. h , Percentage of UMIs mapping to protein-coding genes (top) or intergenic elements (bottom) in gorilla. i , Translational efficiency values (data from ref. 6 ) are plotted for all genes with predominant expression in a given cell type in human. j , Mean of tissue-specificity values (data from ref. 2 ) in opossum. k , Percentage of expressed genes associated with infertility (out of 3,552 knockouts) in mouse. c – h , j , k , Plotted is the mean value per cell. c – k , Superimposed thick black dots indicate medians from biological replicates. Box plots depict the median (centre value); upper and lower quartile (box limits) with whiskers at 1.5 times the interquartile range. Red lines separate somatic (OS, other somatic; ST, Sertoli cells) and germ cells. Data for other studied species are shown in Extended Data Fig. 3 . Full size image Pairwise species comparisons, including downsampling analyses (Extended Data Fig. 2b ), confirm the rapid expression evolution of postmeiotic cell types across amniotes and that gene expression divergence increases with evolutionary time (Fig. 2b ), in accord with the expression phylogeny results (Extended Data Fig. 2a ). It is noteworthy, however, that expression divergence levels are roughly as similar between human and chicken as they are between human and platypus, although the bird lineage diverged 110 million years before the separation of monotremes and therians (that is, eutherians and marsupials). This observation, previously made for whole organs 1 , supports at the cellular level the notion that the conservation of core spermatogenic functions restricts transcriptome divergence. Evolutionary forces We sought to trace the evolutionary forces underlying the rapid evolution of late spermatogenesis. Two non-mutually exclusive patterns of natural selection may account for this observation. First, later stages of spermatogenesis might evolve under weaker purifying selection (that is, reduced functional constraints) and hence be less refractory to change. Second, the greater divergence in later stages might result from stronger positive selection, increasing the rate of fixation of adaptive changes. To investigate patterns of functional constraint during spermatogenesis, we assessed the tolerance to functional mutations of genes 24 used in different spermatogenic stages in humans, which showed a progressive increase of mutational tolerance starting during meiosis and culminating in early spermiogenesis (Fig. 2c and Extended Data Fig. 3a ). Consistently, on the basis of a set of neutrally ascertained mouse knockouts 25 , we found that the percentage of expressed genes associated with lethality decreases during spermatogenesis (Fig. 2d ). Also, in agreement with a progressive reduction in functional constraints towards later spermatogenic stages, we find that the normalized rate of amino acid altering substitutions in coding sequences across primates is higher in late spermatogenesis (Fig. 2e and Extended Data Fig. 3b ), although this increase might additionally reflect a higher proportion of genes under positive selection. Indeed, an examination of the temporal expression pattern of genes whose encoded protein sequences have been shaped by positive selection revealed a notable increase in percentages of positively selected genes used during spermatogenesis, with a peak in rSD (Fig. 2f and Extended Data Fig. 3c ). Because new genes also contribute to evolutionary innovations, we investigated the temporal contribution of recently emerged genes to gene expression programs in germ cells, using an index that combines the phylogenetic age of genes with their expression 2 ( Methods ). This analysis revealed that transcriptomes become younger during spermatogenesis (Fig. 2g and Extended Data Fig. 3d ), indicating that new genes have increasingly more prominent roles in later stages, particularly in rSD, consistent with previous observations 22 , 23 . Previous work based on bulk cell-type analyses in mouse 11 uncovered a transcriptionally permissive chromatin environment during spermatogenesis, in particular in rSD, which was suggested to have facilitated the emergence of new genes during evolution 3 , 9 , 11 . Consistently, we detect in all species considerably increased contributions of intergenic transcripts after meiosis and a concomitant decrease in the contributions of protein-coding genes (Fig. 2h and Extended Data Fig. 3e ). Notably, an analysis of translatome data 6 revealed a decline of translational efficiencies of transcripts during spermatogenesis (reaching a minimum in rSD) in all species (Fig. 2i and Extended Data Fig. 3f ). This decline is consistent with observations from mouse bulk data for a restricted number of cell types 6 and probably mitigates the functional consequences of the concurrent increase in transcriptional promiscuity of the genome. We next explored the reasons underlying the dynamic changes of selective forces and patterns of innovation during spermatogenesis. The breadth of expression across tissues and developmental processes (here referred to as expression pleiotropy) was proposed to represent a key determinant of the types of mutation that are permissible under selection 26 . We, therefore, assessed patterns of expression pleiotropy across spermatogenesis using spatiotemporal transcriptome data for several mammalian organs 2 , which revealed that genes used later in spermatogenesis, in particular those in rSD, have substantially more specific spatiotemporal profiles than genes used earlier in spermatogenesis and in somatic cells (Fig. 2j and Extended Data Fig. 3g,h ). Given that a decrease in expression pleiotropy can explain both a decrease in functional constraints and an increase in adaptation 2 , 26 , we suggest that it is probably a main contributor to the accelerated molecular evolution in late spermatogenesis. In addition, the specific type of selection acting on haploid cells 27 (haploid selection), in which expressed alleles are directly exposed to selection, may have contributed to the exceptionally rapid evolution of rSD. Whereas the tissue- and time-specific late spermatogenic genes, in general, are not essential for viability (Fig. 2d,j and Extended Data Fig. 3g,h , above), we proposed that the specific aforementioned evolutionary forces indicate that many of these genes evolved crucial roles in spermatogenesis. Indeed, we find that the proportion of genes associated with infertility 25 is relatively high in SC and spermatids (especially rSD): higher than in SG and somatic cells (Fig. 2k ). Gene expression conservation and innovation We next sought to trace the individual genes underlying conserved (ancestral) and diverged aspects of germ cells by comparing expression trajectories along spermatogenesis of one-to-one (1:1) orthologous genes across species 2 (Supplementary Fig. 2 and Methods ). For the primates, this analysis revealed roughly 1,700–2,900 genes with conserved expression trajectories across different lineages or species (Fig. 3a and Supplementary Table 5 ). For example, the temporal expression of 1,687 genes is conserved across the seven primates and probably reflects the core ancestral gene expression program of the simian testis. Fig. 3: Evolution of gene expression trajectories along spermatogenesis. a , b , The numbers of changed trajectories (in purple) and conserved trajectories (in olive) are indicated. In a – c , red asterisks indicate the branch for which a trajectory change has been called. For each replicate, mean expression levels across cells of a given cell type were calculated. Marks indicate values for the replicates with the highest and lowest mean expression levels, dots indicate the median of mean expression values of the replicates (three replicates for chimpanzee and opossum; two for human, bonobo, gorilla, macaque, marmoset, mouse, platypus and chicken; one for gibbon). In a – c , dashed vertical lines separate SG, SCs, rSD and eSD. a , Primate trajectories (human, bonobo, chimpanzee, gorilla, gibbon, macaque and marmoset; from top to bottom) based on 4,459 1:1 orthologues. TEX11 and PI4KB are examples of conserved and changed trajectories, respectively. b , Amniote trajectories (human, mouse, opossum, platypus and chicken; from top to bottom) based on 2,927 1:1 orthologues. DMRT1 and IP6K1 are examples of conserved and changed trajectories, respectively. c , RUBCNL expression trajectory from snRNA-seq data along spermatogenesis for human, bonobo, chimpanzee and gorilla (from top to bottom). Available orangutan samples did not allow for the generation of high-quality snRNA-seq data. d , Detection of RUBCNL (each red dot reflects a single transcript) expression in human, chimpanzee and orangutan testis by smISH using RNAScope. Left, seminiferous tubule cross sections counterstained with haematoxylin, and closeups on areas containing SG, SC and spermatids (SD). Right, quantification of RUBCNL expression levels in ten tubules per section ( n = 3 for human, n = 1 for chimpanzee and orangutan). Box plots represent the mean (central value) distribution of staining (dots); upper and lower quartile (box limits) with whiskers at 1.5 times the interquartile range. Full size image By contrast, we detected 635 trajectory changes during the evolution of apes and Old World monkeys (Fig. 3a and Supplementary Table 5 ). For example, 63 and 94 trajectory changes occurred in the human and chimpanzee–bonobo lineages, respectively, since their divergence roughly 7 million years ago. We spatially validated three of these changes using single-molecule RNA in situ hybridization (smISH) for three great apes (human, chimpanzee and orangutan as the outgroup) (Fig. 3c , Extended Data Fig. 4 , Supplementary Fig. 3 , Supplementary Table 6 and Methods ). Our smISH experiments confirm that the expression of the gene RUBCNL , encoding a regulator of autophagy 28 , changed in the human lineage, with a relative reduction of expression levels in spermatids (Fig. 3c,d ). Thus, the role of RUBCNL in autophagy during spermatogenesis 29 is potentially different in humans compared to other Old World anthropoids. We also confirmed the inversion of the expression trajectory of the myosin-encoding gene MYO3B in SG and rSD on the chimpanzee–bonobo lineage (Extended Data Fig. 4a ), as well as the expression increase of ADAMTS17 , a family-member of proteases with key spermatogenic functions 30 , in rSD relative to SG in humans (Extended Data Fig. 4b ). We note that observed quantitative and partly qualitative expression pattern differences between the two complementary data types are expected because of technical differences. We also assessed the conservation of expression trajectories across the three main mammalian lineages and amniotes (Fig. 3b and Supplementary Table 7 ). The 416 genes with conserved expression across all mammals probably trace back to ancestral gene expression programs of the ancestral mammalian testis, whereas 389 genes may represent the core ancestral spermatogenic program of amniotes. A notable example is DMRT1 , which is highly expressed specifically in SG across amniotes and is required for mouse spermatogonial stem cell maintenance and replenishment 31 . In agreement with these highly conserved sets of genes having key roles in spermatogenesis, our analyses of fertility phenotypes 25 unveiled that genes involved in fertility are significantly more conserved in their expression trajectories than genes not associated with fertility (Extended Data Fig. 5b ). Consistently, a Gene Ontology 32 enrichment analysis indicates an involvement of conserved genes in fundamental spermatogenic processes that are typical of the cell type in which they show peak expression (Extended Data Fig. 6a,b ). Thus, genes with conserved trajectories for which spermatogenesis functions remain uncharacterized represent promising candidates for the exploration of fertility phenotypes (Supplementary Tables 5 , 7 and 8 ). Notably, genes with lineage-specific trajectory changes are enriched with broader, typically metabolic, processes (Extended Data Fig. 6c,d ). They are also significantly more tissue- and time-specific than genes with conserved expression (Extended Data Fig. 6e ), which may have facilitated their expression change during evolution because of reduced pleiotropic constraints. However, genes with changed trajectories nevertheless include many genes for which key fertility functions have been described (Supplementary Tables 7 and 8 ). For example, IP6K1 , which elicits infertility when knocked out in the mouse 33 , shows strongly increasing expression towards the end of spermiogenesis in all amniotes except the primates, for which expression is high in SG and then declines (Fig. 3b and Extended Data Fig. 5c ). Thus, the primary function of IP6K1 probably shifted from late to early spermatogenesis during primate evolution. In conjunction with mouse fertility data, we used our data to explore the contribution of new genes (mostly arising from gene duplications) to the evolution of new spermatogenic functions, focusing on the rodent lineage leading to mouse. This analysis revealed the emergence of key spermatogenic genes at different time points during evolution (Extended Data Fig. 7a and Supplementary Table 8 ), such as two rodent-specific retrogenes— D1Pas1 and H2al2a —that originated through RNA-based duplication from parental genes on the X chromosome (Extended Data Fig. 7c ). D1Pas1 is essential for meiosis 34 , whereas H2al2a is essential for genome compaction in late spermatogenesis 35 . Both genes show strongly increasing expression levels in late spermatogenesis (Extended Data Fig. 7b ) and are thus in agreement with new genes contributing to functional roles predominantly during late spermatogenesis (above, Fig. 2g and Extended Data Fig. 3d ). Finally, we used our data to investigate ligand–receptor interactions underlying the communication between Sertoli cells, which have a central role in supporting and controlling spermatogenesis 11 , and germ cells across species ( Methods ). Our cross-species comparisons revealed various conserved known and new ligand–receptor interactions (Extended Data Fig. 8a,b and Supplementary Table 9 ). For example, our data provide evidence that communication of Sertoli cells with SG, SC and rSD occurs in all amniotes through interactions of the cell adhesion molecule CADM1 (ref. 36 ) or between CADM1 (Sertoli cells) and NECTIN3 (SC and rSD) (Supplementary Table 9 ). We note that CADM1 was previously thought to not to be expressed in Sertoli cells 37 , but our data show high expression in this cell type across amniotes, in agreement with human protein atlas data (proteinatlas.org/ENSG00000182985-CADM1/tissue/testis#). Our work also supports the notion that the NECTIN2–NECTIN3 complex mediates the communication of Sertoli cells with spermatids not only in mouse 38 but also in humans 39 (Supplementary Table 9 ). Sex chromosomes Sex chromosomes emerged twice in parallel in mammals from different sets of ancestral autosomes. The therian XY chromosome system originated just before the split of eutherians and marsupials (Fig. 4a ) and hence has evolved largely independently in these two lineages. Around the same time, the monotreme sex-chromosome system arose from a different pair of autosomes and subsequently expanded to five XY pairs 40 . These sex-chromosome formation events entailed substantial remodelling of gene contents and expression patterns due to structural changes and sex-related selective forces 3 . We used our data to systematically assess testicular expression patterns of sex chromosomal genes and their evolution across mammals. Fig. 4: Mammalian sex-chromosome evolution. a , Phylogenetic tree for human, mouse, opossum and platypus (from top to bottom). Arrows show sex systems origination 40 . Illustration of therian (human, top) and monotreme (platypus, bottom) sex chromosomes. Recombining (crosses) PARs in turquoise and SDRs in red. b , Percentages of X-linked genes among testis-specific genes with predominant expression in a given cell type for human, macaque, mouse, opossum and platypus (from left to right; Supplementary Table 10 ). The red dashed line represents the expected percentages of X-linked genes, if testis-specific genes with predominant expression in the different cell types were randomly distributed across the genome. Asterisks indicate significance after two-sided exact binomial testing (Benjamini–Hochberg corrected P values from left to right: 0.000353, 1.18 × 10 −7 , 0.000687, 0.0142, 1.08 × 10 −7 , 3.45 × 10 −8 , 0.000792, 0.0355 and 0.00043). The number of testis-specific X-linked genes enriched in each cell type is indicated above each bar. c , UMAP representation of germ cells (left). Progression of spermatogenesis (grey arrow) and meiotic divisions (black arrows) are indicated. Spermatids identified as X- and Y-bearing are coloured in orange and purple, respectively. Box plots show the median (central value) percentages of X and Y transcripts in X- and Y-bearing spermatids; upper and lower quartile (box limits) with whiskers at 1.5 times the interquartile range. Two-sided Wilcoxon rank-sum tests were performed for statistical comparisons (Benjamini–Hochberg corrected **** P < 2.2 × 10 −16 ). d , X-to-autosome transcript ratios in individual germ cells across spermatogenesis (from left to right). For spermatids, only X-bearing cells are considered. Lines depict generative additive model trend. Red arrows indicate MSCI. c , d , For platypus, X transcripts are dissected according to their location on the X chromosomes ( Methods ). Full size image We first contrasted cell-type specificities of X-linked and autosomal genes, which showed a notable excess (12–60%) of X-linked genes with predominant expression in SG across all eutherians, in agreement with a previous mouse study 17 and includes conserved genes with key spermatogenic functions such as TEX11 (Figs. 3a and 4b , Extended Data Fig. 9a and Supplementary Table 10 ). The higher-resolution primate data showed that the X chromosome is also enriched for genes expressed in leptotene SC (Fig. 4b ), presumably reflecting transcript carry-over from SG, given the global transcriptional silencing of the genome during the leptotene stage 41 (Supplementary Fig. 1b and Supplementary Table 3 ) and the observation that most X-linked genes expressed in leptotene SC are also expressed in differentiated SG (Supplementary Fig. 4 ). Notably, we detected enrichments of genes with SG-specific expression also on the opossum and platypus X chromosomes, thus revealing this to be a shared pattern across mammals (Fig. 4b and Extended Data Fig. 9a,b ). We also uncovered an enrichment of genes with predominant expression in Sertoli cells across X chromosomes (except for platypus, for which Sertoli cells could not be unambiguously distinguished) (Fig. 4b ). Altogether, our observations suggest that sex-related forces favoured the independent accumulation of SG- and Sertoli-specific genes on the X during the evolution of the different sex-chromosome systems. Consistently, autosomes in outgroup species corresponding to the different mammalian X chromosomes (for example, platypus chromosome 6, which is homologous to the therian X) do not show any excess of SG- or Sertoli-expressed genes (Extended Data Fig. 9a,c ), which means that the ancestral autosomes that gave rise to present-day sex chromosomes were not enriched for such genes. Overall, we propose that the accumulation of SG- and Sertoli-specific genes was facilitated by the specific selective environment on the X chromosome, in which male-beneficial mutations are always visible to selection because of the single-copy (hemizygous) status of the X chromosome in males 3 , 17 , 21 . We next sought to separate X- and Y-bearing spermatids to investigate their distinct transcriptomal properties during spermiogenesis. Such an analysis is probably not possible using single-whole-cell transcriptomic data because X and Y spermatids remain connected by cytoplasmic bridges and hence are thought to contain similar cytoplasmic transcript pools 42 . However, our single-nucleus data should afford the separation of X and Y spermatids. Indeed, on the basis of differential X and Y transcript contents ( Methods ), we were able to separate spermatids into distinct X and Y lineages across mammals (Fig. 4c and Supplementary Table 3 ). As expected, our approach failed to separate X and Y spermatids in available human 23 and mouse 43 scRNA-seq datasets (Extended Data Fig. 10 ), supporting the notion of substantial transcript exchange across X- and Y-spermatid cells through cytoplasmic bridges 42 , although this equilibration may not be complete 44 . A differential expression analysis between X and Y spermatids identified, as expected, most sex-chromosome genes, including gametologues (that is, genes with homologous counterparts on X and Y chromosomes), such as the translational regulatory genes DDX3X / DDX3Y (Extended Data Fig. 11 and Supplementary Table 11 ). However, we also found autosomal genes, especially in some species (for example, human), which might reflect trans-regulatory effects associated with the X and Y chromosomes (Extended Data Fig. 11 , Supplementary Table 11 and Methods ). We then assessed gene expression across spermatogenesis separately for X- and Y-linked genes. We traced a substantial dip in X transcript abundances around the pachytene stage of meiosis across therians (Fig. 4d ), which reflects the process of MSCI 13 , a sex-chromosome-specific instance of the general epigenetic phenomenon of meiotic silencing of unsynapsed chromatin 13 (MSUC). An analysis of Y transcripts provides consistent results but at lower resolution owing to the very small number of Y-linked genes (Extended Data Fig. 12a ). Previous work did not find evidence for MSCI in monotremes 45 , suggesting that MSCI originated in the therian ancestor after the separation from the monotreme lineage 45 . We revisited this question in platypus using a new assembly and annotation 46 that includes a detailed definition of sexually differentiated regions (SDRs) versus pseudoautosomal regions (PARs), which are large in monotremes (Fig. 4a ). We proposed that expression signals from the large PARs, which are expected to synapse and hence not to be affected by MSUC, might have prevented the detection of MSCI in previous studies. Indeed, whereas the joint analysis of all platypus X-linked genes only shows a small expression dip around the pachytene stage (Fig. 4d , upper platypus graph), an analysis only of SDR genes reveals a strong reduction of X transcript levels. By contrast, PAR genes show stable expression levels across spermatogenesis (Fig. 4d , lower graph). Moreover, the difference in transcript abundances between SDRs and PARs owing to MSCI is visible for all five platypus X chromosomes (Extended Data Fig. 12b ). Notably, our assessment of the completeness of MSCI across species reveals that platypus MSCI is as complete as that in other species for which there is little or no MSCI escape 23 , 47 , 48 (Extended Data Fig. 12c ). The presence of MSCI at the SDRs in monotremes is consistent with the partial association of platypus sex chromosomes with perinucleolar repressive histone modifications at the pachytene stage 45 . Altogether, our data reveal that efficient MSCI is common to all mammalian sex-chromosome systems, which indicates that the general mechanism of MSUC is an ancestral mammalian feature. Discussion Our analyses uncovered that the previously observed rapid evolution of the testis 1 , 2 , 3 , 4 , 5 , 6 , 7 is driven by an accelerated rate of molecular innovation in late spermatogenesis, in particular in rSD. Our findings suggest a scenario in which the accelerated fixation of regulatory changes, amino acid altering substitutions and new genes during evolution in late spermatogenesis (presumably driven by sperm competition) was facilitated by reduced pleiotropic constraints, a transcriptionally permissive chromatin environment and potentially haploid selection for spermatid genes that are not rendered effectively diploid through transcript exchange through cytoplasmic bridges. In agreement with late spermatogenic stages explaining the rapid evolution of the testis, early spermatogenic cell types and somatic cells show patterns of constraint and innovation similar to those of cell types in the brain (Extended Data Fig. 3a,b,d ), an organ that evolves slowly at the molecular level 1 , 2 , 3 . We note that differences in cell-type abundances across mammals, which are pronounced for the testis 7 (Extended Data Fig. 1f ), presumably also contributed to the rapid gene expression divergence of this organ 6 . Our cross-species comparisons of individual genes revealed temporal expression differences across species, which were probably facilitated by reduced pleiotropic constraints. Our results thus provide an extensive list of candidates whose contributions to the evolution of species-specific spermatogenesis phenotypes can be experimentally scrutinized. We also uncovered conserved expression programs underlying spermatogenic processes ancestral to individual mammalian lineages and mammals as a whole. Further analyses illuminated the role of sex chromosomal genes in spermatogenesis. We found that genes predominantly expressed in SG and Sertoli cells independently accumulated on X chromosomes across mammals and their two sex-chromosome systems during evolution. This suggests that X chromosomes have been shaped by strong male-related selective forces 21 , leading to the emergence of X-linked genes with functional roles in testis cell types in which active transcription is possible. Indeed, in addition to the SG and Sertoli-cell enrichment found here, previous work showed that these selective pressures led to the repeated duplication of genes on the X that facilitated expression after meiosis 15 . Our ability to separate X- and Y-bearing spermatids and the availability of a new platypus genome assembly 46 unveiled that MSCI is a general feature of mammalian sex-chromosome systems, indicating that MSUC was already present in the common mammalian ancestor. Previous work did not find evidence for MSCI in birds 49 , which raises the question whether MSUC arose in the common mammalian ancestor after the split from the reptile lineage roughly 310 million years ago or whether this mechanism was lost in the avian lineage and arose earlier in evolution. The latter scenario would be consistent with the observation of MSCI in invertebrate species 13 . Our data and results, together with the accompanying online resource we developed ( ) to facilitate the exploration of our data, provide an extensive resource for investigating the biology of the testis and associated fertility disorders across mammals. Future studies should seek to complement our snRNA-seq data to overcome its limitations. scRNA-seq data will be valuable for inferring transcriptome patterns unique to the cytoplasm, single-cell full-length transcript data are needed to assess the pronounced isoform diversity of the testis 11 , and single-cell translatome data 50 are required to understand the contribution of post-transcriptional changes 6 to the evolution of spermatogenesis. Methods Data reporting No statistical methods were used to predetermine sample size. The experiments were not randomized and investigators were not blinded to allocation during experiments and outcome assessment. Biological samples and ethics statement We generated snRNA-seq data for adult testis samples from human ( Homo sapiens ), chimpanzee ( Pan troglodytes ), bonobo ( Pan paniscus ), gorilla ( Gorilla gorilla ), lar gibbon ( Hylobates lar ), rhesus macaque ( Macaca mulatta ), common marmoset ( Callithrix jacchus ), mouse ( Mus musculus , strain: RjOrl:SWISS; Janvier Laboratories), grey short-tailed opossum ( Monodelphis domestica ), platypus ( Ornithorhynchus anatinus ) and chicken (red jungle fowl, Gallus gallus ) (Supplementary Table 2 ). In addition, we produced bulk RNA-seq data for chimpanzee, gorilla, gibbon and marmoset from the same individuals. Adult human testis samples used for in situ hybridization experiments were obtained from orchiectomy specimens from three individuals with testicular cancer. Tissue adjacent to the tumour that was devoid of cancer cells and germ cell neoplasia in situ, and tubules with normal spermatogenesis were used. Other adult primate testis tissue was obtained from a western chimpanzee and a Bornean orangutan ( Pongo pygmaeus ). Our study complies with all relevant ethical regulations with respect to both human and other species’ samples. Human samples underlying the snRNA-seq data were obtained from scientific tissue banks ( ) or dedicated companies ( ); informed consent was obtained by these sources from donors before death or from next of kin. The samples used for the RNA in situ hybridization experiments were obtained from the tissue biobank at the Department of Growth and Reproduction (Rigshospitalet, Copenhagen, Denmark) containing orchiectomy specimens from individuals with testicular cancer (The Danish Data Protection Agency, permit number J. no. 2001-54-0906). All patients have given informed consent for donating the residual tissues for research. The use of all human samples for the type of work described in this study was approved by an ethics screening panel from the European Research Council (ERC) and local ethics committees: from the Cantonal Ethics Commission in Lausanne (authorization 504/12); from the Ethics Commission of the Medical Faculty of Heidelberg University (authorization S-220/2017) and from the regional medical research ethics committee of the capital region of Copenhagen (H-16019637). All primates used in this study suffered sudden deaths for reasons other than their participation in this study and without any relation to the organ sampled. The use of all other mammalian samples for the type of work described in this study was approved by ERC ethics screening panels. Nuclei isolation For the samples from therian species we developed a nuclei preparation method that includes fixation with dithio- bis (succinimidyl propionate) (DSP; or Lomant’s Reagent), a reversible cross-linker that stabilizes the isolated nuclei. The method was adapted from protocols used for the fixation of single-cell suspensions 51 and for the isolation of single nuclei from archived frozen brain samples 52 . Tissue pieces weighing roughly 5 mg were homogenized in 100–150 µl mg −1 of prechilled lysis buffer (250 mM sucrose, 25 mM KCl, 5 mM MgCl 2 , 10 mM HEPES pH 8, 1% BSA, 0.1% IGEPAL and freshly added 1 µM DTT, 0.4 U µl −1 RNase Inhibitor (New England BioLabs), 0.2 U µl −1 SUPERasIn (ThermoFischer Scientific)) and lysed for 5 min on ice. The lysate was centrifuged at 100 g for 1 min at 4 °C. The supernatant was transferred to a new reaction tube and centrifuged at 500 g for 5 min at 4 °C. The supernatant was removed and the pellet resuspended in 0.67 vol. (of volume lysis buffer used) of freshly made fixation solution (1 mg ml −1 DSP in PBS) and incubated for 30 min at room temperature. The fixation was quenched by addition of Tris-HCl to a final concentration of 20 mM. The fixed nuclei were pelleted at 500 g for 5 min at 4 °C. The supernatant was removed and the pellet resuspended in 0.67 vol. of Wash Buffer (250 mM sucrose, 25 mM KCl, 5 mM MgCl 2 , 10 mM Tris-HCl pH 8, 1% BSA and freshly added 1 µM DTT, 0.4 U µl −1 RNase Inhibitor, 0.2 U µl −1 SUPERasIn). This was centrifuged at 500 g for 5 min at 4 °C. The supernatant was removed and the pellet resuspended in 0.5 vol. of PBS. Then nuclei were strained using 40 μm Flowmi strainers (Sigma). For estimation of nuclei concentration, Hoechst DNA dye was added and the nuclei were counted using Countess II FL Automated Cell Counter (ThermoFischer Scientific). For platypus and chicken, a similar preparation method was used, but the nuclei were not fixed, given that this protocol gave optimal results for these species (the fixation protocol failed to yield data of adequate quality). In brief, tissue pieces weighing roughly 5 mg were homogenized in 100–150 µl mg −1 of prechilled lysis buffer (250 mM sucrose, 25 mM KCl, 5 mM MgCl 2 , 10 mM Tris-HCl pH 8, 1% BSA, 0.1% IGEPAL and freshly added 1 µM DTT, 0.4 U µl −1 RNase Inhibitor, 0.2 U µl −1 SUPERasin) and lysed for 5 min on ice. The lysate was centrifuged at 100 g for 1 min at 4 °C. The supernatant was transferred to a new reaction tube and centrifuged at 500 g for 5 min at 4 °C. The supernatant was removed and the pellet resuspended in 0.67 vol. of Wash Buffer. This was centrifuged at 500 g for 5 min at 4 °C. The supernatant was removed and the pellet resuspended in 0.5 vol. of PBS. Then nuclei were strained using 40-μm Flowmi strainers (Sigma). For estimation of nuclei concentration, Hoechst DNA dye was added and the nuclei were counted using Countess II FL Automated Cell Counter (ThermoFischer Scientific). Preparation of sequencing libraries For the construction of snRNA-seq libraries, Chromium Single Cell 3′ Reagent Kits (10X Genomics; v.2 chemistry for human, chimpanzee, bonobo, gorilla, gibbon, macaque, marmoset, mouse and opossum; v.3 chemistry for platypus and chicken) were used according to the manufacturer’s instructions. Then 15,000 to 20,000 nuclei were loaded per lane in the Chromium microfluidic chips and complementary DNA was amplified in 12 PCR cycles. Sequencing was performed with NextSeq550 (Illumina) according to the manufacturer’s instructions using the NextSeq 500/550 High Output Kit v.2.5 (75 cycles) with paired-end sequencing (read lengths of read1 26 bp, read2 57 bp; index1 8 bp, roughly 170 to 380 million reads per library for v.2 chemistry; read lengths of read1 28 bp, read2 56 bp Index1 8 bp, roughly 247 to 306 million reads per library for v.3 chemistry) (Supplementary Table 2 ). For bulk RNA-seq data generation, RNA was extracted using the RNeasy Micro kit (QIAGEN). The tissues were homogenized in RLT buffer supplemented with 40 mM DTT. The RNA-seq libraries were constructed using the TruSeq Stranded messenger RNA LT Sample Prep Kit (Illumina) as described in ref. 2 . Libraries were sequenced on Illumina NextSeq550 using a single-end run (read1 159 bp; index1 7 bp) with roughly 24–60 million reads per library (Supplementary Table 2 ). Genome and transcript isoform annotation Given that the quality of genome annotation differs substantially between the studied species and given the specific and widespread transcription of the genome in the testis 11 , we refined and extended previous annotations from Ensembl 53 on the basis of testis RNA-seq data. Specifically, akin to the procedure described in refs. 2 , 6 , we used previous extensive stranded poly(A)-selected RNA-seq datasets 2 , 54 (100 nt, single-end) for human, macaque, mouse, opossum, platypus and chicken, and generated and used stranded poly(A)-selected RNA-seq datasets (159 nt, single-end) for chimpanzee, gorilla, gibbon and marmoset. For each species, we downloaded the reference genome from Ensembl release 87 (ref. 53 ): hg38 (human), CHIMP2.1.4 (chimpanzee), rheMac8 (rhesus macaque), C_jacchus3.2.1 (marmoset), mm10 (mouse), monDom5 (opossum), and galGal5 (chicken); from Ensembl release 96 (ref. 55 ): gorGor4 (gorilla) and Nleu_3.0 (gibbon); and from Ensembl release 100 (ref. 56 ): mOrnAna1.p.v1 (platypus). Raw reads were first trimmed with cutadapt v.1.8.3 (ref. 57 ) to remove adapter sequences and low-quality (Phred score <20) nucleotides, then reads shorter than 50 nt were filtered out (parameters: --adapter=AGATCGGAAGAGCACACGTCTGAACTCCAGTCAC --match-read-wildcards --minimum-length=50 -q 20). Processed reads were then mapped to the reference transcriptome and genome using Tophat2 v.2.1.1 (ref. 58 ) (parameters: --bowtie1 --read-mismatches 6 --read-gap-length 6 --read-edit-dist 6 --read-realign-edit-dist 0 --segment-length 50 --min-intron-length 50 --library-type fr-firststrand --max-insertion-length 6 --max-deletion-length 6). Next, we assembled models of transcripts expressed using StringTie v.1.3.3 (ref. 59 ) (parameters: -f 0.1 -m 200 -a 10 -j 3 -c 0.1 -v -g 10 -M 0.5). Stringent requirements on the number of reads supporting a junction (-j 3), minimum gap between alignments to be considered as a new transcript (-g 10) and fraction covered by multi-hit reads (-M 0.5) were used to avoid merging independent transcripts and to reduce the noise caused by unspliced or incompletely spliced transcripts. We compared the assembled transcript models to the corresponding reference Ensembl annotations using the cuffcompare program v.2.2.1 from the cufflinks package 60 . Finally, we combined the newly identified transcripts with the respective Ensembl gene annotation into a single gtf file. We extended the original Ensembl annotations by 2–61 Mbp with new transcripts and by 23–49 Mbp with new splice isoforms (Supplementary Table 1 ). Raw reads processing CellRanger v.3.0.2 was used for platypus and chicken, and CellRanger v.2.1.1 for the other species in line with the used Chromium chemistry. The CellRanger mkref function was used with default settings to build each species reference from genomic sequences and customized extended annotation files (Supplementary Table 1 ). Given that pre-mRNA transcripts are abundant in nuclei 61 , exons and introns features were concatenated as described in the CellRanger v.2.1.1 documentation for considering intronic and exonic reads for gene expression quantification. The CellRanger count function was used with default settings to correct droplet barcodes for sequencing errors, align reads to the genome and count the number of UMIs for every gene and barcode combination. Identification of usable nuclei We used a combined approach for detection of usable nuclei. This was done to optimally account for the lower RNA content of nuclei compared to the whole cells. Specifically, to identify usable nuclei, we used a knee point-based approach combined with the fraction of intronic reads as a marker of pre-mRNA transcripts (abundant in the nucleus) (Supplementary Fig. 1f ) and MALAT1 (nuclear-enriched long non-coding (lnc)RNA) expression as a marker of nuclei (when present in the genome of a given species). Quality control of filtered cells For each sample independently, high-quality nuclei were selected removing outliers on the basis of the number of UMIs and the percentage of mitochondrial RNA (Supplementary Fig. 1b,c,e ). We created a Seurat 62 object using the Seurat R package v.3.1.4 from the subset raw UMI count table generated by CellRanger corresponding to the usable droplets identified upstream, normalized the data using the NormalizeData function, identified the top 10,000 most variables genes using the FindVariableFeatures function, scaled the data using the ScaleData function, performed the PCA using the RunPCA function and calculated the Louvain clusters using the FindNeighbors (parameters dims = 1:20) and FindClusters (dims = 1:20, resolution = 0.5) functions. To optimally account for the fact that testis cell types have diverse transcriptome characteristics 11 , we filtered out outlier droplets for each cluster independently with values lower than the first quartile (Q1) − 1.5 × IQR (interquartile range) and higher than the third quartile (Q3) + 1.5 × IQR for both the UMI content and the fraction of mitochondrial RNA. Then, we removed potential doublets using doubletFinder_v3 function of DoubletFinder 63 v.2.0.1 (parameters PCs = 1:20, pN = 0.25, nExp = 5% of the total number of cells, identifying pk using paramSweep_v3, summarizeSweep and find.pK functions). Integration of datasets From the previously filtered UMI count tables, we created Seurat objects for every sample independently, normalized the data and identified the top 10,000 most variable genes. Next, for each species independently, we applied the Seurat 62 anchoring approach using FindIntegrationAnchors and IntegrateData functions with 20 principal components to integrate all datasets together into a single Seurat object correcting for the batch effect. For each integrated species-specific Seurat object, we normalized (NormalizeData function) and scaled the data (ScaleData function) and performed a PCA (RunPCA function). Louvain clusters were calculated using FindNeighbors and FindClusters functions (parameters dims = 1:20, 1:20, 1:20, 1:20, 1:20, 1:20, 1:20, 1:17, 1:8, 1:10 and 1:10, and resolution = 0.5, 0.5, 0.2, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.3 and 0.5, respectively, for human, chimpanzee, bonobo, gorilla, gibbon, macaque, marmoset, mouse, opossum, platypus and chicken). The uniform manifold approximation and projection (UMAP) embedding coordinates were calculated using the RunUMAP function (parameters dims = 1:20, 1:20, 1:20, 1:20, 1:20, 1:20, 1:20, 1:17, 1:10, 1:10 and 1:10, and min_dist = 0.3, 0.3, 0.1, 0.1, 0.3, 0.3, 0.3, 0.1, 0.2, 0.3 and 0.6, respectively, for human, chimpanzee, bonobo, gorilla, gibbon, macaque, marmoset, mouse, opossum, platypus and chicken). We note that—consistent with the high correlation between biological replicates (Supplementary Fig. 5a,b )—the data already integrate well before the batch correction (Supplementary Fig. 18c,d ). We also note that key marker genes are expressed in the same integrated areas across replicates when assessing their expression in the different replicates using the integrated object coordinates (Supplementary Fig. 18e ), which supports that the integration is correct. All primate datasets were merged using the LIGER 64 (v.0.5.0) integration tool. A LIGER object was created using the createLiger function based on primate 1:1 orthologues from Ensembl release 87, and normalized with normalize, selectGenes and scaleNotCenter functions with default settings. Then, the joint matrix was factorized using the optimizeALS function ( k = 20) and the quantile normalization was performed with the quantile_norm and default settings. The Louvain clusters were calculated with the louvainCluster function and default settings as well as UMAP coordinates with the runUMAP function (n_neighbors = 100, min_dist = 0.2). Estimation of expression levels and normalization The gene UMI counts per cell were normalized using the Seurat R package and its NormalizeData function. Therefore, the UMI counts of each gene in each cell are divided by the total UMI counts of each cell, multiplied by 10,000 and log transformed. Cell-type assignment We identified the main cell-type populations from the primate integrated, mouse, opossum, platypus and chicken objects independently using known marker genes 65 , 66 mostly from human and mouse and their respective 1:1 orthologues in the other species. CLU marks Sertoli cells; TAGLN and ACTA2 peritubular and smooth muscle cells; CD34 and TM4SF1 endothelial cells; APOE and CD74 macrophages; STAR and CYP11A1 Leydig cells; GFRA1, PIWIL4 (undifferentiated), DMRT1 (differentiated) and STRA8 SG; SYCE1 (leptotene), SYCP1 (zygotene), PIWIL1 (pachytene) , SYCP2, TANK and AURKA SC; LRRIQ1 (early), ACRV1 and SPACA1 (late) rSD; and SPATA3 , NRBP1, PRM1 and GABBR2 eSD. Cell-type assignment was robustly reinforced by complementary analyses and metrics such as UMAP coordinates, pseudotime trajectories, transcriptional activities (UMI counts) and previous knowledge. Pseudotime Pseudotime trajectories were calculated using the slingshot v.1.2.0 R package 67 . We applied the getLineages function with the upstream calculated clusters and UMAP embedding coordinates of the germ cells to obtain connections between adjacent clusters using a minimum spanning tree. We provided the starting and ending clusters on the basis of the previous cell-type assignment with known marker gene expression. Then we applied the getCurves function to the obtained lineages to construct smooth curves and order the cells along a pseudotime trajectory. Pseudotime values are highly variable depending on used tools, thus we ordered the cells one by one on the basis of their pseudotime values and divided their rank by the total number of cells, to obtain evenly distributed values between 0 and 1. Finally, we validated the obtained pseudotime trajectories on the basis of previous cell-type assignments, expression patterns of marker genes and UMAP embedding coordinates. Marker gene identification To precisely identify marker genes along spermatogenesis, we grouped the germ cells into 20 evenly distributed bins along the pseudotime trajectory for each species. Then, we applied the FindAllMarkers function from the Seurat 62 R package (parameter only.pos = T, min.pct = 0.25, logfc.threshold = 0.25, return.thresh = 0.05) of the Seurat v.3.1.4 R package to the 22 groups (20 germline groups, the Sertoli and other somatic cell groups) (Extended Data Fig. 1 and Supplementary Table 4 ). Phylogenetic trees Phylogenetic trees and indicated divergence times (Figs. 1a and 4a ) are based on TimeTree 68 (v.5) ( ). Orthologous gene sets We used four different sets of orthologous genes in our study: (1) comparative analyses involving all 11 amniote species were performed using 4,498 1:1 orthologue genes that are expressed (that is, one UMI in at least three cells of any cell type) across all species (among a total of 8,045 1:1 orthologues). (2) Comparative analyses involving the seven primate species were performed using 8,451 1:1 orthologue genes expressed across all primate species (among 11,948 1:1 orthologues). (3) The comparative Sertoli-germ cell communication analysis was based on mapping 35,186 human testis-expressed genes to 1:1 orthologous genes in the other species (macaque 13,090; mouse 14,302; opossum 10,865 and chicken 10,515). (4) Species-specific analyses were performed using all genes expressed in a given species (roughly 15,000 genes per cell type; Supplementary Fig. 1d ). Orthologous gene sets were extracted from Ensembl 53 using the biomaRt R package v.2.40.5. Global patterns of gene expression differences across mammals Pseudo-bulk samples were generated using the AverageExpression function of the Seurat R package with various groups of cells depending on the pseudo-bulk samples produced in the study. For the analyses presented in Fig. 1b , we performed the PCA of normalized expression in amniote testicular cell types (pseudo-bulks) for each individual based on 4,498 1:1 amniote orthologues. PCA was performed using the prcomp function of the stats R package. For Fig. 1c , we constructed gene expression trees (as described in ref. 1 ) using the neighbour-joining approach, on the basis of pairwise expression distance matrices between corresponding pseudo-bulk samples for the different cell types across species. The distance between samples was computed as 1 − ρ , where ρ is Spearman’s correlation coefficient and was computed using the cor function of the stats R package. The neighbour-joining trees were constructed using the ape R package v.5.3. The reliability of branching patterns was assessed with bootstrap analyses (the 4,498 1:1 amniote orthologues were randomly sampled with replacement 1,000 times). The bootstrap values are the proportions of replicate trees that share the branching pattern of the majority-rule consensus tree shown in the figures (Fig. 1c and Extended Data Fig. 2a ). The total tree length was calculated by removing the intra-species variability between individuals (Fig. 2a ). Evolutionary forces In Fig. 2c , we plotted the median pLI score 69 across expressed genes ( \(\ge \) 1 UMI) in each nucleus. We obtained the pLI scores from ref. 70 . For Fig. 2d , we used a set of neutrally ascertained knockouts consisting in 4,742 protein-coding genes, 1,139 of which are classified as lethal. For each cell, the denominator is the number of genes expressed that were tested for lethality and the numerator the genes among those that resulted in a lethal phenotype. Tested genes for viability and associated phenotype information were downloaded from the International Mouse Phenotyping Consortium 25 . For Fig. 2e (and Extended Data Fig. 3a ), we used the average d N /d S values across 1:1 orthologues in primates. For each cell, the mean d N /d S value is plotted. Conserved 1:1 orthologues across six primates (human, chimpanzee, gorilla, gibbon, macaque and marmoset) as well as their coding and protein sequences were extracted from Ensembl 53 , providing a set of 11,791 protein-coding genes. For each species and orthologue the longest transcript was extracted. Orthologous protein sequences were aligned using clustalo v.1.2.4; then pal2nal v14 was used (with protein sequences alignments and coding sequences as input) to produce codon-based alignments. The codeml software from the PAML package 71 v.4.9 was used to estimate d N /d S ratios. The M0 site model was applied to the orthologue alignments to estimate one average dN/dS ratio per orthologous gene set across species (parameter NSites = 0, model = 0). In Fig. 2f (and Extended Data Fig. 3b ), we plotted the percentage of positively selected genes expressed across nuclei. For each nucleus, the denominator is the number of expressed genes that were tested for signatures of positive selection, and the numerator is the number of genes among those with evidence for positive selection. We used sets of genes previously identified as carrying evidence for coding-sequence adaptation in primates 72 (331 positively selected genes out of 11,170 genes tested) and mammals 73 (544 positively selected genes out of 16,419 genes tested). In Fig. 2g (and Extended Data Fig. 3c ), we plotted the average phylogenetic age of expressed genes across somatic and germ cells. The phylogenetic age of genes is an index that gives greater weight to young new genes (as described in ref. 2 , 74 ). The range of the score differs between species depending on the number of outgroup lineages available (more lineages allowed for more details in the phylogeny) and therefore this index cannot be compared across species, only within species (that is, across cells and cell types). The phylogenetic age of genes was obtained from GenTree ( ) with Ensembl release 69 (ref. 74 ). In Fig. 2h (and Extended Data Fig. 3d ), we plotted the percentage of the cell transcripts originating from protein-coding genes and intergenic elements. Gene biotypes were obtained from Ensembl. Intergenic elements are all elements that are not protein-coding genes (lncRNAs, pseudogenes, pseudogenes and other sequences). In Fig. 2i (and Extended Data Fig. 3e ), normalized log 2 -transformed median expression values across replicates at the transcriptome (e tr ) and translatome (e tl ) layers were used to calculate translation efficiency (TE = log 2 (e tr ) − log 2 (e tl )) in testis (as described in ref. 6 ) from RNA-seq and Ribo-seq data 6 , respectively. Translation efficiency values were calculated at the whole testis level, thus only cell-type-specific genes (for which 60% of their transcripts at the whole testis level are concentrated in a single cell type) were used. Higher values show a more efficient translation of transcripts, whereas lower values indicate relative translational repression. For Fig. 2j (and Extended Data Fig. 3f,g ), we used time- and tissue-specificity indexes of expressed genes across somatic and germ cells in testis. As described in ref. 2 , tissue and time specificity indexes are calculated from RNA-seq data across organs and developmental stages. Both indexes range from 0 (broad expression) to 1 (restricted expression). The indexes were obtained from ref. 2 . For each nucleus, we plotted the median index across expressed genes. In Fig. 2k , we plotted the percentage of genes causing infertility when knocked out (out of 3,252 knockouts, 173 of which caused infertility). Tested genes for infertility and associated phenotype information were downloaded from the International Mouse Phenotyping Consortium database 25 . For each nucleus, the denominator corresponds to the number of genes expressed that were tested for infertility and the numerator to the genes among those that resulted in an infertility phenotype. Gene expression trajectories along spermatogenesis We compared gene expression trajectories along spermatogenesis across primates using human, chimpanzee, bonobo, gorilla, gibbon, macaque and marmoset (based on Ensembl 87 orthologues), and across amniotes using human (as a representative of primates), mouse, opossum, platypus and chicken (based on Ensembl 100 orthologues). To compare robustly expressed genes, we used genes that are expressed in at least 5% of the cells in at least one cluster in all considered species. We used the mfuzz package 75 (v.2.44.0), an unsupervised soft clustering method, to cluster gene expression trajectories along spermatogenesis (eight cell types in primates; four cell types in amniotes) across species using Dmin and mestimate functions to estimate the number of clusters and the fuzzification parameter (Supplementary Fig. 2 ). As described in ref. 2 , we inferred within a phylogenetic framework the probability that there were changes in trajectories along spermatogenesis, that is, that genes changed their cluster assignment in specific branches, using a 5% probability cut-off. Trajectory conservation score We calculated a global trajectory conservation score across species for each 1:1 orthologous gene set. For a given orthologous gene set, this score corresponds to the log-transformed probability that all members fall into the same mfuzz trajectory cluster as: $${\rm{Conservation}}\_{{\rm{score}}}_{g}={\log }_{2}({\sum }_{i\in c}{\prod }_{j\in s}{P}_{g,i,j})$$ where g corresponds to a given orthologous gene set, c to all mfuzz trajectory clusters (1–9 for primates, 1–12 for amniotes), s to all species (human, chimpanzee, bonobo, gorilla, gibbon, macaque and marmoset for primates; human, mouse, opossum, platypus and chicken for amniotes) and P g,i,j to the probability that the gene g of the species j falls into the cluster i . A higher conservation score indicates a greater global trajectory conservation. As a proof of concept, we plotted the trajectory conservation score for conserved and changed trajectories that revealed a significant higher conservation score for conserved trajectories (Extended Data Fig. 5a ). RNA in situ hybridization and expression quantification Fresh testicular tissue was fixed in GR-fixative (7.4% formaldehyde, 4% acetic acid, 2% methanol, 0.57% sodium phosphate, dibasic and 0.11% potassium phosphate, monobasic) overnight (for at least 16 h) at 4 °C, dehydrated and embedded in paraffin. The in situ hybridization experiments were carried out on 4 μm sections mounted on SuperFrost Plus Slides (ThermoFisher Scientific) using the RNAScope 2.5 HD Detection Reagent RED kit according to the manufacturer’s recommendations (Advanced Cell Diagnostics). Briefly, testicular tissue sections were dewaxed in xylene and washed in 100% ethanol followed by treatment with hydrogen peroxide for 10 min. Target retrieval was performed for 15 or 30 min (see Supplementary Table 6 for specifications for each probe and species) using a steamer, followed by treatment with protease plus for 30 min at 40 °C. The slides were hybridized with the target probe (Supplementary Table 6 ) for 2 h at 40 °C followed by a series of signal amplifications (with amplification around 5 for 30 or 60 min). The sections were counterstained with Mayer’s haematoxylin and mounted with Vectamount Permanent Mounting Medium (Vector Laboratories). The negative control probe DapB (a bacterial RNA) was run in parallel with the target probes and showed ≤5% positive cells in each section. For ADAMTS17 and MYO3B , positive (that is, with red dots) rSD and SG were counted. For each section (human n = 3, chimpanzee n = 1, orangutan n = 1), ten tubules were counted using the NDP.viewPlus software (Hamamatsu Photonics). Two independent observers (S.B.W. and K.A.) counted positive and negative rSD and SG. No discrimination in intensity of the dots or the number of dots per cell was performed. Cell-type identification was performed on the basis of nucleus morphology and localization in the tubule. Only SG lining the edge of the tubules were counted (Supplementary Fig. 3a ). The inter-observer variance was found to be 7 and 9% for rSD and SG, respectively. For RUBCNL , quantification of staining intensity was performed with R v.3.6.1 using countcolors 76 (v.0.9.1) and colordistance 77 (v.1.1.1) packages. For each section (human n = 3, chimpanzee n = 1, orangutan n = 1), ten tubules were divided into three parts: area dominated by SG, area dominated by SCs (also containing Sertoli cells) and area dominated by spermatids (no distinction between the different types of spermatid) (Supplementary Fig. 3b ). In each tubular area, the number of cells was counted manually using the NDP.view2Plus software (v.2.8.24). Then, the pixels occupied by red staining were quantified and the expression level for each cell type was calculated by dividing the stained pixels by the number of cells. For each picture, the stained pixels for each cell type were normalized by the total amount of stained pixels. Gene Ontology analysis Enriched terms in the Gene Ontology 32 analyses of genes with conserved and diverged expression trajectories were identified using the goana function of the limma R package, v.3.40.6 (default parameters). Sertoli-germ cell communication analysis We identified ligand–receptor interactions underlying Sertoli-germ cell communications for human, macaque, mouse, opossum and chicken, respectively, using the CellPhoneDB 78 (v.2) approach and recommended parameter settings (parameters method statistical_analysis). To apply CellPhoneDB, which uses a database of human ligand–receptor interactions, to the data for the other species, we mapped human testis-expressed genes to their corresponding 1:1 orthologues in each of these species. Enriched receptor–ligand interactions between two cell types are predicted on the basis of expression of a receptor by one cell type and a ligand by another cell type 78 . Only receptors and ligands expressed in more than 10% of the cells in each cell type were considered for pairwise comparisons between all cell types in the dataset. CellPhoneDB uses empirical shuffling to calculate which ligand–receptor pairs show significant ( P < 0.05) cell-type specificity 78 . Significant interactions across species were illustrated using the R package UpSetR (v.1.4.0). Finally, given potential false positive (and negative) predictions of CellPhoneDB and similar approaches 79 , we consider interactions that are predicted for several species and probably reflect evolutionary conservation (Extended Data Fig. 8 and Supplementary Table 9 ), such as those reported in the main text, to be more reliable than species-specific predictions. Cell-type and testis-specific genes per chromosome Testis-specific genes were obtained from previously generated RNA-seq data 2 of adult organs (RPKM \(\ge \) 1 in testis and RPKM < 1 in brain, cerebellum, heart, kidney and liver). Among these, cell-type-specific genes were studied for each chromosome. Genes with predominant expression in specific somatic cell types were identified using the FindAllMarkers function (parameter only.pos = TRUE, min.pct = 0.05, logfc.threshold = 0.25, return.thresh = 0.05). Predominant expression of genes in specific germ cell types was assigned on the basis of the trajectory analyses (above); that is, predominant expression was assigned on the basis of the cell type in which the expression level of the gene peaks in the trajectory analysis. We then first calculated the percentage of genes located on a given chromosome among all genes in the genome ( x axis of plots in Extended Data Fig. 9a,c ; red horizontal line for X-linked genes in Fig. 4b ). We then contrasted this with the percentage of testis-specific genes with predominant expression in a given cell type ( y axis of plots in Extended Data Fig. 9a,c ; y axis of plots in Fig. 4b for X-linked genes). Finally, the percentage of testis-specific genes per cell type and chromosome was statistically compared to the percentage of genes per chromosome in the genome using exact binomial tests. Classification of X- and Y-bearing spermatids The Y chromosome carries a low number of genes and is missing in some genome assemblies. Thus, we focused on the fraction of X transcripts in spermatids to classify them as X- or Y-bearing cells. For this, we fitted a Gaussian Mixture Model to the data with two components (bimodal distribution) independently for each replicate, using the function normalmixEM of the mixtools (v.1.2.0) R package. The two obtained normal distributions were used to classify X- (higher levels of X transcripts) and Y-bearing (lower levels of X transcripts) spermatids using 95% confidence intervals. Outlier and overlapping cells were not assigned to either category. Finally, we checked that the fraction of Y transcripts was significantly higher in Y-bearing spermatids (Fig. 4c ). Bifurcating UMAPs (Fig. 4c ) were obtained using X- and Y-linked genes in addition to previously identified highly variable genes to perform the PCA associated with the UMAP coordinate calculation. For platypus, X transcripts are separated according to their location on the X chromosomes (that is, PARs and SDRs, respectively, as annotated in a previous study 46 ). In the platypus genome assembly used 46 , X and Y PARs are both assigned to the X chromosome. Thus, reported X transcripts may stem from X SDRs, X PARs or Y PARs, whereas reported Y transcripts only stem from Y SDRs. Illustrations of human and platypus sex chromosomes (with their respective PARs/SDRs) in Fig. 4a are based on previous work 46 , 80 . Transcriptomal differences between X and Y spermatids We identified differentially expressed genes between X- and Y-spermatid populations using the FindMarkers function from Seurat 62 R package (parameters, default). A Wilcoxon rank-sum test was used to calculate P values that were adjusted using Bonferroni corrections for several tests. Only genes that were detected as expressed in at least 10% of cells from either of the two populations were tested. Genes that show, on average, at least a 0.25-fold higher expression (log 2 -scale) in one of the populations, that are at the same time expressed in twice the number of cells in that population and that have an adjusted P value below 0.01 were considered to be differentially expressed. We note that several of the differentially expressed genes, including the most significant cases in human (Extended Data Fig. 11a ), are putatively non-coding, which is noteworthy because lncRNAs are typically nuclear 81 and hence their differential expression levels are unlikely to be offset by transcript exchange between spermatid cells through cytoplasmic bridges. The three most Y-spermatid-specific transcripts are lncRNAs emanating from homologous low copy repeats (that is, segmental duplications) on chromosomes 13 (FAM230C), 21 (XLOC-095504) and 22 (FAM230F) that cause genomic disorders by triggering non-allelic homologous recombination events. They include the FAM230F lncRNA in the q11.2 low copy repeat region on chromosome 22 (22q11.2) that is particularly susceptible to non-allelic homologous recombination-generated deletions that lead to various congenital malformation disorders, including the DiGeorge syndrome, the most frequent microdeletion disorder 82 . MSCI completeness analysis To identify potential MSCI escapee genes, we screened for X-linked genes with a significant increase in transcript abundance from SG to SC stages subject to MSCI, which ensures that potential escape genes are indeed actively transcribed in SC (that is, they do not merely represent genes expressed in SG with stable transcripts still detectable in SC), akin to previous work 23 . Specifically, we identified differentially expressed genes between SC and SG using the FindMarkers function from Seurat 62 R package (parameters, default). A Wilcoxon rank-sum test was used to calculate P values, which were then adjusted using a Bonferroni correction for several tests. Only genes that were detected as expressed in at least 10% of cells from either of the two cell-type populations were tested. Genes showing, on average, at least a 0.25-fold expression difference (log e -scale) between the two groups, and an adjusted P value below 0.05 were considered to be differentially expressed. X-linked genes in SDRs with significantly higher expression in SC than SG were considered to be potential escapees (Extended Data Fig. 12c ). General statistics and plots Unless otherwise stated, all statistical analyses and plots were done in R v.3.6.2 (ref. 83 ). Plots were created using ggplot2 v.3.2.1, tidyverse v.1.3.0, dplyr v.0.8.5, cowplot v.1.0.0 and pheatmap v.1.0.12. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Data availability Raw and processed bulk and snRNA-seq data have been deposited in ArrayExpress with the accession codes E-MTAB-11063 (human snRNA-seq), E-MTAB-11064 (chimpanzee snRNA-seq), E-MTAB-11067 (bonobo snRNA-seq), E-MTAB-11065 (gorilla snRNA-seq), E-MTAB-11066 (gibbon snRNA-seq), E-MTAB-11068 (macaque snRNA-seq), E-MTAB-11069 (marmoset snRNA-seq), E-MTAB-11071 (mouse snRNA-seq), E-MTAB-11072 (opossum snRNA-seq), E-MTAB-11070 (platypus snRNA-seq), E-MTAB-11073 (chicken snRNA-seq) and E-MTAB-11074 (chimpanzee, gorilla, gibbon and marmoset bulk RNA-seq) ( ). All other data are available as Supplementary Information or available upon request. The testis gene expression at the single-nucleus level across the 11 studied species can be visualized using the shiny app we developed: . Code availability Custom scripts used to generate the results reported in the paper and processed data are available at .

Evolutionary pressure across male mammals to guarantee the procreation of their own offspring led to a rapid evolution of the testicle. Bioinformatic studies—conducted by an international team of researchers led by Prof. Dr. Henrik Kaessmann from the Center for Molecular Biology of Heidelberg University—show that this pressure particularly accelerated the evolution of later stages of sperm formation. The aim of these contrastive studies was, for the first time, to decode the genetic regulation of sperm formation in various species of mammals and in human beings, thereby tracing the evolution of this spermatogenesis. At the same time, the researchers were also able to detect genes whose activity had remained unchanged in the course of evolution. Spermatogenesis in the testicle is controlled by a finely coordinated, complex interplay of the activity of different genes—also known as gene expression. Hitherto the understanding of these genetic programs had been largely confined to the mouse. "Consequently little was known about the genetic foundations that constitute the big differences in spermatogenesis across different mammals, both with regard to the number of sperm cells formed and also to their properties," explains Noe Mbengue, a doctoral researcher in Prof. Kaessmann's group "Evolution of the mammalian genome." The Heidelberg scientists have now succeeded in defining the expression of all genes at the level of individual cells during the whole of spermatogenesis for ten different mammals. The organisms they studied represent all major groups of mammals and include humans as well as their closest relatives, great apes. To do this, the researchers used state-of-the-art single-cell genomics technologies. Based on this data they were subsequently able to trace the evolution of spermatogenesis with the aid of bioinformatic comparisons between the different mammals. According to Prof. Kaessmann, these comparative studies uncovered a time-related pattern. "While the genetic programs in the early stages of spermatogenesis are very similar among mammals, in later stages they differ greatly; that means that the rapid evolution of the testicle is a result of major differences in cells during late spermatogenesis," says Dr. Florent Murat, a former postdoc in Henrik Kaessmann's research group and now a group leader at the National Research Institute for Agriculture, Food and Environment (INRAE) in Rennes (France). Further analyses by the scientists revealed genes whose activity had remained unchanged in the course of evolution. They regulate fundamental processes of sperm cell formation that are the same for all mammals. "Hence our data also supplies valuable elements for researching fertility disorders in men," Prof. Kaessmann explains. Finally, the scientists' data enabled them for the first time to distinguish sperm cells that carry either an X or a Y chromosome and thus determine the sex of the offspring. With the aid of this division, the researchers succeeded in systematically studying the gene expression on these sex chromosomes. As these investigations showed, gene expression on the sex chromosomes of all male mammals is downregulated during the maturation division known as meiosis. This mechanism is presumably fundamental for preventing a disadvantageous genetic exchange between the X and Y chromosome during meiosis. The results of the study on the evolution of spermatogenesis across mammals were published in the journal Nature.

10.1038/s41586-022-05547-7

Medicine

How reducing body temperature could help a tenth of all ICU patients

C. Autilio et al. Molecular and biophysical mechanisms behind the enhancement of lung surfactant function during controlled therapeutic hypothermia, Scientific Reports (2021). DOI: 10.1038/s41598-020-79025-3 Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-020-79025-3

https://medicalxpress.com/news/2021-02-body-temperature-tenth-icu-patients.html

Abstract Therapeutic hypothermia (TH) enhances pulmonary surfactant performance in vivo by molecular mechanisms still unknown. Here, the interfacial structure and the composition of lung surfactant films have been analysed in vitro under TH as well as the molecular basis of its improved performance both under physiological and inhibitory conditions. The biophysical activity of a purified porcine surfactant was tested under slow and breathing-like dynamics by constrained drop surfactometry (CDS) and in the captive bubble surfactometer (CBS) at both 33 and 37 °C. Additionally, the temperature-dependent surfactant activity was also analysed upon inhibition by plasma and subsequent restoration by further surfactant supplementation. Interfacial performance was correlated with lateral structure and lipid composition of films made of native surfactant. Lipid/protein mixtures designed as models to mimic different surfactant contexts were also studied. The capability of surfactant to drastically reduce surface tension was enhanced at 33 °C. Larger DPPC-enriched domains and lower percentages of less active lipids were detected in surfactant films exposed to TH-like conditions. Surfactant resistance to plasma inhibition was boosted and restoration therapies were more effective at 33 °C. This may explain the improved respiratory outcomes observed in cooled patients with acute respiratory distress syndrome and opens new opportunities in the treatment of acute lung injury. Introduction Therapeutic hypothermia (TH) is used as an effective method to limit brain injury after certain types of cardiac arrest and perinatal asphyxia, by reducing patient body temperature from 37 to 33.5 °C 1 , 2 . Early studies also suggest some possible benefits in cooled patients suffering of acute respiratory distress syndrome (ARDS) 3 , 4 , 5 , 6 . ARDS accounts for 10% of all cases in intensive care units worldwide. Mortality remains at 30–40% and no effective treatments are available to date 7 . Primary ARDS occurs after a direct insult to the lung, whereas secondary ARDS is triggered by an extra-pulmonary pro-inflammatory event 7 , 8 , 9 . Theoretically, TH may be useful to treat primary but might be harmful in secondary ARDS 10 , 11 . Conversely, during primary ARDS the injury is at first confined to the lung, resulting in lower compliance and more severe surfactant dysfunction than in the secondary form of the syndrome 7 , 8 , 9 . During ARDS, surfactant function is impaired in several ways 12 , 13 . Protein-rich oedema fluid, high levels of cholesterol and secretory phospholipase A2 (sPLA 2 ) can all inhibit surfactant activity directly and enhance the production of reactive oxygen species (ROS). Lung surfactant is dramatically needed to reduce surface tension under the expiration-mediated shrinking of the alveolar air–liquid interface. The precise proportions of dipalmitoylphosphatidylcholine (DPPC), unsaturated phospholipids and hydrophobic surfactant proteins B (SP-B) and C (SP-C) are essential for its function in a healthy lung 12 . However, oedema proteins, such as albumin, increase in ARDS patients 14 and may compete with surfactant membranes for the air–liquid interface. The resulting interfacial steric barrier affects the in vitro surfactant adsorption and the reduction in surface tension under compression 15 . Beyond the physiological range, the drastic rise in cholesterol compared to the saturated lipid amount leads to surfactant membranes fluidification 12 . This impairs the ability of clinical and native surfactants to decrease surface tension upon breathing-like cycles 16 , 17 . Besides, methyl-β-cyclodextrin can ex vivo overcome this cholesterol-mediated inhibition in a mice model of ARDS 18 , 19 . PLA 2 targets and hydrolyses the condensed domains of phosphatidylcholine monolayers with a crumbling effect at the air–liquid interface 20 . Consistently, high levels of the secretory form of the enzyme was detected in ARDS patients along with changes in surfactant phospholipid composition and surface tension properties 21 . As a result, surfactant amount and performance are reduced and the work of breathing increases upon ARDS 13 . Interestingly, cooled animals with different types of lung failure showed better respiratory mechanics and reduced lung tissue inflammation 22 , 23 , 24 . Moreover, reports on cooled asphyxiated neonates with and without neonatal ARDS due to meconium aspiration describe an improvement in surfactant function 25 , 26 , 27 , pulmonary inflammation 27 , oxygenation and some clinical outcomes such as reduction of respiratory support and hospital stay 25 , 28 . The biological mechanisms underlying this benefit are not totally clear. DPPC turnover did not change in cooled neonates 29 , but a reduction in sPLA 2 activity has been reported 30 . However, several biophysical mechanisms may also contribute to surfactant performance under cooling. Drastic changes in body temperature leads to adaptive variations in both surfactant structure and composition to maintain proper interfacial properties in hibernating animals 31 , 32 . Besides, the surface activity of rabbit lung extracts tends to improve decreasing temperature from 40 to 15 °C 33 . It is well known that the organization and physical properties of lipid membranes, like those in lung surfactant, can be influenced by shifting temperatures towards their melting points 34 . Cooling to values below 30 °C affects lung surfactant performance in the presence of plasma compounds 35 . However, the reduction from 37 to 33 °C may have a different effect, as it is still above surfactant melting temperature (≈ 32 °C) 34 . To the best of our knowledge, no detailed in vitro studies have investigated surfactant biophysical properties under TH. Here, we analyse for the first time both the interfacial structure and the composition of lung surfactant films subjected to compression at 33 °C compared with 37 °C. We also describe enhanced surfactant function and resistance to inhibition by plasma under cooling, as a well as a more efficient restoration of function as a consequence of further surfactant supplementation. Finally, we investigate the molecular mechanisms governing this enhanced performance and propose a feasible biophysical model. Results Moderate hypothermia enhances surfactant activity in a concentration-dependent manner The activity of purified porcine surfactant (PS) was tested at different concentrations (1.5, 2.5 and 5 mg/mL) under slow (2 cycles/min) and quick (20 cycles/min) dynamic cycles by Constrained Drop Surfactometer (CDS) and Captive Bubble Surfactometer (CBS), respectively (Fig. 1 ). PS initial adsorption did not change among conditions (Fig. 1 a), whereas dynamic activity significantly depended on sample concentration. Figure 1 Changes in PS activity depending on temperature and concentration under slow and quick compression-expansion cycling. ( a ) CDS experiments under increasing concentration of PS (from 1.5 to 5 mg/mL) at 33 °C and 37 °C. On the left, surface tension values (γ, mN/m) reached by PS 10 s after the buffer deposition on top of the pedestal (initial adsorption). On the right, one typical replicate of γ-∆area isotherms representative of the ten slow compression-expansion cycles (2 cycles/min, shrinking the interface area by 20%, 0.2 cm 2 /min). ( b ) Minimum and maximum surface tensions reached during cycles 1, 5 and 10, testing surfactant at limiting concentration (1.5 mg/mL, n = 5). ( c ) CBS experiments, using a PS concentration of 1.5 mg/mL at 33 °C and 37 °C. On the left, γ values reached during 5 min (300 s) of PS adsorption after injecting material onto the bubble (initial adsorption). In the middle, one representative replicate (n = 3) of γ-∆area isotherms under 10 quick compression-expansion cycles (20 cycles/min). On the right, minimum and maximum surface tensions reached during cycles 1, 5 and 20. Black and light-grey bars represent experiments performed at 33 °C and 37 °C, respectively. Means and standard deviation of at least three replicates are shown. Compression-expansion cycles are depicted in grey scale. Horizontal lines represent statistical comparisons. Two-way ANOVA test followed by post-hoc test = (a) initial adsorption: n. s.; (b) γ min : temperature (p = 0.002, D.F. = 1, F = 11.6) and cycles (p < 0.001, D.F. = 2, F = 30.2), γ max : n. s.; (c) initial adsorption: n. s.; γ min : temperature (p = 0.001, D.F. = 1, F = 20.2) and cycles (p < 0.001, D.F. = 2, F = 40.2), γ max : temperature (p = 0.003, D.F. = 1, F = 13.4) and cycles (p < 0.001, D.F. = 2, F = 180.1). *p < 0.05 and > 0.01, **p ≤ 0.01 and > 0.005, ***p ≤ 0.005. The exact p values of the most relevant post-hoc tests are indicated. The corresponding comparison bars and symbols are highlighted in orange color. PS purified porcine surfactant, γ surface tension, min minimum, max maximum, n. s. not significant, D.F. degrees of freedom. Full size image At limiting concentration (1.5 mg/mL), PS significantly reduced surface tension after 5 slow cycles at 33 °C, whereas it needed 10 cycles at 37 °C (Fig. 1 a,b). This trend was also evident at a PS concentration of 2.5 mg/mL. Using the latter condition, there was a significant difference in γ min comparing cycle 5 and 10 for material tested at 37 °C, but it was not the case for PS assayed at 33 °C (Supplementary Fig. S1 a). However, this trend was not present at higher concentration (5 mg/mL) (Supplementary Fig. S1 b). Similar results were obtained in CBS experiments. At limiting concentration, γ min is significantly lower after 10 quick cycles at 33 °C compared with 37 °C. Consistently, the minimum surface tension was significantly higher along cycles at 37 °C. At the same time, the maximum surface tension significantly rose upon the interfacial dynamics irrespective of the experimental temperature (Fig. 1 c). This is probably related to the highly diluted surfactant used for the experiments and the resulting easy loss of material from the air–liquid interface under compressions. Moderate hypothermia increases the size of condensed domains in surfactant films, promoting the exclusion of unsaturated phospholipids upon breathing The activity and structure of PS were tested in a Langmuir–Blodgett trough (Fig. 2 ) before and after 10 compression-expansion cycles (0.3 cycles/min). PS activity was slightly better at 33 °C compared with 37 °C. Maximum surface pressures were significantly higher (minimum surface tensions significantly lower) at 33 °C compared with 37 °C. This occurred both before and in the first cycle of the experiments (Fig. 2 a,c). Figure 2 Temperature-dependent changes in PS activity, structure and composition in a Langmuir–Blodgett trough. ( a ) One representative replicate (n = 5) of Π-Δarea isotherms before and during 10 slow compression-expansion cycles of the interface (0.3 cycles/min, ≈65 cm 2 /min, reducing the interface area by ≈65%) at 33 °C and 37 °C. Around 50 μg (at 5 mg/mL) of PS was added at the interface before starting the experiment. Cycles 1, 5 and 10 are shown. ( b ) Epifluorescence analysis (n = 2) of PS lateral structure upon slow interface compressions (25 cm 2 /min) before and after 10 quicker compression-expansion cycles (65 cm 2 /min). PS was previously doped with the fluorescent probe BODIPY-PC (1% mol/mol). Upon compression, clusters of condensed lipids (enriched in saturated PL species) excluding the bulky fluorescent probe are visible as black domains against a green background (which represents the mixture of more expanded and disordered lipids and probe). Brilliant green spots, probably representing tridimensionally excluded structures, are also visible. The lateral structure of PS is shown before reaching the equilibrium plateau at several surface pressures. White scale bars, 100 μm. For each image, histogram stretching has been performed to enhance contrast without deleting pixel data. ( c ) On the left, the initial and the maximum surface pressures obtained before and after the first compression (25 cm 2 /min). On the right, the initial surface pressure obtained before starting the experiment together with the maximum and minimum surface pressures during the 1st and 10th compression-expansion cycles (65 cm 2 /min). Mean and SD of 5 replicates are shown. ( d ) Lipidomic analysis of PS proximally associated to the surface during compression of the interface (25 cm 2 /min) at 33 °C or 37 °C. The interface recollected from 5 replicates were pooled per condition and analysed by lipidomic analysis. All lipid species were detectable. Each lipid species was normalized with respect of the total amount of PLs and reported as percentage. Black and light-grey colors represent experiments performed at 33 °C and 37 °C, respectively. Black horizontal lines represent statistical comparisons. Black arrows highlight the variation in PS lipid species at 33 °C compared with 37 °C. Paired t test = Π max before cycles: p = 0.029, t = 3.3, D.F. = 4. Two-way ANOVA test followed by post-hoc test = Π max during cycles: temperature (p = 0.025, D.F. = 1, F = 5.7) and cycles (p < 0.001, D.F. = 1, F = 593); Π min during cycles: temperature (p < 0.001, D.F. = 1, F = 375.5) and cycles (p < 0.001, D.F. = 2, F = 83.6). *p < 0.05 and > 0.01, ***p ≤ 0.005. The exact p values of the most relevant post-hoc tests are indicated. The corresponding comparison bars and symbols are highlighted in orange color. Π surface pressure, max maximum, min minimum, PC phosphatidylcholine, PLs phospholipids, PG phosphatidylglycerol, sat saturated, unsat unsaturated, CHOL cholesterol, puPC polyunsaturated phosphatidylcholine, D.F. degrees of freedom. Full size image Upon cycling, the plateau during expansions was visible at higher surface pressures at 33 °C than 37 °C (Fig. 2 a). This suggested that the compressed phase was more stable at 33 °C while the material tended to relax quicker at physiological temperature. Moreover, under compression of the air–liquid interface, a higher amount of surfactant seems to be excluded at 37 °C than at 33 °C and not replaced during the subsequent expansion. This results in a tendency to increase minimum surface pressures at 33 °C taking the last cycles into account (p = 0.09) (Fig. 2 a,c). Upon compression, and before further cycling, the condensed black domains at the surface films were rounder and less polymorphic than after compression-expansion cycling, as observed under epifluorescence microscopy (Fig. 2 b). This result was detected regardless of the temperature, suggesting an intrinsically fluid character of the ordered phase before cycling, with a reduced contact perimeter with the coexisting surrounding disordered phases. On the contrary, a de-mixing of lipids, together with larger gel condensed phases, were observed after cycling, especially at 33 °C. These phases apparently contained smaller domains that were organized in clusters. Moreover, brilliant fluorescent spots were present at both temperatures and remained visible after cycles, especially at 33 °C. These probe-rich spots are likely areas of three-dimensional exclusion that appear bright due to the light scattering. Their apparent size was reduced after cycling associated with a lighter green liquid-expanded phase. Interestingly, due to the lower lipid miscibility at 33 °C than 37 °C, phase-coexistence was more evident under hypothermia temperature. This leads to larger green spots, presumably excluded from the interface, along with larger and darker condensed domains at 33 °C than 37 °C (Fig. 2 b, Supplementary Fig. S2 ). The composition of PS lipids that were proximally associated to the surface under compression was analysed by LC-HRMS (Supplementary Table S1 ). The percentage of saturated lipids (especially DPPC) and cholesterol increased at 33 °C compared with 37 °C, whereas the amount of unsaturated lipids, especially those with more than 2 double bonds (polyunsaturated phatidylcholine (puPC)), decreased (Fig. 2 d). DPPC proportion, hydrophobic surfactant proteins and steric hindrance of excluded phospholipids contribute to improve surfactant activity at 33 °C Several surfactant lipid mixtures (DPPC, palmitoyloleoylphosphatidylglicerol (POPG), palmitoyloleoylphosphatidylcholine (POPC) and dioleoylphosphatidylcholine (DOPC) at different proportions) with decreasing capability to pack under compression were combined with SP-B alone (1% total protein-to-lipid by weight) or with both SP-B and SP-C (2% total protein-to-lipid by weight). The combination of DPPC, POPC and POPG roughly simulates the composition of a typical lung surfactant in terms of saturated/unsaturated and zwitterionic/anionic phospholipid species. The presence of DOPC brings the contribution of phospholipids (PLs) with a large area per molecule, such as it is the case with polyunsaturated species. The surface properties of each mixture were assayed under quick cycles (20 cycles/min) by CBS (Fig. 3 a,b). No differences among 33 and 37 °C were noticed for PS at 8 mg/mL (data not shown). Thus, only results at physiological temperature (37 °C) for PS are shown and considered as control for optimal biophysical activity (γ min < 2mN/m, γ max ≈ 30 mN/m). Figure 3 Contribution of lipid species and hydrophobic surfactant proteins in the temperature-mediated improvement of surfactant activity. Several surfactant lipid mixtures (DPPC, POPG, POPC and DOPC at different proportions) with decreasing capability to pack under compression were combined with purified porcine surfactant protein SP-B alone (1% w/w) or both SP-B and SP-C (2% total protein-to-lipid by weight). CBS experiments, using a concentration of 8 mg/mL of each lipid-protein mixture, were performed at 33 °C or 37 °C. ( a ) On the left, the mass percent of the lipids used in each experiment is shown. Drawings of the lipids are also shown to illustrate the different space occupied at the interface according to their unsaturation grade. Graphs of initial adsorption show means and SD of three replicates per condition. One representative replicate (n = 3) of γ-∆area isotherms under 10 quick compression-expansion cycles (20 cycles/min, 1.37 atm/s) is also shown. ( b ) Minimum and maximum surface tensions reached upon 20 dynamic cycles (n = 3), comparing each lipid–protein mixture to purified PS (8 mg/mL). Compression-expansion cycles are depicted in grey scale. Means and SD of three replicates are shown. Horizontal lines represent statistical comparisons. Paired t test = (a) γ(300 s) during initial adsorption: p = 0.025, D.F. = 2, t = − 6.2. Two-way ANOVA test followed by post-hoc test = (b) first line, γ min : temperature (n. s.) and composition (p = 0.007, D.F. = 2, F = 8.4), γ max : temperature (p < 0.001, D.F. = 1, F = 26.4) and composition (p < 0.001, D.F. = 2, F = 150.6); second line, γ min : temperature (p < 0.001, D.F. = 1, F = 258.9) and composition (p < 0.001, D.F. = 2, F = 270.3), γ max : temperature (n. s.) and composition (p < 0.001, D.F. = 2, F = 22.4); third line, γ min : temperature (n. s.) and composition (p = 0.005, D.F. = 2, F = 9.4), γ max : temperature (n. s.) and composition (p < 0.001, D.F. = 2, F = 41.1); fourth line, γ min : temperature (p < 0.001, D.F. = 2, F = 168.9) and composition (p < 0.001, D.F. = 2, F = 160.9), γ max : temperature (n. s.) and composition (p < 0.001, D.F. = 2, F = 31.5). *p < 0.05 and > 0.01, **p ≤ 0.01 and > 0.005, ***p ≤ 0.005. The exact p values of the most relevant post-hoc tests are indicated. The corresponding comparison bars and symbols are highlighted in orange color. γ surface tension, DPPC dipalmitoylphosphatidylcholine, POPG palmitoyloleoylphosphatidylglicerol, POPC palmitoyloleoylphosphatidylcholine, DOPC dioleoylphosphatidylcholine, PS purified porcine surfactant, SP-B surfactant protein B, SP-C surfactant protein C, min minimum, max maximum, n. s. not significant, D.F. degrees of freedom. Full size image The presence of both SP-B and SP-C increased the adsorption and dynamic activity of all tested materials regardless of the temperature. Upon dynamic cycling, there was a significant reduction in minimal surface tensions at 33 °C compared with 37 °C when a high percentage of DPPC (65%) with respect to POPG (35%) was present (Fig. 3 a, second line). This difference was more evident in the absence of SP-C, suggesting the importance of both hydrophobic proteins to exclude POPG from the interface under compression. Moreover, although both SP-B and SP-C were present (Fig. 3 a, second line on the left), a trend to significance was observed at 33 °C for the reduction in area that is required to reach minimal surface tension: [13.6 (2.3)% at 33 °C vs 23.8 (2.2)% at 37 °C, p = 0.05]. Together with the DPPC content, the steric hindrance of excluded PLs under compression also enhanced surfactant activity at 33 °C. The difference among temperatures was significant in the presence of DOPC (Fig. 3 a,b, fourth line). Conversely, the lipid-protein mixture containing POPC did not show any temperature-mediated changes in surface-active properties. PLs of this mixture probably did not mix well, leading to a marked variability among replicates in CBS and the co-existence of 2 peaks in Differential Scanning Calorimetry (DSC) (Supplementary Fig. S3 a). Finally, the dynamic properties of each mixture and PS were also compared (Fig. 3 b). Regarding minimum surface tensions, PS activity was comparable to the activity of the mixture containing DPPC/POPG (65/35%, w/w) with SP-B alone or with SP-C. This result was more evident for material tested at 33 °C. Interestingly, the melting temperatures of these 2 mixtures (Supplementary Fig. S3 a,b) were very close to TH: 32.4 (0.12) and 33.9 (0.06) °C. Moderate hypothermia decreases surfactant inhibition by plasma and enhances the restoration activity of therapeutic surfactants The inhibition of PS activity by plasma was tested by CDS upon slow cycling (2 cycles/min) at different surfactant concentrations (1.5, 2.5 and 5 mg/mL), both at 33 and 37 °C (Fig. 4 a). Within each condition, no differences between temperatures in initial adsorption and γ max were noticed. Conversely, the dynamic activity (γ min ) of PS at 1.5 mg/mL was clearly impaired in the presence of plasma (Fig. 4 a, second line). Interestingly, at this concentration, PS capability to produce very low surface tension was reduced with and without plasma at 37 °C, but still functional in the absence of plasma at 33 °C (Fig. 4 a, second line). Moreover, when PS was tested at 2.5 mg/mL it worked properly in the presence of plasma at 33 °C, but it was inactivated at 37 °C (Fig. 4 a, second line). Figure 4 PS inhibition by plasma and restoration at different temperatures. ( a ) CDS experiments of PS are shown at different concentrations at the 2 temperatures (1.5 (n = 4), 2.5 (n = 3) and 5 (n = 3) mg/mL) with and without plasma (68 mg/mL of plasma total proteins (TP)), corresponding to 1.6, 2.6 and 5.2% PLs/plasma TP. Initial adsorption along with minimum and maximum surface tensions upon 10 slow compression-expansion cycles (2 cycles/min, shrinking the interface area by 20%, 0.2 cm 2 /min) are depicted compared to plasma (n = 3) and buffer alone (n = 3). ( b ) Restoration of PS activity after plasma inhibition, with 1.5 mg/mL and 2.5 mg/mL of PS at 33 °C and 37 °C. After inhibiting surfactant, several amounts of further surfactant were supplemented to mimic therapy, using PORα or the lipid-protein mixture [DPPC/POPG (65/35%, w/w) + SP-B/SP-C (2% total protein-to-lipid by weight)]. These supplemented amounts were calculated as the percentage of μg of PLs of therapy with respect to the μg of inhibited PS. As further controls, only buffer was dispensed as a mock therapy for inhibited PS and the restoration therapy was analysed in the presence of plasma alone. Black and light-grey bars represent experiments performed at 33 °C and 37 °C, respectively. Means and standard deviation of at least three replicates are shown. Horizontal lines represent statistical comparisons. Two-way ANOVA test followed by post-hoc test = (a) initial adsorption: temperature (n. s.), PS concentration (p < 0.001, D.F. = 3, F = 139.1) and plasma presence (n. s.); γ min : temperature (n. s.), PS concentration (p < 0.001, D.F. = 3, F = 139.1), plasma presence (n. s.), temperature/PS concentration/plasma presence (p < 0.006, D.F. = 3, F = 4.7); γ max : temperature (n. s.), PS concentration (p < 0.001, D.F. = 3, F = 124.7) and plasma presence (p < 0.001, D.F. = 1, F = 16.9). Paired t test γ min = 33 °C: PS at 1.5 mg/mL with and without plasma (p = 0.046, D.F. = 3, t = − 3.3), 37 °C: PS at 2.5 mg/mL with and without plasma (p = 0.016, D.F. = 3, t = − 5.0). One-way ANOVA test followed by post-hoc test = (b) synthetic lipid mixture: 33 °C (p < 0.001, D.F. = 3, F = 101), 37 °C (p < 0.001, D.F. = 3, F = 22.2); PORα: 33 °C (p = 0.044, D.F. = 5, F = 3), 37 °C (p = 0.004, D.F. = 3, F = 8.6). *p < 0.05 and > 0.01, **p ≤ 0.01 and > 0.005, ***p ≤ 0.005. The exact p values of the most relevant post-hoc tests are indicated. The corresponding comparison bars and symbols are highlighted in orange color. PS purified porcine surfactant, γ surface tension, min minimum, max maximum, DPPC dipalmitoylphosphatidylcholine, POPG palmitoyloleoylphosphatidylglicerol, SP-B surfactant protein B, SP- C surfactant protein C, PORα poractant α, n. s. not significant, D.F. degrees of freedom. Full size image Restoration experiments (Supplementary Fig. S4 ) were then performed, maintaining the same amount of plasma, but using different concentrations of PS depending on the experimental temperature: 1.5 mg/mL at 33 °C and 2.5 mg/mL at 37 °C. The restoration activities of two different materials were studied under these conditions: (1) Poractant α (PORα), a widely used therapeutic surfactant, and (2) DPPC/POPG (65/35, w/w) + SP-B/SP-C (2% total protein-to-lipid by weight), the lipid-protein mixture that had shown the best surface-activity (comparable to PS). Several amounts of these materials were used to mimic a therapy, by adding small volumes at increasing concentrations on top of plasma-inhibited surfactant (Supplementary Table S2 , Fig. S4 ) until reaching restoration. These amounts were calculated as the mass percentage of dispensed PLs to PLs in the inhibited surfactant (e.g. 5 μL of PS at 1.5 μg/μL were used at 33 °C, thus 80% of required therapy means 6 μg of therapy). Since the concentration of surfactant material highly influences its surface-active properties 36 , we decided to use the same therapy concentration for both temperatures. This was performed by dispensing higher volumes at 37 °C compared with 33 °C, still maintaining values up to 4% of the constrained drop volume. As a result, the same percentage (with respect to total PLs in inhibited PS) and concentration of therapy were used regardless of the experimental temperature. Moreover, due to the different starting PS concentration (1.5 and 2.5 mg/mL for 33 and 37 °C, respectively), we focused on the therapy percentage necessary to rescue PS activity for data interpretation. The effects of adding buffer alone to the inhibited PS were also analysed as a reference (Fig. 4 b). Regardless of the therapy used, the amount necessary to restore PS activity was always lower at 33 °C compared with 37 °C. Moreover, the lipid-protein mixture showed better restoration (80–100%), compared to PORα (200%) with a temperature-mediated improvement in the restoration capability. For this material, the restoration dose required was lower at 33 °C than 37 °C (Fig. 4 b). This was probably due to the better exclusion of POPG from the interface under compression at 33 °C, leaving enough DPPC to reduce surface tension. To further investigate this hypothesis, three lipid-protein mixtures with decreasing DPPC/POPG ratios were tested with plasma in CDS at 33 and 37 °C and compared to PORα (Fig. 5 a,b). Interestingly, the lipid-protein mixtures with DPPC ≥ 50% and POPG ≤ 50%, seem to be more resistant to plasma inhibition than PORα regardless of the experimental temperature. On the contrary, when increasing the amount of POPG to 65%, the activity was inhibited at 37 °C but still functional at 33 °C (Fig. 5 b). Figure 5 Temperature-mediated resistance to plasma inhibition, depending on DPPC/POPG ratio. About 200 nL at 40 mg/mL of PORα (n = 5 33 °C, n = 3 37 °C) or the lipid-protein mixtures [DPPC/POPG + SP-B/SP-C (2% total protein-to-lipid by weight)] with increasing percentages of POPG (n = 3 for each mixture and temperature), were dispensed at the interface of plasma and subjected to 4 slow cycles in CDS (2 cycles/min, shrinking the interface area by 20%, 0.2 cm 2 /min). ( a ) One representative replicate of γ-∆area isotherms at the 2 temperatures for each tested material. ( b ) On the left, γ values reached after 10 s of material spreading at the plasma interface (initial spreading). In the middle and on the right, γ min and γ max reached during cycling. Compression-expansion cycles are depicted in grey scale. Black and light-grey bars represent experiments performed at 33 °C and 37 °C, respectively. Means and SD of at least three replicates are shown. Horizontal lines represent statistical comparisons. Two-way ANOVA test followed by post-hoc test = initial adsorption: (n. s.); γ min : temperature (p < 0.001, D.F. = 1, F = 20.5) and composition (p < 0.001, D.F. = 3, F = 11); γ max : temperature (n. s.) and composition (p = 0.003, D.F. = 3, F = 6.8). Paired t test = 65% POPG, γ min at different temperatures: p = 0.027, D.F. = 2, t = − 6. *p < 0.05 and > 0.01, **p ≤ 0.01 and > 0.005, ***p ≤ 0.005. The exact p values of the most relevant post-hoc tests are indicated. The corresponding comparison bars and symbols are highlighted in orange color. DPPC dipalmitoylphosphatidylcholine, POPG palmitoyloleoylphosphatidylglicerol, SP-B : surfactant protein B, SP-C surfactant protein C, PORα poractant α, γ surface tension, min minimum, max maximum, n. s. not significant, D.F. degrees of freedom. Full size image Discussion We found that TH induces several effects on surfactant properties, compared with the physiological temperature: (1) the capability of PS to drastically reduce surface tension under breathing-like conditions is enhanced in a concentration-dependent manner; (2) surfactant structure and composition at the air–liquid interface re-organize differently under compression (expiration): larger DPPC-enriched condensed domains and lower percentages of less active lipids are detected at 33 °C; (3) PS resistance to plasma inhibition is boosted under cooling; (4) in vitro restoration therapies are more effective at 33 than at 37 °C; (5) the higher the DPPC proportion, the better the restoration performance under TH. Up to now, no causal therapies are available for ARDS. Although surfactant inhibition is not necessarily the primary pathogenic factor during ARDS 7 , the protein-rich oedema fluid leaking into the alveoli and the local inflammation make its inactivation a secondary and critical step of the syndrome 13 , 37 , 38 , 39 . This occurs through different biological mechanisms. For instance, several plasma proteins inhibit the proper adsorption of surfactant, creating a steric barrier at the interface (albumin, fibrinogen), fluidifying surfactant membranes (C-reactive protein) or degrading surfactant lipids (lipases, sPLA 2 ) and proteins (proteases) 12 , 13 , 40 . Here, we observed that the initial adsorption of diluted PS in the presence of plasma was similar at 33 °C and 37 °C (Fig. 4 a). However, plasma compounds were better excluded at 33 °C upon interfacial dynamics. Consistently, the γ min was very low regardless of the plasma presence, testing surfactant at 2.5 mg/mL and 33 °C. On the contrary, there was a significant increase in γ min when PS was assayed with plasma at 37 °C (Fig. 4 a, second line). Interestingly, some plasma proteins, such as fibrinogen, can be found into the liquid-expanded phase during lateral compression 41 . Due to the lipid de-mixing, we observed larger condensed (DPPC-enriched) domains at 33 °C and a liquid-expanded phase with larger excluded three-dimensional structures (Fig. 2 b, Supplementary Fig. S2 ). A temperature-dependent appearance of condensed domains during interfacial compression was already described for DPPC and bovine surfactant 42 . Thus, we can speculate that a similar mechanism occurs with plasma compounds. Plasma proteins might be better (at lower pressures) squeezed out from the interface at 33 °C, thus increasing the resistance to plasma inactivation under TH. Moreover, in both adult and meconial ARDS, high cholesterol and polyunsaturated phosphatidylcholine levels were described as a lung response or due to plasma extravasation and meconium presence 25 , 43 , 44 . This may impede surfactant to properly reduce surface tension during expiration. For instance, increasing the amount of cholesterol with respect to DPPC leads to changes in the proper packing of DPPC domains, affecting the capability to rise surface pressures upon lateral compression 45 . Interestingly, the surface elasticity of spread DPPC monolayers at high surface pressures seems to lower under increasing temperatures 46 . This reduces the interfacial activity of material, probably due to the quicker relaxation of DPPC monolayers that lose their capability to sustain high surface pressure. In our experiments, an increased percentage of cholesterol and saturated lipids was detected at the air–liquid interface at 33 °C (Fig. 6 ). Figure 6 Schematic representation of possible candidates involved in the temperature-mediated improvement of surfactant performance and structural changes at the air–liquid interface. Our findings suggest the presence of higher levels of cholesterol and saturated lipids (mainly DPPC) along with reduced amount of unsaturated lipid species at 33 °C compared with 37 °C. This promotes the formation of ordered condensed domains with bigger size at hypothermia temperature compared with 37 °C. Full size image The sterol presence seems to influence the dynamic surface elasticity of DPPC monolayers under compression 47 . However, we still observed a temperature-mediated effect on the lateral organization of DPPC/unsaturated PC domains of PS at the air–liquid interface. This is probably due to the simultaneous increase in both cholesterol and saturated lipids detected at hypothermia temperature. In this regard, we demonstrated formation of larger condensed interfacial domains at 33 °C than at 37 °C during lateral compression (Figs. 2 b, 6 , Supplementary Fig. S2 ). Cholesterol may insert preferentially into ordered structures. Thus, a larger DPPC proportion at the interface at 33 °C (Fig. 2 d) might cause a higher amount of cholesterol to partition from the surrounding disordered phases. These results may suggest a protective role of TH, reducing the cholesterol-mediated inhibition. This might occur by promoting DPPC de-mixing during compression and indirectly the sterol captures into ordered domains at 33 °C. During ARDS, raised levels of sPLA 2 leads to a significant reduction in both surfactant DPPC 13 , 37 , 38 , 39 and PG amounts 21 . At the same time, surfactant proteins decrease 13 and ROS increase, oxidizing the lipid-protein membranes 12 . As a result, the availability of surface-active material is drastically reduced and surfactant complexes become inactivated. Interestingly, considering our data, our findings suggest that TH could be also beneficial under these impaired surfactant conditions: (1) at lower concentration, as in the ARDS, surfactant performance is better at 33 °C than 37 °C (Fig. 1 ), (2) surfactant exhibits higher resistance to plasma inhibition under TH (Fig. 4 ); (3) due to the lower miscibility of lipids at 33 °C, unsaturated lipids (e.g. puPC and DOPC) and plasma components are removed more efficiently from the interface. This occurs with a lower loss of DPPC during the interfacial compression (at expiration) (Figs. 2 , 5 ). This last mechanism (outlined in Fig. 7 ) fits with the first two effects perfectly. A better performance of surfactant at 33 °C with and without plasma is closely linked to the concentration of tested material. Lower concentrations mean lower DPPC amounts at the air–liquid interface, that are even more reduced under compression at 37 °C. Figure 7 Temperature-mediated compression-driven reorganization of surfactant at the air–liquid interface. ( a ) When surfactant adsorbs at the air–liquid interface, the lateral segregation between packed ordered domains (DPPC enriched) and expanded disordered lipids is less evident at 37 °C due to temperature-facilitated mixing. Larger condensed domains are visible at 33 °C. (b ) During compression (at expiration), SP-B and SP-C promote surfactant reorganization at the air–liquid interface, facilitating maximal reduction in surface tension. Areas rich in lipids with lower stability at high pressures, and therefore poorer surface-active properties (unsaturated lipids) are laterally and three-dimensional excluded from the air–liquid interface. Due to lipid de-mixing, the condensed ordered/liquid-expanded disordered phase separation is more evident at 33 °C. This leads to compression-driven preferential exclusion of unsaturated lipids. Conversely, a fraction of DPPC is also lost from the interface at 37 °C because of the temperature-dependent partial mixing. ( c ) At high compression rates of the alveolar interface (at the end of expiration), lipids excluded from the interface are organized in three-dimensional structures that remain below. These structures might be more stable and firmly associated to the surface at 33 °C than 37 °C. Indeed, part of those membranes, containing higher DPPC proportions, may detach from the surface at 37 °C. This probably fuels both catabolism and recycling of surfactant lipids under physiological conditions when new surfactant components continuously adsorb and replenish the lost material at the interface. Conversely, the adsorption of surfactant would be compromised by several plasma compounds when they pass from systemic circulation to the alveolar space. This affects the amount of DPPC available to reduce surface tension at 37 °C and suggests a protective role at 33 °C. SP-B surfactant protein B, SP-C surfactant protein C, γ surface tension, AT-I alveolar type I cell, AT-II alveolar type II cell. Full size image In asphyxiated neonates, the improvement in surfactant activity under TH seems to be significant after 48–72 h 25 . At this time point, two sPLA 2 IIA modulators, SP-A and DOPG, might increase and the activity of the enzyme decreases 30 . This may explain the time-dependent effect on surfactant activity under TH, probably related to the higher DPPC amount at the interface. Finally, we demonstrated an increase in both the endogenous and therapeutic surfactant performance under cooling (Figs. 4 , 5 ). Regardless of the material applied (PORα or lipid-protein mixtures containing different proportions of DPPC and POPG), lower doses were more resistant to plasma inhibition and able to restore PS activity at 33 °C compared with 37 °C. Moreover, the resistance to plasma inactivation was higher at high DPPC ratios regardless of the temperature (Fig. 5 ). Altogether, these results are promising for primary ARDS patients. In those patients, the maximum beneficial effect of exogenous surfactant should be warranted with the minimum dose. Our data show a better plasma resistance for the lipid-protein mixture containing ≥ 50% DPPC when compared to PORα at both 33 °C and 37 °C (Fig. 5 ). Similar observations were recently described for CHF 5633 at physiological temperature, a novel synthetic surfactant that shows promising results in clinical phase II 48 , 49 . CHF 5633 composition is similar to the lipid-protein mixtures employed in our study: 50/50 DPPC/POPG (w/w%) and 1.8% by mass of SP-B/SP-C analogs 48 . Thus, the mechanism involved in plasma resistance upon dynamic cycles may probably be analogous and particularly enhanced at 33 °C. In this line, the lipid-protein mixture with increased POPG/DPPC ratio showed a reduced activity at 37 °C, but was still functional at 33 °C. We speculate that plasma proteins are excluded from the interface together with POPG during compression, presumably in a process lead by surfactant proteins/peptides. The same mechanism might occur under TH, with less DPPC lost from the interface during compression (Fig. 7 ). The lower lipid miscibility at 33 °C may contribute to reduce DPPC mixing into the liquid-expanded phase and its interfacial exclusion under compression. This would explain the good performance of the material at low DPPC ratio (35%) (Fig. 5 ). We acknowledge some study limitations. PS undergoes a temperature-mediated collapse under compression in the Langmuir–Blodgett balance 50 . This prevents analysing surfactant interfacial structure and composition at surface pressures higher than 40 mN/m. In addition, we could not perform a lipidomic analysis after compression-expansion cycles since these experiments last around 30 min and the lipids at the interface may oxidize during this period. Also, experiments using lipid-protein mixtures containing only SP-C could not be carried out in the CBS, because the presence of SP-B is indeed essential for the adsorption and re-adsorption of a significant proportion of lipids at the bubble interface 51 . As previously described, surfactant is resistant to plasma inactivation under quick cycles in the CDS 52 . For that reason, the use of slow cycles was necessary to facilitate the competition of plasma proteins for the air–liquid interface. Finally, we cannot exclude that the lower interfacial activity at 37 °C, especially for the mixture containing high percentages of POPG or DOPC, was influenced by a higher lipid oxidation at increased temperatures 53 . Altogether, we demonstrated that TH influences surfactant structure and composition at the air–liquid interface, increasing the interfacial proportion of DPPC upon compression. This leads to improved lung and therapeutic surfactant activity both under physiological and inhibitory conditions. Our evidences should be added to the multiple effects of TH on human lungs. Methods Materials PS and purified SP-B and C were obtained from porcine bronchoalveolar lavage (BAL) as previously reported 15 (Supplementary Fig. S5 ). The experimental procedures are described in the supplementary information. Porcine plasma was obtained by centrifuging (500× g for 10 min at 4 °C) and pooling blood from 6 pigs with different sexes coming from the same slaughterhouse. Similarly, PS was purified from the BAL pool obtained from the 6 pig lungs. Animals used to recover these materials were healthy and subjected to Vet control according to local regulations (Spanish hygiene rules legislation, law articles 83, 85 and 178). Pigs were sacrificed for food and not for the sole purpose of the study. PS and synthetic PLs concentrations were determined by phosphorus mineralization 54 . Plasma total protein (TP) concentration was measured with the Lowry protein assay 55 . PORα is a widely available surfactant obtained by minced porcine lung and considered to be the most efficacious in neonatal critical care, according to evidence based data and European guidelines 56 , 57 . It was purchased from Chiesi Farmaceutici S.p.A. (Parma, Italy), lyophilized and stored at − 80 °C until use. DPPC, POPG, POPC and DOPC were purchased from Avanti Polar Lipids, Inc. (Alabaster, Alabama, USA). BODIPY-PC was purchased from Molecular Probes (Life Technologies, Carlsbad, California, USA). The same buffer solution containing 5 mM Tris and 150 mM NaCl at pH 7.4 was used for every experiment. Biophysical activity and lateral structure of surfactant films The biophysical activity of both PS and lipid-protein surfactant suspensions was tested at limiting concentrations (1.5, 2, 5 and 8 mg/mL) under slow (2 cycles/min) and quick breathing-like (20 cycles/min) compression-expansion cycling by CDS 58 and CBS 51 (Supplementary Fig. S6 ). Surfactant capability to reduce surface tension was not inhibited during quick cycles in the CDS after bulk replacement with serum 52 . Thus, both plasma inhibition and therapy restoration experiments were performed under slow cycles (Supplementary Fig. S6 , Table S2 ). A Langmuir–Blodgett surface balance was used to study PS lateral structure (Supplementary Fig. S8 ) before and after subjecting material (around 30 μg at 5 mg/mL) at the interface to 10 compression-expansion cycles (65 cm 2 /min). All methods were applied keeping the target temperature (33 °C or 37 °C) constant along the experiments. Details about protocols and techniques are described in supplementary information. Composition of interfacial films and thermotropic properties The temperature-dependent changes in lipid composition of PS films under compression at the air–liquid interface were analysed in a Langmuir–Blodgett balance (Supplementary Fig. S8 ). Briefly, PS (around 50 μg, 5 mg/mL), pre-heated to 33 °C or 37 °C, was deposited drop by drop at the air–liquid interface of the trough filled with buffer at the same temperature. After 10 min of equilibration (≈ 25 mN/m surface pressure), surfactant films were compressed, reducing interfacial area at a velocity of 25 cm 2 /min and simultaneously transferred onto a glass slide. Immediately after drying, the glass slide was rinsed with chloroform/methanol (2:1 v/v) to collect lipids from the film. Five experiments per condition were pooled in a single sample for lipidomic analysis by liquid chromatography-high resolution mass spectrometry (LC-HRMS) 43 . Each lipid class was expressed as molar % of the total PLs. DSC was performed on both PS and lipid/protein suspensions, testing materials at 3 mg/mL as described earlier 48 and in supplementary information. Statistics Data were expressed as mean (standard deviation). Results comparisons were conducted using One-Way ANOVA followed by post-hoc Tukey’s test or Two-Way ANOVA followed by post-hoc paired and unpaired T-test when appropriate. Details are described in figure captions. Correlation analysis was performed using Spearman (ρ) coefficient. Analyses were carried out with Sigma Plot (v. 11, Systat Software, San Jose, USA) and IBM SPSS Statistics 25 (v.25, IBM Corp., Armonk, NY, USA). p < 0.05 was considered to be significant. Change history 05 May 2021 A Correction to this paper has been published:

A tenth of all intensive care unit patients worldwide, and many critical patients with COVID-19, have acute respiratory distress syndrome (ARDS). Therapeutic hypothermia, an intentional cooling of the body, has been suggested as a way to improve ARDS. New research by Chiara Autilio and colleagues in the lab of Jesus Perez-Gil at the Complutense University of Madrid shows not only how therapeutic hypothermia works in the lungs at the molecular level, but also why it could be successfully applied to ARDS. Autilio and her colleagues' work was published in Nature Scientific Reports in January 2021 and will be presented on Tuesday, February 23 at the 65th Annual Meeting of the Biophysical Society to be held virtually. Inside our lungs, surfactant is a molecular mixture that is essential for breathing. Premature babies are sometimes born without yet having developed surfactant and require emergency surfactant replacement treatments in order to breathe. But surfactant is also inactivated and broken up in adults with lung injuries or inflammation. Because therapeutic hypothermia, a cooling of the body to about 33°C (or 91°F), has been used to improve breathing for some premature babies and for some kinds of cardiac arrest in adults, and some early studies have shown a benefit for ARDS, Autilio and colleagues wanted to know if cooling could impact surfactant. They looked at the physics of isolated surfactant in their lab, and Autilio said, "unexpectedly, we found an improvement in surfactant activity at 33°C." The team found that at 33°C, the surfactant had lower surface tension, which could make it easier for oxygen to enter the lungs. They also found that the lower tension changed the activity of the molecules in the surfactant, which prevented surfactant from being disrupted by blood molecules, which can occur during lung injury. Their results indicate that "using therapeutic hypothermia could help people with acute respiratory distress syndrome to breathe." There are currently clinical trials underway in other labs, testing therapeutic hypothermia as a treatment for severe breathing problems associated with COVID-19, according to Autilio. And in the Perez-Gil lab at the Complutense University, "we are working to create a surfactant for adults, a surfactant that could work in the context of acute respiratory distress syndrome," says Autilio.

10.1038/s41598-020-79025-3

Medicine

Music training may not make children smarter after all: new study

Cognitive and academic benefits of music training with children: A multilevel meta-analysis, Sala and Gobet, Memory & Cognition 2020. DOI: 10.3758/s13421-020-01060-2

http://dx.doi.org/10.3758/s13421-020-01060-2

https://medicalxpress.com/news/2020-07-music-children-smarter.html

Abstract Music training has repeatedly been claimed to positively impact children’s cognitive skills and academic achievement (literacy and mathematics). This claim relies on the assumption that engaging in intellectually demanding activities fosters particular domain-general cognitive skills, or even general intelligence. The present meta-analytic review ( N = 6,984, k = 254, m = 54) shows that this belief is incorrect. Once the quality of study design is controlled for, the overall effect of music training programs is null ( \( \overline{g} \) ≈ 0) and highly consistent across studies ( τ 2 ≈ 0). Results of Bayesian analyses employing distributional assumptions (informative priors) derived from previous research in cognitive training corroborate these conclusions. Small statistically significant overall effects are obtained only in those studies implementing no random allocation of participants and employing non-active controls ( \( \overline{g} \) ≈ 0.200, p < .001). Interestingly, music training is ineffective regardless of the type of outcome measure (e.g., verbal, non-verbal, speed-related, etc.), participants’ age, and duration of training. Furthermore, we note that, beyond meta-analysis of experimental studies, a considerable amount of cross-sectional evidence indicates that engagement in music has no impact on people’s non-music cognitive skills or academic achievement. We conclude that researchers’ optimism about the benefits of music training is empirically unjustified and stems from misinterpretation of the empirical data and, possibly, confirmation bias. Working on a manuscript? Avoid the common mistakes Introduction It has been claimed that music fosters children’s cognitive skills and academic achievement. Learning to play the violin or the piano, to recognize pitches, and to keep the beat are often presented as effective cognitive enhancement tools (e.g., Jaušovec & Pahor, 2017 ). However, the idea that practicing cognitively demanding tasks may lead to domain-general cognitive enhancement is in stark contrast with empirical evidence in cognitive science and educational psychology. In fact, while human cognition has been shown to be malleable to training, transfer of skills appears to be limited to the training domain and, at best, other similar domains. It is customary to distinguish between two broad categories of transfer: near transfer and far transfer (Barnett & Ceci, 2002 ). Whereas near transfer – i.e., the transfer of skills within the same domain – is sometimes observed, far transfer – i.e., the transfer of skills across two distant domains – is rare or possibly inexistent (Melby-Lervåg, Redick, & Hulme, 2016 , Sala et al., 2019a ). Moreover, when it does occur, transfer of skills is often limited to the degree to which the two domains (source and target) share contents. For example, even transfer of skills within subspecialties of the same discipline seems to be limited. In fact, performance significantly worsens when experts engage in certain subspecialties of their field of expertise. For example, chess masters who are asked to recall or find the best move in positions coming from chess openings that do not fall into their repertoire exhibit a drastic (about 1 SD) reduction in performance (Bilalić, McLeod, & Gobet, 2009 ). This so-called curse of specificity has been recently defined as one of the fundamental particles in the standard model of human cognition (Sala & Gobet, 2019 ). Researchers involved in cognitive training do not deny that between-domain, or even within-domain, transfer is hard to trigger. Nonetheless, they claim that it is possible to induce far transfer by engaging in domain-specific cognitively demanding activities that boost domain-general cognitive skills; those skills, in turn, are supposed to generalize across many different domains (e.g., academic proficiency; Strobach & Karbach, 2016 ). At a neural level, this generalization is thought to be enabled by the activation of shared brain structures that are common to the practiced activity (e.g., music) and other core cognitive skills (e.g., fluid intelligence, working memory, and language; Moreno et al., 2011 ). In other words, domain-general cognitive enhancement and far transfer are believed to be by-products of domain-specific training (Taatgen, 2016 ). With respect to music, three main hypotheses have been formulated to explain why playing it should lead to broad cognitive benefits. To begin with, music might directly impact on general intelligence rather than on some particular cognitive skills (Schellenberg, 2004 ). This idea is consistent with the vast amount of correlational evidence showing that musicians tend to outperform non-musicians in a variety of cognitive tests. Examples include memory (Sala & Gobet, 2017a ; Talamini, Altoè, Carretti, & Grassi, 2017 ), fluid and general intelligence (Ruthsatz, Detterman, Griscom, & Cirullo, 2008 ; Schellenberg, 2006 ), attention (Saarikivi, Putkinen, Tervaniemi, & Huotilainen, 2016 ), and phonological processing (Forgeard et al., 2008 ). The same pattern of results occurs in academic skills. In fact, music skills appear to be related to better reading abilities (Anvari, Trainor, Woodside, & Levy, 2002 ), and music engagement is a predictor of overall academic achievement (Wetter, Koerner, & Schwaninger, 2009 ). Another possible link connecting music engagement and cognitive enhancement might be working memory (WM). Multimodal cognitively demanding activities are thought to strengthen WM capacity (Diamond & Ling, 2019 ; Morrison & Chein, 2011 ), which, in turn, enhances fluid intelligence and learning (Jaeggi et al., 2008 ). Music training is one such activity (Saarikivi, Huotilainen, Tervaniemi, & Putkinen, 2019 ). Simply put, the putative broad benefits of music training would stem from a boost in domain-general WM capacity rather than general intelligence. Finally, music training might positively impact on one’s sound perception and, consequently, phonological processing and even reading skills (Patel, 2011 ; Tierney & Kraus, 2013 ). This hypothesis is upheld by the fact that numerous brain structures and neural patterns are shared by music skills and language (for a review, see Jäncke, 2009 ). Interestingly, improved reading skills may also facilitate the acquisition of new skills and therefore enhance people’s IQ performance (Ritchie & Tucker-Drob, 2018 ). This further mechanism would again be consistent with the overall idea that music training conveys multiple cognitive and academic benefits. Experimental evidence The theories just described imply that music training causes cognitive enhancement and improvement in academic performance. However, correlational evidence gathered in natural groups is not sufficient to establish a causal link. In the last few decades, dozens of experimental trials have been carried out to examine a potential causal link between music training and improved cognitive/academic performance. Researchers in this field have reached inconsistent conclusions. While most of them have expressed optimism about the benefits of music training (e.g., Barbaroux, Dittinger, & Besson, 2019 ; Nan et al., 2018 ; Tierney, Krizman, & Kraus, 2015 ), others have found this enthusiasm unjustified (e.g., Kempert et al., 2016 ; Rickard, Bambrick, & Gill, 2012 ). Like in many other fields in the social sciences, meta-analyses have been carried out to resolve such controversies. The only comprehensive meta-analytic review performed so far about the benefits of music training is that by Sala and Gobet ( 2017b ). This meta-analysis – which includes 38 studies, 118 effect sizes, and 3,085 participants – found an overall effect of \( \overline{d} \) = 0.16. It also highlighted that the impact of music training on cognitive skills and academic performance was a function of the quality of the study’s experimental design. Specifically, the magnitude of the music-induced effects was significantly smaller (around zero) in those studies implementing active controls and random allocation of the participants to groups. Two meta-analyses examined a subset of studies (Cooper, 2019 ; Gordon, Fehd, & McCandliss, 2015 ), and drew somewhat more positive implications for the cognitive and educational benefits of music teaching. Gordon et al. ( 2015 ) reviewed 12 studies ( n = 901) assessing the effects of music training on language-related skills. The overall effect was small but significant ( \( \overline{d} \) = 0.20). Analogously, Cooper ( 2019 ) analysed 21 studies ( n = 1,767) and found an overall effect size of \( \overline{g} \) = 0.28 across several measures of cognitive ability (measures related to academic achievement were not included because they were considered too different from cognitive ability). Interestingly, the effect was maintained in studies employing active controls ( \( \overline{g} \) = 0.21). The present meta-analysis Despite the less than encouraging evidence, dozens of new experimental investigations have been carried out in recent years, including the two largest randomized control trials (RCTs) in this field (Aleman et al., 2017 ; Haywood et al., 2015 ). Once again, the claims about the effectiveness of music training have been inconsistent across studies (e.g., James et al., 2019 ; Lukács & Honbolygó, 2019 ; Nan et al., 2018 ). We thus ran a meta-analysis including both old and new experimental studies to establish (a) which claims are justified, (b) what are the sources of heterogeneity across studies, and (c) which of the theories predicting that music training enhances cognitive and academic skills are corroborated/refuted. Beyond being relatively dated, the previous meta-analyses suffer from several technical limitations. First, no multilevel modeling was employed. Multilevel modeling is necessary to adjust standard errors when a certain degree of statistical dependence is present in the data (i.e., effect sizes nested in studies). Footnote 1 Also, some of the effect sizes were incorrectly calculated because of a mistake in the reporting of the results in one of the primary studies (Rickard et al., 2012 ; personal communication). Both issues probably inflated the amount of between-study true heterogeneity, which tended to bias meta-analytic model estimates. In addition, the presence of a non-negligible amount of unexplained true heterogeneity (as in both Sala & Gobet, 2017b , and Cooper, 2019 ) makes the overall effect sizes hard to interpret because the sources of between-study variability remain hidden. Finally, no thorough sensitivity analysis was performed (e.g., outlier analysis and multiple publication bias analysis). In brief, such suboptimal modeling choices produce biased estimates. The present meta-analytic review aims to correct these problems and to update the findings of the music-training literature. The current meta-analysis also carries out Bayesian analyses that compare the support for the null and alternative hypotheses, and relies on a larger number of studies (19 new studies) and therefore a larger number of participants (an increase from about 3,000 to about 7,000, compared to Sala & Gobet, 2017b ). Since the number of participants and effect sizes are more than double, the current meta-analysis has a much higher power than the 2017 meta-analysis. Method Literature search A systematic search strategy was implemented (Appelbaum et al., 2018 ). Using the following Boolean string (“music” OR “musical”) AND (“training” OR “instruction” OR “education” OR “intervention”), we searched through ERIC, PsycINFO, and ProQuest Dissertation & Theses databases to find studies that reported music training programs. We retrieved 3,044 records. Inclusion criteria Five inclusion criteria were applied: 1) The study was experimental in nature and implemented a cognitively demanding music-training program (e.g., learning to play instruments, Kodály method, etc.). No correlational or ex-post facto studies were included. 2) The study included at least one control group that isolated the variable of interest (i.e., music training). 3) The study included non-music-related cognitive tests or academic outcomes. 4) The study included participants aged between 3 and 16 years with no previous formal music experience or clinical condition. 5) The study reported sufficient data to calculate the effect sizes. Alternatively, the author(s) had to provide the necessary data. We searched for eligible articles through 1 December 2019. When the data reported in the study were insufficient to calculate the effect sizes or important details about the study design were unclear, we contacted the corresponding authors by email ( n = 11). We received three positive replies. We found 54 studies, conducted from 1986 to 2019, that met the inclusion criteria (reported in Appendix A in the Supplemental Online Materials). Nineteen of these studies had never been included in any previous meta-analysis. The studies included 254 effect sizes and a total of 6,984 participants. Thus, compared to the previous most comprehensive meta-analysis in the field (i.e., Sala & Gobet, 2017b ), the number of both effect sizes and participants was more than doubled. The studies originally evaluated for inclusion but eventually excluded are reported in Appendix B in the Supplemental Online Materials. The procedure is described in Fig. 1 . Fig. 1 Flow diagram of the search strategy Full size image Moderators We assessed six moderators based on the previous meta-analyses in the literature: 1) Baseline difference (continuous variable): The standardized mean difference between the experimental and control groups at pre-test. This moderator was added to evaluate the amount of true heterogeneity accounted for by pre-post-test regression to the mean. It thus aimed at ruling out potential confounding effects of this statistical artifact. 2) Randomization (dichotomous variable): Whether the children were randomly allocated to the groups. 3) Type of controls (active or non-active; dichotomous variable): Whether the music training group was compared to another novel activity (e.g., dancing); no-contact groups and business-as-usual groups were treated as “non-active.” This moderator thus controlled for potential placebo effects. 4) Age (continuous variable): The mean age of the study’s participants. A few studies did not report the participants’ mean age. In these cases, the participants’ mean age was obtained from the median or the school grade. 5) Outcome measure: The effect sizes were grouped into four broad groups based on the Cattell-Horn-Carroll taxonomy (McGrew, 2009 ): non-verbal ability (e.g., fluid reasoning [Gf], mathematical skills [Gq], and spatial skills [Gv]); verbal ability (e.g., vocabulary and reading skills [Gc], phonological processing [Grw]); memory (e.g., short-term/working-memory tasks [Gsm]); and speed (e.g., processing speed [Gs] and inhibition tasks [Gt]). The inter-rater agreement was κ = 1. We also examined this moderator without grouping these eight categories into the four groups. Finally, since some primary studies employed academic tests (e.g., Haywood et al., 2015 ), we examined whether the effect sizes related to cognitive skills were greater than those related to academic achievement (as suggested by Cooper, 2019 ). The latter category included all those effect sizes that were obtained from academic tests of literacy and mathematics (a subset of the Gc and Gq groups). All the other effect sizes fell into the cognitive group. 6) Duration of training: The total number of hours, weeks, and sessions of the training program in each study. These three variables were tested separately because they were collinear (i.e., they measured the same construct with three different metrics). Effect size calculation The effect sizes were calculated for each eligible outcome measure in the primary studies. Hedges’s g s – an adjusted standardized mean difference – were calculated with the following formula: $$ g=d\times \left(1-\frac{3}{\left(4\times N\right)-9}\right) $$ (1) with $$ d=\frac{\left({M}_{e\_ post}-{M}_{e\_ pre}\right)-\left({M}_{c\_ post}-{M}_{c\_ pre}\right)}{S{D_{pooled}}_{pre}} $$ (2) where M e_post and M e_pre are the mean of the experimental group at post-test and pre-test, respectively, M c_post and M c_pre are the mean of the control group at post-test and pre-test, respectively, SD pooled_pre is the pooled pre-test SDs in the experimental group and the control group, and N is the total sample size. The sampling error variances were calculated with the following formula: $$ Va{r}_g=\left(\frac{N_e-1}{N_e-3}\times \left(\frac{2\times \left(1-r\right)}{r_{xx}}+\frac{d_e^2}{2}\times \frac{N_e}{N_e-1}\right)\times \frac{1}{N_e}+\frac{N_c-1}{N_c-3}\times \left(\frac{2\times \left(1-r\right)}{r_{xx}}+\frac{d_c^2}{2}\times \frac{N_c}{N_c-1}\right)\times \frac{1}{N_c}\right)\times {\left(1-\frac{3}{\left(4\times N\right)-9}\right)}^2 $$ (3) where r xx is the test-retest reliability of the test, N e and N c are the sample sizes of the experimental group and the control group, respectively, d e and d c are the within-group standardized mean differences of the experimental group and the control group, respectively. Finally, r is the pre-post-test correlation (Schmidt & Hunter, 2015 ; pp. 343–355). The pre-post-test correlations and test-retest coefficients were rarely provided in the primary studies. Therefore, we assumed the reliability coefficient ( r xx ) to be equal to the pre-post-test correlation (i.e., no treatment by subject interaction was postulated; Schmidt & Hunter, 2015 ; pp. 350–351), and we imposed the pre-post-test correlation to be r xx = r = .600. When the study implemented an only-post-test design (i.e., no pre-test assessment) we used the following formulas for effect size and sampling error variance, respectively: $$ g=\frac{M_{e\_ post}-{M}_{c\_ post}}{S{D_{pooled}}_{pre}}\times \left(1-\frac{3}{\left(4\times N\right)-9}\right) $$ (4) $$ Va{r}_g=\frac{N-1}{N-3}\times \frac{4}{N}\times \left(1+\frac{d^2}{8}\right)\times {\left(1-\frac{3}{\left(4\times N\right)-9}\right)}^2 $$ (5) Finally, in a few cases, t - and F -values were used to calculate d (for the details, see the Supplemental Online Materials). Modeling approach Robust variance estimation (RVE) with correlational weights was employed to perform the intercept and meta-regression models (Hedges, Tipton, & Johnson, 2010 ; Tanner-Smith, Tipton, & Polanin, 2016 ). RVE has been designed to model nested effect sizes (i.e., extracted from the same study). Two indexes were used to report the models’ between-cluster true (i.e., not due to random error) heterogeneity: τ 2 , which indicates the absolute amount of true heterogeneity; and I 2 , which indicates the percentage of true heterogeneity. In addition, we manipulated the within-study effect-size correlation ( ρ ) assumed by the RVE models to test the sensitivity of the results to this parameter. We performed these analyses with the Robumeta R package (Fisher, Tipton, & Zhipeng, 2017 ). Publication bias We examined publication bias with two methods: Duval and Tweedie’s ( 2000 ) trim-and-fill analysis and Vevea and Woods’ ( 2005 ) selection models. The trim-and-fill method estimates whether some smaller-than-average effect sizes have been suppressed from the literature and calculates an adjusted overall effect size and standard error. This analysis was conducted after averaging the statistically dependent effects using Cheung and Chan’s ( 2014 ) approach. We employed the L0 and R0 estimators designed by Duval and Tweedie ( 2000 ). Vevea and Woods’ ( 2005 ) selection models estimate publication bias and calculate an adjusted overall effect size (but no standard error) by assigning to p -value ranges different weights. In other words, the method assumes that the probability of an effect not to be suppressed is a function of its p -value. As recommended by Pustejovsky and Rodgers ( 2019 ), the weights used in the publication bias analyses were not a function of the effect sizes (for more details, see Appendices C and D in the Supplemental Online Materials). We performed these analyses with the Metafor R package (Viechtbauer, 2010 ). True heterogeneity and sensitivity analysis Explaining between-study true heterogeneity is one of the main goals of meta-analysis. While small to null true heterogeneity indicates that between-study differences are merely an artifact of random error (Schmidt, 2010 ), large amounts of true heterogeneity suggest that more than one true effect is present in the data. Moreover, true heterogeneity reduces the statistical power of meta-analytic models, tends to artificially inflate overall effect sizes in asymmetric distributions, and sometimes produces biased publication-bias adjusted estimates (Cheung & Chan, 2014 ; Henmi & Copas, 2010 ; Schmidt & Hunter, 2015 ; Stanley, 2017 ). Investigating the sources of true heterogeneity is thus essential to make the results more interpretable and accurate. Therefore, beyond running meta-regression analysis, we performed a two-step sensitivity analysis. First, we excluded three studies that, probably due to lack of random allocation or small sample sizes, reported unusually high between-group differences (≈ 1 SD) in the participants’ baseline IQ (Patscheke, Degé, & Schwarzer, 2019 ; Roden, Kreutz, & Bongard, 2012 ; Roden, Grube, Bongard, & Kreutz, 2014 ). That is, these three studies were included in the main analysis but removed from the sensitivity analysis. Such large baseline differences make any findings hard to interpret and may introduce noise in the data. Second, we ran Viechtbauer and Cheung’s ( 2010 ) influential case analysis. This method evaluates whether some effect sizes exerted an unusually strong influence on the model’s parameters such as the amount of between-study true heterogeneity ( τ 2 ). Those effect sizes that inflated true heterogeneity were excluded. Bayesian analysis A vast quantity of data regarding cognitive-training programs has been collected in the last 15 years. For example, Sala et al.’s ( 2019a ) second-order meta-analysis estimates that more than 20,000 participants have undergone cognitive-training programs such as music training, videogame training, and WM training. This previous evidence can be employed to establish a set of distributional assumptions (informative priors) in the Bayesian framework. The distribution of the effect sizes was assumed to be normal. Based on Sala et al.’s ( 2019a ) second-order meta-analysis, we expected the mean effect size to be null (prior \( \overline{g} \) = 0) in models including active control groups and slightly positive (prior \( \overline{g} \) = 0.150) in models including passive controls groups. The prior for the standard deviation was the same in all the models ( SD g = 0.500). The true heterogeneity parameter ( τ ) was assumed to have a half-Cauchy distribution (centered on 0 and scale γ = 10) in all the models. No further prior was used for other moderators. We thus estimated the Bayes factors ( BF s) for two sets of competing hypotheses for \( \overline{g} \) and τ . First, we compared the alternative hypothesis H1: \( \overline{g} \) ≠ 0 with the null hypothesis H0: \( \overline{g} \) = 0. Second, we compared the alternative hypothesis H1: τ > 0 with the null hypothesis H0: τ = 0. BF s > 1 indicated support for H1, while BF s < 1 indicated support for H0. In line with common guidelines, H1 was considered as substantially supported only if BF > 3 (i.e., H1 three times more likely to be true than H0; e.g., Dougherty, Hamovitz, & Tidwell, 2016 ). Analogously, H0 was substantially supported only if BF < 0.333 (i.e., H0 three times more likely to be true than H1). Since the priors were conditional to the type of controls employed by the primary study (as indicated by Sala et al., 2019a ), these analyses were carried out after running moderator analysis. The analyses were carried out with the bayesmeta R package (Röver, 2017 ). Results Descriptive statistics The mean age of the samples was 6.45 years. The median age was 5.90, the first and third quartiles were 5.03 and 7.85, and the mean age range was 3.50–11.59. The mean Baseline difference was -0.038, the median was 0, the first and third quartiles were -0.210 and 0.141, and the range was -1.058–0.844. The mean duration of training was 53.37 h (range 2.00–507.00, median 30.00), 29.29 weeks (range 3.00–117.00, median 26.00), and 53.43 sessions (range 6.00–195.00, median 30.00). The descriptive statistics of the categorical moderators are reported in Table 1 . Table 1 Number of studies and effect sizes sorted by categorical moderators Full size table Main analyses The overall effect size of the RVE intercept model was \( \overline{g} \) = 0.184, SE = 0.041, 95% confidence interval (CI) [0.101; 0.268], m = 54, k = 254, df = 38.36, p < .001, τ 2 = 0.041, I 2 = 43.16%. Different values of within-study effect-size correlation ( ρ ) did not significantly affect the results ( \( \overline{g} \) range 0.184– 0.185, τ 2 = 0.041). The random-effect (RE) model (with Cheung and Chan’s correction) yielded very similar estimates: \( \overline{g} \) = 0.176, SE = 0.037, p < .001, τ 2 = 0.033. Overall, the results showed a small and moderately heterogeneous overall effect of music training on cognitive and academic outcomes. The results were not affected by modeling choices (i.e., ρ values and procedure for modeling nested data). Baseline difference and Type of controls were the only two statistically significant moderators ( p = .031 and p = .035, respectively) and accounted for part of the true heterogeneity ( τ 2 = 0.038, I 2 = 34.87%). Age was not significant ( p = .403), neither was Allocation ( p = .518). No significant differences were found across the four broad groups of outcome measures (all p s ≥ .624; Holm’s correction for multiple comparisons), nor across the more fine-grained categorization (eight levels, all p s ≥ .362), and there was no difference between cognitive skills and measures of academic achievement ( p = .981). Duration of training was not significant either ( p = .266, p = .952, and p = .662 for hours, weeks, and sessions, respectively). Type of controls Since Type of controls was statistically significant, we performed the analyses on the two sub-samples separately. In those studies that implemented non-active controls, the results showed a small and moderately heterogeneous overall effect of music training on cognitive and academic outcomes. The overall effect size was \( \overline{\boldsymbol{g}} \) = 0.228, SE = 0.045, 95% CI [0.137; 0.320], m = 41, k = 144, df = 30.1, p < .001, τ 2 = 0.042, I 2 = 43.11%. Different values of within-study effect-size correlation (ρ) did not affect the results ( \( \overline{\boldsymbol{g}} \) = 0.228, τ 2 = 0.042). The RE model provided similar results, \( \overline{\boldsymbol{g}} \) = 0.201, SE = 0.041, p < .001, τ 2 = 0.023. Again, the results were not affected by modeling choices. Also, some evidence of a small publication bias was found. The trim-and-fill retrieved no missing study with the L0 estimator. Five missing studies were retrieved with the R0 estimator, and the adjusted estimate was \( \overline{\boldsymbol{g}} \) = 0.170, 95% CI [0.064; 0.276]. Vevea and Woods’ ( 2005 ) selection model calculated a similar estimate ( \( \overline{\boldsymbol{g}} \) = 0.119). Finally, the Bayes factors confirmed these findings. BF g was greater than 730,000, indicating that \( \overline{\boldsymbol{g}} \) was far more likely to be non-null (H1: \( \overline{\boldsymbol{g}} \) ≠ 0) than null (H0: \( \overline{\boldsymbol{g}} \) = 0). Regarding the model’s true heterogeneity, BF τ was greater than 5,000, again indicating that τ was far more likely to be positive than null. In those studies that implemented active controls, the results showed a near-zero and slightly heterogeneous overall effect of music training on cognitive and academic outcomes. The overall effect size was \( \overline{g} \) = 0.056, SE = 0.058, 95% CI [-0.069; 0.182], m = 23, k = 110, df = 12.6, p = .350, τ 2 = 0.025, I 2 = 23.10%. Different values of within-study effect-size correlation ( ρ ) did not significantly affect the results ( \( \overline{g} \) range 0.054–0.057, τ 2 range 0.023–0.025). The results were robust to the type of modeling approach employed. In fact, the RE model provided similar results, \( \overline{g} \) = 0.090, SE = 0.060, p = .136, τ 2 = 0.032. Some evidence of a small publication bias was found, suggesting that the unbiased overall effect size is essentially null. No missing study was retrieved with the L0 estimator, whereas the R0 estimator identified four missing studies and the adjusted estimate was \( \overline{g} \) = -0.020, 95% CI [-0.183; 0.142]. The selection model estimate was \( \overline{g} \) = 0.039. The Bayes factors were BF g = 1.231 and BF τ = 0.044. These results showed that – as indicated by the publication-bias-corrected estimates – \( \overline{g} \) was not convincingly more likely to be non-null than null ( BF g < 3), and that τ was approximately 23 times more likely to be null than positive. The latter finding confirms that the low observed true heterogeneity ( τ 2 = 0.025, I 2 = 23.10%) is very likely to be spurious. Sensitivity analyses This section replicated the analyses after excluding the three studies reporting large baseline IQ differences across the groups and implemented Viechtbauer and Cheung’s ( 2010 ) influential case analysis to explain the model’s residual true heterogeneity (if any). The overall effect size of the RVE intercept model was \( \overline{g} \) = 0.166, SE = 0.041, 95% CI [0.083; 0.249], m = 51, k = 235, df = 34.9, p < .001, τ 2 = 0.036, I 2 = 40.62%. Different values of within-study effect-size correlation ( ρ ) did not significantly affect the results ( \( \overline{g} \) range 0.165–0.166, τ 2 range 0.035–0.036). The random-effect (RE) model provided similar estimates: \( \overline{g} \) = 0.149, SE = 0.035, p < .001, τ 2 = 0.024. Baseline difference and Type of controls were again the only two statistically significant moderators ( p = .017 and p = .003, respectively) and accounted for part of the true heterogeneity ( τ 2 = 0.029, I 2 = 29.70%). Therefore, the results were pretty much the same as in the main analyses so far. Non-active controls When non-active controls were used, the overall effect size was \( \overline{g} \) = 0.226, SE = 0.045, 95% CI [0.133; 0.319], m = 40, k = 139, df = 29.2, p < .001, τ 2 = 0.041, I 2 = 42.96%. Different values of within-study effect-size correlation ( ρ ) did not significantly affect the results ( \( \overline{g} \) = 0.226, τ 2 = 0.041). The RE model provided similar results, \( \overline{g} \) = 0.200, SE = 0.041, p < .001, τ 2 = 0.024. Five effect sizes were found to be significantly inflating the true heterogeneity. After excluding these effect sizes, the overall effect size was \( \overline{g} \) = 0.181, SE = 0.042, 95% CI [0.093; 0.268], m = 39, k = 134, df = 21.9, p < .001, τ 2 = 0.018, I 2 = 24.92%. Similar results were obtained with the RE model, \( \overline{g} \) = 0.161, SE = 0.037, p < .001, τ 2 = 0.013. Finally, in order to investigate the sources of the unexplained true heterogeneity ( τ 2 = 0.018, I 2 = 24.92%), a moderator analysis was run. Randomization was the only statistically significant moderator ( p = .042) and explained nearly all the true heterogeneity ( τ 2 = 0.005, I 2 = 7.61%). Therefore, the observed true between-study heterogeneity in the studies employing non-active controls was accounted for by a few extreme effect sizes and the type of allocation of participants to the groups. For non-randomized studies, the overall effect sizes were \( \overline{g} \) = 0.246, SE = 0.049, 95% CI [0.140; 0.352], p < .001; for randomized studies, the relevant statistics were \( \overline{g} \) = 0.064, SE = 0.065, 95% CI [-0.116; 0.244], p = .381. Thus, when random allocation was employed, the overall effect was near-zero. Publication bias analysis: Studies without randomization With the studies that did not implement any randomization of participants’ allocation, the trim-and-fill analysis retrieved two missing studies with the L0 estimator (adjusted estimates \( \overline{g} \) = 0.211, 95% CI [0.095; 0.328]). Three missing studies were retrieved with the R0 estimator (adjusted estimates \( \overline{g} \) = 0.189, 95% CI [0.068; 0.310]). Vevea and Woods’ ( 2005 ) selection model calculated a more conservative estimate ( \( \overline{g} \) = 0.126). Thus, a small amount of publication bias was still detected. The Bayes factors were BF g = 217.840 ( \( \overline{g} \) far more likely to be non-null than null) and BF τ = 0.021 ( τ nearly 50 times more likely to be null than positive). While confirming that the overall effect size of music training in non-randomized samples and passive controls is positive (yet small), these results showed that no between-study true heterogeneity was present in the data. Publication bias analysis: Studies with randomization Regarding the randomized samples, all the publication bias analyses estimated a substantially null overall effect. The trim-and-fill analysis retrieved six and ten studies with the L0 and R0 estimators, respectively (adjusted estimates \( \overline{g} \) = 0.009, 95% CI [-0.095; 0.113] and \( \overline{g} \) = -0.034, 95% CI [-0.131; 0.063]). Vevea and Woods’ ( 2005 ) selection model yielded a similar estimate ( \( \overline{g} \) = -0.002). The Bayes factors were BF g = 0.257 and BF τ = 0.025. Therefore, the Bayes factors provided compelling evidence that both \( \overline{g} \) and τ are more likely to be null than non-null (approximatively 4 and 40 times, respectively). Active controls Turning our attention to the studies implementing active controls, the overall effect size was \( \overline{g} \) = -0.021, SE = 0.032, 95% CI [-0.109; 0.068], m = 20, k = 96, df = 4.2, p = .558, τ 2 = 0, I 2 = 0%. Different values of within-study effect-size correlation ( ρ ) did not affect the results ( \( \overline{g} \) = -0.021, τ 2 = 0). The RE model provided similar results, \( \overline{g} \) = -0.010, SE = 0.035, p = .787, τ 2 = 0. Since this model showed no true heterogeneity and null overall effects, no publication bias analysis was performed. The Bayes factors largely favored the null hypothesis ( BF g = 0.063 and BF τ = 0.006). The null hypothesis was approximatively 16 times and 180 times more likely than the alternative hypothesis for \( \overline{g} \) and τ , respectively. In brief, all the analyses showed that the overall effect in studies implementing active controls is null and homogeneous across studies (i.e., \( \overline{g} \) = 0, τ 2 = 0). Discussion This meta-analytic review investigated the impact of music training on children’s cognitive skills and academic achievement. The overall impact of music training programs on cognitive and academic outcomes is weak and moderately heterogeneous ( \( \overline{g} \) = 0.184, SE = 0.041, τ 2 = 0.041, I 2 = 43.16%). The inspection of true heterogeneity shows that there is an inverse relationship between the studies’ design quality and magnitude of the effect sizes. Specifically, those studies using active controls or implementing random assignment report homogeneous null or near-zero effects ( \( \overline{g} \) = -0.021–0.064, τ 2 ≤ 0.005). Conversely, a small overall effect size is observed in those studies employing neither active controls nor random assignment ( \( \overline{g} \) = 0.246). The results of the Bayesian analyses corroborate the conclusions that the unbiased effect of music training on cognitive and academic skills is null and highly consistent across the studies (i.e., \( \overline{g} \) = 0 and τ 2 = 0). No other study features (e.g., age, duration of training, and outcome measure) seem to have any influence on the effect sizes – not even the outcome measures. In particular, contrary to Cooper’s ( 2019 ) hypothesis, there was no difference between cognitive skills and academic achievement (literacy and mathematics), which means that it is justifiable to pool the two outcomes together, as was done for example in Sala and Gobet ( 2017b ). Altogether, these results indicate that music training fails to produce solid improvements in all the examined cognitive and academic skills equally. Finally, only a low amount of publication bias is observed in the models (about 0.100 standardized mean difference at most), which is in line with the near-zero effect sizes estimated. The results are summarized in Table 2 . Table 2 Overall effects in the meta-analytic models Full size table These findings confirm and extend the conclusions of the previous comprehensive meta-analysis in the field (Sala & Gobet, 2017b ). Along with re-establishing the fundamental role of design quality in affecting the experimental results, the present meta-analysis has succeeded in explaining all the observed true heterogeneity. We can thus conclude that these findings convincingly refute all the theories claiming that music training causes improvements in any domain-general cognitive skill or academic achievement (e.g., Moreno et al., 2011 ; Patel, 2011 ; Saarikivi et al., 2019 ; Tierney & Kraus, 2013 ). In fact, there is no need to postulate any explanatory mechanism in the absence of any genuine effect or between-study variability. In other words, since there is no phenomenon, there is nothing to explain. More broadly, these results establish once again that far transfer – due to the very nature of human cognition – is an extremely rare occurrence (Gobet & Simon, 1996 ; Sala & Gobet, 2019 ). Beyond meta-analytic evidence It is worth noting that other researchers have reached the same conclusions using different methodologies. To begin with, Mosing, Madison, Pedersen, and Ullén ( 2016 ) have investigated the relationship between music training and general intelligence in twins. Notably, music-trained twins do not possess a higher IQ than non-music-trained co-twins. This study thus suggests that engaging in music has no effect on people’s IQ. Swaminathan, Schellenberg, and Khalil ( 2017 ) show that music aptitude, rather than the amount of music training, predicts fluid intelligence in a sample of adults. This finding upholds the idea that the correlation between intelligence and engagement in music is mediated by innate (as opposed to trained) music skills. Similarly, Swaminathan, Schellenberg, and Venkatesan ( 2018 ) demonstrate that the correlation between amount of music training and reading ability in adults disappears when domain-general cognitive skills are controlled for. These findings corroborate the hypothesis according to which the observed correlation between music training and particular domain-general cognitive/academic skills is a byproduct of previous abilities. Once pre-existing differences in overall cognitive function are ruled out, the correlation disappears (Swaminathan & Schellenberg, 2019 ). Therefore, there is no reason to support the hypothesis that music training boosts cognition or academic skills. Rather, all the evidence points toward the opposite conclusion, that is, that the impact of music training on cognitive and academic skills is null. Finally, the failure of music-training regimens to induce any generalized effect is mirrored by findings in other cognitive-training literatures. For instance, WM training does not enhance children’s domain-general cognitive skills or academic achievement (Aksayli, Sala, & Gobet, 2019 ; Melby-Lervåg et al., 2016 ; Sala & Gobet, 2020 ). The same applies to action and nonaction videogame training and brain training (Duyck & Op de Beeck, 2019 ; Kassai, Futo, Demetrovics, & Takacs, 2019 ; Libertus et al., 2017 ; Lintern & Boot, 2019 ; Sala et al., 2019a ; Sala, Tatlidil, & Gobet, 2018 , 2019b ; Simons et al., 2016 ). The perception of music training effectiveness is biased It is our conviction that, while the data show a consistent picture, the narrative that has been built around music training is substantially distorted. For example, Schellenberg ( 2019 ) has shown how correlational evidence is often used by scholars to incorrectly infer causal relationships between engagement in music and non-music outcomes. Correlation is notoriously insufficient to establish causal links between variables, which makes Schellenberg’s ( 2019 ) findings quite concerning. Interestingly, this problem appears to be particularly severe in neuroscientific studies. The overall interpretation of the results reported in the primary studies is another example of the extent to which authors sometimes misread the empirical evidence presumably supporting music training. For instance, Barbaroux et al.’s ( 2019 ) study does not implement any type of controls, which makes their results uninterpretable. Tierney et al. ( 2015 ) report non-significant and inconsistent effects on language-related outcomes between a music training group and an active control group. However, this study is not experimental because the participants were recruited after they had chosen what activity to take part in (i.e., self-selection of the sample). (This is why, incidentally, this study is not included in the present meta-analysis.) Despite representing very little evidence in favor of a causal link between music training and improved cognitive/academic skills, the study has gained a considerable amount of attention in news outlets and among researchers in the field (top 5% in Altmetric). In the same vein, Nan et al. ( 2018 ) have found no significant effect of music training on any two music-related measures and no effect at all on the examined non-music outcome measures. (The paper reports a barely significant effect [ p = .044] in an auditory task that is obtained with an ANOVA performed on the mean pre-post-test gains. This is a well-known incorrect practice that inflates Type I error rates.) Therefore, this study corroborates the idea that the impact of music training on cognitive/academic skills is slim to null. Nonetheless, both the authors and several news outlets provide an over-optimistic, if not utterly incorrect, view of the benefits of music training (e.g., McCarthy, 2018 ). By contrast, the two largest randomized controlled trials in the field have been either somewhat ignored (Aleman et al., 2017 ) or nearly completely overlooked (Haywood et al., 2015 ) by researchers involved in music training (and news outlets). Both studies report no effect of music training on any cognitive or academic skills. Neither of them makes any overstatement about the benefits of music training on any domain-general cognitive or academic skill. It is thus apparent that if all the results are considered and correctly interpreted, the whole music training literature depicts a very consistent mosaic. What is mixed is how the same findings are described by different scholars (Schmidt, 2017 ). Conclusions and recommendations for future research This meta-analysis has examined the experimental evidence regarding the impact of music training on children’s non-music cognitive skills and academic achievement. The ineffectiveness of the practice is apparent and highly consistent across studies. Moreover, recent correlational studies have confirmed that music engagement is not associated with domain-general cognitive skills or academic performance. Two alternative potential avenues involving music activities may be worth some interest. First, music may be beneficial for non-cognitive constructs in children such as prosocial behavior and self-esteem (e.g., Aleman et al., 2017 ). These possible advantages are not likely to be specific to music, though. In fact, any enticing and empowering activity may improve children’s well-being. Second, elements of music instruction (e.g., arithmetical music notation) could be used to facilitate learning in other disciplines such as arithmetic (Azaryahu, Courey, Elkoshi, & Adi-Japha, 2019 ; Courey, Balogh, Siker, & Paik, 2012 ; Ribeiro & Santos, 2017 ). Too few studies have been conducted to reach a definite conclusion. Nonetheless, this approach is undoubtedly more likely to succeed than the music training programs reviewed in this meta-analysis. In fact, while the latter program regimens have tried and failed to reach cognitive enhancement via music training, the former methodology tries to convey domain-specific knowledge by focusing on domain-specific information. This type of near transfer is notoriously much easier to achieve (Gobet, 2016 ; Gobet & Simon, 1996 ). Notes Here, the adjective “multilevel” broadly refers to any technique that allows the researcher to correctly model multivariate data (i.e., effect sizes) nested within studies. We do not intend to imply that the specific modeling method used in the present meta-analysis is superior to other methods. Rather, we simply highlight the necessity of adopting multilevel/multivariate techniques in order to produce accurate results.

Music training does not have a positive impact on children's cognitive skills, such as memory, and academic achievement, such as maths, reading or writing, according to a study published in Memory & Cognition. Previous research trials, carried out to examine a potential causal link between music training and improved cognitive and academic performance, have reached inconsistent conclusions, with some suggesting that there may be a link between music training and better cognitive and academic performance and others finding little effect. Researchers Giovanni Sala at Fujita Health University, Japan and Fernand Gobet at the London School of Economics and Political Science, UK examined existing experimental evidence regarding the impact of music training on children's non-music cognitive skills and academic achievement. The authors re-analyzed data from 54 previous studies conducted between 1986 and 2019, including a total of 6,984 children. They found that music training appeared to be ineffective at enhancing cognitive or academic skills, regardless of the type of skill (such as verbal, non-verbal, speed-related and so on), participants' age, and duration of music training. When comparing between the individual studies included in their meta-analysis, the authors found that studies with high-quality study design, such as those which used a group of active controls—children who did not learn music, but instead learned a different skill, such as dance or sports—showed no effect of music education on cognitive or academic performance. Small effects were found in studies that did not include controls or which did not randomize participants into control groups (ones that received different or no training) and intervention groups (ones that received music training). Giovanni Sala, the lead author said: "Our study shows that the common idea that 'music makes children smarter' is incorrect. On the practical side, this means that teaching music with the sole intent of enhancing a child's cognitive or academic skills may be pointless. While the brain can be trained in such a way that if you play music, you get better at music, these benefits do not generalize in such a way that if you learn music, you also get better at maths. Researchers' optimism about the benefits of music training appears to be unjustified and may stem from misinterpretation of previous empirical data." Fernand Gobet, the corresponding author added: "Music training may nonetheless be beneficial for children, for example by improving social skills or self-esteem. Certain elements of music instruction, such as arithmetical music notation could be used to facilitate learning in other disciplines." The authors caution that too few studies have been conducted to reach a definitive conclusion about possible positive effects of music education on non-academic or cognitive characteristics. Alternative potential avenues involving music activities may be worth exploring.

10.3758/s13421-020-01060-2

Medicine

Study signals need to screen genes for stem cell transplants

Human pluripotent stem cells recurrently acquire and expand dominant negative P53 mutations, Nature (2017). nature.com/articles/doi:10.1038/nature22312 Journal information: Nature

http://nature.com/articles/doi:10.1038/nature22312

https://medicalxpress.com/news/2017-04-screen-genes-stem-cell-transplants.html

Abstract Human pluripotent stem cells (hPS cells) can self-renew indefinitely, making them an attractive source for regenerative therapies. This expansion potential has been linked with the acquisition of large copy number variants that provide mutated cells with a growth advantage in culture 1 , 2 , 3 . The nature, extent and functional effects of other acquired genome sequence mutations in cultured hPS cells are not known. Here we sequence the protein-coding genes (exomes) of 140 independent human embryonic stem cell (hES cell) lines, including 26 lines prepared for potential clinical use 4 . We then apply computational strategies for identifying mutations present in a subset of cells in each hES cell line 5 . Although such mosaic mutations were generally rare, we identified five unrelated hES cell lines that carried six mutations in the TP53 gene that encodes the tumour suppressor P53. The TP53 mutations we observed are dominant negative and are the mutations most commonly seen in human cancers. We found that the TP53 mutant allelic fraction increased with passage number under standard culture conditions, suggesting that the P53 mutations confer selective advantage. We then mined published RNA sequencing data from 117 hPS cell lines, and observed another nine TP53 mutations, all resulting in coding changes in the DNA-binding domain of P53. In three lines, the allelic fraction exceeded 50%, suggesting additional selective advantage resulting from the loss of heterozygosity at the TP53 locus. As the acquisition and expansion of cancer-associated mutations in hPS cells may go unnoticed during most applications, we suggest that careful genetic characterization of hPS cells and their differentiated derivatives be carried out before clinical use. Main Somatic mutations that arise during cell proliferation and are then subject to positive selection are major causes of cancer and other diseases 6 . Acquired mutations are often present in just some of the cells in a sample, and can therefore be detected in next generation sequencing data from their presence at allelic fractions less than 50% 5 , 7 . We reasoned that the analysis of sequencing data from hES cells might reveal previously unappreciated mosaic mutations and mutation-driven expansions acquired during hES cell culture. This approach would complement previous studies describing culture-derived chromosomal-scale aneuploidies and megabase-scale copy number variants (CNVs) in hPS cells 1 , 8 , 9 . To this end, we collected and performed whole-exome sequencing (WES) of hES cell lines that were listed on the registry of hES cell lines maintained by the US National Institutes of Health (NIH) ( Fig. 1a, b ) and were able to obtain, bank and sequence 114 independent hES cell lines ( Fig. 1c–e ). We selected cell lines at low to moderate passage (P) numbers (mean P18, range P3–P37) and cultured them in a common set of growth conditions for an average of 2.7 ± 0.7 (± s.d.) passages (range 2–6 passages) before banking and sequencing ( Fig. 1f, g ). As hES-cell-derived differentiated cells are currently being evaluated in clinical trials for their safety and utility in a range of diseases such as macular degeneration 10 , we also obtained genomic DNA from an additional 26 independent hES cell lines that had been prepared under good manufacturing practice (GMP) conditions for potential clinical use ( Fig. 1a, c, e, g ). We performed WES of these 140 hES cell lines from 19 institutions to a mean read depth of 79.7 ± 0.1 (± s.e.m.) (range, 57 to 115 for UM4-6 to UM78-2) ( Fig. 1h ). Further details on cell line acquisition and selection are in Supplementary Table 1 and Methods . Figure 1: Acquisition and WES of 140 hES cell lines. a , Schematic workflow for hES cell line acquisition and sequencing. b , c , 114 hES cell lines were obtained, banked ( b ), and analysed by WES along with 26 GMP-prepared cell lines ( c ). d , 45 hES cell lines were excluded owing to use restrictions. e , 140 hES cell lines were banked and/or sequenced (see also Supplementary Table 1 and Methods ). f , HES cells were minimally cultured before banking and sequencing. g , Cumulative passage number of hES cells was moderate. h , WES coverage for sequenced hES cell lines. IRB, institutional review board; MTA, material transfer agreement; PGD, pre-implantation genetic diagnosis. PowerPoint slide Full size image To identify acquired mutations, we analysed the sequencing reads at high-quality, high-coverage heterozygous sites across the exome. To eliminate most inherited polymorphisms from consideration, we restricted this search to variants found in only 1–2 of the cell lines sequenced and in fewer than 0.01% of the individuals sampled by the Exome Aggregation Consortium (ExAC Database) 11 . The allelic fractions at which most remaining variants were represented among the sequence reads for any one hES cell DNA sample followed a binomial distribution, reflecting statistical sampling around the 50% level expected of inherited alleles ( Fig. 2a ); at a much smaller set of sites, variant alleles were present at lower fractions ( Fig. 2a, b ). We applied a statistical test to identify variants for which the observed allelic fractions were unlikely ( P < 0.01 by binomial test) to have been generated by random sampling of two equally present alleles. This search identified 263 candidate mosaic variants, of which 28 were predicted to have a damaging or disruptive effect on gene function ( Supplementary Table 2 ). Figure 2: Identification of recurrent, cancer-associated TP53 mutations in hES cells. a , Some heterozygous variants are present at low allelic fractions (boxed left) in hES cells. b , c , Likely mosaic variants ( P < 0.01, binomial test, red shading), include six mutations in TP53 ( b , Supplementary Table 3 ) that are rare in ExAC (<0.0001) ( c ). d , The four affected P53 residues are commonly mutated in human tumours. e , On a crystal structure of P53 bound to DNA, the affected residues map to the DNA binding domain and include arginine (R) residues that directly interact with DNA. f , The residues mutated in hES cells disrupt DNA binding by P53. PowerPoint slide Full size image The only gene affected by more than one such mutation was the tumour suppressor gene TP53. We identified six TP53 mutations in five unrelated hES cell lines ( Supplementary Table 3 ), including a GMP-prepared cell line (MShef10) that carried two distinct TP53 variants (G245S and R248W). These six missense mutations, although rare (<0.01%) in the general population 11 ( Fig. 2c ), mapped to the four residues most frequently disrupted in human cancer 12 , 13 , 14 ( Fig. 2d , Supplementary Table 3 ). As P53 is mutated in approximately 50% of tumours 15 , coding mutations in these four residues are associated with a substantial fraction of human cancer disease burden. Each of the six mutations involved a cytosine residue of a CpG dinucleotide and may therefore involve a highly mutable site 16 . On a crystal structure of the human P53 protein in complex with DNA, each of the mutations mapped to the DNA-binding domain of P53 (ref. 17 ) ( Fig. 2e, f ). Mutations at these positions are associated with cancer and act in a dominant negative fashion to diminish P53-mediated regulation of apoptosis, cell cycle progression, and genomic stability 18 . Individuals with germ-line mutations at these residues develop Li–Fraumeni syndrome, an autosomal dominant disease with a lifetime cancer risk of nearly 100% 19 . In these patients, tumours can arise at any age and can affect most tissues, including the brain, bones, lung, skin, soft tissues, adrenal gland, colon, stomach, and blood 20 . To independently test the hypothesis that the inactivating TP53 mutations were acquired, we developed droplet digital PCR (ddPCR) assays to count the abundance of each allele at the four TP53 mutation sites ( Fig. 3a, b , Supplementary Table 4 ). Analysis of genomic DNA derived from the 140 hES cell lines confirmed that all six mutations identified by WES were indeed mosaic, with allelic fractions ranging from 7–40%, suggesting their presence in 14–80% of cells in culture ( Fig. 3c ). We did not identify additional cell lines carrying mutations at these positions, suggesting that such mutations were either absent or present at allelic fractions below the sensitivity of the assay (approximately 0.1%) 21 . These findings demonstrate that each of the TP53 mutations identified in hES cells was an acquired mutation and that cells with the mutation had come to represent a significant fraction of cells in affected lines. Figure 3: TP53 mutations in hES cells are mosaic and confer strong selective advantage. a , ddPCR assay schematic. b , Representative ddPCR data showing droplets containing the reference allele (grey), mutant allele (red), both alleles (pink), or neither allele (black). c , Estimated fraction of mutant cells (red) in affected hES cell lines. d , Mutant allelic fraction rapidly increases during standard hES cell culture. Error bars depict s.e.m. and numbers indicate replicate wells. Similar results were observed in replicate experiments ( Extended Data Fig. 1 ). Note further allele-fraction expansion (after P17) for WA26, probably involving LOH. e , Model of the role of P53 in both cancer and stem cell biology. PowerPoint slide Full size image We next asked whether the cells harbouring these TP53 mutations expanded their representation within the hES cell population across passages. We re-obtained early passage vials for hES cell lines that were mosaic for TP53 mutations (CHB11 at P22, and WA26 at P13), thawed a fresh vial of ESI035 at P36, and analysed the genomic DNA from the frozen vial and at each subsequent passage to test for changes in mutant allelic fraction. In each of the three hES cell lines, TP53 mutant alleles increased in representation over passages ( Fig. 3d ) in all but one experiment, suggesting that TP53 mutations conferred a strong selective advantage (approximately 1.9-fold per passage) under routine culture conditions ( Fig. 3e , Extended Data Fig. 1 , Supplementary Table 5 ). To confirm that this selective advantage was conferred by TP53 mutations and not by CNVs at chr20q11.21 (refs 1 , 2 , 3 ), we analysed all 140 hES cell lines using single nucleotide polymorphism (SNP) arrays and found that none of the lines carrying TP53 mutations also carried the chr20q CNV ( Supplementary Table 6 ). Our results are consistent with a model in which the routine culture of hPS cells selects for mutations that inactivate P53, resulting in the rapid clonal expansion of such mutations when they occur. Indeed, it has previously been reported that loss of P53 activity facilitates the reprogramming of somatic cells to pluripotency 22 , 23 and promotes hPS cell survival and proliferation 24 , suggesting a prominent role for P53 in regulating self-renewal in hPS cells ( Fig. 3e ). To test the reproducibility of our observations and to explore the effects of P53 mutations in additional contexts, we screened for TP53 mutations in publically available RNA sequencing (RNA-seq) data from 251 hPS cell samples in 57 published studies, corresponding to 13 hES cell and 104 human induced PS cell (hiPS cell) lines ( Fig. 4a–c , Supplementary Table 7 ). The relatively high expression of TP53 in hPS cells provided sufficient read depth for allelic counting and allowed us to identify nine instances of eight disparate point mutations in TP53 . These eight variants were all distinct at the nucleotide level from those we had previously seen by WES. However, like the mutations ascertained by WES, each of these variants led to missense substitutions in the DNA-binding domain of P53 ( Fig. 4d–f , Supplementary Table 3 , Extended Data Fig. 2 ). When we considered both WES and RNA-seq datasets, we identified four codons that were recurrently mutated in hPS cells: R181, G245, R248, and R273. Notably, the commonly used WA09 (H9) hES cell line manifested four distinct TP53 mutations (P151S, R181H, R248Q, and R267W) in different laboratories, further demonstrating that the mutations arose during cell culture ( Fig. 4h ). Figure 4: A substantial fraction of hPS cells in published studies harbour TP53 mutations. a – e , Published RNA-seq data show that 7 out of 117 (6%) unique hPS cell lines harbour P53 mutations. f , Combined DNA-seq and RNA-seq analysis reveals 12 out of 252 (5%) distinct cell lines affected by 15 TP53 mutations ( Supplementary Table 3 ). g – i , P53 mutant WA01 was seen in three studies ( g ), WA09 acquired four distinct TP53 mutations in three groups ( h ), and WIBR3 lost all normal copies of TP53 after gene editing ( i ). j , k , TP53 mutant cells could be differentiated ( j , k ), and expanded relative to wild-type cells ( k ). Values are the mean of 2–3 replicate samples, and error bars depict s.e.m. l , m , Model of TP53 mutation enrichment during hPS cell culture ( l ) or during clonal bottlenecks ( m ). Organ., gut organoid; NEC, neuroectodermal cell; Terato., Teratoma. PowerPoint slide Full size image Of the 15 instances of TP53 mutations observed by either WES or RNA-seq, the percentage of mutant reads suggested that 10 were mosaic and that 3 had reached fixation (allelic fraction of 50%). Surprisingly, TP53 mutations in two cell lines, WA09 (R248Q) and WIBR3 (H193R), were present in 80% ± 3% and 100% of reads, respectively ( Supplementary Table 3 ). These findings were consistent with the excess allelic fraction observed during the culture of WA26 ( Fig. 3d ) and suggested the presence of additional mutational mechanisms affecting mutant TP53 allelic fraction. Indeed, we observed loss of heterozygosity (LOH) of a large telomeric domain along chromosome 17 including the TP53 locus ( Extended Data Fig. 3 ) that was almost complete in a gene-targeted derivative of WIBR3 ( Fig. 4i ) and was partial in WA09, consistent with the observed high fraction (80%) of mutant TP53 reads. These results suggest that follow-on LOH after an initial TP53 point mutation confers additional selective advantage. We next determined whether TP53 mutations affect cell differentiation or affect the survival of differentiated cells. To this end, we examined studies in which there was RNA-seq data for both hES cells and their differentiated progeny. Cell lines with substantial fractions of TP53 mutant cells could readily form teratomas, gut epithelial cells 25 ( Fig. 4i ), neuroepithelial cells 26 ( Fig. 4j ), and pancreatic polyhormonal cells 27 ( Fig. 4k ). Notably, a mosaic G245C mutation in TP53 expanded in allelic fraction over the course of differentiation 27 , suggesting a continued selective advantage in differentiating mutant cells. Together, our analyses indicate that researchers have unknowingly and routinely used hPS cells that harbour cancer-related missense mutations in TP53 , sometimes accompanied by LOH. These findings have practical implications for the use of hES cells in disease modelling and transplantation medicine. The fact that we observed TP53 mutations among both hES cells and hiPS cells cultured with a wide variety of media, substrates, and passaging methods ( Extended Data Fig. 4 ) suggests that new culture conditions should be explored to reduce the selective pressure for TP53 mutations. We also suggest regular genetic testing of hPS cells, particularly before and after stressful interventions such as gene editing or single-cell cloning that force hPS cell populations through bottlenecks ( Fig. 4l, m ). Our specific findings here suggest that the P53 pathway could be an immediate focus for these genetic tests. A comprehensive ascertainment of recurrent culture-acquired mutations will require the analysis of still-larger collections of stem cell lines by both exome and whole-genome sequencing. Our findings also demonstrate that sequencing can detect potentially harmful mutations in differentiated cell preparations derived from hPS cells, providing an opportunity to increase the safety of cell replacement therapies for conditions ranging from diabetes to Parkinson’s disease. Clinical trials with hPS-cell-derived materials had recently been halted owing to the discovery of undisclosed mutations 28 , though such trials have since resumed. We join others 29 in suggesting that hPS cells and their derivatives be subjected to genome-wide analyses at several key steps: during initial cell line selection; as part of the characterization of a master bank of hPS cells; and as an end-stage release criterion before the transplantation of the hPS cell-derived cellular product. Importantly, although TP53 mutations recurred at detectable fractions in several cell lines, most lines (around 95%) were free of detectable TP53 mutations despite having spent extensive time in culture. Regenerative medicine remains a viable and exciting goal that is more likely to succeed as potential pitfalls, like the one we report here, are identified and addressed. Methods Data reporting No statistical methods were used to predetermine sample size. All cells were tested for mycoplasma and included for analysis only upon testing negative. The identity of all cell lines was confirmed by whole-exome sequencing and SNP array analysis. The experiments were not randomized and the investigators were not blinded to allocation during experiments and outcome assessment. hES cell acquisition As a source of hES cells for this study, we focused on those that had been voluntarily listed by research institutions on the registry of hES cell lines maintained by the US National Institutes of Health (NIH) ( ). As of 8 July 2015, a total of 307 hES cell lines were listed on this registry. Of these, we requested viable frozen stocks of the 182 lines annotated to be available for distribution and to lack known karyotypic abnormalities or disease-causing mutations. During our effort to obtain these cell lines, we found that 45 were subject to overly restrictive material transfer agreements that precluded their use in our studies and 11 could not be readily obtained as frozen stocks owing to differences in human subjects research regulations between the US and the UK. Nine cell lines were unavailable upon request or were overly difficult to import, and three could not be cultured despite repeated attempts. Further details on the availability of cell lines can be found in Supplementary Table 1 . The generation of hES cells used in this study was previously approved by the institutional review boards (IRBs) of all providing institutions. Use of the hES cells for sequencing at Harvard was further approved and determined not to constitute Human Subjects Research by the Committee on the Use of Human Subjects in Research at Harvard University. hES cell culture A protocol for the adaptation of hES cell lines from diverse culture conditions can be found at Protocol Exchange 30 . In brief, we considered that different laboratories employ different methods to culture hES cells, raising the question of how best to thaw and culture the cell lines we obtained from multiple sources. Traditionally, hES cells are maintained on gelatinized plates and co-cultured with replication-incompetent mouse embryonic fibroblast (MEF) feeder cells in tissue culture medium containing knockout serum replacement (KOSR). More recently, hES cells have been cultured on a substrate of cell-line-derived basal membrane proteins known by the trade names of Matrigel (BD Biosciences) or Geltrex (Life Technologies), in mTeSR1 (ref. 31 ), E8 (ref. 32 ) or similar in the absence of feeder cells. In previous work, we found that a medium containing an equal volume of KOSR-based hES cell medium (KSR) and mTeSR1 (STEMCELL Technologies) (KSR–mTeSR1) robustly supports the pluripotency of hES cells undergoing antibiotic selection during the course of gene-targeting experiments under feeder-free conditions 33 . To minimize stress to hES cells previously cultured and frozen under diverse conditions, cell lines were thawed in the presence of 10 μM Y-27632 (DNSK International) into two wells of a 6-well plate, one of which contained KSR–mTeSR1 on a substrate of Matrigel, and the other containing KOSR-based hES cell medium on a monolayer of irradiated MEFs. After 24 h, Y-27632 was removed and cells were fed daily with the aforementioned media in the absence of any antibiotics. All cultures were tested for the presence of mycoplasma and cultured in a humidified 37 °C tissue culture incubator in the presence of 5% CO 2 and 20% O 2 . Colonies of cells with hES cell morphology and with a diameter of approximately 400 μm were transferred into KSR–mTeSR1 medium containing 10 μM Y-27632 on a substrate of Matrigel by manual picking under a dissecting microscope. Cells with differentiated morphology were removed from plates by aspiration during feeding. Once cultures consisting of cells with homogeneous pluripotent stem cell morphology had been established, they were passaged by brief (2–10 min) incubation in 0.5 mM EDTA in PBS followed by gentle trituration in KSR–mTeSR1 medium containing 10 μM Y-27632 and re-plating. Once cultures had reached approximately 90% confluence in one well of a six-well plate, they were passaged with ETDA onto a Matrigel-coated 10 cm plate. Upon reaching approximately 90% confluence, cell lines were dissociated with EDTA as described above and banked for later use in cryoprotective medium containing 50% KSR–mTeSR1, 10 μM Y-27632, 10% DMSO, and 40% fetal bovine serum (HyClone). A subset of hES cell lines ( Supplementary Table 1 ) were passaged enzymatically with TrypLE Express (Life Technologies), expanded onto two 15 cm plates, and frozen down in 25 cryovials. Whole-exome sequencing and genotyping Cell pellets of approximately 1–5 million cells were generated from banked cryovials of research-grade hES cell lines, or were obtained directly from institutions providing GMP-grade hES cell lines. Cell pellets were digested overnight at 50 °C in 500 μl lysis buffer containing 100 μg ml −1 proteinase K (Roche), 10 mM Tris (pH 8.0), 200 mM NaCl, 5% w/v SDS, 10 mM EDTA, followed by phenol:chloroform precipitation, ethanol washes, and resuspension in 10 mM Tris buffer (pH 8.0). Genomic DNA was then transferred to the Genomics Platform at the Broad Institute of MIT and Harvard for Illumina Nextera library preparation, quality control, and sequencing on the Illumina HiSeq X10 platform. Sequencing reads (150 bp, paired-end) were aligned to the hg19 reference genome using the BWA alignment program. Genotypes from WES data for the cell lines were computed using best practices from GATK software 34 compiled on 31 July 2015. Sequencing quality and coverage were analysed using Picard tool metrics. Cross sample contamination was estimated using VerifyBamID (v1.1.2) 35 , and none was detected. Data from each cell line were independently processed with the HaplotypeCaller walker and further aggregated with the CombineGVCFs and GenotypeGVCFs walkers to generate a combined variant call format (VCF) file. Genotyped sites were finally filtered using the ApplyRecalibration walker. To determine whether lines with or without acquired TP53 mutations showed other chromosomal aberrations or smaller regional changes in copy number, additional genotyping of the 140 hES cell lines was performed using a custom high density SNP array (‘Human Psych array’) that contains more than half a million SNPs across the genome. CNVs larger than 500 kb were identified using the PennCNV (v1.0.0) 36 tool ( ). All CNVs were manually reviewed and are shown in Supplementary Table 6 . Mosaic variant analysis To identify candidate mosaic variants, a table of heterozygous variants was generated from the VCF ( Supplementary Table 2 ). To limit the frequency of false positive calls due to sequencing artefacts and PCR errors, variants were included if they had a variant read depth of at least 10, if they were either flagged as a ‘PASS’ site or were not reported in the Exome Aggregation Consortium (ExAC) database 11 , and if they were not located in regions of the genome with low sequence complexity, common large insertions and segmental duplications, as described by Genovese and colleagues 5 . Multiallelic sites were split, left-aligned, and normalized. The resulting list of 2.1 million ‘high-quality heterozygous variants’ was further refined to include sites that were covered by at least 60 unique reads and had a high confidence variant score (‘PASS’) as ascertained by GATK’s Variant Quality Score Recalibration software (840,222 variants). To exclude common inherited variants, we selected variants present in less than 0.01% of the (ExAC) control population and restricted our analysis to only singleton or doubleton variants (9,490 variants present in 1–2 of the 140 samples). Coverage was calculated by summing reference and alternate allele counts for each variant. Allelic fraction was calculated by dividing the alternate allele count by the total read coverage (both alleles) of the site. Although the allelic fraction of inherited heterozygous variants is expected to be 50%, reference capture bias (a tendency of hybrid selection to capture the reference allele more efficiently than alternative alleles) causes the actual expected allele fraction for SNPs and indels to be closer to 45% and 35%, respectively 5 . To account for these technical biases, we used a binomial test with a null model centred at 45% allelic fraction for inherited SNPs and 35% for inherited indels. Variants for which this binomial test was nominally significant ( P < 0.01) were deemed to be candidate mosaic variants. The nominal P -value threshold of 0.01 was chosen as an inclusive threshold in order to screen sensitively for potentially mosaic variants, at the expense of also capturing false positives for which low allelic fractions represented statistical sampling fluctuations. For this reason, we considered it important to further evaluate putative mosaic variants by independent molecular methods that deeply sample alleles at the nominated sites ( Fig. 3 ). A more stringent computational screen based on a P -value threshold of 1 × 10 −7 identified three of the six TP53 variants, and TP53 was also the only gene with multiple putatively mosaic variants in this screen. We also identified all high quality heterozygous variants that passed the inclusive statistical threshold of ( P < 0.01) in our binomial test and could potentially be mosaic ( n = 36,396). These data are included in Supplementary Table 2 . Variant annotation was performed using SnpEff with GRCh37.75 Ensembl gene models. Variants with moderate effect were classified as damaging by a consensus model based on seven in silico prediction algorithms 37 . Assessment of TP53 mutation frequency in cancer We turned to the ExAC database 11 that compiles the whole-exome sequences of over 60,000 individuals to assess the frequency at which the amino acid residues we observed to be mutated in some hES cells were affected in the general population. We then consulted the COSMIC 12 ( ), ICGC 13 ( ), and IARC P53 (ref. 14 ) ( ) databases and plotted the percentage of tumours carrying a mutation in each codon ( Fig. 2d , Extended Data Fig. 2b ). Molecular modelling of P53 protein To visualize the spatial location of the amino acid residues affected by TP53 mutations observed in hES cells by WES on the P53 protein, we downloaded the 1.85 Angstrom X-ray diffraction-based structure file from the Research Collaboratory for Structural Bioinformatics Protein Data Bank (file 2AHI) and built the model protein/DNA system (chain IDs D, G, and H) to visualize the secondary structure of a P53 monomer complexed to DNA as a ribbon diagram. DNA was illustrated as a space-filling model. Water molecules were discarded when building the wild-type model and minimized in two steps using the AMBER 16 package 38 . Affected residues were indicated as space-filling model superimposed on the ribbon diagram of P53 and highlighted in blue (wild-type) or red (mutated) without consideration of how the mutations might affect the secondary or tertiary structure of the protein. Measurement of TP53 variant allele fraction by ddPCR We assayed the allelic fraction of the four distinct TP53 mutations identified by WES ( Supplementary Table 3 ) in the 140 hES cell lines by droplet digital PCR (ddPCR). Each ddPCR analysis incorporated a custom TaqMan assay (IDT). Assays were designed with Primer3Plus and consisted of a primer pair and a 5′ fluorescently labelled probe (HEX or FAM) with 3′ quencher (Iowa Black with Zen) for either the control (reference) or mutant (alternative) base for each identified P53 variant ( Supplementary Table 4 ). Genomic DNA from each hES cell line was analysed by ddPCR according to the manufacturer’s protocol (BioRad). The frequency of each allele for a given sample was estimated first by Poisson correction of the endpoint fluorescence reads 21 . These corrected counts were then converted to fractional abundance estimates of the mutant allele and multiplied by two to determine the fraction of cells carrying the variant allele. Longitudinal hES cell culture and calculation of TP53 mutation expansion To assess how the allelic fraction of TP53 mutations might change over time in culture, hES cell lines CHB11 (passage 22 or 25), WA26 (passage 13 or 15), and ESI035 (passage 36 in two separate experiments) were serially passaged in mTeSR1 media (STEMCELL Technologies) at a density of approximately 30,000 cells cm −2 in the presence of 10 μM Y-27632 on the day of passaging. Cells were fed daily with mTeSR1 and passaged with Accutase (Innovative Cell Technologies Inc.) at approximately 90% confluence. To monitor changes in allelic fractions, genomic DNA from cells at the indicated passages were analysed by ddPCR. To calculate the relative expansion rate of mutant relative to wild-type cells, we applied the following formula: where R 0 is defined as the ratio of (variant positive cells)/(variant negative cells) after some number of starting passages and R 1 and R 2 represent the aforementioned ratios measured on the same sample at T 1 and T 2 > T 1 passages respectively. From this equation, the estimation of variant positive cells after T passages from starting ratio R 0 can be defined as R 0 e gT . Note that this equation estimates the relative growth rate of cells carrying the variant allele with a round of passaging as unit of time, with both relative survival and growth being incorporated. These data are included in Supplementary Table 5 . For the subsequent calculation of the earliest passage at which these mutations might have become detectable, the detection thresholds ( R 0 ) for WES and ddPCR was assumed to be 0.1 (10 / 100 reads) and 0.001 (1 per 1,000 droplets), respectively. RNA sequencing analysis In order to identify TP53 mutations in hPS cells, we analysed 256 publicly available high-throughput RNA sequencing samples of hPS cells from the SRA database 39 ( ). Data accession numbers for SRA (and GEO, where applicable) are provided in Supplementary Table 7 . 5 of these 256 samples were not considered further as they were from single cells rather than cell lines. Following sequence alignment to the hg19 human reference genome with Tophat2 (ref. 40 ), single nucleotides divergent from the reference genome were identified using GATK HaplotypeCaller 34 . As sufficient sequencing depth is required to deduce sequence mutation, a threshold of 25 reads per nucleotide was set. Under this criterion, 43 samples (40 hES cell lines and 3 hiPS cell lines) had a missense mutation in TP53 . 10 of the 40 hES cell samples (WA09) carried two separate mutations ( Supplementary Table 7 ). Upon the identification of cell lines carrying mutant reads, RNA sequencing data from studies containing differentiated samples were included for analysis. Loss of heterozygosity analysis In order to evaluate TP53 alleles, we assessed the level of polymorphism by calculating the ratio between the minor and major alleles across chromosome 17. So as to minimize sequencing noise and errors, we included SNPs covered by more than 10 reads and that are located in the dbSNP build 142 database 41 . The resulted wig files were then plotted using Integrative Genomics Viewer (IGV) 42 ( Extended Data Fig. 4 ). In order to quantify the difference in polymorphism between samples, we converted the wig files to BigWig using UCSC Genome Browser utilities 43 and summed the allelic ratios between the distal part of the short arm of chromosome 17 (17p), the proximal side of this arm and the long arm of chromosome 17 (17q). The allelic ratio sum was then divided by the region’s length (bp), which resulted in the proportion of SNPs, followed by one-sided Z -score test for two population proportion to compare between the chromosome 17 areas within each sample. Whereas most samples with mutations in TP53 showed a comparable, non-significant rate of polymorphic sites along the chromosome, WIBR3 samples with H193R mutations and WA09 samples with both P151S and R248Q mutations had a significantly different proportion ( P < 0.001) of polymorphic sites, in the distal part of the short arm of the chromosome (first 16 × 10 6 base pairs), including the TP53 site. Unlike the three mutant WIBR3 samples, the wild-type WIBR3 sample had a normal distribution of polymorphic sites with no significant difference between the short and long arms. Data availability Sequence data from cell lines listed on the NIH hES cell registry have been deposited in the NCBI database of Genotypes and Phenotypes (dbGaP) under accession number phys001343.v1.p1 (at ). Sequence data from the remaining cell lines reported in our study have been deposited at the European Genome-phenome Archive (EGA), which is hosted by the EBI and the CRG, under accession number EGAS00001002400 (at ). Change history 04 May 2017 New references 29 and 30 were added, and subsequent citations were renumbered accordingly.

Regenerative medicine using human pluripotent stem cells to grow transplantable tissue outside the body carries the promise to treat a range of intractable disorders, such as diabetes and Parkinson's disease. However, a research team from the Harvard Stem Cell Institute (HSCI), Harvard Medical School (HMS), and the Stanley Center for Psychiatric Research at the Broad Institute of MIT and Harvard has found that as stem cell lines grow in a lab dish, they often acquire mutations in the TP53 (p53) gene, an important tumor suppressor responsible for controlling cell growth and division. Their research suggests that genetic sequencing technologies should be used to screen for mutated cells in stem cell cultures, so that cultures with mutated cells can be excluded from scientific experiments and clinical therapies. If such methods are not employed it could lead to an elevated cancer risk in those receiving transplants. The paper, published online in the journal Nature on April, 26, comes at just the right time, the researchers said, as experimental treatments using human pluripotent stem cells are ramping up across the country. "Our results underscore the need for the field of regenerative medicine to proceed with care," said the study's co-corresponding author Kevin Eggan, an HSCI Principal Faculty member and the director of stem cell biology for the Stanley Center. Eggan's lab in Harvard University's Department of Stem Cell and Regenerative Biology uses human stem cells to study the mechanisms of brain disorders, including amyotrophic lateral sclerosis, intellectual disability, and schizophrenia. The research, the team said, should not discourage the pursuit of experimental treatments but instead be heeded as a call to screen rigorously all cell lines for mutations at various stages of development as well as immediately before transplantation. "Our findings indicate that an additional series of quality control checks should be implemented during the production of stem cells and their downstream use in developing therapies," Eggan said. "Fortunately, these genetic checks can be readily performed with precise, sensitive, and increasingly inexpensive sequencing methods." With human stem cells, researchers can recreate human tissue in the lab. This enables them to study the mechanisms by which certain genes can predispose an individual to a particular disease. Eggan has been working with Steve McCarroll, associate professor of genetics at Harvard Medical School and director of genetics at the Stanley Center, to study how genes shape the biology of neurons, which can be derived from these stem cells. McCarroll's lab recently discovered a common, precancerous condition in which a blood stem cell in the body acquires a pro-growth mutation and then outcompetes a person's normal stem cells, becoming the dominant generator of his or her blood cells. People in whom this condition has appeared are 12 times more likely to develop blood cancer later in life. The study's lead authors, Florian Merkle and Sulagna Ghosh, collaborated with Eggan and McCarroll to test whether laboratory-grown stem cells might be vulnerable to an analogous process. "Cells in the lab, like cells in the body, acquire mutations all the time," said McCarroll, co-corresponding author. "Mutations in most genes have little impact on the larger tissue or cell line. But cells with a pro-growth mutation can outcompete other cells, become very numerous, and 'take over' a tissue. We found that this process of clonal selection - the basis of cancer formation in the body - is also routinely happening in laboratories." To find acquired mutations, the researchers performed genetic analyses on 140 stem cell lines—26 of which were developed for therapeutic purposes using Good Manufacturing Practices, a quality control standard set by regulatory agencies in multiple countries. The remaining 114 were listed on the NIH registry of human pluripotent stem cells. "While we expected to find some mutations in stem cell lines, we were surprised to find that about five percent of the stem cell lines we analyzed had acquired mutations in a tumor-suppressing gene called p53," said Merkle. Nicknamed the "guardian of the genome," p53 controls cell growth and cell death. People who inherit p53 mutations develop a rare disorder called Li-Fraumeni Syndrome, which confers a near 100 percent risk of developing cancer in a wide range of tissue types. The specific mutations that the researchers observed are "dominant negative" mutations, meaning, when present on even one copy of P53, they are able to compromise the function of the normal protein, whose components are made from both gene copies. The exact same dominant-negative mutations are among the most commonly observed mutations in human cancers. "These precise mutations are very familiar to cancer scientists. They are among the worst P53 mutations to have," said Sulagna Ghosh, a co-lead author of the study. The researchers performed a sophisticated set of DNA analyses to rule out the possibility that these mutations had been inherited rather than acquired as the cells grew in the lab. In subsequent experiments, the Harvard scientists found that p53 mutant cells outperformed and outcompeted non-mutant cells in the lab dish. In other words, a culture with a million healthy cells and one p53 mutant cell, said Eggan, could quickly become a culture of only mutant cells. "The spectrum of tissues at risk for transformation when harboring a p53 mutation include many of those that we would like to target for repair with regenerative medicine using human pluripotent stem cells," said Eggan. Those organs include the pancreas, brain, blood, bone, skin, liver and lungs. However, Eggan and McCarroll emphasized that now that this phenomenon has been found, inexpensive gene-sequencing tests will allow researchers to identify and remove from the production line cell cultures with concerning mutations that might prove dangerous after transplantation. The researchers point out in their paper that screening approaches to identify these p53 mutations and others that confer cancer risk already exist and are used in cancer diagnostics. In fact, in an ongoing clinical trial that is transplanting cells derived from induced pluripotent stem cells (iPSCs), gene sequencing is used to ensure the transplanted cell products are free of dangerous mutations.

nature.com/articles/doi:10.1038/nature22312

Earth

New insight into the relationship between slow slip events and the build-up and release of tectonic strain

Hiroki Kawabata et al, Strain accumulation and release associated with the occurrence of SSEs in the subduction zones of the Japanese Islands, Scientific Reports (2023). DOI: 10.1038/s41598-023-28016-1 Journal information: Scientific Reports

https://dx.doi.org/10.1038/s41598-023-28016-1

https://phys.org/news/2023-02-insight-relationship-events-build-up-tectonic.html

Abstract We investigated the spatiotemporal changes in strain associated with the occurrence of slow slip events (SSEs) in the subduction zones of the Japanese Islands and compared the spatial distribution of both the amount of strain accumulated for the period before and during the SSEs release using time series data from the Global Navigation Satellite System (GNSS). In this study, four SSEs were analysed: the Tokai long-term SSE (2000–2005), the Boso-Oki short-term SSE (2007), and the Bungo Channel long-term SSEs (2009–2011 and 2018–2019). As a result, we found strong negative correlations for all four dilatations before and during SSE occurrence. For these dilatations, we estimated the amount of strain released at the time of occurrence of the SSE relative to that accumulated during the period prior to the SSE. The result indicates that not all the accumulated strain before the SSEs was released when the SSEs occurred. Moreover, it is highly likely that the strain released by SSE is not only due to the strain accumulation just below the SSE occurrence region, but also due to the strain accumulation on the shallower plate boundary, which is a seismogenic zone for a future megathrust earthquake. Introduction Interplate earthquakes are considered to occur when the accumulated strain at the plate interface between a continental plate and a subducting oceanic plate exceeds a certain threshold, resulting in the release of the accumulated strain. Crustal deformation caused by the interaction between oceanic and continental plates is active in the Japanese Islands. Sagiya et al. 1 investigated the strain rates (principal strain, dilatation, and maximum shear strain) of the Japanese Islands during 1997–1999 after removing coseismic steps and found that many regions are subject to contraction. In southwestern Japan, the oceanic Philippine Sea (PHS) plate is subducting in a northwesterly direction beneath the continental Amurian plate along the Suruga Trough to the Nankai Trough (Fig. 1 ). On the plate boundary, in addition to ordinary interplate earthquakes, interplate aseismic slow slip events, known as long-term slow slip events (hereafter referred to as L-SSEs), take place in the Tokai region and the Bungo Channel at durations of several months to several years. Figure 1 Tectonic map of the study area. The black barbed lines represent the plate boundary, and the triangles indicate the direction of motion of the subducting plate with respect to the plate being subducted (Bird 28 ). The white arrows represent the velocity vectors of the Philippine Sea plate with respect to the Amurian plate (DeMets et al. 22 ). The map was created by using the Generic Mapping Tools (GMT) 27 (version: GMT4.5.7, URL link: ). Full size image L-SSEs with durations of several years have occurred in the Tokai region from 2000 to 2005 and from 2013 to 2016 and in the Bungo Channel from 1996 to 1997, 2003 to 2004, 2009 to 2011, and 2018 to 2019. Ozawa et al. 2 and Miyazaki et al. 3 reported southeastward slip as the L-SSE, which is distinct from the northwestward steady crustal movement, in the Tokai region from 2000 to 2005. Kobayashi and Hashimoto 4 found strain changes in the area east of Lake Biwa from the latter half of 2000, which is considered to be due to the L-SSE in the Tokai region. Ozawa et al. 5 , 6 , Yoshioka et al. 7 , Nakata et al. 8 , Seshimo and Yoshioka 9 , and Agata et al. 10 estimated the slip distribution of the L-SSEs in the Bungo Channel that occurred from 2009 to 2011 and from 2018 to 2019. Short-term SSEs (hereafter referred to as S-SSEs), with a duration of several tens of days, have also occurred off the Boso Peninsula, where the PHS plate is subducting beneath the North American plate along the Sagami Trough with recurrence intervals of several years. According to Ozawa et al. 11 , Hirose et al. 12 , and Fukuda 13 , displacements caused by the Boso-Oki S‒SSEs were observed in the Boso Peninsula for approximately 10 days from approximately 10 August 2007. The Boso-Oki S‒SSE was included in this study because the displacement changes identified in GNSS time series data are more distinct than those of S-SSEs in other regions, and its recurrence interval is well known (e.g., Fukuda 13 ). Although many studies on L-SSEs and S-SSEs in the Japanese Islands have been conducted, a comprehensive relationship between the accumulation and release of strain before and during the occurrence of SSEs, respectively, has yet to be conducted. Strain accumulation and release are fundamental phenomena in seismology, make it very important to investigate them. One of the major characteristics of SSEs is that the interval between the end of the previous SSE and the onset of the next SSE is well known. This makes it possible to directly compare the amount of strain accumulated before the SSE with the amount of strain released during the SSE. On the other hand, although the earthquake cycles of large earthquakes are generally considered to range from several decades to several thousand years, GNSS data are currently available for only at most 25 years. Therefore, it is difficult to know in detail the accumulation of strain during the period between earthquakes and to directly compare the amount of strain accumulated before and that released by a large earthquake. Therefore, in this study, for these SSEs, we calculated the strain fields before and during the occurrence of SSEs and investigated the relationship between the accumulation and release of strain caused by one cycle of the SSEs based on their spatial distribution and the amount of strain accumulation and release using GNSS time series data. Results The Tokai L-SSE (1 July 2000–30 June 2005), the Boso-Oki S‒SSE (9 August 2007–23 August 2007), and the two Bungo Channel L-SSEs (24 November 2009–6 February 2011 and 20 April 2018–26 May 2019) were the four SSEs analysed in this study. The former and the latter Bungo Channel L-SSE are hereafter referred to as the Bungo Channel L-SSE (1) and the Bungo Channel L-SSE (2), respectively. We selected the Tokai L-SSE and the Boso-Oki S‒SSE that occurred before the Tohoku earthquake (Mw 9.0) (11 March 2011) to avoid the influence of its postseismic effects. The Bungo Channel L-SSEs (1) and (2) are selected because many observation stations are available. For the Tokai L-SSE and the Bungo Channel L-SSE (1), (2), we used the results of previous studies (Suito and Ozawa 14 ; Yoshioka et al. 7 ; Seshimo and Yoshioka 9 ) as a reference, dividing one year into every 0.1 year to estimate each L-SSE. We estimated the period from the onset to the end of the L-SSE. Since the occurrence time of the Boso-Oki S‒SSE is very short compared to other L-SSEs, we determined the onset and end dates of the S-SSEs, by checking the analysis results of Fukuda 13 and the actual GNSS time series data. The number of observation stations used for each SSE and the respective analysis periods before and during the occurrence of SSE are tabulated in Table S1 . The spatial distributions of GNSS stations used in the analysis of the four SSEs are shown in Fig. S1 a–d. In this study, the displacement fields for the four SSEs were obtained by GNSS time series analysis, and various strain fields (dilatation, maximum shear strain, and principal strain) were then obtained by the method of Shen et al. 15 as described in the “ Methods ” section. For dilatation and maximum shear strain, correlation coefficients between the amount of strain accumulated before the SSE and the amount of strain released during the occurrence of SSE were calculated. Calculation points were the only metric used in spots where the amount of displacement during the SSE exceeded a threshold (Table S2 ). The threshold values for each SSE were obtained using the trial-and-error method, such that the negative correlation of dilatation, which specifically represents the accumulation and release of strain, is large. The distribution of calculation points used to calculate the correlation coefficient for each SSE is shown in Fig. S2 a–d. For both dilatation and maximum shear strain, the stronger the negative correlation is, the greater the amount of strain release during the occurrence of SSE at locations where a large amount of strain accumulation before the SSE took place. This suggests that there is a strong relationship between strain accumulation and strain release. In the following subsections, we show the results of the Bungo Channel L-SSE (1), (2). As for the results of the Tokai L-SSE and the Boso-Oki S‒SSE, please refer to Text S1 and S2 , respectively in Supplementary Information. Bungo channel L-SSE (1) The displacement, dilatation, maximum shear strain, and principal strain during the Bungo Channel L-SSE (1), which occurred from 25 November 2009 to 6 February 2011, are shown in Fig. 2 a–d for the period before the L-SSE and are shown in Fig. 2 e–h during the occurrence of the L-SSE. GNSS stations mainly in the Shikoku region show a northwestward displacement field before the L-SSE, while those in the Kyushu region show a gradual spatial change to the southwest. This may be partly due to the spatial change in plate interactions along the Nankai Trough (Sagiya et al. 1 ). Figure 2 Spatial distributions of displacement, dilatation, maximum shear strain, and principal strain before (7 August 2004–24 November 2009) and during (25 November 2009–6 February 2011) the Bungo Channel L-SSE (1). ( a )–( d ) represent before the L-SSE, and ( e )–( h ) denote during the L-SSE. ( a ), ( e ) Spatial distribution of displacement. Black arrows show horizontal displacements during the analysis period. The ellipse at the tip of the black arrow represents the 1σ error ellipse. ( b ), ( f ) Spatial distribution of dilatation. The blue and red areas represent contraction and expansion, respectively. ( c ), ( g ) Spatial distribution of maximum shear strain. The green and red areas show compressive and tensile fields, respectively. ( d ), ( h ) Spatial distribution of principal strain. The black outwards and inwards arrows show tension and compression, respectively. The map was created by using the Generic Mapping Tools (GMT) 27 (version: GMT3.4.6, URL link: ). Full size image The dilatation showed a characteristic contraction and expansion before and during the L-SSE, respectively, approximately 32.8° N on the south side of the Bungo Channel, and the maximum change in dilatation during the L-SSE occurrence was \(8.2 \times 10^{ - 7}\) . For the maximum shear strain, negative and positive values are defined in this study. Positive values represent lateral deformation when the tensile component is dominant, while negative values denote lateral deformation when the compressive component is dominant. The region with large maximum shear strain was identified near the same region for dilatation during the L-SSE, and the values shifted from negative before the L-SSE to positive during the L-SSE. However, the location of the large strain release during the L-SSE moved slightly east to the west of Cape Ashizuri from the location of the large strain accumulation before the L-SSE. At the time of the L-SSE, the positive values were larger near the west side of Cape Ashizuri than the location where the expansion was identified in the dilatation. With regard to principal strain, before the L-SSE, compression in the northwest‒southeast direction was dominant from western Shikoku to the Bungo Channel, but during the L-SSE, tension in the north‒northeast-south‒southwest direction was identified at approximately 32.8° N, where large changes in the dilatation and maximum shear strain can be seen. At other locations, no characteristic principal strain was observed. Comparing the amount of strain accumulated before the L-SSE and the amount of strain released during the L-SSE, the latter was generally smaller. Next, correlation coefficients were obtained for dilatation and maximum shear strain before and during the L-SSE (Fig. 3 a and b). Fig. S2 c shows the 48 calculation points used with a displacement of 2.5 cm or greater during the L-SSE. The correlation coefficient for dilatation was − 0.82, which is a strong negative correlation, while that for maximum shear strain was − 0.20, showing that the correlation was not as negative as that for dilatation. In the correlation diagram for the maximum shear strain (Fig. 3 b), a positive correlation can be seen in the upper left area. This is because the release of maximum shear strain during the L-SSE occurred slightly east of where the maximum shear strain had accumulated before the L-SSE, as described above. As a result, the maximum shear strain did not show a stronger negative correlation than the dilatation. In conclusion, the Bungo Channel L-SSE (1) indicates that as for dilatation, the release of strain during the L-SSE is greater at locations where a greater accumulation of strain is identified before L-SSE. Figure 3 Correlation diagram between the amount of accumulated strain before the Bungo Channel L-SSE (1) (7 August 2004–24 November 2009) and the amount of change in strain during the L-SSE occurrence (25 November 2009–6 February 2011). The values of dilatation and the maximum shear strain at the calculation points shown in Fig. S2 c are used. The straight line represents the linear regression line. ( a ) Dilatation. ( b ) Maximum shear strain. Full size image Bungo channel L-SSE (2) The displacement, dilatation, maximum shear strain, and principal strain for the Bungo Channel L-SSE (2) that occurred from 20 April 2018 to 26 May 2019 are shown in Fig. 4 a–d for the period before the L-SSE and in Fig. 4 e–h during the L-SSE. The GNSS stations in western Shikoku show the northwesterly displacement field, while those in the southeastern Kyushu region show a gradual spatial change to the southwesterly direction, as in the previous L-SSE (1). The displacements for the accumulation period of the L-SSE (2) (Fig. 4 a) were larger than those of the L-SSE (1) (Fig. 2 a) because of the longer analysis period before L-SSE (2). During L-SSE (2), both the direction and magnitude of displacement were similar to those of L-SSE (1). Figure 4 Spatial distributions of displacement, dilatation, maximum shear strain, and principal strain before (7 February 2011–19 April 2018) and during (20 April 2018–26 May 2019) the Bungo Channel L-SSE (2). The expressions of the figures are the same as those in Fig. 2 a–h. The map was created by using the Generic Mapping Tools (GMT) 27 (version: GMT3.4.6, URL link: ). Full size image The results of the dilatation of L-SSE (2) were also similar to those of L-SSE (1), with characteristic contraction and expansion before and during L-SSE at approximately 32.8° N, respectively, and a maximum amount of dilatation of approximately \(8.8 \times 10^{ - 7}\) during the occurrence of L-SSE. The region with a large maximum shear strain was identified at the time of the L-SSE (2), and the values turned from negative before the L-SSE to positive during the L-SSE. As in the case of L-SSE (1), the location of the large strain release during L-SSE (2) shifted slightly east of the location of the large strain accumulation before L-SSE. At the time of the L-SSE (2), positive values on the west side of the Bungo Channel were larger than positive values around the Bungo Channel. However, the location of the spatial boundary between the positive area and negative area on the west side of Cape Ashizuri was slightly shifted to the west during L-SSE (2) than during L-SSE (1). The compressive principal strain was dominant in the northwest‒southeast direction before L-SSE (2), whereas during L-SSE, extension in the north‒northeast-south‒southwest direction was dominant. Comparing the amount of accumulated strain before the L-SSE with the amount of strain released during the L-SSE, the latter was smaller than the former. Figure 5 a and b show the correlations for dilatation and maximum shear strain, respectively, for L-SSE (2). Fig. S2 d shows 37 used calculation points with a displacement of 2.8 cm or greater during the L-SSE. As with the L-SSE (1), the dilatation showed a strong negative correlation with a correlation coefficient of − 0.84. That of the maximum shear strain is − 0.59, indicating a negative correlation similar to that of the Boso-Oki S‒SSE (Text S2 ). However, Fig. 5 b does not show a clear downwards trend to the right, as in the case of dilatation. As with L-SSE (1), this is because the location of the large maximum shear strain moved slightly to the east at the time of the L-SSE compared to that before the L-SSE, and the relationship between the accumulation and release of maximum shear strain is considered to be small. These results indicate that for the Bungo Channel L-SSE (2), a large amount of expansion was released when the L-SSE occurred, where the accumulation of contraction was also large before the occurrence of L-SSE in the dilatation. Figure 5 Correlation diagram between the amount of strain accumulated before the Bungo Channel L-SSE (2) (7 February 2011–19 April 2018) and the amount of change in strain during the L-SSE occurrence (20 April 2018–26 May 2019). The values of the dilatation and the maximum shear strain at the calculation points shown in Fig. S2 d are used. The expressions of the figures are the same as those in Fig. 3 a and b. Full size image Discussion Estimation of the number of years of strain accumulation before each SSE From the results described in the previous section, strong negative correlations in the dilatation were identified for Tokai L-SSE (Text S1 ), Boso-Oki S‒SSE (Text S2 ), and Bungo Channel L-SSEs (1) and (2) (Table S3 ). In this section, we estimate the number of years of strain accumulation for these four SSE dilatations with large negative correlations. Here, the number of years of strain accumulation before a SSE is defined as the number of years required to accumulate an amount of strain equivalent to the change in strain released by the SSE. The following Eqs. ( 1 ) and ( 2 ) were used to estimate the number of years of strain accumulation \(T\) . For this calculation, the same calculation points were used as those used for the calculation of the correlation coefficients. $$\begin{array}{*{20}c} {RMS = \sqrt {\frac{{\mathop \sum \nolimits_{i = 1}^{n} \left( {\left| {e_{i}^{eq} } \right| - t\left| {e_{i} } \right|} \right)^{2} }}{n}} } \\ \end{array}$$ (1) $$\begin{array}{*{20}c} {T = tT_{0} } \\ \end{array}$$ (2) where \(e_{i}^{eq}\) is the amount of released dilatation during the SSE at the \(i\) -th calculation point, \(e_{i}\) is the amount of strain accumulated at the \(i\) -th calculation point in the period before the SSE, n is the number of calculation points, and the value of \(t\) that minimizes the RMS value is obtained by the least-squares method. In Eq. ( 1 ), the smaller the value of RMS is, the better the absolute value of the released strain \(e_{i}^{eq}\) and the product of the absolute value of the accumulated strain \(e_{i}\) and \(t\) match. In Eq. ( 2 ), \(T_{0}\) is the period before the occurrence of each SSE set in this study, and the accumulated years \(T\) were obtained by taking the product of the value of \(t\) when the value of RMS was at its minimum in Eq. ( 1 ). The optimal value for the number of years of strain accumulation was searched by changing the values of \(t\) in the range from 0.01 to 10.00 at intervals of 0.01. Table 1 shows the estimated number of years of strain accumulation for each SSE. For the Boso-Oki S‒SSE and the Bungo Channel L-SSEs (1) and (2), for which the end of the previous SSE is known, the estimated accumulation years are 1.7 years for the Boso-Oki S‒SSE, 1.4 years for the Bungo Channel L-SSE (1), and 2.1 years for the Bungo Channel L-SSE (2). Considering that the analysis periods before these SSEs were 4.7, 5.3, and 7.2 years, respectively, the accumulation years for the Boso-Oki S‒SSE, Bungo Channel L-SSE (1), and Bungo Channel L-SSE (2) were all approximately 30% of the total years. This indicates that approximately 30% of the strains accumulated prior to the SSEs being released during the SSEs. Table 1 Years of strain accumulation at each SSE. Full size table For the Tokai L-SSE, the estimated accumulation period was 4.7 years, which is longer than 3.5 years of the analysis period before the L-SSE. This longer period is because the installation of GNSS stations in the Tokai region started on 21 March 1996, and we were unable to observe L-SSEs that ended before that date in the region. However, there were reports that crustal deformations equivalent to L-SSE were observed in the Tokai region before 2000, without using GNSS data. Fujii 16 pointed out that, based on the measured side lengths in the Tokai region, an extension different from the constant east–west compression was observed in the area from Lake Hamana to Omaezaki in approximately 1974. Ohtake and Asada 17 reported unusual crustal deformation in the Omaezaki area in 1976 based on observations of tidal level changes, radon concentrations in water, and strain gauges operated by the Japan Meteorological Agency. Kimata et al. 18 stated that L-SSEs occurred from 1978 to 1983 and 1987 to 1991 based on the analysis of laser ranging and levelling. Kobayashi and Yoshida 19 noted that similar changes occurred in 1980–1982 and 1988–1990 based on their analysis of tidal records, suggesting that L-SSEs may have occurred. Yamamoto et al. 20 reported that an L-SSE occurred between 1988 and 1990 based on an analysis of tilt change. Based on these previous studies, we can conjecture that crustal deformations similar to the 2000–2005 L-SSE occurred in the Tokai region in the late 1970s and late 1980s, with a recurrence interval of approximately 10 years, although there was some variance during the period. If the end year of the previous L-SSE that took place before the L-SSE analysed in this study was assumed to be 1990, the period before the L-SSE was approximately 10 years. For the Tokai L-SSE that initiated in 2000, considering that the accumulation period of dilatation was 4.7 years, approximately 50% of the strain accumulated before its occurrence was released when the L-SSE occurred. The relationship between slip caused by SSEs and interplate coupling In this subsection, we discuss the relationship between interplate coupling and release due to slip at the plate boundary based on the above strong spatial relationship in the dilatation between the accumulation before the SSE and release during the SSE at the Earth’s surface. Here, we consider the Bungo Channel L-SSE as an example, where approximately 30% of the accumulated strain was released during the L-SSE. According to Seshimo and Yoshioka 9 , in many areas of the plate boundary beneath the Bungo Channel, long-term interplate coupling is more dominant than release caused by slip associated with L-SSEs. In terms of the displacement field at the Earth’s surface, from Figs. 2 a and e and 4 a and e, the amount of displacement in the northwest direction before the L-SSE is much larger than the amount of displacement during the L-SSE. To explain the result of 30% strain release relative to the amount of strain accumulation, we discussed the relationship between interplate coupling (or slip deficit) and forwards slip related to strain release at the plate boundary. In this discussion, we attempted to find a model that can explain the displacement fields observed in the Bungo Channel during the accumulation and release periods by forward modelling, using a static fault model in a semi-infinite homogeneous perfect elastic body (DC3D) (Okada 21 ). Here, we used the accumulation and release periods of the 2018–2019 Bungo Channel L-SSE, with a displacement of 3.3 cm/yr for the accumulation period at the GNSS station closest to the Bungo Channel (Fig. 6 a). For interplate coupling, the slip deficit in the direction of plate convergence by DeMets et al. 22 was given on a rectangular fault plane at the plate boundary placed directly beneath the Bungo Channel. For strain release, forward slip in the opposite direction to that of DeMets et al. 22 was given on a rectangular fault on the plate interface directly beneath the Bungo Channel. Figure 6 Displacement and various strain fields at the Earth’s surface calculated using the rectangular fault model beneath the Bungo Channel and its shallower plate interface. Red rectangles 1 to 3 represent the horizontal projections of the assumed fault planes (see Table 2 ). The green dot in ( a ) is the Sukumo station (021059), which is the closest to the Bungo Channel. The map was created by using the Generic Mapping Tools (GMT) 27 (version: GMT3.4.6, URL link: ). Full size image As a result, we found that the observed displacement field was not well reproduced by providing a slip deficit only immediately beneath the Bungo Channel. However, by also providing the interplate coupling of the shallower side, namely, the seismogenic zone, in addition to the coupling beneath the Bungo Channel, we were able to reproduce the observed displacement field. This is consistent with the results of Yokota et al. 23 and Noda et al. 24 , who estimated relatively strong interplate coupling on the shallow side of the plate boundary. Thus obtained displacement and various strain fields are shown in Fig. 6 a–d. The slip deficit rate is 6.2 cm/yr in both the Bungo Channel and the shallower seismogenic zone, which indicates a coupling ratio of 1.0 compared with the convergence rate by DeMets et al. 22 . This implies acoupling of 6.2 (cm/yr) × 7.2 (yr) ≈ 44.6 cm during the period of strain accumulation prior to the L-SSE occurrence. On the other hand, for the strain release process, a forward slip of 28.0 cm was given, which was very close to the observed displacement field associated with the L-SSE. The displacement field and various strain fields are shown in Fig. 6 e–h. Each fault parameter used for these calculations is also tabulated in Table 2 . Table 2 Fault parameters for the three rectangular faults shown in Fig. 6 . Full size table Therefore, by comparing 44.6 cm during the accumulation period and 28.0 cm during the release period, the ratio of release to accumulation is approximately 60%, which means that 60% of the interplate coupling at the plate boundary beneath the Bungo Channel and its shallow seismogenic zone slipped beneath the Bungo Channel. This value is large compared to the dilatation release rate of 30%. However, this value is likely due to the smaller limited forward slip area beneath the Bungo Channel compared to the large strongly coupled area at the plate interface both in the shallow seismogenic zone and the Bungo Channel, resulting in the larger value of the forward slip/slip deficit ratio. We also confirmed that the ratio of strain accumulation and release at this time was calculated to be approximately 30% of the total release, so this amount of slip appears to be reasonable. In this case, the correlation coefficient in dilatation became -0.93, which is slightly larger than the value obtained in the strain analysis. However, this may be due to neglecting the northwestwards to southwards displacements in the Kyushu region before the L-SSE occurrence (Figs. 2 a, 4 a) in the observed value. This effect may have reduced the correlation coefficient to approximately − 0.8 for the observed value. In addition, the poor correlation of the maximum shear strain (Fig. 6 c, g) and why the principal strain is not in the opposite direction before and during the L-SSE occurrence (Fig. 6 d, h) can be explained by forward slip associated with the L-SSE only beneath the Bungo Channel at the time of strain release. Therefore, we attempted to reproduce the strain field in the Bungo Channel using DC3D and found that the observed values were best explained when a coupling ratio of 1.0 was applied to both the shallow part of the plate boundary and beneath the Bungo Channel. This finding suggests that interplate coupling in the shallow part may also contribute significantly to the accumulation of strain in and around the Bungo Channel. Conclusions In this study, the spatial distribution of the strain field (dilatation, maximum shear strain, and principal strain) during the periods before and during the Tokai L-SSE, the Boso-Oki S‒SSE, and the Bungo Channel L-SSE (1) and (2) were obtained by analysing GNSS time series data. The significant results obtained in this study are summarized as follows: 1. In the Tokai L-SSE, the Boso-Oki S‒SSE, and the Bungo Channel L-SSE (1) and (2), in the dilatation, larger accumulation and release were observed on the Earth's surface in the area where the SSE occurred before and at the time of occurrence, respectively. In all the SSEs, the spatial distribution of the strain accumulation region before the SSE and the strain release region at the time of the SSE were in good agreement. 2. For the Boso-Oki S‒SSE and the Bungo Channel L-SSE (1) and (2), approximately 30% of the dilatation was released compared to the amount of strain accumulated before the SSE. In the Tokai L-SSE, approximately 50% of the strain release was identified when the strain accumulation period before the L-SSE was assumed to be 10 years. For the Bungo Channel L-SSE (2), the period required for strain accumulation was increased due to the 2011 Tohoku-Oki earthquake and the 2016 Kumamoto earthquake (Text S3 ). This increase indicates that some of the strain that had been accumulating before the Bungo Channel L-SSE (2) was released by the occurrences of both earthquakes. 3. The meaning of the ratio of strain release to accumulation at the Earth's surface was considered in terms of ratio of the forward slip to slip deficit at the plate boundary by forward modelling. As a result, we found that approximately 60% of the interplate coupling at the plate boundary beneath the Bungo Channel and its shallow seismogenic zone was released by forward slip only beneath the Bungo Channel. Methods Data processing of time series data In this study, we used the daily coordinate positioning values (F5 solutions) of GNSS Earth Observation Network System (GEONET) stations provided by the Geospatial Information Authority of Japan (GSI), and the following data processing was performed for the time series. Displacement is defined as the displacement relative to a distant reference station that is not affected by an SSE to confirm consistency with the direction of plate convergence. For the case of the Bungo Channel L-SSE (1) and (2), we used six reference stations, following Yoshioka et al. 7 and Seshimo and Yoshioka 9 . Except for the Tokai L-SSE, the strain accumulation period before the SSE was defined as the period from the end of the previous SSE to just before the onset of the target SSE, and the period during the SSE was defined as the period from the onset of the SSE to its end. For the Tokai L-SSE, the starting date of the period before the SSE was set as 1 January 1997 because the GNSS time series data did not include the period of the previous Tokai L-SSE. This is because the observations by the GNSS stations operated by the GSI started on 21 March 1996. In this study, only the east–west and north–south components of horizontal displacement time series data were used. The analytical procedure of GNSS time series data is described below, following the method of Abe and Yoshioka 25 . Note that the GNSS time series data are processed differently before and during the occurrence of an SSE: We investigated the accumulation of strain due to all crustal deformation, including steady crustal deformation and coseismic steps, whereas the release of strain was due to only the occurrence of the SSE. GNSS time series data include steps due to antenna exchanges, coseismic steps, annual and semiannual variations, variations due to SSEs, and common-mode error. In this study, to obtain the strain field due to crustal deformation, including an SSE, we first corrected the steps due to antenna exchanges from the original time series. The steps were corrected by taking the difference in the averages of the previous and next 10 days when a step of 0.5 cm or more was detected in the east–west or north–south component at a certain observation station. Next, curve fitting was performed on the corrected GNSS time series data using Eq. ( 3 ). $$y\left( t \right) = a_{0} + \mathop \sum \limits_{i}^{n} a_{i} T_{i} \left( t \right) + b\sin \left( {\frac{2\pi t}{T}} \right) + c\cos \left( {\frac{2\pi t}{T}} \right) + d\sin \left( {\frac{4\pi t}{T}} \right) + e\cos \left( {\frac{4\pi t}{T}} \right) + \mathop \sum \limits_{k = 1}^{{N_{eq} }} f_{k} H\left( {t - t_{k}^{eq} } \right)$$ (3) $$\begin{gathered} T_{1} \left( t \right) = t \hfill \\ T_{2} \left( t \right) = 2t^{2} - 1 \hfill \\ \begin{array}{*{20}c} \begin{gathered} T_{3} \left( t \right) = 4t^{3} - 3t \hfill \\ \vdots \hfill \\ \end{gathered} \\ \end{array} \hfill \\ \end{gathered}$$ (4) where \(t\) is the elapsed time from the analysis start date, \(y\left( t \right)\) is the approximate curve, and \(T\) = 1 year. The first and second terms on the right-hand side of Eq. ( 3 ) represent crustal deformation, and the second term in Eq. ( 3 ), \(T_{i} \left( t \right)\) , is a Chebyshev polynomial, as shown in Eq. ( 4 ). \(n\) is varied from 1 to 20, and the optimal value of \(n\) was determined using the AIC minimization principle (Akaike 26 ). The third to sixth terms on the right-hand side of Eq. ( 3 ) denote the annual and semiannual variations. The seventh term represents coseismic steps caused by large earthquakes. \(t_{k}^{eq}\) is the date of the \(k\) th earthquake, and \(H\left( {t - t_{k}^{eq} } \right)\) is a Heaviside step function. The unknown parameters \(a_{0}\) , \(a_{i}\) , \(b\) , \(c\) , \(d\) , \(e,\) and \(f_{k}\) were determined by the least-squares method. The residuals of the east–west and north–south components were obtained between the approximate curves and the step-corrected GNSS time series data. The residuals were obtained at each observation station and averaged over all observation stations used in the analysis except the reference stations to obtain the common-mode error. The common-mode error was removed from the GNSS time series data corrected for the steps due to antenna exchanges before the SSE, and then the curve fitting was performed again using Eq. ( 3 ). Likewise, the time series data at the time of an SSE were corrected for the steps due to antenna exchanges, and the curve fitting was performed again using Eq. ( 3 ). In this case, the unknown parameters for the annual and semiannual variations were fixed, using the obtained values calculated in the analysis before the SSE because the analysis period was much shorter than that before the SSE. The same analysis was performed for the reference stations, which were used only for the purpose of plotting the displacement field before and during the SSE, and the average GNSS time series data of the reference stations were subtracted from the GNSS time series data of each observation station. By accurately estimating the difference between the GNSS time series data at the start and end dates of each analysis, we used the above curve fitting. Analysis of the strain field We analysed the strain field according to the coordinate system of the International Terrestrial Reference Frame 2014 (ITRF2014), and the time series data at the reference stations were not used. In this study, the method of Shen et al. 15 was used to calculate the strain field. To obtain the strain field, the calculation points were placed on a grid interval of 0.1° over the area including the GNSS observation stations used, and the strain value was obtained at each of these points. According to Shen et al. 15 , we obtained the strain field, as described below. $$\begin{array}{*{20}c} {\left( {\begin{array}{*{20}c} {U_{x}^{i} } \\ {U_{y}^{i} } \\ \end{array} } \right) = \left( {\begin{array}{*{20}c} 1 & {\quad 0} & {\quad \Delta x_{i} } & {\quad \Delta y_{i} } & {\quad 0} & {\quad \Delta y_{i} } \\ 0 & {\quad 1} & {\quad 0} & {\quad \Delta x_{i} } & {\quad \Delta y_{i} } & {\quad - \Delta x_{i} } \\ \end{array} } \right)\left( {\begin{array}{*{20}c} {u_{x} } \\ {u_{y} } \\ {e_{xx} } \\ {e_{xy} } \\ {e_{yy} } \\ \omega \\ \end{array} } \right) + \left( {\begin{array}{*{20}c} {\varepsilon_{x}^{i} } \\ {\varepsilon_{y}^{i} } \\ \end{array} } \right)} \\ \end{array}$$ (5) $$\begin{array}{*{20}c} {\varepsilon_{x,y}^{i} = \sigma_{x,y}^{i2} \exp \left( {\frac{{\Delta x_{i}^{2} + \Delta y_{i}^{2} }}{{D^{2} }}} \right)} \\ \end{array}$$ (6) Although Shen et al. 15 used the equation to obtain the strain rate, in this study, the total strain accumulation and release for each analysis period were obtained. Equation ( 5 ) was used to obtain the value of strain at the calculation points. \(x\) and \(y\) are the east–west and north–south components, respectively, with east and north as positive directions. \(i\) is the observation point, \(U_{x}^{i}\) and \(U_{y}^{i}\) are the displacements at the GNSS station, \(\Delta x_{i}\) and \(\Delta y_{i}\) are the distances between the GNSS station and the calculation point, \(u_{x}\) and \(u_{y}\) are the displacements at the calculation points, \(\omega\) is the rotation of the rigid body, \(e_{xx}\) , \(e_{yy}\) , and \(e_{xy}\) are the strains at the calculation points to be obtained, \(e_{xx}\) and \(e_{yy}\) are the normal strains, and \(e_{xy}\) is the shear strain. The second term on the right-hand side, \(\varepsilon_{x}^{i}\) and \(\varepsilon_{y}^{i}\) , represents the observation error and decay over distance and is obtained by Eq. ( 6 ). In Eq. ( 6 ), \(\sigma_{x}^{i}\) and \(\sigma_{y}^{i}\) represent the observation error, which is given by the obtained standard deviation when determining the displacement at the GNSS station, performing the curve fitting using Eq. ( 3 ). \(D\) represents the Distance Decaying Constant (DDC), and Table S4 shows values of the distance decaying constants used for each SSE. In this study, for each calculation point, GNSS observation points within a radius of \(2D\) centred on the calculation point were used. We used the strains ( \(e_{xx}\) , \(e_{yy}\) , and \(e_{xy}\) ) to obtain the dilatation, maximum shear strain, and principal strain. $$\begin{array}{*{20}c} {e_{d} = e_{xx} + e_{yy} } \\ \end{array}$$ (7) $$\begin{array}{*{20}c} {e_{s} = \sqrt {\frac{{\left( {e_{xx} - e_{yy} } \right)^{2} }}{4} + e_{xy}^{2} } } \\ \end{array}$$ (8) $$\begin{array}{*{20}c} {e_{1,2} = \left( {e_{d} \pm e_{s} } \right)/2} \\ \end{array}$$ (9) Equations ( 7 ) and ( 8 ) are used to determine the dilatation and maximum shear strain, \(e_{d}\) and \(e_{s}\) , respectively. The dilatation is a strain that represents an increase or decrease in an area, with a positive value indicating expansion and a negative value indicating contraction. The maximum shear strain represents the degree of lateral deformation and varies in magnitude depending on the axis taken, but the maximum value is obtained. From the values of dilatation and maximum shear strain, the principal strains \(e_{1}\) and \(e_{2}\) are obtained using Eq. ( 9 ). The principal strain is expressed in terms of the strain in the two principal axes, and the direction of the principal axes, \(\theta\) , is determined by the following Eq. ( 10 ). $$\begin{array}{*{20}c} {\theta = \frac{1}{2}\tan^{ - 1} \left( {\frac{{2e_{xy} }}{{e_{xx} - e_{yy} }}} \right)} \\ \end{array}$$ (10) where \(\theta\) is the azimuth of \(e_{2}\) with north as the reference azimuth \(0^\circ\) and positive values in the clockwise direction. Although the maximum shear strain originally takes only positive values, in this study, the maximum shear strain is redefined to take negative values to see the relationship between strain accumulation and release. From the two principal strains \(e_{1}\) and \(e_{2}\) , when \(\left| {e_{1} } \right| \ge \left| {e_{2} } \right|\) , $$\begin{array}{*{20}c} {e_{s} = e_{1} - e_{2} } \\ \end{array}$$ (11) and when \(\left| {e_{1} } \right| < \left| {e_{2} } \right|\) , $$\begin{array}{*{20}c} {e_{s} = e_{2} - e_{1} } \\ \end{array}$$ (12) Equation ( 11 ) represents the maximum shear strain at the calculation points where the tensile component of the principal strain dominates, where the maximum shear strain is positive. Equation ( 12 ) represents the maximum shear strain at the calculation points where the compressive component of the principal strain is dominant and negative. The above three types of strains were calculated for the period before and during the occurrence of each SSE. Data availability We used the daily coordinate data of GNSS stations provided by the Geospatial Information Authority of Japan (GSI) ( ) for displacement, the electronic reference point maintenance work list ( ), and JMA's seismic intensity database ( ). The figures were created using the Generic Mapping Tools (version: GMT 3.4.6, URL link: ) by Wessel and Smith 27 .

The Japanese archipelago is actively undergoing seismic shifts due to interactions between the oceanic plate and the continental plate. At the plate boundaries located directly beneath areas of Japan (especially the Bungo Channel, Tokai and Boso-Oki regions), slow slip events (SSEs) occur, which involve gradual aseismic slipping taking place at a recurrence interval of several years. However, it is still not clear exactly how tectonic strain is accumulated and released in association with SSEs in these regions. A thorough understanding of SSEs would improve our knowledge of how megathrust earthquakes occur, leading to better forecasts. To shed some light on this, Kobe University's Kawabata Hiroki and Professor Yoshioka Shoichi (Research Center for Urban Safety and Security) used GNSS time series data to investigate the relationship between strain accumulation and release before and after the occurrence of SSEs in the Bungo Channel, Tokai and Boso-Oki regions. The results for all regions investigated revealed that not all of the accumulated strain is released when an SSE occurs. In addition, this release only occurs at the plate interface directly underneath. In contrast to this, it is highly likely that strain build-up not only happens in the SSE area but also in the shallow part of the plate boundary; in other words, the zone where a future megathrust earthquake could potentially occur. This new understanding is expected to contribute towards better earthquake forecasts in the future. These results were published in Scientific Reports. Main points An SSE is a phenomenon believed to be related to the occurrence of powerful megathrust earthquakes. Consequently, SSEs have received a lot of attention in recent years.This research study analyzed past slow slip events that occurred in the following areas of Japan: the Bungo Channel, the Tokai Region and Boso-Oki.A fundamental phenomenon that causes earthquakes is the accumulation and subsequent release of tectonic strain. Comprehensive research (that encompasses slow slip events and strain accumulation/release) is therefore considered vital to illuminate the onset mechanism behind megathrust earthquakes (such as the highly devastating Nankai Trough megathrust earthquake, which is forecasted to occur in the near future). SSEs background SSEs occur at the interface of tectonic plates, located underground below areas including the Bungo Channel, the Tokai Region and Boso-Oki. SSEs involve gradual aseismic slipping on the plate interface at a recurrence interval of several years. Plates sliding against each other in this manner is caused by the release of strain that has accumulated at the plate boundary. This movement is one of the fundamental causes of earthquakes. Many research studies have been conducted on SSEs up until now, however they tend to focus on one type. Studies have not examined different durations of SSEs across multiple regions from the perspective of strain accumulation and release. SSEs are linked to the occurrence of megathrust earthquakes. Therefore, it is vital to investigate SSE-related strain accumulation and release in order to illuminate the causal mechanism behind megathrust earthquakes. This study analyzed past SSEs that have occurred in the Bungo Channel, Tokai and Boso-Oki regions of Japan. GNSS time series data and models The researchers used GNSS time series data provided by the Geospatial Information Authority of Japan to analyze the amount and spatial distribution of strain accumulation and release before and after SSE occurrence. They analyzed the following events: a Tokai Region long-term slow slip event (L-SSE) that occurred between 2000 and 2005, a short-term slow slip event (S-SSE) that occurred in Boso-Oki in 2007, and two Bungo Channel L-SSEs that occurred in 2009-2011 and 2018-2019, respectively. These results confirmed that there was significant strain accumulation and release on the surface of the ground before and after SSE occurrence, respectively. Regardless of region, less than half of the strain accumulated beforehand was released when the SSE occurred. In addition, the researchers also created models of plate interface adhesion and slippage for the Bungo Channel L-SSEs that could explain the analysis results well. The models revealed that the impact of SSEs is not limited to where they occur but is also connected to strain accumulation at the epicenter of megathrust earthquakes in shallower areas than the plate interface. Slow slip events differ from people's typical image of an earthquake as they cannot be directly felt. However, they are intrinsically related to the occurrence of megathrust earthquakes, as indicated by this study. SSEs, especially those occurring in the Bungo Channel and Tokai regions, are considered important phenomena that could provide an early warning for earthquakes, including Tokai earthquakes and the devastating Nankai Trough megathrust earthquake (which is likely to occur in the near future). It is hoped that investigating the relationship between the strain accumulation and release that accompanies SSEs in various regions throughout the Japanese archipelago will lead to new insight into the mechanism behind megathrust earthquakes as well as the state of interplate coupling.

10.1038/s41598-023-28016-1

Earth

Climate change means land use will need to change to keep up with global food demand, researchers say

T.A.M. Pugh et al. Climate analogues suggest limited potential for intensification of production on current croplands under climate change, Nature Communications (2016). DOI: 10.1038/ncomms12608 Journal information: Nature Communications

http://dx.doi.org/10.1038/ncomms12608

https://phys.org/news/2016-09-climate-global-food-demand.html

Abstract Climate change could pose a major challenge to efforts towards strongly increase food production over the coming decades. However, model simulations of future climate-impacts on crop yields differ substantially in the magnitude and even direction of the projected change. Combining observations of current maximum-attainable yield with climate analogues, we provide a complementary method of assessing the effect of climate change on crop yields. Strong reductions in attainable yields of major cereal crops are found across a large fraction of current cropland by 2050. These areas are vulnerable to climate change and have greatly reduced opportunity for agricultural intensification. However, the total land area, including regions not currently used for crops, climatically suitable for high attainable yields of maize, wheat and rice is similar by 2050 to the present-day. Large shifts in land-use patterns and crop choice will likely be necessary to sustain production growth rates and keep pace with demand. Introduction Between 1960 and 2000, global cereal production doubled, primarily through intensification of agriculture on current croplands and by the breeding of more productive crop varieties 1 . It is projected that global food production needs to increase by another 60–110% by 2050, to keep up with anticipated increases in human population and changes in diet 2 , 3 . Increases of this magnitude are possible through aggressive intensification in areas that are currently far below their potential 4 . However, several challenges exist relating to our ability to overcome economic, societal and environmental impediments to intensification 3 , 4 , 5 . To address these challenges effectively, and with a long-term outlook, it is necessary to understand how climate change is likely to impact the agricultural sector. Current expectations are that changes in climate over the next century will tend to decrease yields at lower latitudes and increase yields at northern latitudes 6 , 7 , but no consensus exists on the magnitude, timing and exact location of these changes. Assessments of future food production under climate change have to-date relied on synthesis of crop-climate assessments at field and/or regional scales using models which vary from highly process-based to relatively empirical in formulation 6 , 8 . Both approaches can give important insights, but suffer also from significant limitations. Empirical models implicitly capture the effects of all relevant processes over the period and location for which they were parameterized, but they are often poorly tested for the conditions likely to become the future norm under climate change 8 , 9 , and are typically limited in their spatial coverage. Process-based models account explicitly for physiological mechanisms, and can thus capture emergent behaviour under novel conditions, but generally suffer from missing processes, such as heat stress and the coupling of transpiration to leaf temperature 8 , 10 . Many research groups now develop and operate global-scale crop-climate models, but the necessary data for model development, parameterization and evaluation is often not available at this scale 11 , and in most cases, global-scale assessments fail to account for farmer adaptation to climate change 7 , 9 . As a result, these models vary strongly in both the magnitude and overall direction of their yield projections, an effect which is especially pronounced across the tropics and arid regions 7 . This uncertainty impedes understanding the scope of the societal challenge of climate change impacts on agriculture, particularly in developing countries where agriculture has larger shares in gross domestic product. Furthermore, large uncertainties in yield projections hinder our ability to assess the potential of various climate-change adaptation versus mitigation options; future changes in the productivity of different cropland areas will affect future land use, including societal decision-making on whether to, for example, reforest or intensify an area of cropland, or to extend cropland coverage 12 , 13 , 14 . We use here a data-driven approach, utilizing observation-based estimates of current maximum-attainable yield with existing technology (hereafter referred to as attainable yield, see Methods section for definition) 4 and climate analogues 15 , 16 , to assess the vulnerability of yields of the three major global cereal crops, wheat, maize and rice to climate change. Using future climate projections from five General Circulation Models (GCMs) that contributed to the Coupled Model Intercomparison Project Phase 5 (CMIP5) as part of the Fifth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC) 17 , we identify present-day analogues of future climatic conditions, across the current global cropland area, with medium- (time window of 2041–2059, henceforth mid-century) and long-term (2081–2099, henceforth end-of-century) perspectives. Analogue climates are defined based on the accumulated annual sums of growing-degree-days (GDD) and precipitation, although we also test other variable combinations and seasonal averaging periods (Methods section). We make the assumptions that, given no changes in the global cultivar pool, the attainable yields of crops are purely a function of climate, and that the most effective yield-enhancing managements available are either spatially and temporally invariant, or also a function of climate; for instance, irrigation is unlikely to be widely available in very dry climates (neglecting possible finite fossil groundwater supplies or large rivers). Based on these premises, a yield which can be attained in the present day under a given climate can also be attained in the future, under an analogous climate, but at a different location. Areas with fundamentally unsuitable soils are avoided either by restricting projection of our results to current cropland area, or by masking with soil suitability information (Methods section). Our approach is thus independent of crop modelling methodologies typically used to project future yields, although, as for crop models, the results remain subject to uncertainties in terrain and soil type, particularly at the sub-grid scale. Further, it has important added value as, in addition to considering climatic effects on crop yields, it implicitly considers adaptation of management and crop type in line with current technology available for a given climatic environment. The technique allows us to (i) assess the change in attainable crop yield across the current global cropland area for each of maize, wheat and rice, (ii) identify areas where current dominant crops are likely to no longer be suitable in future, and (iii) to explore potential new growth locations for these crops. We find widespread reductions in attainable yields of major cereals across a large fraction of current cropland by ca. 2050, but find that the total global land area climatically suitable for these crops remains largely unchanged, suggesting that in the absence of substantial advances in technology, it will be necessary to rethink crop choice and land-use patterns to achieve substantial production growth rates in the future. Our results not only improve confidence in the projections of vulnerability of the yields of cereal crops to climate change, but also point towards needs for knowledge on new technologies, and changes in management and genotypes. Results Effects on current cropland By mid-century, following a strong climate change scenario (Representative Concentration Pathway, RCP, 8.5; Methods section) we find that most current wheat, maize and rice areas are within climatic conditions under which these crops are also cultivated today ( Supplementary Fig. 1 ), although the analogues may be drawn from very distant locations. Analogues in the tropics, however, deteriorate greatly by end-of-century, especially for rice, as climates begin to emerge which are inconsistent with present-day cultivation of these crops. The results are qualitatively similar to results for previous assessments in which climate analogues were applied to investigate changes in biodiversity and natural vegetation, and which also showed poor present-day analogues for end-of-century climates in the tropics with strong climate change 15 , 16 . We find that large areas of current cropland are projected to experience strong reductions in attainable yields of major cereal crops over the next century ( Fig. 1 ), indicating persistent vulnerability to climate change in the absence of significant advances in agricultural technology beyond that currently available. Our results are robust to the choice of GCM used to simulate climate, with agreement on the sign of yield change for at least four out of five GCM climates for the dominant spatial features in Fig. 1 ( Supplementary Figs 2–7 ). Reductions are particularly seen in the number of grid-cells within the current harvested area that are projected to have a high attainable yield for maize and wheat ( Fig. 1c,f ). Vulnerable areas, which we define here as those showing a reduction in attainable yield >10% relative to the present-day reference data set 4 and those which have no present-day analogue, vary by crop, but include some of the major regions of global food production (see below). Based only on areas where our results agree for at least four out of five GCM climates, our results indicate that already by mid-century 43, 28 and 40% of current global harvested area 18 for wheat, maize and rice, respectively, is located in regions in which yields are vulnerable to climate change. In these regions, attainable yields are projected to either decline in the new climate, or no present-day analogue for the expected future climate can be found ( Table 1 ). Vulnerable areas are not just located in low-producing regions; crops grown on areas designated as vulnerable by mid-century correspond to 43, 33 and 37% of the current global production of maize, wheat and rice respectively, increasing to ca. 74, 56 and 65% by end-of-century. If we consider all vulnerable areas indicated by at least one GCM climate, the vast majority of wheat, maize and rice production would be located in vulnerable areas by mid-century ( Supplementary Table 1 ). Figure 1: Change in attainable yield from the reference period to 2041–2059 and 2081–2099. Maps show percentage changes in attainable yield for for maize ( a , b ), wheat ( d , e ) and rice ( g , h ). Areas where yields change by 10% or less are marked in magenta. Areas with no present-day climate analogue are marked in grey (very few grid-cells). Yield changes are only shown for grid-cells which have a present-day climate analogue within the current harvested area of each crop, and where the current area devoted to that crop within the grid-cell exceeds 1,000 ha (ref. 35 ). Panels c , f and i show how the distribution of attainable yields at the grid-cell level is projected to evolve between the reference period (magenta line) and end-of-century (red line). Loss of area under the probability density curves is equivalent to the number of grid-cells for which no valid climate analogue can be found. Present-day attainable yields are obtained from ref. 4 . Full size image Table 1 Portion of current global production and harvested area in vulnerable or no-analogue zones for at least four out of five GCM climates. Full size table The spatial consequences of these changes in attainable yield for agricultural production can be appreciated by considering the effect on yields of the dominant crop at each location. Reductions in attainable yield of the dominant crop (by area) of >10% are projected across much of the eastern United States, Europe, Sub-Saharan Africa and temperate South America, and much of Eastern-Asia ( Fig. 2 ). With the exception of regions with very strong management limitations, reductions in attainable yields are expected to generally imply reduction in actual yields unless measures can be taken to compensate by closing yield gaps (defined here as the difference between actual and maximum-attainable yields, given current technology). Highly developed countries, where yield gaps are already very small 4 , may thus face difficulties in sustaining current production without new technological interventions to increase attainable yield (for example, breeding novel crop cultivars). Such interventions could prove particularly challenging for maize and rice for which there has been little or no change in the genetic yield potential over recent decades 19 . Figure 2: Yield vulnerability of dominant crop. Maps show grid-cells where the dominant crop by area is projected to undergo reductions in attainable yield of at least 10% by 2041–2059 ( a ) and 2081–2099 ( b ) relative to the reference period. The dominant crop is indicated by colour, with darker shades indicating that the attainable yield is maintained or increased, and lighter shades indicating areas in which it undergoes a decrease >10%. Only grid-cells where the sum of cropped area of wheat, maize and rice exceed 1,000 ha are shown 18 . Full size image The effect of these changes in attainable yield on efforts to intensify cereal production will be particularly marked; there will be a greatly reduced capacity to intensify crop production beyond current levels through efforts to close the yield gap on existing croplands. It has been estimated that closing yield gaps in current climates could increase global production of maize, wheat and rice by 67, 100 and 52%, respectively, on their current harvested areas 4 . Based on end-of-century climate analogues, we estimate these potential increases over year 2000 yields to be only 14, 62 and 6% ( Fig. 3 ), an average 57% reduction in global attainable yield growth potential. Intensification of production on current croplands therefore appears highly unlikely to be able to meet growing global demand over the next decades. Figure 3: The effect of climate change on production increases through agricultural intensification. The potential increases in global production of maize, wheat and rice that could be achieved by closing the yield gap (difference between actual and attainable yields) on current cropland are plotted for baseline, mid-century, and end-of-century climates (light green bars). For reference, these bars are imposed on top of actual production for ca. year 2000 (dark green bars), as calculated based on actual yields from Mueller et al . 4 and cropland area from Monfreda et al . 18 . Full size image Although climate change stands to reduce production potential in many parts of the world, it nevertheless brings new opportunities in some regions. We identify large areas of current croplands where attainable yields increase substantially under climate change, with enhancements often exceeding 50% ( Fig. 1 ). These areas are most prominent for wheat in the northern mid-latitudes, and include the large concentration of cropland in the central Canadian wheat belt, as well as western and central Russia. Assuming yield depressions from mismanagement, pests, diseases and other stresses are no higher in the target region than in the analogue region, this implies substantial increases in actual yields across these regions as a result of climate change. Changes in climatically suitable areas for cropping Climate change can also increase the production potential of crops to levels of economic significance in regions which have hitherto been essentially uncropped. By expanding our analysis beyond the current harvested area of crops, we show that a large belt of land at high northern latitudes develops climates which are suitable for the production of wheat or maize with attainable yields by mid-century that are at least equivalent to the median attainable yield for that crop across current harvested area ( Fig. 4 ); the area of land poleward of the 30th parallels climatically suitable for attainable yields above this threshold increases from 1.8 × 10 9 to 2.1 × 10 9 ha for maize and from 1.6 × 10 9 to 1.7 × 10 9 ha for wheat ( Supplementary Table 2 ). Notably, the area of land climatically suitable for rice expands strongly to mid-century in the extratropics, but decreases in the tropics. Under RCP 8.5, the size of climatically high-yielding areas in the tropics collapses by end-of-century for all three crops, highlighting the importance of avoiding such a strong climate-change scenario. The nature of these results is qualitatively consistent with simple suitability indices based upon GDD alone, and with results from global crop models ( Supplementary Note 1 , Supplementary Figs 8-10 ). Nonetheless, whilst this global-scale analysis is indicative of increased climatic suitability for cropping in many areas that are not principally unsuitable from large-scale soil constraints 20 (Methods section), we caution that because of its global scale it does not take account of local variations in soils or terrain properties which might adversely affect yields (although it is worth noting that poor soils can be remediated 21 ). It also says nothing of whether it is economically viable to put in place the measures to achieve such yields in these locations. Ultimately, climatic suitability must be qualified by considering local-scale drivers. Figure 4: Areas with climate suitable to provide attainable yields greater than present-day median attainable yields. Present day median attainable yields are 8.1 t ha −1 for maize, 4.6 ton ha −1 for wheat and 5.2 ton ha −1 for rice. Note that climatically suitable does not necessarily mean economically or socially viable. Grid-cells are coloured black, purple, or blue when analogues produce attainable yields over the crop-specific thresholds for four out of five GCMs. Black shows grid-cells (0.5° × 0.5°) meeting the yield criteria which are currently heavily cropped (>50,000 ha harvested area over all crop types). Purple colours show grid-cells which are currently lightly cropped (1,000–50,000 ha harvested area). Grid-cells which have a current harvested area <1,000 ha are coloured blue. Red grid-cells show cropped areas which had an attainable yield over the threshold in the present day, but fell below this threshold in the future. White areas do not obtain the yield threshold. Panels a , d and g show the current situation 4 , b , e and h the situation in 2041–2059 based on climate analogues from the ensemble of climate models, and c , f and i the same for 2081–2099. Current cropland areas are taken from Monfreda et al . 18 . Areas with soils classed as fundamentally unsuitable for cropping over at least 90% of the 0.5° grid-cell are masked out in light grey, whilst areas with very low precipitation relative to the typical growing conditions of that crop are masked out in dark grey (Methods section). Full size image Influence of climate change scenarios Although we base our results on a strong climate-warming scenario (RCP 8.5), large changes in attainable yield are apparent already by mid-century, when warming projections between RCPs have not yet strongly diverged 22 . Consequently, our results for mid-century are qualitatively similar, although of slightly lower magnitude, when calculated for the low climate change RCP 2.6 scenario ( Supplementary Fig. 11 ). In fact, estimated mid-century impacts on attainable yield, both positive and negative, are much stronger than most of the projections of actual yield synthesized in the IPCC Fifth Assessment Report 6 . This reflects the fact that when climate is the main limiting factor for yields, then changes in climate will have a greater effect, and implies that greater intensification likely brings greater vulnerability of yields to climate change. This is in line with previous findings that weather variability often is the main driver of yield variability in highly productive areas, but not necessarily in areas of low productivity 23 . Comparison against crop models The pattern of attainable yield changes in Fig. 1 are very similar to those reported for actual yields in previous assessments 24 , with the exception of the positive changes in West Africa and India at mid-century, which we discuss further below. We further compare the changes in attainable yield produced by our data-driven method, with those of an ensemble of gridded global crop models recently used to assess the impact of climate change on crop yields as part of the ISI-MIP project 7 ( Supplementary Figs 12 and 13 ). The simulations we compare against here are conceptually similar to the climate analogue approach in that they use fixed atmospheric CO 2 mixing ratios [CO 2 ] from 2000 onwards, however they differ in the fact that most models use fixed present-day fertiliser levels rather than unlimited nutrients, and in that the simulations allow unlimited irrigation, which may not be possible in some low-rainfall areas. Although the model-based climate-impact ensemble projections have a broad spread, the qualitative pattern of yield changes is very similar to the analogue method, with yield losses generally in the tropics and arid areas and increases in high latitudes. Notable differences between the model ensemble and climate analogue method arise for maize and rice in India and tropical West Africa. Here the models tend to show yield decreases, whilst the analogue approach suggests increases, however, we have lower confidence in yield projections in these locations as they show sensitivity to the analogue variable choice ( Supplementary Note 2 ). The models are also much more pessimistic regarding changes in wheat yields in the mid-latitudes, and especially across central-Asia. These more pessimistic projections from the models may result from limited or absent consideration of climate adaptation 7 . In reality, farmers will change their preferred cultivars, growing seasons and other management actions, as climate evolves. The analogue method implicitly accounts for cultivar adaptation within the bounds of the pool of globally-used cultivars, and for changes in growing seasons and other managements, possibly leading to positive changes in attainable yield in some regions where strongly negative effects have previously been expected. As there are many differences in methodology between the crop models and our analogue method, it is not possible to definitively attribute these different regional responses to adaptation. However, the comparison nonetheless demonstrates that inclusion of adaptation within the bounds of current technology does not fundamentally alter the global picture of yield change over the 21st century. Further comparison with models is provided in Supplementary Note 3 . Source of analogues A feature of the analogue method is that it identifies present day regions whose agricultural practices and crop varieties may hold lessons for other regions in the future, and thus where efforts should be made to maximize technology and knowledge transfer, that is, transferring existing crop varieties from the analogue to the target location, although we do not imply that the analogue approach can be used to derive explicit technology transfer pathways. It is not feasible to catalogue all potentially important analogue locations in a global analysis. However, we show some examples in Supplementary Fig. 14 , and summarize the general situation using the mean latitudinal distance from which analogues are drawn ( Supplementary Table 4 ). Mid-century analogues for cropland poleward of the 30th parallels are drawn, in the mean, from regions 18°, 13° and 22° latitude closer to the equator for maize, wheat and rice, respectively, and change relatively little between mid- and end-of-century, as analogues have already largely reached the extent of current harvested area of the crop in question by mid-century. Discussion Our analogue-based method only considers climate-induced changes in yield, but increased [CO 2 ] has been found to directly enhance photosynthesis in C3 crops, whilst also promoting greater water use efficiency across all higher plants 25 . These [CO 2 ] effects may offset some of the strong negative effects of climate change on attainable yields. A meta-analysis of field experiments suggests a yield enhancement of ca. 8% for non-water-stressed C3 crops for a doubling of [CO 2 ] from pre-industrial levels, although this number disguises huge variability in responses 26 . For an RCP 8.5 scenario this would imply a yield increment due to [CO 2 ] of +6% by 2050, and +18% by 2100, relative to 2000, although the response may be larger in some water-limited regions. Changes of this magnitude only marginally reduce the large vulnerable areas of wheat and rice by mid-century, although by end-of-century they would reduce substantially vulnerable areas of wheat and rice in East Asia ( Supplementary Fig. 15 ). Attainable yields of maize, a C4 crop, are not expected to be affected by increased [CO 2 ], although actual yields of all crops may be substantially enhanced in areas where water stress is limiting growth 25 , 27 . Simulations from the ISI-MIP ensemble show enhancements of wheat and rice yield due to CO 2 fertilization ranging from 4 to 80% ( Supplementary Table 3 ), illustrating the huge uncertainty in projections of CO 2 fertilization, however, direct comparison is complicated by the differing assumptions on nutrient availability and climate adaptation between the models (see Supplementary Note 3 and Supplementary Figs 16 and 17 ). The analogue method employed here has a number of limitations and uncertainties. The approach assumes that management practices for maximizing yield can be directly transferred between grid-cells. In most cases we expect this to be true, as socio-economic limitations do not apply in the context of attainable yields. However, there may be some circumstances where this concept fails due to specificity of managements to terrain or soil types. In areas with low precipitation, the method will tend towards more positive projections if analogues are drawn from grid cells in which sources of irrigation water are disconnected from local precipitation, that is, regions with large fossil groundwater abstractions or supplied by large rivers. These projected yields remain attainable, but water resources will not allow them to be reached on a large scale unless the grid cell for which yields are being projected also has access to such a water resource. We find no clear evidence of our projections for current cropland area being notably influenced by this effect, but if they were the effect would be to make our conclusions of vulnerability more conservative. It is very relevant, however, for projections of suitable future cropland area. As a result, we have marked such areas in Fig. 4 , and excluded them from the other calculations in this analysis (Methods section). Importantly, our projected changes in attainable yields must not be taken as being indicative of changes in rainfed yields in regions where irrigation water is potentially available, but is not utilized on socio-economic grounds; the relation between attainable and rainfed yields is non-linear. Although we mask fundamentally unsuitable soils (Methods section), we do not include variations in soil quality across current cropland area in our analogues. Soil is an important determinant of crop productivity, and thus some of the increases in yields suggested by our method may be limited by poor soil quality. There is a particularly high uncertainty in assessing the importance of soil quality in high northern latitudes, where soil core data are very scarce 28 . Here there is a risk that the analogue method tends to overestimate attainable yields on unimproved soils. We suggest, however, that because properties such as soil nutrient content, organic carbon and pH can be remediated through amendments and management, soil quality is generally a limitation on actual, rather than on attainable, yields. Our conclusions are robust to the choice of analogue variables, with simulations including killing degree days (degree-days above a crop-specific threshold for heat damage) and radiation, excluding precipitation, or limiting to current growing seasons, giving extraordinarily consistent results for yields, despite some variation in the location of the analogue climates ( Supplementary Note 2 ). We contend that early investment in regions that are likely to be climate analogues in the future could serve a dual benefit, both helping to alleviate any deficiencies in food production in these regions today, whilst also serving to inform, and thereby speed-up, climate adaptation in other locations in the future. This is particularly important when complications to the simple ‘transfer’ perspective arise where for instance tropical analogues are found for temperate grid-cells, and differences in, for example, day-length requirements may require identification and transferal of adapted germplasm, a process with potentially long lead-times 19 , 26 . However, while the climate analogue approach could be a (biophysically based) tool to assess technology transfer, the specific boundaries of such an analogue framework would need to be evaluated with observations of cross-location performance of crop and soil management technologies before it could be operationally applied in this context. Even with the assumption of adaptation within current technology inherent in our approach, climate change is projected to decrease production potential, and therefore likely actual production, from major cereal crops across the tropics and much mid-latitude cropland already by the middle of the 21st century. Our results thus provide an independent line of evidence qualitatively supporting, and thereby increasing confidence in, projections from crop models. In the absence of successfully breeding new crop cultivars to suit warmer climates, attempts to intensify crop production across much of existing cropland are likely to be frustrated by climate change, requiring compensation by other means. We conclude that much more efficient and sustainable production increases are likely to be achieved by altering crop choice and the location of cropland itself, to best take advantage of changes in maximum-attainable yields resulting from climate change. However, we caution that socio-economic constraints, along with any unmanageable aspects of local terrain and soil, will influence whether the climatic potential calculated here can be realised; detailed local-scale assessments will thus be necessary to assess this. Such land-use alterations do not necessarily imply extensification (that is, increasing the area under cultivation). However, changing the distribution of agricultural area might in some regions lead to dramatic changes in land use that will face political, social and cultural impediments if, for instance, some regions need to become net food importers or to change traditional diets. A drastic shift in distribution of cropland would likely also have very serious consequences for other ecosystem services such as carbon storage provided by non-agricultural lands where such a reconfiguration may take place 29 , although these consequences might be mitigated by the regeneration of areas of abandoned cropland in the long-term. Rather than being able to rely principally on continued intensification on existing croplands, as is often a major assumption of economic models 13 , 14 , the need to increase crop production is therefore likely to present a major land-use, trade and economic dilemma over the coming decades. Furthermore, for achieving attainable yields projected by the analogue approach, our results imply the need to transfer technologies and knowledge over large distances, demonstrating that there is yet another important dimension in the agricultural challenge that requires a global perspective; not only trade and markets, but also production knowledge needs to be globalized. Methods Attainable yields data set Present-day maximum-attainable yields used in this analysis are obtained from the data set created by Mueller et al . 4 . Mueller et al . divided harvested area of each crop into 100 equal-area bins based on GDD and annual precipitation. They then calculated the attainable yield as the 95th percentile of yields within each climate bin, after exclusion of outliers. As such, attainable yields represent a maximum level consistently achievable in a climatic zone, given current technology, and not physiological potential yields. Identification of climatic analogues We define climate analogues using an approach previously used to identify novel and disappearing climates and changes in vegetation carbon under climate change 15 , 30 . The method is modified here for use with crops. Difference between climates is calculated using Standardised Euclidean Distance ( SED ), where μ k,ref is the mean of climate variable k over the reference period (1981–1999), and μ k,fut the mean over the future period (2041–2059 or 2081–2099), σ k,ref is the s.d. of the interannual variability over the reference period. The letter i refers to the primary grid-cell, and j to any other grid-cell which i is being compared against. Temperature and water availability are fundamental constraints on crop growth. Since growing seasons can be adapted to changing climate, we consider annual (rather than seasonal) sum of GDD (base temperatures, T b , of 0, 8 and 5 °C for wheat, maize and rice, respectively 31 ) and mean precipitation as climatic variables k (1) and k (2). This maintains consistency with the methods used to derive the attainable yield data set 4 . Precipitation is included because areas with very low rainfall may not have the option of irrigation (that is, the attainable yield is a rainfed yield). Seasonal timing of precipitation is not considered critical, as water may be stored for irrigation later in the year. More discussion of the effect of analogue variable choice is provided in Supplementary Note 2 and Supplementary Fig. 18 . SED is calculated individually for every crop type. For every 0.5° × 0.5° grid-cell, i , SED is calculated between the future climate in that grid-cell, and reference climates in all grid-cells, j , including at least 1,000 ha of that crop 18 . We note that harvested area has expanded substantially for some crops since Monfreda et al . 18 was published 32 , but this data set remains the de facto standard and maintains consistency with Mueller et al . 4 , from whom we have taken the present-day attainable yield data. As changes in cropped area are much smaller than for harvested area, this should have minimal influence on our results. For any given grid-cell i , the grid-cell j under reference period climate which returned the minimum value of SED was chosen as the best present-day climate analogue for the future climate at location i . The attainable yield for the crop in question, as calculated by Mueller et al . 4 for ca. year 2000, from the chosen reference-period grid-cell j was then taken as the attainable yield for that crop in grid-cell i . In the results presented herein, we used the mean of the projected yield across the five best analogues to each grid-cell i . Note that when we test the production potential of grid-cells outside the current harvested area of a crop ( Fig. 4 ), we only allow analogues to be drawn from within the current harvested area. Values of SED and changes in attainable yield were calculated individually for bias-corrected daily climate data from each of 5 Global Climate Models (GCMs): IPSL-CM5A, HadGEM2-ES, GFDL-ESM2M, MIROC-ESM-CHEM, NorESM2 (ref. 33 ). Presented values represent the mean of results from those 5 GCMs, unless stated otherwise. Calculations presented herein were made based on RCP 8.5 (ref. 34 ), but were also performed for RCP 2.6 to test sensitivity to moderate climate change. Calculation of threshold SED To capture values of SED which are so high as to indicate that there is no climate in the reference period (within the current harvested area of the crop in question) from which to draw an analogue to the future climate, we calculate a threshold value of SED , SED t , above which the analogue is considered too weak for conclusions regarding attainable yield to be drawn. To calculate SED t , we randomly draw two sets of 19 individual years from the period 1981 to 2010 and assign one as climate ‘ref’ and the other as climate ‘fut’, and then calculate SED as per Equation E1. We thereby calculate for each grid-cell a value of SED for the case when climates in each location are drawn from the same distribution, which we term SED D . We then define SED t as the characteristic maximum value of SED D (excluding outliers) by calculating its 95th percentile across the current harvested area of each crop. We thus follow the premise that for an analogue to be valid for projecting crop yield, the value of SED for that analogue must be similar to the SED calculated between two climates drawn from the same distribution. Taking the mean value of SED t per crop across the 5 GCM climates generates values of 0.7, 0.7 and 1.0 for maize wheat and rice, respectively. For simplicity, we take the most conservative of these, SED t =0.7, as our baseline, however the projections of attainable yield are extremely robust to this choice ( Supplementary Fig. 19 ). This definition of SED t closely reflects the calculated gradients in this variable between typically cropped and non-cropped areas ( Supplementary Fig. 1 ). We also use these two sets of 19 random years from the same climate distribution to show that the climate analogue method calculates only minimal changes in yield when the climate is unchanged ( Supplementary Note 4 , Supplementary Fig. 20 ). Analysis of climatic potential for cropland expansion In considering areas of potential cropland expansion ( Fig. 4 and Supplementary Table 2 ), areas with soil fundamentally unsuitable for cropping were masked out based on the GAEZ v3.0 database 20 : we combined soil suitability indices at a resolution of 10 arc minutes for excess salts, oxygen availability, rooting conditions, toxicities and workability, and excluded any areas that were not marked as minimal or moderate constraints in any of the five indices. Conditions relating to nutrients were neglected, as it is considered that fertilization would be required to achieve attainable yields in any case. To mask out areas with very low rainfall drawing analogues with regions with substantial fossil groundwater or non-local irrigation water sources (that is, water sources independent of grid-cell annual mean precipitation), we calculated the 5th percentile of the distribution of reference period annual mean precipitation across the current rainfed area of each crop 35 . We then excluded from the cropland expansion analysis all grid-cells with annual mean precipitation below this level. Crop modelling Comparisons against Global Gridded Crop Models (GGCMs) used data from 6 GGCMs (LPJ-GUESS, LPJmL, pDSSAT, PEGASUS, EPIC, GEPIC) 7 . These GGCMs were run with the same climate data used for the SED calculations herein for the period 1971–2099, and simulated maize, wheat and rice at 0.5° × 0.5° resolution for all land grid-cells (no rice simulations from PEGASUS). GGCM set-ups were not harmonized except for climate and [CO 2 ]. Simulations were made both with transient [CO 2 ], and with [CO 2 ] fixed to 369 ppmv from year 2000 onwards. Here we compare our climate-analogue-derived attainable yields with modelled yields under full irrigation. Data availability The authors declare that all data supporting the findings of this study are available within the article and its Supplementary information files or are available from the corresponding author upon request. Additional information How to cite this article: Pugh, T. A. M. et al . Climate analogues suggest limited potential for intensification of production on current croplands under climate change. Nat. Commun. 7:12608 doi: 10.1038/ncomms12608 (2016).

A team of researchers led by the University of Birmingham warns that without significant improvements in technology, global crop yields are likely to fall in the areas currently used for production of the world's three major cereal crops, forcing production to move to new areas. With a worldwide population projected to top nine billion in the next 30 years, the amount of food produced globally will need to double. A new study led by researchers from the University of Birmingham shows that much of the land currently used to grow wheat, maize and rice is vulnerable to the effects of climate change. This could lead to a major drop in productivity of these areas by 2050, along with a corresponding increase in potential productivity of many previously-unused areas, pointing to a major shift in the map of global food production. The study, published today in Nature Communications, uses a new approach combining standard climate change models with maximum land productivity data, to predict how the potential productivity of cropland is likely to change over the next 50-100 years as a result of climate change. The results show that: Nearly half of all maize produced in the world (43%), and a third of all wheat and rice (33% and 37% respectively), is grown in areas vulnerable to the effects of climate changeCroplands in tropical areas, including Sub-Saharan Africa, South America and the Eastern US, are likely to experience the most drastic reductions in their potential to grow these cropsCroplands in temperate areas, including western and central Russia and central Canada, are likely to experience an increase in yield potential, leading to many new opportunities for agriculture While the effects of climate change are usually expected to be greatest in the world's poorest areas, this study suggests that developed countries may be equally affected. Efforts to increase food production usually focus on closing the yield gap, i.e. minimising the difference between what could potentially be grown on a given area of land and what is actually harvested. Highly-developed countries already have a very small yield gap, so the negative effects of climate change on potential yield are likely to be felt more acutely in these areas. Our model shows that on many areas of land currently used to grow crops, the potential to improve yields is greatly decreased as a result of the effects of climate change," says lead researcher and University of Birmingham academic Dr Tom Pugh. "But it raises an interesting opportunity for some countries in temperate areas, where the suitability of climate to grow these major crops is likely to increase over the same time period." The political, social and cultural effects of these major changes to the distribution of global cropland would be profound, as currently productive regions become net importers and vice versa. "Of course, climate is just one factor when looking at the future of global agricultural practices," adds Pugh. "Local factors such as soil quality and water availability also have a very important effect on crop yields in real terms. But production of the world's three major cereal crops needs to keep up with demand, and if we can't do that by making our existing land more efficient, then the only other option is to increase the amount of land that we use."

10.1038/ncomms12608

Medicine

New star rating system helps people make informed decisions about diet and healthy habits

Christopher Murray, The Burden of Proof studies: assessing the evidence of risk, Nature Medicine (2022). DOI: 10.1038/s41591-022-01973-2 IHME Website: vizhub.healthdata.org/burden-of-proof/ Global burden of disease study: www.healthdata.org/gbd/2019 Journal information: Nature Medicine

https://dx.doi.org/10.1038/s41591-022-01973-2

https://medicalxpress.com/news/2022-10-star-people-decisions-diet-healthy.html

Abstract Exposure to risks throughout life results in a wide variety of outcomes. Objectively judging the relative impact of these risks on personal and population health is fundamental to individual survival and societal prosperity. Existing mechanisms to quantify and rank the magnitude of these myriad effects and the uncertainty in their estimation are largely subjective, leaving room for interpretation that can fuel academic controversy and add to confusion when communicating risk. We present a new suite of meta-analyses—termed the Burden of Proof studies—designed specifically to help evaluate these methodological issues objectively and quantitatively. Through this data-driven approach that complements existing systems, including GRADE and Cochrane Reviews, we aim to aggregate evidence across multiple studies and enable a quantitative comparison of risk–outcome pairs. We introduce the burden of proof risk function (BPRF), which estimates the level of risk closest to the null hypothesis that is consistent with available data. Here we illustrate the BPRF methodology for the evaluation of four exemplar risk–outcome pairs: smoking and lung cancer, systolic blood pressure and ischemic heart disease, vegetable consumption and ischemic heart disease, and unprocessed red meat consumption and ischemic heart disease. The strength of evidence for each relationship is assessed by computing and summarizing the BPRF, and then translating the summary to a simple star rating. The Burden of Proof methodology provides a consistent way to understand, evaluate and summarize evidence of risk across different risk–outcome pairs, and informs risk analysis conducted as part of the Global Burden of Diseases, Injuries, and Risk Factors Study. Main Exposure to different risk factors plays an important role in the likelihood of an individual developing or experiencing more severe outcomes from certain diseases, such as high blood pressure increasing the risk of heart disease or not having access to a safe water source increasing the risk of diarrheal diseases 1 . Understanding and quantifying the relationship between risk factor exposure and the risk of a subsequent outcome is therefore essential to set priorities for public policy, to guide public health practices, to help clinicians advise their patients and to inform personal health choices. Consequently, information on risk–outcome relationships can be used in the formulation of many types of public policies, including national recommendations on diet, occupational health rules, regulations on behavior such as smoking in public places, and guidance on appropriate levels of taxes and subsidies. As new evidence is continuously being produced and published, the systematic and comparable assessment of risk functions is a dynamic challenge. Up-to-date assessments of risk–outcome relationships are essential to, and a core component of, the Global Burden of Diseases, Injuries, and Risk Factors Study (GBD) comparative risk assessment (CRA) 1 , 2 , 3 , which aims to help decision-makers understand the magnitude of different health problems. Evidence on risk–outcome relationships comes from many types of studies, including randomized controlled trials (RCTs), cohort studies, case-control studies, cross-sectional analyses, ecological studies and animal studies. Each study type has characteristic strengths and weaknesses. For example, RCTs are the most robust method for dealing with confounding but are often conducted with strict inclusion and exclusion criteria, meaning that trial participants are unlikely to be fully representative of the general population, as well as being done over relatively short durations 3 , 4 , 5 . Case-control studies are well suited for understanding the risks linked to rare outcomes but may be subject to recall bias for past exposure 6 , 7 . Animal studies are widely used in evaluating the risks of consumer products and environmental risks but may not be generalizable to humans 8 . Study design and analysis impacts causal interpretation and understanding of the results 9 . When synthesizing evidence from different studies, strong assumptions—usually that of a log-linear relationship between risk and exposure—are often made to increase the mathematical tractability of the analysis 10 , 11 , 12 . Between-study heterogeneity—that is, disagreement in study-specific inferred relationships between risk exposure and outcome—is quantified in meta-analytic summaries, and has some effect on fixed-effects variance estimates, but is not otherwise used in the overall assessments of the uncertainty in risk–outcome relationships 12 , 13 . Risk factors associated with comparatively modest increases in the hazard are often questioned because of the potential for residual confounding 14 . Given the very mixed evidence landscape, it is perhaps not surprising that there are so many controversies in the literature 15 , 16 , 17 , 18 . While evidence is often heterogeneous, the need for clear guidance has led national advisory groups and international organizations to use expert committees to evaluate the evidence and formulate recommendations. The biggest advantage of expert groups is their ability to carefully consider nuances in the available evidence, but they are inherently subjective. For instance, expert groups across subfields of health science weight types of evidence differently, and even groups of experts within the same subfield may arrive at divergent conclusions. These expert groups often use meta-analyses of the available evidence, such as those produced by the Cochrane Collaborations 19 , as an input to their deliberations. Even Cochrane Reviews, however, allow authors to use a range of methodologies and approaches to studies on risk of bias, limiting comparability across risk–outcome pairs 19 . Tools have been produced to help standardize consideration of evidence, such as Grading of Recommendations, Assessment, Development and Evaluations (GRADE 20 , 21 ), but while very helpful, they cannot be implemented algorithmically. No quantitative assessment of the evidence can or should substitute completely for expert deliberation, but a quantitative meta-analytic approach that addresses some of the issues identified by GRADE and others could be a useful input to international and national expert committee considerations. Here, we propose a complementary approach, in which we quantify the mean relationship (the risk function) between risk exposure and a disease or injury outcome, after adjusting for known biases in the existing studies. Unlike existing approaches, our approach does not force log-linearity in risk functions or make additional approximations, such as midpoint approximations for ranges or shared reference groups 22 , 23 , 24 . To quantify the effect of bias, we considered risk of bias criteria that inform GRADE 20 , 21 , Cochrane Reviews 19 and evidence-based practice, and consulted widely outside of the Institute for Health Metrics and Evaluation, including with clinicians, physicians, medical and public health researchers and national health policy-makers (for example, former Ministers of Health). We encoded these variables that are used to assess risk of bias as potential study-level bias covariates within the proposed meta-analytic framework. This approach complements GRADE and Cochrane Reviews, which require analysts to assess and flag risks of bias. We then developed the burden of proof risk function (BPRF), which complements the mean risk and is defined as the smallest level of excess risk (closest to no relationship) that is consistent with the data. To aid interpretation of the results, we classify risk–outcome pairs into five categories (star ratings of one to five) based on the average magnitude of the BPRF. To illustrate this approach to assessing risk–outcome relationships, we provide four selected examples, showing both weak and strong risk–outcome relationships. Results Overview To support estimation of the BPRF, we developed a meta-analytic approach that addresses a number of issues that have previously limited interpretations of the available evidence. This approach relaxes the assumption that the relative risk of an outcome increases exponentially as a function of exposure, standardizes the assessment of outliers, explicitly handles the range of exposure in a study in both the ‘alternative’ groups (numerator) and ‘reference’ groups (denominator) of a relative risk, tests for systematic bias as a function of study design using automatic covariate selection, and quantifies between-study heterogeneity while adjusting for the number of studies. Using unexplained between-study heterogeneity and accounting for small numbers of studies, we estimate the BPRF as the 5th (if harmful) or 95th (if protective) quantile risk curve closest to the null (relative risk equal to 1). We flag evidence of the small-study effect (significant association between mean effect and standard error) as an indicator of potential publication or reporting bias. We evaluated the BPRF for 180 risk–outcome pairs in the GBD CRA framework. To simplify communication, we then computed the associated risk–outcome score (ROS) for each pair by averaging the BPRF across a relevant exposure interval and converted each ROS into a star rating from one to five. One star refers to risk–outcome pairs where a conservative interpretation of the evidence—accounting for all uncertainty including between-study heterogeneity—may suggest there is no association and two–five stars refers to risk–outcome pairs where a conservative interpretation of the evidence may suggest that, for harmful effects, average exposure increases excess risk relative to the level of exposure that minimizes risk from 0 to 15% (two stars; weak evidence of association), from >15 to 50% (three stars; moderate evidence of association), from >50 to 85% (four stars; strong evidence of association) and >85% (five stars; very strong evidence of association), and for protective effects, decreases excess risk relative to no exposure from 0 to 13% (two stars), from >13 to 34% (three stars), from >34 to 46% (four stars) and >46% (five stars). The corresponding ROS thresholds for both harmful and protective risks are <0 for one star, >0–0.14 for two stars, >0.14–0.41 for three stars, >0.41–0.62 for four stars and >0.62 for five stars. Of the 180 risk–outcome pairs investigated, 40 risk–outcome pairs were given a one-star rating, 72 pairs were given a two-star rating, 46 were given a three-star rating, 14 were given a four-star rating and 8 were given a five-star rating (Table 1 ). Here, we present results from each step of the evaluation process for four risk–outcome pairs to demonstrate how our methodology can be applied to pairs across the ROS spectrum and across a range of available study types and risk curve shapes, varying levels of between-study heterogeneity, and varying numbers of data points and studies. These four pairs also allow us to demonstrate how policy-makers should interpret our findings for both strong and weak risk–outcome relationships. Table 1 BPRF and ROS ranges associated with each star rating, and number of risk–outcome pairs assigned to each star rating Full size table Smoking and lung cancer (five stars) We used a standardized approach to search for and extract data from published studies on the relationship between pack-years smoked and the log relative risk of lung cancer, resulting in 371 observations from 25 prospective cohort studies and 53 case-control studies (three of them nested) reported from 1980 onwards (Fig. 1 ; step 1 in Methods ) 25 . The studies spanned a wide range of pack-years of smoking, from nearly one to over 112 pack-years. We found the 15th percentile of exposure in the reference group to be zero pack-years (and the 85th percentile of exposure among exposed groups in the cohort studies to be 50.88 pack-years (Fig. 1a,b ). On average, we found a very strong relationship between pack-years of smoking and log relative risk of lung cancer (step 2 in Methods ). At 20 pack-years, the mean relative risk (an effect size measure) was 5.11 (95% uncertainty interval (UI) 1.84–14.99), and at 50.88 pack-years (85th percentile of exposure) it was 13.42 (2.63–74.59) (Fig. 1b and Supplementary Table 1 ). The relationship is not log-linear, with declining effects of further pack-years of smoking, particularly after 40 pack-years. In the analysis of bias covariates (step 3 in Methods ), we adjusted data from studies that did not adjust for more than five confounders, including age and sex. There is enormous heterogeneity in the reported relative risk for lung cancer across studies (Fig. 1b ; step 4 in Methods ). In trimming 10% of observations, we identified observations both above and below the cloud of points, which we excluded (step 5 in Methods ). The mixed-effects models fit the data, that is, the reported uncertainty together with estimated between-study heterogeneity covers the estimated residuals, as Fig. 1c demonstrates. Even taking the most conservative interpretation of the evidence—the 5th quantile risk function including between-study heterogeneity, or the BPRF—smoking dramatically increases the risk of lung cancer (Fig. 1a,b ). There is evidence of potential reporting or publication bias (Fig. 1c ). The BPRF suggests that smoking in the range of the 15th–85th percentiles of exposure raises the risk of lung cancer by an average of 106.7%, for an ROS of 0.73 (step 6 in Methods ). These findings led us to classify smoking and lung cancer as a five-star risk–outcome pair. Fig. 1: Smoking and lung cancer. a , Log relative risk function for smoking and lung cancer. b , Relative risk function for smoking and lung cancer. c , A modified funnel plot for smoking and lung cancer showing the residuals (relative to 0) on the x axis and the estimated standard deviation (s.d.) that includes reported s.d. and between-study heterogeneity on the y axis. Full size image Systolic blood pressure and ischemic heart disease (five stars) We extracted 189 observations from 41 studies (39 RCTs, 1 cohort and 1 pooled cohort) quantifying the relationship between systolic blood pressure (SBP) and ischemic heart disease (Fig. 2 ) 26 . We included RCTs designed to compare the health effects of different levels of blood pressure. Head-to-head trials of drug classes or combinations not designed to achieve different levels of SBP were excluded. We calculated the 15th percentile of exposure in the cohorts and trials to be an SBP of 107.5 mm Hg and the 85th percentile to be 165 mm Hg (Fig. 2a,b ). The relationship is close to log-linear, although it appears to flatten out and deviate from the log-linear assumption over an SBP of 165 mm Hg (though the data are sparse over this level). An SBP of 140 mm Hg had a mean relative risk of ischemic heart disease of 2.38 (2.17–2.62) compared to 100 mm Hg, while an SBP of 165 mm Hg had a mean relative risk of 4.48 (3.81–5.26) compared to 100 mm Hg. (Fig. 2b and Supplementary Table 2 ). Trimming removed 10% of outlying observations with high relative risk at SBP levels between 125 and 180 mm Hg and low relative risk at SBP levels between 130 and 175 mm Hg (Fig. 2b ). In the analysis of bias covariates, we found that none had a significant effect. Because the RCTs and cohorts are very consistent and because there are many consistent studies within each type, between-study heterogeneity is small (Fig. 2a,b ). While there is little asymmetry in the funnel plot (Fig. 2c), we found statistically significant evidence of small-study bias using an Egger’s regression (Egger’s regression P value <0.05). Given the small between-study heterogeneity, the BPRF suggests that SBP in the range from the 15th to 85th percentile of exposure raises the risk of ischemic heart disease by an average of 101.36%, for an ROS of 0.70. These findings led us to classify SBP and ischemic heart disease as a five-star risk–outcome pair. Fig. 2: Systolic blood pressure and ischemic heart disease. a , Log relative risk function for systolic blood pressure and ischemic heart disease. b , Relative risk function for systolic blood pressure and ischemic heart disease. c , A modified funnel plot for systolic blood pressure and ischemic heart disease showing the residuals (relative to 0) on the x axis and the estimated standard deviation (s.d.) that includes reported s.d. and between-study heterogeneity on the y axis. Full size image Vegetable consumption and ischemic heart disease (two stars) Figure 3 summarizes the cohort data on vegetable consumption and ischemic heart disease using 78 observations from 17 cohort studies 27 . The relationship is not log-linear. We found that on average, vegetable consumption was protective, with the relative risk of ischemic heart disease being 0.81 (0.74–0.89) at 100 grams per day vegetable consumption compared to 0 grams per day (Supplementary Table 3 ). Incrementally higher levels of exposure are associated with less steep declines in relative risk compared to those at lower levels of exposure (Fig. 3b ). For this pair, trimming removed one observation that suggested a weaker protective effect size than the mean estimate, and seven observations that suggested a stronger protective effect than the mean estimate. Including between-study heterogeneity expanded the UI only slightly (Fig. 3a,b ), suggesting strong agreement between studies. In the analysis of bias covariates, three were found to have a significant effect: incomplete confounder adjustment, incidence outcomes only and mortality outcomes only. The funnel plot (Fig. 3c ) shows that after trimming, residual standard error (reflecting both study data variance and between-study heterogeneity) is within the expected range of the model. While there is little asymmetry in the funnel plot (Fig. 3c ), we found statistically significant evidence of small-study bias using an Egger’s regression (Egger’s regression P value = 0.044). The BPRF suggests that vegetable consumption in the range of the 15th to the 85th percentile lowers risk of ischemic heart disease by 12.10% on average (ROS of 0.13). This leads to vegetable consumption and ischemic heart disease being classified as a two-star pair. Fig. 3: Vegetable consumption and ischemic heart disease. a , Log relative risk function for vegetable consumption and ischemic heart disease. b , Relative risk function for vegetable consumption and ischemic heart disease. c , A modified funnel plot for vegetable consumption and ischemic heart disease showing the residuals (relative to 0) on the x axis and the estimated standard deviation (s.d.) that includes reported s.d. and between-study heterogeneity on the y axis. Full size image Unprocessed red meat and ischemic heart disease (two stars) We identified 43 observations from 11 prospective cohort studies on unprocessed red meat and ischemic heart disease (Fig. 4 ) 28 . At an exposure of 50 grams per day, the mean relative risk is 1.09 (0.99–1.18) compared to 0 grams per day, and at 100 grams per day, it is 1.12 (0.99–1.25) (Fig. 4b and Supplementary Table 4 ). In the analysis of bias covariates, we found that none had a significant effect. Trimming removed five observations that reported extreme values across the range of red meat consumption. There is no visual evidence or finding of potential publication or reporting bias (Fig. 4c ). Fig. 4: Unprocessed red meat consumption and ischemic heart disease. a , Log relative risk function for unprocessed red meat consumption and ischemic heart disease. b , Relative risk function for unprocessed red meat consumption and ischemic heart disease. c , A modified funnel plot for unprocessed red meat consumption and ischemic heart disease showing the residuals (relative to 0) on the x axis and the estimated standard deviation (s.d.) that includes reported s.d. and between-study heterogeneity on the y axis. Full size image For unprocessed red meat and ischemic heart disease, the exposure-averaged BPRF is 0.01, essentially on the null threshold (Fig. 4a ), equating to an ROS of 0.01, with a corresponding increase in risk of 1.04%. These findings led this risk–outcome pair to be classified as a (nominal) two stars, on the threshold between weak evidence and no evidence of association for the risk–outcome pair. Model validation To validate key aspects of the meta-regression tool, we ran detailed simulation experiments (step 7 in Methods ). We found that the approach proposed in this study outperformed existing approaches, particularly for non-log-linear relationships (Fig. 5 and Extended Data Figs. 1 – 6 ). Discussion Using a meta-analytic approach built using open-source tools, we estimated both the mean risk function and the BPRF for 180 risk–outcome pairs and assigned them a star rating based on the strength of the evidence (indicated by ROS that aggregate BPRF across standard exposure ranges) and severity of the risk. We achieved this by capturing the shape of the relationship between exposure and the risk of an outcome, detecting outliers using robust statistical methodology (trimming), testing and correcting for bias related to study design, and estimating between-study heterogeneity, adjusted for the number of studies. The BPRF is the level of elevated risk for a harmful factor (or the level of reduced risk for a protective factor) based on the most conservative (closest to null) interpretation compatible with the available evidence. It is a reflection of both the magnitude of the risk and the extent of the uncertainty surrounding the mean risk function. The four examples in the results section demonstrate the range of evidence, between-study heterogeneity and mean relative risks across risk–outcome pairs, and how these factors impact the BPRF and star rating. Importantly, only 22 of 180 pairs received a four- or five-star rating (12.22%), whereas 112 received a one- or two-star rating (62.22%). The BPRF and associated star ratings, as well as the background rates of burden for the outcomes of concern, are intended to be useful for informing individual choices on risk exposure. For example, harmful risk–outcome pairs with four- and five-star ratings are associated with an increase in risk of more than 50% for the exposed (and more than a 34% decrease in risk for protective risks), even based on the most conservative interpretation of the evidence. For these risks, the mean effect size is often much higher. Harmful risk–outcome pairs with three stars have average increases in risk ranging from more than 15% to 50% (and a decrease of at least 13–34% for protective risks), even in the BPRF, and may be much higher depending on the individual level of risk exposure. Further, some risks have high star ratings for multiple outcomes, such as high systolic blood pressure increasing risk of ischemic heart disease and stroke, and smoking increasing risk of lung cancer, aortic aneurysm, peripheral artery disease, laryngeal cancer and other pharynx cancer (all five-star pairs), which should be considered when making individual decisions around risk exposure. Conversely, individuals can reasonably pay less attention to risks with a one-star rating. These may be real risks with small but meaningful benefits for individuals if their exposure is reduced, but the existing evidence is too limited to make stronger conclusions. Of course, individual choice should also be informed by the background risk of an outcome for an individual and the totality of risk–outcome pairs associated with a risk; a five-star relationship for a rare outcome may not be something that an individual would choose to act on, whereas three-star ratings for one risk and a set of common outcomes may warrant more action. While the general public and committees formulating guidelines on individual behaviors—such as recommended diets—should pay attention to the star ratings, policy-makers should consider the impact of all risk–outcome pairs, not only those with high star ratings. These higher-star relationships should reassure decision-makers that the evidence supporting a risk factor is strong, but it would be unwise for decision-makers to ignore all one- and two-star risk–outcome pairs. The precautionary principle implies that public policy should pay attention to all potential risks. Lower star rating risk–outcome pairs may turn out to be null as evidence accumulates, but it is unlikely that a set of one-star risks will all turn out to be null. Public policy to address risks, even those where the BPRF indicates that risk is small or even nonexistent, will, on average, improve health. At the same time, investing in more widespread data collection for pairs with lower star ratings will reduce uncertainty and allow policy-makers to be more strategic in addressing potential risks (as star ratings may go up or down with more evidence). For example, due to very high heterogeneity between studies, a conservative interpretation of the available evidence suggests that there is weak to no evidence of an association between red meat consumption and ischemic heart disease. There is, therefore, a critical need for more large-scale, high-quality studies on red meat consumption so policy-makers can make better-informed decisions about how to prioritize policies that address this potential risk. Moreover, public policy should pay attention not only to the risk functions that are supported by evidence but the prevalence of exposure to those risks. For example, a two-star risk with high prevalence of exposure could pose a greater risk at the population level than a five-star risk with low prevalence of exposure. The GBD CRA 1 , 2 , 3 provides a framework for incorporating the BPRF, the prevalence of exposure and the background rates of specific outcomes to help policy-makers evaluate the importance of risk–outcome pairs across the full range of star ratings. In the future, risk–outcome pairs with one- and two-star ratings should be investigated further through more robust, well-powered research, especially for those risks where exposure and outcome are common, so policy-makers and individuals alike can better understand whether there is a real association between risk and outcome. The BPRF and associated star ratings have immediate applications for GBD and its users. For GBD 2020, 180 risk–outcome pairs have so far been analyzed using this approach. The remaining risk–outcome pairs will be evaluated using this meta-analytic approach in subsequent GBD rounds. Since different users will be interested in the GBD results focusing on certain star rating categories, we have developed online visualization tools ( ) that allow users to filter results by star rating. Providing dynamic tools with this capability will empower users with different thresholds for considering risk–outcome pairs and will allow broader audiences to access this information. These tools are intended to fill in a gap in the landscape of risk assessment accessibility and transparency. The standard approach to estimate the relationship between a risk and outcome has been to compute the mean across the universe of studies. We believe, however, that it is useful to report both the mean risk function and the BPRF, and that the more conservative interpretation may be more appropriate, particularly for exposures associated with small increases in risk, because of the risk of residual confounding. By including between-study heterogeneity in the uncertainty estimation and using this estimated uncertainty to compute a 5th or 95th quantile risk function (our BPRF), our risk assessment accounts for results that vary drastically across studies even after correcting for biases due to study design. This highlights the importance of accounting for unexplained between-study heterogeneity when estimating uncertainty and significance testing. In particular, when the BPRF spans zero (that is, the risk is one star), a conservative interpretation of the evidence is consistent with no association between the risk and the outcome. We argue that the field should eventually move to incorporating between-study heterogeneity into significance testing of the mean function. Our meta-analytic approach uses splines to estimate the shape of the risk function without imposing a functional form such as log-linearity, and can be widely applied to other risk–outcome pairs not included in this analysis. This flexibility is an important strength of our approach because many risk–outcome pairs do not have a log-linear relationship. When there are strong threshold effects, log-linear risk functions can exaggerate risk at higher exposure levels and obfuscate important detail at lower exposure levels. This more flexible approach helps identify the true shape of the risk function. Previously, the main challenge had been that if the assumption of log-linearity is relaxed, the level of risk exposure matters, so comparisons between an exposed group and a reference group needed to take into account the range of exposure in each group. We dealt with this problem directly by integrating the risk function over a range of exposures and including this mechanism in the likelihood. Our approach may be of use for meta-analyses in many areas, even if the analyst is only interested in the mean function. The model validation analysis (step 7 in Methods ; Extended Data Figs. 1 – 7 ) demonstrates that our approach captures non-log-linear functions with significantly greater accuracy compared to existing dose–response meta-regression tools while still capturing log-linear relationships when present. The advantage in accuracy over existing tools increases in data-sparse cases that have fewer studies and observations. Our approach is also robust to bias covariates, such as study type, since we explicitly test for the impact of these covariates using a Lasso framework, and then adjust the estimated risk curves using covariates that are found to be statistically significant. The proposed framework uses robust methodology (trimming) to make the approach robust to outlying observations. Sensitivity analyses (step 8 in Methods ; Extended Data Figs. 8 , 9 and Supplementary Table 5 ) show that trimming 10% of the data makes estimation more stable and reliable. Trimming automatically focuses on unexplained large errors, and in practice does not remove ‘gold standard’ data points. The intention of this approach to evaluating risk–outcome associations is to complement the existing GRADE and Cochrane Review frameworks for assessing evidence and making recommendations. Our suite of meta-analyses address many of the limitations of GRADE and Cochrane by focusing on study design factors whose impact can be assessed quantitatively from the body of studies informing the analysis rather than requiring analysts to assess and flag risks of bias. Thus, we believe our approach contributes important information for expert deliberation. This framework has a number of limitations. First, there are qualitative considerations about study design and execution that may be hard to capture in a set of structured risk of bias covariates. Our framework for adjusting for study design is necessarily limited to observable study design characteristics. Further, while our approach offers a rigorous way to combine results from different studies and different types of studies, fundamentally discordant evidence or types of evidence, such as chemical experiments, do not lend themselves to direct inclusion in this framework. Second, the trimming approach requires a user-specified level of outliers, and although fitting 90% of the data works well in practice, automated techniques to estimate the potential level of outliers in the dataset would strengthen the approach. Third, while our approach can explicitly test for potential publication or reporting bias related to the association of reported effect sizes and standard errors, other types of publication bias are more challenging to evaluate, namely when studies are more consistent with each other than expected by chance. In these cases, the Fisher information matrix approach is still helpful, particularly when the number of studies is small, because it is guaranteed to provide a quantile of heterogeneity even if the heterogeneity parameter is estimated to be 0. Fourth, our study bias covariates cannot fully capture and correct for bias if all or even the vast majority of the input studies are biased. Fifth, including pooling studies in the meta-regression, though extremely useful in providing robust estimates of mean effects, may artificially decrease our estimates of between-study heterogeneity because many of these studies do not publish measures of between-study heterogeneity across cohorts. We partially addressed this issue by using the Fisher information matrix approach to estimate the plausible between-study heterogeneity, but more emphasis on between-study heterogeneity in future pooling studies would allow users to better interpret the generalizability of the mean effects. Sixth, to avoid overfitting bias covariates, we used Gaussian priors on the bias covariates so that these relationships were only detected when there were sufficient studies supporting the estimate of the bias. Alternative priors could increase or decrease the biases that are detected. Seventh, we estimated the BPRF and translated this into a star rating for each risk–outcome pair. There may be ranges of exposure within which there is a marked increase in the BPRF, but over most of the range, the increase in risk is small. Giving different star ratings to different ranges of exposure would, however, add a further degree of complexity that we sought to avoid. Eighth, we had no direct way of introducing or including animal studies and are thus agnostic to that evidence category. We developed a data-driven meta-analytic approach, using open-source computational tools cited in this study, that identifies the shape of the risk–outcome relationship and robustly quantifies between-study heterogeneity after correcting for bias correlated with attributes of study design. We used this risk function to estimate both the mean relationship and BPRF for 180 risk–outcome pairs. The BPRF provides the most conservative interpretation of the severity of risk based on the available evidence. Using the BPRF, we classified risk–outcome relationships into five categories based on the strength of the Burden of Proof relationship. This standardized tool cannot address every nuance in the interpretation of the available data but can quantify a wide range of dimensions previously addressed in more subjective and qualitative ways, particularly in conjunction with information on risk exposure prevalence and outcome burden. We intend to update these risk functions over time so that they reflect the latest available evidence, including adding new risk–outcome relationships as new evidence is published. The BPRF and associated star ratings improve the field of comparative risk assessment and increase the transparency of expert deliberation of human health risks. The star ratings can be used to assess risk and inform individual and policy-level decisions around risk exposure prevention, public health guidance and personal health choices. Fig. 5: Model validation for the data-rich scenario. a , Estimated risk curves across 100 realizations for all methods for non-log linear risks. b , Estimated risk curves across 100 realizations for all methods for log linear risks. Mrtool is the method used in this paper. Dosresmeta_1stage_ncs refers to the Dosresmeta package with a natural cubic spline, while Dosresmeta_1stage_qs refers to the same tool with a quadratic spline. Dosresmeta_2stage refers to the the 2 stage approach in Dosresmeta. Metafor refers to a standard package that assumes a log linear relationship. (See ’Model validation’ for more details). Full size image Methods Overview Our meta-analytic approach followed six main steps: (1) search and extract data from published studies using a standardized approach; (2) estimate the shape of the exposure versus relative risk relationship, integrating over the exposure ranges in different comparison groups and avoiding the distorting effect of outliers; (3) test and adjust for systematic biases as a function of study attributes; (4) quantify remaining between-study heterogeneity while adjusting for within-study correlation induced by computing the relative risks for several alternatives with the same reference, as well as the number of studies; (5) assess evidence for small-study effects to evaluate a potential risk of publication or reporting bias; and (6) estimate the BPRF, quantifying a conservative interpretation of the average risk increase across the range of exposure supported by the evidence to compute the ROS, and map the ROS into five categories of risk. Zheng and colleagues 29 published the technical developments required to implement this approach, which are also disseminated using open-source Python libraries 30 , 31 . We validated the model through simulation studies, and then applied our meta-analytic approach to 180 risk–outcome pairs. We present our findings for four pairs that demonstrate a range of risk relationships (smoking and lung cancer, systolic blood pressure and ischemic heart disease, vegetable consumption and ischemic heart disease, and red meat consumption and ischemic heart disease). This study complies with the Guidelines on Accurate and Transparent Health Estimate Reporting recommendations (Supplementary Table 6 ) 32 . Searching for and extracting published data following a standardized protocol For each risk–outcome pair, we used standard search strings or Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to identify data from research databases and other sources. Data were then extracted from studies that met inclusion criteria using uniform extraction procedures. We mapped all outcomes in the data to GBD cause categories to produce a standard set of outcomes. When necessary, this process included mapping a standard set of other, often narrower, outcomes to GBD causes. When the lower end of exposure levels was not reported, we used the 15th percentile of known exposure in microdata sources; when the upper end was not reported, we used the width of the adjacent exposure interval; and when neither end was reported, we used the GBD global exposure distribution. See Supplementary Tables 7 – 10 for detailed information on the included data sources for each example risk–outcome pair. For a full description of how each of the four selected risk–outcome pairs were defined and measured, how data inputs were identified and assessed for eligibility, the summary results from input studies and other information on conducting systematic reviews for each risk, see Dai et al. 25 , Razo et al. 26 , Stanaway et al. 27 , and Lescinsky et al. 28 . Detailed PRISMA checklists are also included in the supplementary information of each of these articles. Estimating the shape of the risk–outcome relationship Most classic epidemiological analyses of dose–response risk relationships have either assumed the relationship between risk and outcome to be log-linear or converted continuous exposure variables into dichotomous exposure categories. This assumption simplifies the analysis considerably. Unfortunately, although assuming a log-linear relationship is analytically convenient and allows for the use of simple open-source tools 10 , it is not necessarily biologically or clinically plausible (see step 7 in Methods for more details). For some risks, such as smoking, the log relative risk of the outcome flattens at higher exposures. For others, the log relative risk curves are J-shaped. We therefore chose to estimate the shape of the relationship directly from the data using a regularized spline. Nonlinear modeling of dose–response relationships brings new challenges that are not present in the log-linear case. Rather than simply using midpoints of the data, we need to account for interval exposures in reference and alternative exposure groups. Moreover, the observation mechanism that accounts for this level of detail is nonlinear. Model stability becomes an important problem to make the approach systematically applicable across a broad range of cases. Accounting for outliers in the data becomes more important. Finally, capturing between-study heterogeneity in a tractable and stable way is important. Our approach is sensitive to the level of and difference in exposure, and explicitly handles the lack of common reference groups and exposure ranges for reference and alternative groups that are present in a vast majority of risk–outcome analyses. The statistical approach includes four aspects that make it useful for estimating risk–outcome relationships. Basis splines, measurement mechanism and shape constraints We used a Bayesian regularized spline to obtain the general shape of the nonlinear relationship. Basis splines represent nonlinear curves as linear combination of recursively generated basis elements 33 . The basis elements were recursively generated using piecewise smooth polynomials, and were roughly localized to certain regions of the exposure variable in the data. Most of the time, quadratic or cubic polynomials were used, often with linear tails in the presence of sparse data. This approach allowed the common restricted cubic spline and constraints on the shape of the relationship (including nondecreasing and nonincreasing). Given basis functions f 1 , …, f k and coefficient vector β = ( β 1 ,..., β k ), the final curve is obtained as a β -linear combination $${\rm{signal}} = \beta _1f_1 + \cdots + \beta _kf_k.$$ Specifically, for any given exposure ( x ), the prediction using the spline model is given by $${\mathrm{signal}}\left( x \right) = \beta _1f_1\left( x \right) + \cdots + \cdots + \beta _kf_k\left( x \right) = \left\langle {\mathbf{{X}},\beta } \right\rangle$$ (1) where X is a vector containing ( f 1 ( x ), …, f k ( x )). Derivatives and integrals of splines can likewise be expressed as linear combinations of spline coefficient β . For additional details about B -splines see Zheng et al. 29 Many studies of dose–response relationships report relative risks between categories defined by intervals of consumption. In mathematical notation, these observations are given by $$y_{ij} = \frac{{\frac{1}{{d_{ij} - c_{ij}}} {\smallint }_{c_{ij}}^{d_{ij}} f\left( x \right){\mathrm{d}}x}}{{\frac{1}{{b_{ij} - a_{ij}}} {\smallint }_{a_{ij}}^{b_{ij}} f\left( x \right){\mathrm{d}}x}},$$ (2) where y ij is the reported relative risk corresponding to measurement j in study i , [ a ij , b ij ] delineates the reference group exposure interval, and [ c ij , d ij ] delineates the alternative group exposure interval. When f ( x ) is represented using a spline, each integral is a linear function of β similar to equation ( 1 ). The model (equation ( 2 )) is then a ratio of linear functions, $$y_{ij} = f_{ij}\left( \beta \right): = \frac{{\left\langle {X_{ij}^1,\beta } \right\rangle }}{{X_{ij}^2,\beta }}.$$ (3) with the log relative risk given by $$\ln \left( {y_{ij}} \right) = \ln \left( {\left\langle {X_{ij}^1,\beta } \right\rangle } \right) - \ln\left( {\left\langle {X_{ij}^2,\beta } \right\rangle } \right).$$ (4) Equation ( 4 ) is a nonlinear function of the spline coefficients β . When studying dose–response relationships, we allow for shape constraints of the inferred mean response. For example, for some harmful risks, such as smoking and air pollution, we assume the relative risk is monotonically increasing with exposure. To regularize the splines, and capture biologically plausible limits, we also allow a maximal derivative constraint, which is similar to penalizing total variation or limiting the spline degree. To introduce each of these constraints, we used the fact that derivatives of splines are linear functions of spline coefficients, similar to equation ( 1 ). Monotonicity We imposed monotonicity constraints using several linear inequality constraints based on exemplar exposures. Given an exemplar exposure x i , the requirement that the slope of the spline at exposure x i , be non-negative can be formulated as $$\left\langle {X_{ij},\beta } \right\rangle \ge 0$$ for a computed vector X i . Linear inequality constraints were strictly enforced by the optimization solver used to fit the model, see Zheng et al. 29 . Robust trimming strategy To make the estimation of the overall relationship insensitive to potential outlying studies or observations within studies, we applied a robust, likelihood-based statistical approach—least trimmed squares (LTS) 34 —to our mixed-effects models 29 . The goal of robust statistical methods is to ensure that estimates are robust to outlying observations. Trimming approaches form a subclass of robust statistical methods, and LTS was originally developed in the context of linear regression 35 . LTS works by classifying observations into a majority of inliers and minority of outliers while simultaneously fitting the model with respect to which the inlier/outlier classification is made. Compared with other robust approaches, such as M-estimators 36 , trimming methods are more effective in limiting influence than outliers, and have a high breakdown point 37 , that is, the proportion of the data that can be arbitrarily corrupted before the estimator becomes invalid. Trimming estimators have been applied to a broad range of problems, from linear regression 34 to high-dimensional sparse regression and general machine learning problems 38 . In the context of mixed-effects models, trimming methods are far and away the most effective robust tools currently available for meta-analysis 29 . In practice, the approach requires only a specified inlier proportion, which was set to 90% across all examples, that is, we fit the 90% most self-coherent data points. Using this approach, we trimmed 10% of the observations as part of the model fitting process, simultaneously discovering and fitting the most self-coherent 90% of the observations 29 . Numerical studies in data-rich cases have shown that quality of estimation is unaffected by trimming, even when there are no outliers in the data 38 . In the meta-analytic regime, the 90% level is a heuristic that balances the sparsity of available data with the need to improve estimates in the presence of outliers. This step also substantially decreased the number of risk–outcome pairs with evidence of residual publication or reporting bias. Spline ensemble To make the risk function estimates robust to knot placement, we created 50 models based on random knot placement samples. Spline estimates depend on the choice of spline parameters, including spline degree, number of knots and knot placement. To mitigate the effect of spline parameter selection on results, we developed an ensemble approach over knot placement, so that the modeler only had to specify the spline degree and number of knots. Given the degree and number of knots, we automatically sampled a set of knot placements for a feasible knot distribution. For each knot placement, we fit a spline (including nonlinear measurements, shape constraints and trimming), evaluated each resulting model by computing its fit and curvature, and aggregated the final model as a weighted combination of the ensemble. Sampling knots from simplex We prefixed a minimal set of rules that describe a feasible set from which to sample knots, and uniformly sample from this set. Given a number of knots, the rules specify feasible ranges for each knot and feasible gaps between knots. Given an interval \(\left[ {t_0,t_k} \right]\) delimited by terminal knots (which are always the minimum and maximum of the data), the feasible region of the interior knots \(t_1, \ldots ,t_{k - 1}\) is denoted by $$t_i \in \left[ {a_i,b_i} \right],\quad {\mathrm{for}}\,i = 1, \ldots ,k - 1,\quad t_i - t_{i - 1} \in \left[ {c_i,d_i} \right]\quad {\mathrm{for}}\quad i = 1, \ldots ,k.$$ We enforced the rules $$a_i \ge t_0,\,b_i \le t_k,\quad c_i \ge 0,\quad {\sum} {c_i \le t_k - t_0.}$$ The set of knot placements that satisfy these four rules form a closed polyhedron (a volume in high-dimensional space delineated by hyperplanes). We calculated the vertices of the polyhedron using the double description method in ref. 39 , and uniformly sampled knot placements from within the polyhedron. Each knot placement yielded a model, fit using the trimmed constrained spline approach. Ensemble performance evaluation Once the ensemble was created, we scored the resulting risk curves using two criteria: model fit (measured using the log-likelihood) and total variation (measured using the highest order derivative). These scores balanced competing objectives of fit and generalizability. Once we had these scores, denoted as s 1 and s 2 , we normalized them to the range [0,1]: $$v_i = \frac{{s_i - {{{\mathrm{min}}}}\left( {s_i} \right)}}{{\max \left( {s_i} \right) - {{{\mathrm{min}}}}\left( {s_i} \right)}}$$ and applied a logistic transformation. The transformation was used to make the scoring meaningful even in the presence of spurious curves in a large ensemble. We then multiplied the scores to down-weight models that are low under either criterion (fit or total variation). The final weights are normalized to sum to 1. Using a weighted combination of these metrics, we weighted the 50 models to create the ensemble model. New nonlinear covariates For risk–outcome pairs with nonlinear relationships, we evaluated exposure levels, since this information matters for non-log-linear pairs. To do this, we took advantage of the spline model and directly captured the typical data-generating mechanism. Specifically, we used the final model that we had estimated using the robust spline ensemble to generate a nonlinear dose–response curve, which we encoded into new nonlinear ‘signal’ covariates that were later used to enable linear mixed-effects analyses. Once the nonlinear estimation was complete, the log relative risk for each data point was a function of four parameters: $$F\left( {a_{ij},b_{ij},c_{ij},d_{ij}} \right) = \frac{{\frac{1}{{d_{ij} - c_{ij}}}\mathop {\smallint }\nolimits_{c_{ij}}^{d_{ij}} \hat f\left( x \right){\mathrm{d}}x}}{{\frac{1}{{b_{ij} - a_{ij}}}\mathop {\smallint }\nolimits_{a_{ij}}^{b_{ij}} \hat f\left( x \right){\mathrm{d}}x}}$$ where \(\hat f\) is the nonlinear function obtained by estimating spline coefficients \(\hat \beta\) , see equation ( 4 ), \(\left[ {a_{ij},b_{ij}} \right]\) delineates the reference group exposure interval and \(\left[ {c_{ij},d_{ij}} \right]\) delineates the alternate group exposure interval. We produced two new nonlinear covariates for fixed and random effects. The new nonlinear fixed-effects covariate, denoted signal f , is given by $${\mathrm{signal}}_{ij}^{\mathrm{f}} = F\left( {a_{ij},b_{ij},c_{ij},d_{ij}} \right).$$ (5) The new nonlinear random effect covariate, denoted by signal r , is given by $${\mathrm{signal}}_{ij}^{\mathrm{r}} = F\left( {t,t,c_{ij},d_{ij}} \right),$$ (6) where t denotes a fixed reference, for example, the theoretical minimum risk exposure level. Using this innovation, we implemented further stages of analysis using linear mixed-effects modeling. In particular, at the end of the nonlinear stage we fit the linear mixed-effects model using only the new nonlinear covariates: $$y_{ij} = {\mathrm{signal}}_{ij}^{\mathrm{f}}\beta _s + {\mathrm{signal}}_{ij}^{\mathrm{r}}u_i + {\it{\epsilon }}_{ij}$$ (7) where \({\it{\epsilon }}_{ij}\approx N\left( {0,\sigma _{ij}^2} \right)\) are known by each observation, β s is a scalar linear covariate multiplier on the signal f covariate, and u i is a random study-specific slope on the signal r covariate with unknown variance γ . The posterior for β s in equation ( 7 ) was used as a reference for the prior in bias covariate selection, described in step 3 of the Methods . The relative risk between two exposure groups is a ratio of integrals of the spline across two specified intervals, so we used this exact nonlinear mechanism to inform the fit 29 . Data from studies usually compare outcome rates in one exposure alternative group to those in a separate reference group. For example, in diet cohort studies, it is common to compare the highest quartile of exposure to the lowest quartile of exposure. In trials of anti-hypertensives, the comparison of outcome rates is between the level of blood pressure in the intervention group and the level in the control group. Cohort studies of BMI often report rates in one range of BMI to a variety of reference groups, such as 20–21, 20–25 or 23–25. Finally, for our visualizations (Figs. 1 a – 4a ), we plotted each data point with x value at the midpoint exposure of the alternative group, and y value corresponding to the sum of the log relative risk and estimated curve evaluated at the midpoint of the reference group. These visualizations allow the standard assessment of fit quality, with a perfect fit corresponding to the estimated nonlinear relationship passing through the data. Testing for bias across different study designs and characteristics Following the approach of the GRADE criteria 40 , we quantified common sources of bias across six domains: representativeness of the study population, exposure, outcome, reverse causation, control for confounding and selection bias. In the illustrative cases presented here, these variables were quantified for each study during the study extraction phase. For the set of studies on a risk–outcome association, we tested systematic variation as a function of these risk of bias variables through meta-regression. We converted the dose–response relationship identified in step 1 into a new signal covariate, effectively linearizing the non-log-linear relationship. For each bias covariate x (coded as an indicator variable), we defined a corresponding interaction covariate (that is, an effect modifier): $$y_{ij} = {\mathrm{signal}}_{ij}^{\mathrm{f}} \times \left( {\beta _s + x_{ij}^1\beta _1 + \cdots + x_{ij}^k\beta _k} \right) + {\it{\epsilon }}_{ij}$$ that modified the slope of the signal covariate. We then tested risks of bias of the effect modifiers through linear meta-regression. To be included, every bias covariate must have some studies that are the gold standard (that is, at the standard of the best studies that have been conducted) for that covariate, otherwise it is not possible to incorporate it into the regression framework. Further, in considering potential covariates, we enforced that every categorical covariate had at least two studies in each category. Since bias covariates were already study specific, we only considered the fixed-effects model in bias covariate selection. We used a robust approach to test for bias that limited the risk of overinterpreting differences with limited numbers of studies. We used the Lasso 41 , 42 approach—which augments the least squares loss typically solved in a linear regression by penalizing the sum of absolute values of the bias covariate multipliers—to obtain a ranked list of bias covariates using the following equation: $$\begin{array}{l}\mathop {{\min }}\limits_\beta \mathop {\sum }\limits_{i,j} \frac{1}{{2\sigma _{ij}^2}}\left( {y_{ij} - {\mathrm{signal}}_{ij}^{\mathrm{f}}\times\left( {\beta _s + x_{ij}^1\beta _1 + \cdots + x_{ij}^k\beta _k} \right)} \right)^2\\ + \frac{1}{2}\beta ^T{{\varSigma }}^{\left\{ { - 1} \right\}}\beta + \lambda \left\| \beta \right\|_1\end{array}$$ (8) where β contains specifically bias covariate multipliers, Σ is a diagonal matrix linked to the posterior on β s from the basic linear model equation ( 7 ), and the term \(\lambda \left\| \beta \right\|_1\) penalizes the sum of the absolute values, pushing the bias covariate multipliers β to 0, with a strength determined by λ (ref. 42 ). We then selected bias covariates based on their Lasso ranking, starting with a high value of λ . We then added the selected covariates to the linear meta-regression model one at a time, following the Lasso ranking. To make the selection stable in the face of sparse data, we tested for significance of covariates using a Gaussian prior that biased all bias coefficients to 0 with a strength proportional to the posterior of the main dose–response relationship. If the coefficients were significant, they stayed in the model as the process continued. We terminated the process when the last added bias covariate was no longer significant after accounting for ‘signal’ and any higher-ranked covariates already in the model. We predicted the risk function using the values of the included bias covariates that reflected the preferred level of the covariate, such as the highest level of control for confounding. Supplementary Tables 11 – 14 provide study-specific information on study quality by risk–outcome pair. Quantifying between-study heterogeneity and accounting for heterogeneity, uncertainty and small numbers of studies Estimation of between-study heterogeneity is an important aspect of meta-analysis. It reflects the variation between studies and consistency across literature. After the selection procedure, we fit a final linear mixed-effects model that included the signal and selected bias covariates. Division by a common referent in the typical measurement mechanism induces correlation, specifically an intercept shift in log relative risk space; we therefore used a random intercept in the mixed-effects model to account for this induced within-study correlation. To capture the between-study heterogeneity, we used a study-specific random slope with respect to the signal model so that the random effect for each study effectively scaled the nonlinear relative risk curve. Formally, we fit a linear mixed-effects model of the form $$y_{ij} = {\mathrm{signal}}_{ij}^{\mathrm{f}} \times \left( {\beta _s + x_1\beta _1 + \cdots + x_k\beta _k} \right) + {\mathrm{signal}}_{ij}^{\mathrm{f}}u_i + {\it{\epsilon }}_{ij}$$ where \({\it{\epsilon }}_{ij}\approx N\left( {0,\sigma _{ij}^2} \right)\) are the reported observation standard errors, and u i are random effects with a common unknown variance, $$u_i\approx N\left( {0,\gamma } \right).$$ Parameters β and γ were estimated simultaneously using maximum likelihood; see Zheng et al. 29 for more details. We used the same prior on bias covariates in this analysis as we used in equation ( 8 ), that is \(\beta \approx N\left( {0,\varSigma } \right)\) . For log-linear relative risks, this modeling choice reduced to the classic analysis, where the random slope with respect to exposure was equivalent to the random intercept for log-linear relative risk. To account for the small-studies problem—that is, in the setting of small numbers of studies, between-study heterogeneity ( γ ) can easily be underestimated 43 , and in particular the estimate may be 0 when too few studies are available—we quantified the uncertainty in heterogeneity estimation 13 . This estimate allowed a quantile of the heterogeneity parameter to be used, increasing the robustness of the estimate against the small-study problem. Among several alternatives in the literature 44 , 45 , we used the Fisher information matrix (FIM) 44 to estimate the uncertainty of the heterogeneity. The FIM is weakly dependent on observed data, but is sensitive to the nonlinear relationship, selected bias covariates, reported standard errors and the number of studies. The final UIs we report are composed of two components: (1) posterior uncertainty corresponding to fixed effect β s and (2) 95% quantile of γ , which depends on the estimate of γ and the estimate of the variance of γ using the inverse of Fisher information. The sensitivity analysis shows that small sample size alone did not have a significant effect on the BPRF. Evaluating potential for publication or reporting bias A significant association between mean effect and standard error may indicate potential for publication or reporting bias, or methodological differences between large and small studies, which likewise lead to biased results. Publication bias is an important issue in meta-analysis 46 , and a formal test is typically done in addition to visual inspection of the funnel plot to decrease the chances of flagging apparent bias due to chance alone. In the proposed approach, we checked whether the standard errors were significant predictors of the observations in the presence of the signal and bias covariates. To detect publication bias, we used a data-driven approach known as Egger’s Regression 47 . The approach detects if there is a significant correlation between the residuals and their standard errors. When Egger’s Regression failed to detect significant evidence of publication bias, we terminated the process. While we identified these pairs as having potential for publication or reporting bias, we followed the general literature and did not incorporate any correction to the risk function based on this finding. Estimating the BPRF The combined uncertainty for the mean, estimated between-study heterogeneity, and 95th quantile of the between-study heterogeneity obtained from the FIM estimate were used to generate a BPRF. The BPRF is defined as either the 5th (for harmful risks) or 95th (for protective risks) quantile curve closest to the line of relative risk equal to 1 (the null), and can be interpreted as the smallest harmful or protective effect at each level of exposure consistent with the available evidence. In the range of exposures defined by the 15th and 85th percentiles of exposure levels observed for each risk across available studies, the ROS is defined as the signed value of the average log BPRF. For example, a log BPRF of 0.4 for a harmful risk (where null = 0) and of –0.4 for a protective risk would both have an ROS of 0.4 because the magnitude of the log relative risk is the same. In contrast, for risk–outcome pairs with a BPRF opposite the null from the mean risk (that is, the BPRF suggests that the relationship is opposite of the expected relationship—a BPRF below 1 for a harmful risk and above 1 for a protective risk), ROS would be calculated as negative. This definition is symmetric for harmful and protective risks since the null corresponds to a log relative risk of 0. The ROS provides a single summary of the log relative risk in the range of exposure supported by the available studies, with a higher positive ROS always corresponding to a stronger relationship and a negative ROS corresponding to the situation where the available evidence fails to reject the null. For example, ROS can be negative when between-study heterogeneity is large and the relative risk function close to 1. We tested alternative ranges of exposure such as the 10th and 90th percentiles and the 5th and 95th percentiles, and the correlation of the resulting ROS with the 15th and 85th percentiles across 180 risk–outcome pairs evaluated in GBD 2020 was 0.984 and 0.979, respectively. All risk–outcome examples presented here reflect continuous dose–response relationships. However, the risk score concept was extended to the binary risk–outcome pairs among the full 180 pairs in the analysis. The analysis for binary pairs was simpler than for continuous pairs because relative risks are comparisons between exposed and unexposed groups. The lower envelope of the log relative risk was defined analogously to the continuous risk as the 5% quantile for the effect size that included effect size uncertainty and between-study heterogeneity obtained using the 95% quantile of γ . To account for the fact that continuous risk–outcome analysis averages over an exposure domain, we modified the binary ROS. In the log-linear situation under basic assumptions, the averaging process reduced the score by a factor of two compared to a binary group definition that did not account for exposure. To make the continuous and binary scores comparable, we therefore divided the binary ROS by two. To guide policy-makers and research funders when making broader comparisons across risk–outcome relationships, we converted the ROSs into star-rating categories. In this schema, one-star risks are those for which the ROS is negative and therefore the risk–outcome pair is not significant in the BPRF framework, indicating that a conservative interpretation of the evidence fails to find a significant association. We further divided the positive ROSs into ranges 0.0–0.14, >0.14–0.41, >0.41–0.62 and greater than 0.62, and assigned each range a star rating from two (0.0–0.14) through five (>0.62). Under a conservative interpretation of the evidence, exposure to a harmful risk in the average range of exposure increases the risk of the disease outcome by less than 0% for one-star pairs, 0–15% for two-star pairs, >15–50% for three-star pairs, >50–85% for four-star pairs and greater than 85% for five-star pairs compared to no exposure. Likewise, exposure to a protective risk in the average range of exposure decreases the risk of the subsequent outcome by less than 0% for one-star pairs, 0–13% for two-star pairs, >13– 34% for three-star pairs, >34–46% for four-star pairs and greater than 46% for five-star pairs compared to no exposure. Model validation Method comparison We validated key aspects of the meta-regression model using detailed simulation experiments. To check the accuracy of estimating non-log-linear dose–response relationships, uncertainty and associated BPRFs, we simulated three scenarios: (1) many studies with many data points per study (30 studies, each with 4–9 observations), (2) many studies with few data points per study (30 studies, each with 1–4 observations) and (3) few studies with few data points per study (10 studies, each with 1–4 observations). In each scenario, we simulated log-linear and non-log-linear ground truth data risk functions, with three levels of between-study heterogeneity, characterized by \(\gamma \in \left\{ {0.0,0.01,0.02} \right\}\) . For each of these 18 combinations, we generated 100 dataset realizations to ensure that summary metrics accounted for stochastic error, for a total of 1,800 simulations. For exposures, we used a beta distribution, supported on \(x \in \left[ {0,1} \right]\) , with density proportional to $$x^{ \alpha - 1}(1 - x)^{\beta - 1}$$ with parameters α = 1 and β = 3. Using this exposure distribution, we were more likely to sample smaller exposures, giving wider range exposures in the tail of the distribution. For each simulation, we compared the results of the approach developed here with results obtained using existing approaches with available open-source tools: log-linear meta-analysis implemented in the metafor package 10 , a meta-analysis package for linear models; and two-stage 22 and one-stage 23 approaches for dose–response meta-analysis, both implemented in the dosresmeta package 24 . The metafor package assumes log-linear models. We used midpoint approximations to obtain data points and used weighted least squares to summarize data to have one measure per study, as needed by metafor. The dosresmeta two-stage approach first fits a spline model for each study, then performs a meta-analysis on the estimated coefficients. We used a quadratic polynomial (the simplest model that approximates a quadratic spline) in all examples, since the complexity is limited by the necessity of fitting a model for each study. We used standard midpoint approximations as this method does not allow range exposure integration. The dosresmeta one-stage approach pools the data and does a random spline meta-analysis. Here, we compared two types of splines, the quadratic spline and the natural cubic spline, both available to practitioners who use the tools. For the proposed approach presented in this paper, we used a quadratic spline with linear tails, the ratio model and heterogeneity estimation with uncertainty obtained from the FIM. The results were tabulated across simulations to get aggregate accuracy for both mean risk function and BPRF estimation (measured using root mean squared error). Figure 5 previews the results corresponding to the first (data-rich) scenario, to highlight the advantages of the proposed methods for non-log-linear risks. When the relationship is log-linear, our proposed approach and the metafor package (tailored for log-linear models) show somewhat better performance than competing approaches using more complex spline models. That said, the metafor package’s advantages decreased (particularly for the BPRF estimation) with increased heterogeneity. For non-log-linear relationships, our proposed approach very substantially improves on available methods, getting uniformly better performance across all scenarios and simulated between-study heterogeneity. Modeling the risk as log-linear (necessary to apply metafor) failed to capture salient features. Competing spline-based alternatives were unable to account for range exposure data, which resulted in underestimation of the mean curve. Due to the wide ranges and data sparsity at the tail, one-stage methods struggled to control the tail behavior of the splines used to estimate the risk functions. See Extended Data Figs. 1 and 2 for detailed results from the simulation of scenario one. For the log-linear experiments in the second scenario (30 studies, 1–4 observations each), all approaches were comparable; in this sparser setting data, controlling spline tail behavior made a larger difference than in scenario one. The metafor package was on par with the proposed approach, and both were competitive compared to the one-stage and two-stage methods. For non-log-linear pairs, the proposed approach performed substantially better than all alternatives. Modeling the risk as log-linear failed to capture salient features. Competing spline-based alternatives were unable to account for interval exposure data comparing different references and alternative groups across studies. See Extended Data Figs. 3 and 4 for detailed results from the simulation of scenario two. The findings from the third scenario (10 studies, 1–4 observations each) were in line with those from the second scenario. In this data-sparse case, splines and polynomials in one-stage and two-stage methods became more unstable. Full results for scenario three are given in Extended Data Figs. 5 and 6 . Heterogeneity estimation simulation We validated the utility of the FIM quantile in correcting the well-known problem of underestimating between-study heterogeneity. We generated ten scenarios parametrized by the number of studies in the dataset, ranging across 10, 20 and up to 100. For each scenario, we generated 500 realizations, for a total of 5,000 simulations. Using the quantile obtained from the FIM reduced the bias in heterogeneity estimates. See Extended Data Fig. 7 for full results. Sensitivity analyses We conducted sensitivity analyses comparing the dose–response curves obtained (1) using the fixed-effects model versus the mixed-effects model and (2) with and without trimming. For the first sensitivity analysis, we removed the random effects from the fixed-effects model in the last step of our estimation process. Results are shown in Extended Data Fig. 8 . There were some differences in the estimated levels of risk, and little to no difference in the shape of the risk curve between fixed- and mixed-effects of the mean risk curve. For the second sensitivity analysis, we compared results without trimming with those after trimming the 10% least coherent data points. The results are shown in Extended Data Fig. 9 and Supplementary Table 7 . We found that trimming generally stabilized the estimation, helping to guard the results against spurious observations. Trimming also decreased the estimated between-study heterogeneity, since, by definition, trimming removes points that are least coherent with the majority of the data (as judged using the model that is fit). Without trimming, ROSs and star ratings were generally lower, and more influenced by small numbers of outlying observations. Statistical analysis Analyses were carried out using R v.3.6.1, Python v.3.8 and Stata v.17. To validate key aspects of the meta-regression model used in this analysis, the following packages were used: metafor (R package available for download at ) and dosmesreta (R package available for download at ). Statistics and reproducibility The study was a secondary analysis of existing data involving systematic reviews and meta-analyses. No statistical method was used to predetermine sample size. As the study did not involve primary data collection, randomization, blinding and data exclusions are not relevant to this study, and as such, no data were excluded and we performed no randomization or blinding. We have made our data and code available to foster reproducibility. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The findings from this study were produced using data available in the published literature. Study sources and citations for each risk–outcome pair can be downloaded using the ‘download’ button on each risk curve page at . Study characteristics for all input data used in the analyses for the four example risk–outcome pairs are also provided in Supplementary Tables 7 – 10 . Code availability All code used for these analyses is publicly available online ( ). This includes code for the meta-regression engine, the model specification interface, both parts of the data processing, and risk-specific custom code, as appropriate.

A new set of meta-analyses clarifies the often complex and contradictory health guidance linking certain diets, behaviors, and conditions to illness. The analyses, conducted by researchers from the Institute for Health Metrics and Evaluation (IHME) at the University of Washington's School of Medicine, were published today in Nature Medicine. IHME analyzed the strength of the evidence for 180 pairs of risk factors and health outcomes—such as smoking and lung cancer, diet low in vegetables and type 2 diabetes, and high systolic blood pressure and ischemic heart disease. The findings are presented in an easy-to-understand star rating system showing the strength of evidence for each link. The new star rating system aims to help people make personal health decisions, inform health policy, and guide future research. "There has been extensive research on the links between various risks and health outcomes, but findings are often very different across studies," explained Dr. Christopher Murray, Director of the Institute for Health Metrics and Evaluation and a lead author of the study. "One of the goals of this new star rating system is to clear up confusion and help consumers make informed decisions about diet, exercise, and other activities that can affect their long-term health." In many areas, IHME found that the link between a risk factor and a health outcome was weaker than some might believe. Nearly two-thirds of the risk-outcome pairs investigated—112 out of 180—received only a one- or two-star rating. These include widely publicized pairings such as diet high in unprocessed red meat and ischemic stroke (one star). In other cases, IHME's analysis confirmed widely held consensus. Eight risk-outcome pairs received a five-star rating, including smoking and lung cancer and high systolic blood pressure and ischemic heart disease. A list of the star ratings, including a data visualization tool, can be found on IHME's website. Additional star ratings will be added in the near future. The analysis takes into account both the magnitude of risk shown by studies to date, as well as the consistency of findings between those studies. The star ratings are based on the most conservative interpretation of the available evidence, to limit the impact of error or bias in the underlying data. A one-star rating indicates that there may be no true association between the behavior or condition and the health outcome. Two stars indicates the behavior or condition is at least associated with a 0-15% change in the likelihood of a health outcome, while three stars indicates at least a 15-50% change, four stars indicates at least a 50-85% change, and five stars indicates a more than 85% change. For example, the five-star rating for smoking and lung cancer means that smoking increases the likelihood of developing or dying of lung cancer by more than 85%. At the other end of the scale, the one-star rating for red meat and ischemic stroke means that there may not be an association—in this case, because studies of this link have produced inconsistent results. Notable ratings from the study include: Credit: Institute for Health Metrics and Evaluation "In addition to helping consumers, our analysis can guide policymakers in developing health and wellness education programs, so that they focus on the risk factors with the greatest impact on health," said Dr. Emmanuela Gakidou, Professor of Health Metrics Sciences at the Institute for Health Metrics and Evaluation and a lead author of the study. "Health researchers can also use this analysis to identify areas where current evidence is weak and more definitive studies are needed." The IHME researchers also note that while the meta-analytical approach employed by this study should not replace expert deliberation, it can provide useful input for expert committees and advisory groups making formal health policy recommendations. IHME's analysis, which draws from the landmark Global Burden of Disease study, which marks 30 years this year, will be updated on a regular basis. As a result of constantly evolving research, the star ratings may change as more data becomes available. This is particularly the case for pairings with low star ratings due to limited or contradictory research. On the other hand, high star ratings are unlikely to change significantly because the evidence is already strong.

10.1038/s41591-022-01973-2

Biology

New mechanisms regulating neural stem cells

Germán Camargo Ortega et al. The centrosome protein AKNA regulates neurogenesis via microtubule organization, Nature (2019). DOI: 10.1038/s41586-019-0962-4 Journal information: Nature

http://dx.doi.org/10.1038/s41586-019-0962-4

https://phys.org/news/2019-02-mechanisms-neural-stem-cells.html

Abstract The expansion of brain size is accompanied by a relative enlargement of the subventricular zone during development. Epithelial-like neural stem cells divide in the ventricular zone at the ventricles of the embryonic brain, self-renew and generate basal progenitors 1 that delaminate and settle in the subventricular zone in enlarged brain regions 2 . The length of time that cells stay in the subventricular zone is essential for controlling further amplification and fate determination. Here we show that the interphase centrosome protein AKNA has a key role in this process. AKNA localizes at the subdistal appendages of the mother centriole in specific subtypes of neural stem cells, and in almost all basal progenitors. This protein is necessary and sufficient to organize centrosomal microtubules, and promote their nucleation and growth. These features of AKNA are important for mediating the delamination process in the formation of the subventricular zone. Moreover, AKNA regulates the exit from the subventricular zone, which reveals the pivotal role of centrosomal microtubule organization in enabling cells to both enter and remain in the subventricular zone. The epithelial-to-mesenchymal transition is also regulated by AKNA in other epithelial cells, demonstrating its general importance for the control of cell delamination. Main We searched the developing mouse cerebral cortex for candidate regulators of SVZ generation and expansion 3 , 4 , 5 . We chose to examine AKNA because levels of Akna mRNA correlate with the time at which the subventricular zone (SVZ) is generated (that is, low at embryonic day (E)11, high at E14 and low at E18), and because neural stem cells (NSCs) isolated at the peak of SVZ generation have higher levels of Akna when transitioning to basal progenitors 3 (Extended Data Fig. 1a, b ). To investigate the function of AKNA, we first generated monoclonal antibodies, which were validated by RNA interference (Extended Data Fig. 1c–f , Methods). Given the annotation of AKNA as an AT-hook transcription factor 6 , we were surprised to find specific immunofluorescence signals at centrosomes (Fig. 1a , Extended Data Fig. 1f–m ) in different cell types of various species (Extended Data Figs. 1 – 4 , 8 , 10 ), which was confirmed by BAC-transgenic cell lines (Extended Data Fig. 1h ) and fractionation of cell lysates (Extended Data Fig. 1n, o ). Ultrastructural analysis (Extended Data Fig. 2a–c ) showed that AKNA largely localized to the distal part of the subdistal appendages (SDAs) of the mother centriole in interphase; some AKNA was also found at the proximal ends of centrioles and along microtubules (Extended Data Fig. 2a–d , Supplementary Video 1 ). Recombinant AKNA protein decorated in vitro-reconstituted microtubules (Extended Data Fig. 2e , Supplementary Video 2 ). The centrosomal localization of AKNA is not dependent on microtubules or cargo-motors (Extended Data Fig. 2f–h ); however, it is dependent on centrioles, as AKNA is absent upon deletion of SAS4 7 (Extended Data Fig. 2i ) and the SDA protein ODF2 (Extended Data Fig. 2j, k ). AKNA localizes to the centrosome through its carboxyterminal part—a region that was omitted from its first description 6 (see Supplementary Discussion )—as AKNA(1–1080), which lacks the C-terminal 324 amino acids, is distributed in the cytoplasm but is still able to bind to microtubules (Extended Data Fig. 2l–o ). Similar to most SDA-associated proteins 8 , AKNA dissociates from centrosomes during M-phase without proteolytic degradation (Extended Data Figs. 1 h panel 2 , 3a, b ) and reassembles at the centrosomes during late telophase and early G1 phase (Supplementary Video 3 ); this dissociation and reassembly is regulated by phosphorylation (Extended Data Fig. 3c, d ). Thus, AKNA is a novel integral component of centrosomal SDAs and binding to microtubules, which raises the question of what its role is during development. Fig. 1: AKNA is a centrosome component that regulates NSC delamination. a , Immunostaining of E14 cortex (CTX) cells, showing AKNA at the mother centriole together with ODF2 (distal and subdistal appendages) and GT335 (cilia and centrioles) ( n = 4). b , E14 cortex sections showing AKNA + subsets of centrosomes. Pericentrin + (PCNT + )AKNA + centrosomes (blue arrows) at the ventricular zone (VZ) (panel 3); many AKNA + centrosomes in the SVZ and intermediate zone (IZ) (panel 2); and most centrosomes AKNA − (red arrows) in the cortical plate (CP) (panel 1). V, ventricle. n = 12. c , d , g , h , Confocal micrographs illustrating binning and the distribution of control ( c , g ), Akna shRNA no. 1 (shAkna 1) ( d ), electroporated GFP + cells at 2 d.p.e. and AKNA overexpression (OE) at 1 d.p.e. ( h ). e , f , i , j , Quantifications as indicated in panels. e , i , Mean ± s.e.m. as transparent band in the same colour. f , j , Box plots show median, quartiles (box) and range (whiskers). e , Control shRNA, n = 6; Akna shRNA no. 1, n = 5; Akna shRNA no. 2 (shAkna 2), n = 5 embryos. * P ≤ 0.05, ** P ≤ 0.01 (exact P values (from top to bottom) 0.00796, 0.00432, 0.00865, 0.02217, 0.03174, 0.01731, 0.00432, 0.00865 and 0.00865). f , Control shRNA, n = 5; Akna shRNA no. 2, n = 5 embryos. NS, not significant, ** P ≤ 0.01 (exact P values (from left to right) 0.00793, 0.30952 and 0.00793). i , GFP (control), n = 9; AKNA overexpression, n = 11; AKNA(1–1080), n = 4 embryos. ** P ≤ 0.01, *** P ≤ 0.001 (exact P values (from top to bottom) 0.00779, 0.00147 and 0.00008). j , GFP (control), n = 4; AKNA overexpression, n = 4 embryos. NS, not significant, * P ≤ 0.05 (exact P values (from left to right), 0.02857, 0.48571 and 0.02857). e , f , i , j , Two-sided Mann–Whitney U test. Scale bars, 2.5 μm ( a ), 20 μm ( b ), 50 μm ( c , d , g , h ). Source data Full size image AKNA + centrosomes showed their highest signal at E14 (compare Fig. 1b with Extended Data Fig. 4a–c ), were most frequent in the SVZ (Fig. 1b , Extended Data Fig. 4c ), and there was almost no AKNA signal in the cortical plate (in which neurons differentiate). In NSCs at the ventricular surface, only a fraction of centrosomes were AKNA + (Fig. 1b , Extended Data Fig. 4d–f ). These AKNA + centrosomes were in PAX6 + TBR2 + differentiating NSCs 9 , which is consistent with higher levels of Akna mRNA in this population 9 (Extended Data Fig. 4g, h ). Thus, AKNA exhibits an unprecedented subtype specificity, being restricted to the centrosomes of differentiating NSCs and basal progenitors in forebrain development. Knockdown of AKNA, mediated by short hairpin (sh)RNA, (Extended Data Fig. 1d ) via in utero electroporation (IUE) in E13 cortices led to the retention of GFP + cells in the ventricular zone with more self-renewing (PAX6 + TBR2 − ) NSCs relative to the control conditions, under which many GFP + cells had left the ventricular zone and SVZ by two days post-electroporation (d.p.e.) (Fig. 1c–f , Supplementary Video 4 ). This phenotype was observed with two different shRNAs, was rescued by co-electroporating AKNA (Fig. 1e , Extended Data Fig. 5a, b ) and was independent of p53-mediated cell death (Extended Data Fig. 5c–j ). Thus, AKNA loss of function impairs NSC delamination, and leads to the retention of more NSCs in the ventricular zone. Conversely, AKNA overexpression induced a fast delamination, with most GFP + cells being found in the SVZ at 1 d.p.e. (Fig. 1g–i ); this resulted in more basal progenitors (PAX6 − TBR2 + ) and NEUN + neurons and a concomitant decrease in NSCs (PAX6 + TBR2 − ), as compared to controls (Fig. 1j , Extended Data Fig. 5k ). The increased differentiation that occurred upon AKNA overexpression in vivo did not occur in vitro (Extended Data Fig. 5l ), and is therefore due to the altered niche that is encountered upon delamination in vivo. Live imaging in cortical slices showed that AKNA overexpression accelerated NSC delamination by a retraction of the apical processes without cell division (28% ( n = 33) compared to 3% ( n = 165) in controls) (Supplementary Video 5 ). Thus, AKNA-induced delamination occurs in interphase, during which AKNA localizes at centrosomes. Consistent with this, overexpression of AKNA(1–1080) hardly increased delamination (Fig. 1i ), which indicates the importance of the centrosomal localization of this protein. Given the function of SDAs 10 in microtubule organization, we tested whether AKNA regulates microtubule organization using nocodazole-based microtubule regrowth assays (Extended Data Fig. 6a ). shRNA-mediated AKNA knockdown in primary cortical cells resulted in a significant reduction in the number of cells with centrosomal microtubule regrowth, and of microtubule-fibre length (Fig. 2a–c ). Independently of centrosomes, microtubule asters emanated from ectopic AKNA bodies upon AKNA overexpression (Fig. 2d ) in an apparently size-dependent manner (Extended Data Fig. 6b ). Importantly, AKNA efficiently recruited components of the γ-tubulin ring complex (γTuRC), such as TUBG and GCP4, as well as CKAP5 (also known as ch-TOG) (the mammalian homologue of the microtubule nucleator and polymerase XMAP215 11 ), but not other centrosome- and microtubule-associated proteins (Extended Data Fig. 6c, d ); this demonstrates the specificity of γTuRC and microtubule nucleation factor recruitment, and how AKNA affects microtubule nucleation and growth. AKNA also recruits and co-immunoprecipitates with SDA proteins that are involved in microtubule organization such as EB1 (also known as MAPRE1), DCTN1 (also known as P150-Glued), dynein and ODF2 12 , 13 (Extended Data Fig. 6e, f ). AKNA is thus a novel regulator of the centrosome, controlling microtubule organizing centre (MTOC) activity specifically in differentiating NSCs. Notably, AKNA does not affect cilia formation or length in NSCs, despite its localization at SDAs (Extended Data Fig. 7a–d ). Fig. 2: AKNA regulates microtubule organization in NSCs. a , Images of E14 cortical cells transfected with shRNA control or Akna shRNA no. 2, in nocodazole-based microtubule repolymerization assays. The yellow dashed line indicates a cell in which microtubules did not grow upon AKNA knockdown. b , Box plot, median, quartiles (box) and range (whiskers), showing reduced number of cells with centrosomal microtubule (MT) regrowth after AKNA knockdown (shRNA control, n = 4; Akna shRNA no. 2, n = 4 independent experiments, P = 0.02857). c , Dot and violin plot showing reduced length of microtubule. Dots represent single measurements, mean ± s.e.m in bold in the centre. shRNA control, n = 110; Akna shRNA no. 2, n = 92 microtubule endpoints from 4 independent experiments each ( P = 0.00007). d , AKNA overexpression clusters organize microtubules independently of centrosomes (see also Extended Data Fig. 6b ) ( n = 6). e , Most GFP + electroporated cells lose their integration into the apical cadherin belt at the ventricle (V) 18 h after AKNA overexpression ( n = 3 embryos). f , Line graph illustrating the distribution of GFP + cells 24 h after IUE. Mean ± s.e.m. as transparent band in the same colour. GFP (control), n = 9; AKNA overexpression, n = 11; AKNA and E-cadherin (E-CAD) overexpression, n = 3 embryos. GFP (control) and AKNA overexpression is based on the same data as shown in Fig. 1i . * P ≤ 0.05, ** P ≤ 0.01, *** P ≤ 0.001 (exact P values (from top to bottom) 0.02597, 0.00779, 0.00909 and 0.00008). g , h , En face view of ventricular endfeet in E14 cortex delineated by ZO-1 ( g ), quantified in h , showing the inverse correlation of AKNA immunofluorescence intensity and endfoot size. Transparent band depicts the best-fit curve with the 95% confidence interval ( n = 76 cells from 4 embryos). i , Dot and violin plot (dots represent individual cells, mean ± s.e.m. in bold in the centre) depicting the reduced size of the apical endfeet after AKNA overexpression. GFP (control), n = 76 cells; AKNA overexpression, n = 14 cells; n = 4 embryos each; P = 0.00339. b , c , f , i , Two-sided Mann–Whitney U test. Scale bars, 5 μm. Source data Full size image Interestingly, CAMSAP3—a microtubule minus-end binding protein—was recruited to AKNA + foci (Extended Data Fig. 6e ). Consistent with CAMSAP proteins anchoring non-centrosomal microtubules at epithelial junctions 14 , CAMSAP3 was enriched at the NSC junctions at the ventricular surface (Extended Data Fig. 7e ), and cadherin levels at the apical surface were reduced after AKNA overexpression (Fig. 2e ). However, maintaining high levels of cadherin with AKNA and E-cadherin co-IUE was not sufficient to interfere with the AKNA-induced delamination process (Fig. 2f ). At the junctions surrounding the apical endfoot at the ventricular surface, microtubule and actin belts mediate abscission 15 . High levels of AKNA at the centrosome correlated with a smaller size of apical endfeet, delineated by ZO-1 immunostaining (Fig. 2g, h ), and AKNA overexpression increased the number of cells with small endfeet (Fig. 2i ). Thus, AKNA regulates the delamination process by increasing centrosomal MTOC activity, weakening junctional complexes and promoting contraction of the apical endfoot. As AKNA also binds directly to microtubules, we monitored microtubule dynamics by EB3–GFP live imaging in cortex slices after IUE in vivo. EB3–GFP comets grew mostly in a basal direction in NSCs after GFP IUE, as previously shown 16 ; however, after AKNA overexpression, comets displayed more random directions (Extended Data Fig. 8a, b , Supplementary Videos 6 , 7 ), which indicates a change in the orientation of microtubule growth and a repositioning of the MTOC toward non-apical positions (see schematic in Extended Data Fig. 7j ). AKNA overexpression also had minor effects on the speed of EB3–GFP comets (Extended Data Fig. 8c ). To determine possible direct effects of AKNA on microtubules, we reconstituted microtubules in a centrosome-free environment in vitro. The presence of AKNA affected growth velocity but not microtubule lifetime (Extended Data Fig. 8d–f , Supplementary Video 2 ). Moreover, once a depolymerizing microtubule encountered a site of AKNA enrichment, the depolymerization velocity was significantly reduced (Extended Data Fig. 8g , Supplementary Video 2 ), which demonstrates that AKNA affects microtubule dynamics in vivo and in vitro. These dynamics are important for delamination, as stabilizing microtubules using Taxol in vivo blocked the AKNA-induced delamination (Extended Data Fig. 8h ). As the delamination from the ventricular zone resembles the epithelial–mesenchymal transition (EMT) 17 , 18 , 19 , we asked whether AKNA is more generally relevant for EMT by monitoring normal murine mammary gland (NMuMG) epithelial cells during EMT induced by transforming growth factor beta-1 (TGFβ1) 20 , 21 . Centrosomal AKNA levels are low in untreated NMuMG cells, but upregulated early in EMT (Extended Data Fig. 8i–k ). As expected 14 , NMuMG cells show largely non-centrosomal microtubule polymerization before EMT (Supplementary Video 8 ), and AKNA overexpression increased microtubule nucleation at the centrosome (Extended Data Fig. 8l , Supplementary Video 9 ). Knockdown of AKNA using small interfering (si)RNAs (Extended Data Fig. 8m ) had no effect on the expression of core EMT transcription factors Twist1 , Zeb2 , Snai1 and Snai2 during EMT (Extended Data Fig. 8i ), but led to ZO-1 retention at cell junctions and counteracted the TGFβ1-induced disassembly of these junctions (Fig. 3a–d , Extended Data Fig. 8o, p ). AKNA knockdown also attenuated the rearrangement of actin fibres from the junctions to stress fibres (Fig. 3a–c ), showing the importance of AKNA in fully dissolving these junctional complexes during EMT. Immunostaining confirmed that AKNA is sufficient to significantly reduce CAMSAP3 at the junctional interface closest to AKNA + foci upon overexpression, but not CAMSAP3 at more-distant junctions (Fig. 3e, f , Extended Data Fig. 8q ). Conversely, AKNA knockdown resulted in the retention of CAMSAP3 at cell junctions after TGFβ1-induced EMT (Fig. 3g , Extended Data Fig. 8r ). Thus, AKNA mediates the remodelling of junctional complexes by recruiting CAMSAP3 from junctional microtubules to the centrosome, thereby destabilizing junctional cadherins 22 and enabling cells to detach during EMT. Fig. 3: EMT progression requires AKNA to dissolve cell junctions. a – c , Micrographs of NMuMG cells in untreated ( a ) and in TGFβ1-treated, EMT-inducing conditions with control siRNA (siControl) ( b ) or Akna siRNA (siAkna) ( c ). ZO-1 is retained at cell junctions, and actin filaments (phalloidin) redistribute less frequently from junctions to cytoplasm, upon AKNA knockdown during EMT. d , Dot and box plot showing ZO-1 signal at the cell junction (position 0) (untreated, n = 60 cells; TGFβ1-treated and control siRNA, n = 55 cells; TGFβ1-treated and Akna siRNA, n = 76 cells from 3 independent experiments). NS, not significant, *** P ≤ 0.001 (exact P values (from left to right) 5.2 × 10 −6 , 0.93189 and 0.00004). f , g , Dot and box plots showing significant reduction in CAMSAP3 at the cell junction close to AKNA foci after AKNA overexpression (position 1 in the micrograph in e ) but not in EMT with Akna siRNA. In f , untransfected, n = 26 cells; AKNA overexpression: n = 22 cells, from 3 independent experiments. P = 0.00003. In g , untreated, n = 42 cells; TGFβ1-treated and control siRNA, n = 51 cells; TGFβ1-treated and Akna siRNA, n = 50 cells, from 3 independent experiments. NS, not significant, *** P ≤ 0.001 (exact P values (from left to right) 0.00086, 0.75728 and 0.00095). d , f , g , Two-sided Mann–Whitney U test; box plots show median, quartiles (box) and range (whiskers). Scale bars, 10 μm. Source data Full size image Besides the role of AKNA in the EMT-like delamination of NSCs, we aimed to determine its function within the SVZ, the region with highest levels of AKNA (only 10% of centrosomes were AKNA + above the SVZ in the cortical plate where neurons are located) (Extended Data Fig. 4e ). Accordingly, neurons isolated in vitro showed a greater degree of non-centrosomal microtubule polymerization (Extended Data Fig. 9a, b ), as compared to NSCs. To counteract the reduction in levels of AKNA above the SVZ, we expressed Akna cDNA under the late basal progenitor and neuron-specific doublecortin ( Dcx ) promoter (Extended Data Fig. 9c ) at E13. This resulted in many GFP + cells remaining below the cortical plate, as compared to control (Fig. 4a–d ), without affecting proliferation at two days post-electroporation (Fig. 4e ) or neuronal differentiation (Extended Data Fig. 9d–f ). Conversely, Dcx -promoter-driven Akna microRNAs (miRNAs) did not affect cells that left the SVZ and reached normal positions within the cortical plate, similar to controls (Extended Data Fig. 9g, h ). Thus, the downregulation of AKNA that occurs physiologically (Fig. 1b , Extended Data Fig. 4d, e, g ) facilitates the transition of neurons from the SVZ to the cortical plate. Fig. 4: AKNA regulates retention of cells within the SVZ. a , b , Confocal micrographs illustrating binning and the distribution of electroporated cells expressing GFP ( a ), Akna and GFP, or Akna (1–1080) and GFP ( b ) under the control of the Dcx promoter at 5 d.p.e. c , Line graph illustrating the distribution of GFP + cells after IUE (mean ± s.e.m. as transparent band in the same colour). Dcx -GFP, n = 5; Dcx-Akna , n = 4; Dcx-Akna (1–1080), n = 4 embryos. * P ≤ 0.05 (exact P values (from top to bottom) 0.02857, 0.02857, 0.01587, 0.01587, 0.02857 and 0.01080). d , e , Dot and box plots showing the fraction of GFP + cells retained below the cortical plate labelled by TBR1 at 5 d.p.e. ( d , Dcx -GFP, n = 5; Dcx-Akna , n = 3 embryos, P = 0.03571) or in cell cycle (KI67 + ) at 2 d.p.e. ( e , Dcx -GFP, n = 4; Dcx-Akna , n = 4 embryos; not significant ( P = 0.48571). f , Dot and violin plot showing the migration speed of control and AKNA-knockdown neurons in slices at 2 d.p.e. (dots represent individual cells, mean ± s.e.m. in bold) ( Dcx -miRNA(neg) (sequence predicted not to target any known vertebrate gene), n = 34; Dcx -miRNA( Akna ), n = 26 cells, from 3 independent experiments; not significant ( P = 0.67299)). g – j , Box plots illustrating the morphological transitions of control and AKNA-knockdown cells ( g , h , n = 4 embryos each condition) or after AKNA overexpression ( i , j , n = 3 embryos in each condition). B, bipolar; M, multipolar; PB, persisting bipolar; PM, persisting multipolar. In g – j , NS, not significant, * P ≤ 0.05, ** P ≤ 0.01. Exact P values (from left to right): 0.02460 and 0.02460 ( g ); 0.75566, 0.93038 and 0.71203 ( h ); 0.00385 and 0.00385 ( i ); 0.99877, 0.71153 and 0.75539 ( j ). c – f , Two-sided Mann–Whitney U test; g – j , two-sided Student’s t -test. Box plots show median, quartiles (box) and range (whiskers). Scale bars, 50 μm. Source data Full size image To determine at which step AKNA levels are critical in the transition from multipolar SVZ cells to bipolar neurons migrating into the cortical plate 23 (schematic in Extended Data Fig. 7i ), we performed live imaging in slices after AKNA overexpression or knockdown in vivo. Control and knockdown cells had similar migratory speeds during locomotion (Fig. 4f ), but knockdown cells transitioned faster from the multipolar SVZ morphology to the bipolar morphology of migrating neurons (Fig. 4g, h ). Conversely, many cells with AKNA overexpression retained a multipolar morphology and migrated less frequently (Fig. 4i, j , Supplementary Video 10 ). Thus, AKNA is key not only in bringing cells into the SVZ but also in retaining them there. These data provide further support for the suggestion that the switch to bipolar morphology and radial neuronal migration requires a switch from a centrosomal to more non-centrosomal MTOC via endogenous downregulation of AKNA. Indeed, the mutant, non-centrosome-localizing form of AKNA (that is, AKNA(1–1080)) expressed under the Dcx promoter failed to retain cells in the SVZ (Fig. 4a–c ), which highlights the importance of the centrosomal location of AKNA in this process of morphology switching and migration. In neurons, microtubules originate preferentially from other non-centrosomal compartments 24 , allowing axon and dendrite formation 25 , 26 , and have more stable detyrosinated and acetylated microtubules 27 , 28 . AKNA counteracts these neuronal hallmarks by promoting centrosomal microtubule nucleation, faster microtubule growth, and more dynamic and less detyrosinated microtubules (Extended Data Fig. 8m, n ). Therefore, the downregulation of AKNA is required for neurons to migrate into the cortical plate. Thus, microtubule organization differs profoundly for cell migration into the SVZ, which requires AKNA and a centrosomal MTOC, versus migration out of the SVZ, which requires the loss of AKNA. During the delamination process, AKNA orchestrates the disassembly of junctional complexes as well as apical endfoot constriction. As microtubule stabilization by Taxol blocked the AKNA-induced delamination, microtubule dynamics and rearrangement are critical. Moreover, the centrosomal localization of AKNA is important, as the non-centrosomal form, AKNA(1–1080), showed impaired delamination but could still bind microtubules. AKNA increased centrosomal microtubule nucleation by recruiting γTuRC and CKAP5 11 . This—together with the low-level direct binding of AKNA to microtubules—explains why AKNA gain of function both promotes nucleation and inhibits depolymerization. AKNA recruits CAMSAP3 to the centrosome. This microtubule minus-end binding protein otherwise tethers microtubules to the adherens junctions 14 contributing to adherens junction stability 22 and cell–cell attachment 29 . Thus, AKNA orchestrates the delamination process by increasing centrosomal microtubule nucleation and recruiting nucleation factors and minus-end stabilizers, thereby destabilizing microtubules at the adherens junctions and mediating constriction of the apical endfoot. The recruitment of AKNA to the centrosome is post-translationally regulated; SOX4 regulates Akna mRNA in EMT 21 and the generation of basal progenitors 30 (Extended Data Fig. 10a–c ). Regulating AKNA levels orchestrates delamination and retention in the SVZ—this process is particularly relevant in species with highly proliferative cells 2 in an enlarged SVZ (Extended Data Fig. 10d–r ). Thus, the function of AKNA in ontogeny may have profound relevance in phylogeny. Reporting summary Further information on research design is available in the Nature Research Reporting Summary linked to this paper. Data availability The data generated and analysed in this study are included in the article and its Supplementary Information . Full gel blots can be found in the Extended Data Figures or Supplementary Fig. 1 . Any other data in this study are available from the corresponding author upon reasonable request.

The use of stem cells to repair organs is one of the foremost goals of modern regenerative medicine. Scientists at Helmholtz Zentrum München and Ludwig Maximilian University of Munich (LMU) have discovered that the protein Akna plays a key role in this process. It controls, for example, the behavior of neural stem cells via a mechanism that may also be involved in the formation of metastases. The study was published in the renowned journal Nature. The research team led by Prof. Dr. Magdalena Götz, director of the Institute for Stem Cell Research (ISF) at Helmholtz Zentrum München and Chair of Physiological Genomics of the LMU Biomedical Center, wanted to identify the factors that regulate the maintenance or differentiation of neural stem cells. To this end, the scientists isolated neural stem cells, which either self-renew and generate additional neural stem cells or differentiate. "We found that the Akna protein is present in higher concentrations in stem cells that generate neurons," explains ISF researcher German Camargo Ortega, first author of the study together with Dr. Sven Falk. "Our experiments showed that low levels of the Akna protein cause stem cells to remain in the stem cell niche, whereas higher levels stimulate them to detach from the niche, thus promoting differentiation," the author continues. The scientists were surprised to discover the position of the protein − namely at the centrosome, an organelle in the cell's interior that acts as chief architect for the organization of the cytoskeleton and regulates cell division. "We discovered that an incorrect sequence was originally published for this protein," Sven Falk reports. "However, our work clearly showed that Akna is located directly at the centrosome." The researchers were able to show that Akna recruits and anchors microtubules at the centrosome. This weakens the connections to neighboring cells, and promotes detachment and migration from the stem cell niche. Akna (here in magenta) is a novel centrosome component regulating the interaction with the cytoskeleton. Credit: Helmholtz Zentrum München "Our experiments show that this function also plays an important role in a process known as epithelial-to-mesenchymal transition, or EMT for short," explains the study leader Magdalena Götz. "In this process, cells detach from a cluster, proliferate and begin to migrate. This occurs, for example when stem cells migrate to form new neurons, but it can also be harmful in disease, for example when cancer cells leave a tumor to form metastases elsewhere in the body. "The novel mechanism that we identified by studying the function of Akna therefore appears to play a key role in a broad range of medically relevant processes." In the next step, the research team plans to investigate the role of Akna in other stem cells and in the immune system.

10.1038/s41586-019-0962-4

Medicine

As we develop, the brain connects lessons learned differently

Margaret L. Schlichting et al, Developmental differences in memory reactivation relate to encoding and inference in the human brain, Nature Human Behaviour (2021). DOI: 10.1038/s41562-021-01206-5 Journal information: Nature Human Behaviour

http://dx.doi.org/10.1038/s41562-021-01206-5

https://medicalxpress.com/news/2021-11-brain-lessons-differently.html

Abstract Despite the fact that children can draw on their memories to make novel inferences, it is unknown whether they do so through the same neural mechanisms as adults. We measured memory reinstatement as participants aged 7–30 years learned new, related information. While adults brought memories to mind throughout learning, adolescents did so only transiently, and children not at all. Analysis of trial-wise variability in reactivation showed that discrepant neural mechanisms—and in particular, what we interpret as suppression of interfering memories during learning in early adolescence—are nevertheless beneficial for later inference at each developmental stage. These results suggest that while adults build integrated memories well-suited to informing inference directly, children and adolescents instead must rely on separate memories to be individually referenced at the time of inference decisions. Main Young adults reactivate memories when they encounter new related experiences. Such reactivation can facilitate memory integration, whereby related events experienced at different times are stored as overlapping memory traces 1 , 2 , 3 . Memory integration thus promotes forming links between memories that extend knowledge beyond direct observation. Memory integration in adults relies on hippocampus (HPC) and medial prefrontal cortex (PFC) 2 , 4 , 5 , 6 , 7 and has been shown to benefit behaviours such as inferential reasoning, which requires simultaneous consideration of multiple memories 3 . For example, when asked to derive a relationship that has not been directly observed but must be inferred across several prior events, adults benefit from having previously connected (or integrated) their memories at encoding 5 , 6 , 8 , 9 , 10 . However, inferential reasoning can also be accomplished via an alternative mechanism in which memories for the original experiences are stored separately to be later recombined 11 , 12 , 13 when making the inference itself. Such a retrieval-based mechanism depends on memories for the individual events but importantly does not rely on them having been integrated at encoding. It has been suggested that young children’s inference ability 14 , 15 arises predominantly from a retrieval-based mechanism. For example, children struggle disproportionately with reasoning given their memory performance 16 , 17 , and they are insensitive to manipulations designed to promote integration during encoding 18 . Both of these findings are consistent with the idea that reasoning during childhood relies primarily on operations engaged during inference itself. Given that inferential reasoning is a predictor of academic success 19 , 20 and the real possibility that children approach it in a fundamentally different way, it is crucial that we understand the neural mechanisms supporting its improvement throughout development. We suggest that how memories for related experiences are formed will depend on both (1) the refinement of HPC-based mechanisms that support the ability to flexibly reactivate neocortical representations of related memories during new learning 21 , 22 and (2) the frontal mechanisms available to mediate conflict among reactivated memories, the results of which will ultimately influence how memories are formed in HPC 23 , 24 , 25 . Importantly, we suggest that reactivation is a necessary but not sufficient condition for memory integration to occur, as additional mechanisms must be engaged after memory reactivation to ultimately link related memories according to their overlapping features. Our overarching hypothesis is that, developmentally speaking, reactivation and integration will emerge in succession, paralleling the maturation of HPC 26 , 27 , 28 , 29 , 30 and its PFC connections 31 , 32 , respectively. As such, there will be a period—specifically, adolescence—during which individuals reactivate but do not integrate. Both HPC and PFC along with their interconnecting pathways show a protracted developmental trajectory continuing into (at least) adolescence 26 , 27 , 31 , with PFC being particularly late to mature 33 . We suggest that the emergence of the first step necessary for an adult-like integration mechanism—namely, memory reactivation during new encoding—would therefore require that HPC retrieval is flexible enough to allow for the reinstatement of related memories during a similar but not identical (that is, partially overlapping) new experience 34 , 35 . Such flexible retrieval may not mature until around 10 years of age 25 , 36 , thereby preventing any top-down influence on HPC codes and leaving inference in children to be carried out entirely to the time of decision itself. Along with the maturation of HPC retrieval mechanisms is expected to come a greater likelihood of memory reinstatement in adolescence; and yet, we suggest that such reactivation will nevertheless continue to have a different behavioural consequence than it does in adults. In particular, we suggest that reactivation of past memories in adolescence may yield memory competition that is resolved by suppression and an ultimate de-emphasis of the relationships among memories. We suggest that such a phenomenon in our task may be due to the combined influence of at least two factors. First, there is general maturation of top-down control networks during adolescence that has been linked to improvements in higher-order cognitive abilities 37 , 38 , 39 , 40 . Some have even observed that adolescence may be a peak period for top-down control of memory behaviours, with adolescents showing enhanced engagement of lateral PFC relative to adults 41 . Such increased control of memory among adolescents in our task may correspond to an enhanced tendency to suppress related, interfering memories 25 , 32 . Consistent with this idea, past work in rodents has highlighted adolescence as a unique time during which previous memories are suppressed during new, similar experiences 42 , 43 . Second, there are developmental changes in the HPC 26 , 27 , 28 , 29 , 30 that may lead to differences in memory representation. In terms of overall memory quality, adolescence is associated with increases in precision 44 , 45 , greater richness of episodic detail 46 and enhancements in recollective quality 47 —all potentially attributable to changes in HPC encoding 41 , 48 , 49 , 50 , 51 , 52 . Given these findings, we expect that adolescents will be nearly adult-like in their ability to remember individual experiences. However, they may not yet have the ability to flexibly link across related experiences due to the nature of HPC development: namely, that posterior HPC (pHPC) matures earlier than anterior 26 , 27 , 28 . Informed by past functional studies also showing greater reliance on pHPC in children and adolescents than in adults 53 , we therefore suggest that adolescent memory will accordingly reflect precise pHPC representations 4 , 54 , 55 , 56 (the granularity of which has been shown to increase over this developmental window 57 ) rather than integrated ones 4 , which engage later-developing HPC 27 and medial PFC 31 mechanisms. Together, the increasing availability of control mechanisms enabling memory suppression in tandem with hippocampal biases towards separate storage of related memories may ultimately yield representations emphasizing the unique features of individual experiences 58 , 59 . Importantly, such memories can nevertheless support successful inference 1 , 4 , 8 , as they may be particularly easy to access and recombine during the decision. Here, we test these hypotheses in a functional magnetic resonance imaging (fMRI) study in typically developing children, adolescents and adults. We anticipated nonlinearity in the developmental trajectory 60 , such that adolescents would rely on memory mechanisms for inference that are distinct from those used by either children or adults. We also underscore that the maturation of memory-based inference probably unfolds through gradual change in the availability of these different neural signatures rather than an abrupt transition between mechanisms with development, consistent with an ‘overlapping waves’ perspective more typically discussed in the context of overt strategy 61 . As such, here we characterized development continuously from childhood through early adulthood. Results Integration of new memories in a pair learning task Eighty-six participants aged 7 to 30 years performed an associative inference task ( Methods ) 4 , 5 , 6 , 7 , 8 , 12 , 16 , 62 , 63 . Stimuli (faces, scenes and objects) were organized into groups of three—termed ABC triads—and presented to the participants as overlapping AB and BC pairs that repeated in alternation along with non-overlapping control pairs 6 , 63 (Fig. 1a ). This design allowed us to use an fMRI pattern classification approach (multivoxel pattern analysis (MVPA) 64 ) to decode reinstatement of the related C content type—which was either a face or a scene, depending on the triad—to test for the predicted developmental differences in flexible retrieval. We hypothesized that while adolescents and adults would reactivate related memories during new encoding, children under age 10 36 would not, despite the additional encoding opportunities afforded by repetition. We further reasoned that should related memories be successfully brought to mind, there may nevertheless be lingering developmental differences in the way that conflict between memories is resolved. In particular, we predicted that reactivation in adolescence would be uniquely associated with both an upregulation of control regions implicated in memory suppression during later repetitions of overlapping pairs and impeded performance due to added competition. In contrast, we predicted that reactivation would be behaviourally advantageous in adults, who may instead integrate. Fig. 1: Experimental task. a , Each of four study–test cycles began with participants studying pairs for a later memory test. The timeline depicts every stimulus presentation (the pairs are in colour; the baseline is in grey) for one run. Overlapping (AB and BC) and non-overlapping (labelled NOL in the figure) pairs were blocked by type and jittered within each block to enable analysis at both the block and trial levels. AB pairs were our main trials of interest, with non-overlapping pairs serving as a content-matched baseline. Half of the AB pairs were each associated with a familiar scene (top) or face (bottom). The faces and scenes in the real experiment were images from popular movies and TV shows; they are replaced with uniquely coloured silhouettes in the figure for copyright reasons. The fill colours of the scene and face silhouettes indicate different identities. Moreover, the objects have been replaced with similar photos from the Bank of Standardized Stimuli 133 , 134 in all panels. b , After studying, the participants made self-paced inference judgements in which they first indicated the category (character) and then the identity (lime-green face, representing a particular character identity) of the C item indirectly related to the probe (cupcake). Foils (incorrect options) were other items from the same run that had occurred in the same position, condition and run (that is, C items were foiled by Cs that were members of different triads; different character identities are represented by different fill colours across the three options). c , Participants then completed a self-paced three-alternative forced-choice memory test for all studied pairs from the preceding study run. As in the inference test, the foils in the pair memory test were other items from the same run that had occurred in the same position, condition and run (again represented by different fill colours for the face and scene silhouettes). Images in a – c reproduced with permission from refs. 133 , 134 . d , After the tests, the participants received feedback about their memory performance before moving on to the next study with a new set of pairs. Specifically, the participants saw their previously selected avatar climbing a staircase, with the distance moved proportional to their memory performance. The participants’ avatars continued to climb the staircase over four study–test cycles to earn bonus pay. Full size image After each study run, the participants completed self-paced inference and memory tests (Fig. 1b,c ). In the inference test, the participants were asked to link A and C items that were indirectly related through their common association with B at both the general (category) and specific (item) levels. Hereafter, to limit the influence of guessing, we consider inferences as correct only if the participant made the correct selection for both the category-level and item-level judgements. Pair memory was assessed only at the specific, item level. Importantly, all participants were aware of and had practiced both memory and inference tests before beginning the first study to reduce the influence of age-related differences in strategic approach to the task. The participants were highly accurate on both the inference (mean, 83.08%; range, 20.83–100%; 95% confidence interval (CI), (78.73, 87.43)) and pair memory (94.67%, 56.94–100%, (92.68, 96.66)) tests, with a developmental trajectory characterized by rapid improvements at younger ages followed by plateau at ceiling in adolescence (Fig. 2 ). We then interrogated whether memory varied across direct pair types (AB, BC and non-overlapping) to quantify differential encoding of overlapping versus non-overlapping pairs (Fig. 2a ). There were significant effects of both age ( χ 2 (3) = 34.11, P < 0.001) and trial type ( χ 2 (2) = 7.09, P = 0.03; the interaction was not significant, χ 2 (6) = 10.13, P = 0.12), with trial type effects primarily driven by worse performance for BC pairs. Notably, this difference may be a function of stimulus type rather than overlap per se, as only BC pairs contain an object with a face or scene as opposed to two objects. Age was also related to response times (RTs) on correct trials, with adolescents and adults being faster than children (Fig. 2b ; age effect: χ 2 (3) = 55.68, P < 0.001). There was no significant main effect of condition ( χ 2 (2) = 2.00, P = 0.37), but there was a significant age-by-condition interaction ( χ 2 (6) = 13.80, P = 0.03) such that children showed the smallest RT difference among conditions. While speculative, one possibility for the relatively smaller difference in RT across direct pair types for children is that they do not encode the overlapping AB and BC pairs in a way that reflects their shared relationships; rather, children may treat overlapping the same as non-overlapping pairs and encode them in pattern-separated memories 58 , 65 . Fig. 2: Task performance. a , b , Performance (accuracy) ( a ) and RT for correct trials ( b ) in the pair memory test (types AB, BC and non-overlapping) as functions of age. There were significant main effects of both age ( χ 2 (3) = 34.11, P < 0.001) and trial type ( χ 2 (2) = 7.09, P = 0.03) on accuracy, but no significant interaction ( χ 2 (6) = 10.13, P = 0.12; 7,944 trials). For RTs, there was a significant effect of age ( χ 2 (3) = 55.68, P < 0.001) and an age-by-condition interaction ( χ 2 (6) = 13.80, P = 0.03); there was no significant main effect of condition ( χ 2 (2) = 2.00, P = 0.37; 7,535 trials). c , Memory and inference performance (accuracy) as a function of age. There were significant main effects of test type (memory or inference; χ 2 (1) = 34.28, P < 0.001) and age ( χ 2 (3) = 31.35, P < 0.001) as well as an age-by-test-type interaction ( χ 2 (3) = 10.30, P = 0.02; 10,592 trials). d , RT as a function of age. There were significant main effects of age and trial type, such that RT decreased over development ( χ 2 (3) = 54.69, P < 0.001) and memory was faster than inference ( χ 2 (1) = 16.65, P < 0.001; there was no significant interaction, P = 0.27; 9,749 trials). For all panels, we used (generalized) linear mixed effects models to assess whether age, test trial type or their interaction was associated with accuracy and RT on individual trials. In all charts, the lines and bands depict model predictions and 95% CIs derived from the better-fitting models including age; the dots depict individual participant means (for accuracy) and medians (for RT) by condition. For all panels, N = 86 participants. Full size image We next compared developmental improvements in memory (collapsed across direct pair type) with those in inference. Accuracy (Fig. 2c ) was higher on the pair memory test than on the inference test ( χ 2 (1) = 34.28, P < 0.001), with the magnitude of this difference decreasing with development (age-by-test-type interaction: χ 2 (3) = 10.30, P = 0.02; there was also a main effect of age, χ 2 (3) = 31.35, P < 0.001). This result replicates our previous findings in a different sample 16 and highlights that while the task was within the abilities of all ages—that is, performance was well above chance for all trial types across the entire age range (Fig. 2a,c , confidence bands)—younger participants disproportionately struggled with inference. These behavioural findings suggest developmental differences in how participants approach inferential reasoning; however, further neural mechanistic insight into the specific source of the age-related differences requires the fMRI approach that we turn to next. Identifying developmental differences in memory reactivation One important clue as to when memories are being combined to make an AC inference—that is, during encoding in preparation for an upcoming decision or, conversely, later during the decision itself—might stem from how reactivation unfolds across repeated experiences with the same, overlapping associations. We hypothesized that reactivation changes over repetition would be particularly diagnostic for understanding how memory mechanisms in adolescents differ from those in both children and adults. On the basis of prior work 6 , we hypothesized that reactivation of overlapping memories would increase across repetition in adults, promoting integration and inference (Fig. 3c , coral dots). Adolescents, in contrast, may reactivate overlapping memories during initial repetitions but then resolve this competition by accentuating the differences between overlapping memories in subsequent encounters with related pairs (Fig. 3c , magenta dots) 59 . This adolescent pattern would further differ from that of children, who we predicted would show no significant reactivation at all (Fig. 3c , purple dots). Such a result would be evidence of a developmental pattern in which adolescence is more than a stage between childhood and adulthood—rather, participants in this group may engage a fundamentally unique, adolescent-specific mechanism due to their particular neuromaturational state 60 . Fig. 3: Perception and memory reactivation decoding analyses. a , For both perception and memory analyses, an MVPA classifier was trained on patterns of activation from a visual localizer task that occurred after, and used separate stimuli from, the main memory experiment. As in Fig. 1 , the stimuli were replaced with uniquely coloured silhouettes for copyright reasons. fMRI patterns were extracted from VTC, and the classifier was trained to discriminate scene (pink) from face (orange) viewing. The boundary discriminating between the categories is depicted as a line separating face and scene viewing trials in a two-dimensional space. b , Perception decoding approach and results. Using cross-validation, classifiers were trained on n − 1 localizer runs as in a and applied to the held-out n th run, where n is the number of localizer runs included for a given participant. Classification performance (accuracy; y axis) was high and did not significantly differ with age ( x axis; model comparison: AIC base = 1,581.5, AIC age = 1,500.2, F (3,76) = 1.37, P = 0.26; see Methods and Supplementary Information for the details). Note that one outlier (age = 8.04 yr) was identified as showing accuracy that was not reliably above chance and was >4 s.d. below the mean (open circle); because such low performance on the training dataset precludes our ability to interpret the results of any application to a different task, the data from this participant were excluded from all subsequent analyses. N = 81 participants are shown in the figure. c , Memory decoding approach. Left, the classifier trained on all localizer runs was applied to the fMRI study task patterns. We summarized the classifier evidence (probabilities) across AB trials by computing a single reactivation index per participant and repetition, which was defined as face minus scene evidence for AB F trials plus scene minus face evidence for AB S trials (that is, the interaction term). A reactivation index reliably above zero indicates that classifier evidence depends on trial type. Right, predictions for reactivation as a function of repetition for children (purple; no significant reactivation for any repetition), adolescents (magenta; reactivation on repetition two followed by potential drop-off on repetition three) and adults (coral; significant reactivation on repetitions two and three). The objects shown in the figure are from the Bank of Standardized Stimuli 133 , 134 . Images reproduced with permission from refs. 133 , 134 . Full size image We sought to address these hypotheses by first training an MVPA classifier to identify patterns of activation in anatomically defined ventral temporal cortex (VTC) associated with face versus scene viewing (Fig. 3a,b ). Classifier cross-validation performance was well above chance (one-sample t -test versus 0.5; mean = 89.42%, t (79) = 177.75, P < 0.001, Cohen’s d = 19.87, 95% CI = (88.4, 90.4)), demonstrating our ability to discriminate between face and scene viewing on the basis of VTC activation patterns. Perhaps more importantly, age did not explain significant variance in classifier accuracy (model comparison using the Akaike information criterion (AIC): AIC base = 1,581.5, AIC age = 1,500.2, F (3,76) = 1.37, P = 0.26), such that we were similarly able to decode perception of face versus scene stimuli across the age range. Related memories are reactivated during encoding We next applied our trained MVPA classifier to fMRI patterns from the pair study task to decode the contents of memory (Fig. 3c ). For each fMRI study pattern, the classifier returned continuous values reflecting the probability that it was associated with face processing or scene processing. Importantly, the participants were always viewing two objects during this task; however, the related content was either a face or a scene depending on the condition. For each participant, we generated a reactivation index for which values significantly above zero represent reliable reactivation of the related (more than the unrelated) content type during AB study. Repetition one serves as a baseline, as AB study occurs prior to encountering any overlapping (BC) face or scene content. Across the group, irrespective of age, there was statistically significant reactivation on the second (one-sample t -test versus 0; mean = 0.13, t (83) = 4.54, P < 0.001, Cohen’s d = 0.50, 95% CI = (0.07, 0.18)) and third (mean = 0.08, t (83) = 2.63, P = 0.01, d = 0.29, 95% CI = (0.02, 0.13)) repetitions but not the first (mean = −0.002, t (83) = −0.08, P = 0.94, d = 0.009, 95% CI = (−0.05, 0.05)). Reactivation indices were also significantly greater on both the second ( t (83) = 3.55, P < 0.001, d = 0.39, 95% CI = (0.06,0.20)) and third ( t (83) = 2.15, P = 0.03, d = 0.24, 95% CI = (0.01,0.15)) repetitions than on the first. These results suggest that on average, the participants showed neural evidence of reactivating the associated content type after it had been introduced. We next turn to assessing developmental differences in this signature. Development of reactivation reveals neural mechanistic shift We hypothesized that the transition into adulthood would be accompanied by an increased tendency to form integrated memories that link related experiences during encoding. A mature integration mechanism would predict that reactivation, once it occurs, would be maintained or elevated across repetitions 6 , and that such reactivation would be beneficial for subsequent inferential reasoning. Conversely, an active differentiation mechanism—whereby the commonalities across memories are detected yet de-emphasized during later encoding opportunities 4 , 59 , 66 —might yield initial reactivation that diminishes on subsequent repetitions, while in parallel, lateral PFC control systems ramp up to aid in interference resolution. Of note, such representations emphasizing the unique aspects of related memories may be used to make successful inferences while also supporting a host of other detail-oriented memory behaviours. We thus expected the direction of change in reactivation over repetitions—that is, whether reactivation increased (integration) or decreased (differentiation) from repetitions two to three—to vary with age. Importantly, brain imaging is required to achieve a direct quantification of such processes (potentially occurring outside of awareness) without influencing participants’ strategies. Consistent with our hypothesis, we found that age explained additional variability in reactivation scores above and beyond repetition alone (AIC base = 24.24, AIC age = 23.31, χ 2 (9) = 18.93, P = 0.03). There was a significant age-by-repetition interaction ( χ 2 (6) = 16.55, P = 0.011), demonstrating that how reactivation unfolded across repeated learning experiences was related to development (Fig. 4a ; the results were similar when excluding statistical outliers (‘ Outlier exclusions ’); interaction: χ 2 (6) = 16.22, P = 0.013). Fig. 4: Memory reactivation decoding results. a , Blockwise decoding results showing reactivation ( y axis) as a function of age plotted continuously ( x axis). Age significantly improved the model fit beyond the base model including only repetition (AIC base = 24.24, AIC age = 23.31, χ 2 (9) = 18.93, P = 0.03); in the better-fitting model with age, there was a significant age-by-repetition interaction ( χ 2 (6) = 16.55, P = 0.011). Repetition one serves as a baseline, as AB study occurs prior to encountering any overlapping (BC) face or scene content, and is not plotted here for the sake of simplicity. The coloured inverted triangles along the x axis indicate age points at which model predictions are shown in the subsequent panels. Adults maintain reactivation of related content across encoding repetitions, whereas adolescents show reactivation only on repetition two (light purple line; not significant on repetition three, shown in dark purple). Children show no significant evidence of reactivation. b , Model predictions from a visualized at four age points (10, 15, 20 and 25 years) across all three repetitions (light to dark). The asterisks denote age points and repetitions for which the model predictions are significantly above 0, indicating reliable reactivation according to the better-fitting model. The plots in a and b , represent 252 observations across N = 84 participants. c , Applying our classifier to individual trials rather than blocks yielded reactivation scores associated with each repetition of each specific pair. We found evidence for developmental differences (ages are shown by line colour and correspond to b ) in the direction of the within-participant relationship between reactivation change from repetitions two to three ( x axis) and subsequent inference performance ( y axis; interaction: χ 2 (3) = 8.13, P = 0.043). Specifically, while adults (coral and orange) were more likely to get an inference decision correct when reactivation increased from repetitions two to three (>0 on the x axis), children (10 yr, purple) showed the opposite pattern—reactivation decreases (<0 on the x axis) were associated with a greater probability of correct inference at younger ages. There was also a main effect of age, such that inference accuracy was greater for older than for younger participants ( χ 2 (3) = 20.60, P < 0.001). The figure displays model predictions at specific, user-defined age points; however, within the model, age was treated continuously. The plot in c represents 2,528 observations across N = 84 participants. Full size image Inspecting the resultant model fit curves showed that in late childhood and adolescence, there was a reliable decrease in reactivation from repetitions two to three. In fact, reactivation on repetition three did not exceed chance levels until mid-adolescence (16.09 years old), while repetition two reactivation emerged earlier (10.11 years old)—consistent with a differentiation scheme in this age range. The adolescent pattern contrasted with that in young adults ages 20 and older, who demonstrated the predicted integration signature, in which reactivation was above chance across both repetitions two and three. Consistent with prior observations of limited retrieval flexibility in children 36 , we saw no statistically significant evidence of reactivation before age 10; moreover, children of this age showed significantly less reactivation than adolescents. Model predictions can also be visualized at four age points in Fig. 4b . Together, these results show that there are fundamental shifts in the neural mechanism engaged during encoding of related memories. Children may not take advantage of commonalities across memories at all (as associations are never co-activated in the brain), while adolescents actively differentiate these experiences; in contrast, adults may tend to build up integrated representations that span related events. Reactivation variability across memories relates to behaviour We found developmental differences in the tendency to reactivate related memories during new learning. However, we also know that there exists variability at the specific pair level, such that integration and differentiation strategies can be used for distinct memories within a single individual 4 , 67 . We next leveraged this within-person variability to ask whether one type of encoding mechanism might be behaviourally advantageous for a given age—and critically, whether which mechanism is most beneficial might shift over development. In particular, under a differentiation mechanism, reactivation that is originally high and then drops could reflect initial memory co-activation followed by later suppression and strengthening of the individual AB and BC associations; such a memory scheme could support successful inference at retrieval. In contrast, inference through integration would suggest that pairs showing reactivation enhancements over time should be most likely to be correct. We hypothesized that, if there are developmental differences in how overlapping memories are represented and used to support inference, they might become apparent when looking at how reactivation is related to performance within individuals, on a trial-by-trial basis. We asked whether variability in the degree to which reactivation changed from repetitions two to three for a particular pair was associated with the probability of making a correct inference. Consistent with our hypothesis, we found a significant interaction between age and reactivation change score in inference accuracy (Fig. 4c ; interaction: χ 2 (3) = 8.13, P = 0.043; effect of age: χ 2 (3) = 20.60, P < 0.001; model comparison: AIC base = 1,773.1, AIC age = 1,761.8, χ 2 (6) = 23.36, P < 0.001). The nature of the interaction was such that among young adults, increasing reactivation from repetitions two to three was associated with a higher probability of making a correct response compared with when reactivation declined. In contrast, younger participants showed the opposite pattern: correct inferences were more likely on those trials showing reactivation decreases. The results were similar, albeit no longer exhibiting a statistically significant interaction, when outliers were removed (interaction: χ 2 (3) = 7.68, P = 0.053; effect of age: χ 2 (3) = 17.36, P < 0.001; model comparison: AIC base = 1,709.2, AIC age = 1,700.8, χ 2 (6) = 20.43, P = 0.0023). The ability to benefit from reactivation during encoding (that is, show a positive slope on the reactivation change–accuracy relationship) thus seems to emerge in early adolescence, sometime between 10 and 15 years of age. Of note, this age range is approximately the same as the one over which, on average, the participants showed reactivation initially on repetition two that declined on repetition three. This finding suggests that adolescents are engaging in a mechanism that is fundamentally different from adults—yet, it is one that does confer behavioural advantage. Reactivation impacts frontoparietal and hippocampal activation We found that changes in the level of ventral visual stream (that is, VTC) reinstatement of previously stored memories over repetitions predicts performance on a trial-by-trial basis—but that critically, the nature of this relationship changes over development. We next asked whether reactivation in VTC during repetition two was associated with subsequent (repetition three) neural engagement. In other words, where in the brain is VTC reactivation associated with later activity levels? We reasoned that reactivation might be resolved differently in the brain depending on one’s developmental stage. In particular, reactivation in the face of an inability to integrate should drive increased engagement of brain regions involved with memory suppression and interference resolution, such as inferior frontal gyrus (IFG). In tandem, one might expect initial reactivation to be associated with decreased later engagement of regions reflecting memory reinstatement, such as parietal cortex 24 , 68 , which would be further consistent with a suppression mechanism. We asked this question using a voxelwise general linear model (GLM) in which reactivation for a particular trial on repetition two (mean-centred within participants) was included as a parametric regressor in predicting fMRI activity on repetition three. (Importantly, we restricted our consideration of the relationship between reactivation and engagement to those measures observed on different repetitions. Beyond the theoretical reasons explained above, this choice also ensures independence of our measures and thus reduces the likelihood of spurious relationships reflecting more general neural fluctuations.) Clusters therefore represent regions for which there is a significant correspondence between the degree of reactivation on the preceding (second) repetition and activity during the final (third) repetition. The only region to show a significant effect across the group was left pHPC. This region showed a reliably negative relationship—that is, more reactivation on repetition two was associated with less pHPC activation on repetition three (intercept: F (1,83) = 14.92, P < 0.001). However, this effect was not significantly related to age (Fig. 5a ; P = 0.39), consistent with observations that pHPC matures early, showing signs of being structurally developed in early childhood 27 . Fig. 5: fMRI activation varies as a function of reactivation on preceding study repetition. a , pHPC showed a significant negative relationship between reactivation on repetition two and engagement on repetition three, such that greater within-participant evidence for reactivation in VTC was associated with less engagement of pHPC on the subsequent encoding experience (cluster size: 16 voxels, significant within HPC anatomical region of interest (ROI); F (1,83) = 14.92, P < 0.001). The nature of this relationship did not significantly differ over development ( P = 0.39). b , In parietal cortex, there was a negative relationship between initial reactivation and subsequent engagement that was unique to the child and young adolescent ages (LH effect of age: F (3,80) = 5.07, P = 0.003; RH effect of age: F (3,80) = 5.35, P = 0.002). In other words, the negative relationship present in the children was attenuated over development to the point of being absent in adults. The effects were similar in the LH (147 voxels) and the RH (142 voxels; both L and R clusters are significant at whole-brain, grey-matter level). c , IFG also showed developmental effects (LH effect of age: F (3,80) = 6.82, P < 0.001; RH effect of age: F (3,80) = 4.59, P = 0.005), with a positive reactivation–engagement relationship in children and younger adolescents (especially in the RH; 35 voxels, significant within IFG anatomical ROI) giving way to a negative relationship in older adolescents and adults (especially in the LH; 26 voxels, significant within IFG anatomical ROI). For all panels, the regions were selected for showing either a main effect of reactivation ( a ) or a reactivation–engagement relationship that differed with age ( b , c ). For all panels, N = 84 participants. COPE, contrast of parameter estimate. Full size image We additionally tested for regions in which the relationship changed with development (positively or negatively) by including age as a parametric regressor in the group-level statistical models. This analysis revealed two significant regions: bilateral parietal cortex (Fig. 5b ) and bilateral IFG (Fig. 5c ). In parietal cortex, a negative reactivation–engagement relationship in children and young adolescents was attenuated to no relationship in adults, consistent with the notion that only younger participants will suppress the internally generated content on subsequent encounters (left hemisphere (LH) effect of age: F (3,80) = 5.07, P = 0.003; right hemisphere (RH) effect of age: F (3,80) = 5.35, P = 0.002) 25 . Such an interpretation relates to the role of parietal cortex in reinstating high-fidelity memory representations in a way that is behaviourally relevant and influenced by top-down goals 24 , 68 . Reduced memory reinstatement at later points during study might be particularly advantageous for those memories that were initially reactivated most strongly, reflected in the fact that they are associated with relatively less parietal engagement on repetition three. IFG (bilaterally) showed the opposite pattern: relationships were positive in children and young adolescents but negative in adults (LH effect of age: F (3,80) = 6.82, P < 0.001; RH effect of age: F (3,80) = 4.59, P = 0.005). Such a result is in line with the interpretation that younger participants upregulate regions associated with interference resolution in response to high initial reactivation—perhaps aiding with active disambiguation of related memories 4 . The results were similar after removing outliers (parietal—LH effect of age: F (3,78) = 3.44, P = 0.021; RH effect of age: F (3,78) = 4.11, P = 0.009; IFG—LH effect of age: F (3,78) = 5.39, P = 0.002; RH effect of age: F (3,79) = 3.97, P = 0.011). Discussion We show that developmental differences in memory mechanisms influence how individuals of different ages make inferences about related episodes. Notably, we found that early adolescence was a unique period marked by initial reinstatement of memories during learning followed by later suppression—a signature consistent with differentiation of overlapping memories at this point in development. In contrast, adults showed enhanced reactivation consistent with integration at encoding, while children showed no significant evidence of reactivating at all and may store memories separately. These different memory mechanisms conferred age-specific behavioural advantages for inference: while suppressing reactivation benefitted those at the younger ages, enhancement was associated with correct inferences among adults. Interestingly, these differences emerged despite all participants being fully aware of the task structure and upcoming inference, thereby reducing the possibility that age-related differences in detecting overlap would be driving our effects. However, one limitation is that we did not assess the influence of overt strategy in this task; as such, whether younger learners can engage an integration mechanism when explicitly instructed to do so—or whether, as we would predict, their neural system acts as a fundamental limitation on this ability—remains an open question. That children under 10 years failed to reactivate memories during learning is consistent with prior work suggesting immaturity of HPC retrieval mechanisms before this age 36 , 69 . Importantly, this lack of memory reactivation was observed in the context of our ability to decode perception at all ages, during both the localizer (Fig. 3b ) and BC encoding trials in the main memory task (Supplementary Fig. 5a ). While this finding may seem incompatible with past work showing that children can benefit from learning new information that relates to their prior knowledge 70 , 71 , we suggest that this may be explained by certain features of our task that we designed to tap HPC mechanisms. A number of studies have shown subtle changes in HPC structure that continue into adulthood 16 , 26 , 28 , 48 , 72 and parallel behavioural gains in associative, detailed and recollective memory behaviours 49 , 73 , 74 —that is, those that depend on HPC 75 , 76 . We suggest that reactivation in our task requires a level of retrieval flexibility probably not present in children 36 , who might be less apt to bring to mind a related memory (here, BC associations) when confronted with a similar but not identical new experience (AB). In particular, the related experiences in our task are by design partially overlapping, meaning they (for example, A B ) provide only a partial match to the to-be-retrieved trace ( B C). Additional features of our task such as the relatively limited amount of encoding experience and large number of arbitrary pairs relative to related paradigms 14 might have further decreased the likelihood that children would reactivate related memories while encoding overlapping events. Future work will be needed to understand how this mechanism scales up to explain how more well-established, complex knowledge structures formed over extended experience may scaffold new learning in children 77 , 78 . This lack of reactivation in children would mean that two memories for overlapping experiences are formed in the same way as those for two non-overlapping experiences 58 , because related memories are never co-activated. Such a mechanism is consistent with our behavioural results, in which RTs did not differ for AB versus non-overlapping pairs, suggesting neither facilitation nor interference as a result of overlap. It would thus follow that the separately encoded but related memories for AB and BC associations would be stored and then separately accessed and recombined when faced with the AC inference decision; this hypothesis might suggest that retrieval-phase rather than encoding-phase neural signatures in the youngest children would be most related to inference success. One limitation of this work is that we are not able to test this hypothesis directly because we did not acquire fMRI data during the inference test; therefore, it remains an interesting question for future study. Adolescents showed evidence of initial reactivation, consistent with a level of HPC retrieval flexibility surpassing that of children. However, it is important to note that we measured reinstatement at the category level, reading out patterns in VTC as the product of HPC operations; we did not quantify the reinstatement of particular memories in HPC directly, which would require a different experimental design. With this limitation in mind, our results nevertheless converge with recent evidence showing adult-like HPC retrieval signatures in 13- and 14-year-olds 79 . More broadly, we saw evidence in adolescents for a unique neural encoding mechanism that differed from those of both children and adults. Of note, our results suggest that the adolescent period is a distinct stage 60 of memory development—not simply an intermediate step between childhood and adulthood as has been suggested in prior behavioural reports 16 , 17 . Combining our controlled behavioural task with an fMRI decoding approach, we have been able to characterize memory reactivation over development to reveal this insight into the adolescent brain. Our task included overlapping pairs that allowed us to ask how encoding and retrieval interact to influence memory formation in development. In adolescents, initial reactivation was followed by a notable drop back to baseline during a subsequent learning experience. Further reasoning that high levels of reactivation would elicit competition among memories and differential engagement of control regions (particularly among adolescent learners), we found that greater reactivation was associated with increased IFG engagement in children and adolescents, accompanied by decreased recruitment of parietal cortex. These findings align well with IFG-guided suppression of memories activated in parietal cortex in this age range 25 and are broadly consistent with prior work highlighting developmental differences in controlled aspects of memory in general 80 , 81 and IFG in particular 41 , 82 that track age-related memory improvements. Our findings go beyond prior work to provide a key mechanistic example of how controlled encoding operations might contribute not only to the quality of memories stored but also to their contents and organization. We suggest that in adolescents, reinstatement followed by subsequent study will weaken the connections between memories, as has been suggested previously 59 , leading to memories for related experiences becoming more distinct from one another than two unrelated experiences 4 , 58 , 66 across repetitions 4 , 66 , 83 , 84 . We propose that differentiated representations are beneficial to inferential reasoning in adolescents and yet simultaneously require that they are engaging a fundamentally different mechanism from adults—namely, one in which they recombine memories at retrieval 85 . Such a proposal is in line with previous work on the development of reasoning, which highlights that ongoing maturation of controlled retrieval processes (selecting individual task-relevant memories) supported by IFG 86 and frontoparietal connections 87 underlies the performance gains observed into adolescence. Here, we extend these ideas by jointly incorporating memory and reasoning components in our task, highlighting that in addition to these retrieval differences, there is important developmental change in memory organization due to ongoing maturation of complex encoding mechanisms. Our approach thus links memory with reasoning literatures to show how traditionally conceptualized memory mechanisms guide how knowledge is organized—and therefore ultimately constrain how we might use knowledge to make flexible decisions. The ability to make decisions that span multiple memories is a critical component of behavioural flexibility. In children, this ability is related to academic achievement 19 , 20 , underscoring that the importance of understanding developmental change in this mechanism goes well beyond the lab. Here, we provide neural support for previous suggestions that children do not store memories with respect to their shared content, and we extend this framework into an understudied period of memory development to uncover an adolescent-specific neural phenomenon. Our results directly linking memory operations to the later ability to reason about those memories represent an important step towards bridging these literatures. More directly, these data suggest that child, adolescent and adult learners may rely on different mechanisms to achieve maximal behavioural flexibility—an idea that might be tested in future research and educational settings. Methods Participants All experimental procedures were approved by the Institutional Review Board at the University of Texas at Austin. One hundred and twenty-five volunteers ranging in age from 6 to 30 years (actual range, 6.41–29.33) made up the cross-sectional sample who participated in a behavioural screening session (‘ Experiment overview ’). Adult participants provided informed consent, and permission was obtained from one or more parents or guardians of minor participants (that is, individuals under the age of 18 years). Minors additionally provided informal assent. The participants were compensated US$10 per hour for the first session (mock scanner) and US$25 per hour for the second session (MRI); they also had the opportunity to earn an additional US$5–US$15 in bonus pay based on performance during the MRI session (‘ Memory task ’, ‘ Motivational interlude ’ section). Of this initial group of 125 volunteers, 97 returned to the lab for the MRI session. Reasons for exclusion prior to the MRI session were: opted out or otherwise unable to schedule the scan session ( N = 6 minors and 8 adults (18 years or older)); had a Child Behavior Checklist Total Problems Score ( N = 5 minors) or Symptom Checklist 90-Revised Global Severity Index ( N = 4 adults) in the clinical range; left-handedness ( N = 1 minor); had contraindication(s) to MRI ( N = 2 adults); and diagnosed with a psychiatric condition or learning disability ( N = 2 minors). No participants scored below our inclusion threshold for IQ (>2 s.d. below the mean of FSIQ-2). Of those 97 participants who were scanned, 11 were excluded from all further analyses for the following reasons: did not provide at least two fMRI runs of the encoding task due to terminating the session early ( N = 3 minors) or excessive motion, defined as fewer than two encoding runs with less than one-third of the time points exceeding our framewise motion threshold (see below; N = 6 minors); incidental finding ( N = 1 minor); and technical difficulties with data acquisition ( N = 1 minor). The final sample reported here includes 86 individuals whose ages on the date of MRI ranged from 7.16 to 29.42 years. Minors made up most of our sample ( N = 65 participants), and efforts were made to achieve approximate sex balance among both minors (35 females) and adults ( N = 21 total participants; 11 females). Our target sample size was 21 in each of four age bands: children 7–10 years, younger adolescents 11–14.5 years, older adolescents 14.5–17 years and adults 18–30 years. Power analyses using data 4 from a similar task showed that N = 21 participants would yield 80% power to detect the behavioural effect of a within-participant manipulation of integration at the group level (Cohen’s d = 0.56); as such, we recruited participants until we had a minimum of 21 per band that could be included in our primary reactivation analysis. Our sample size also aligns with prior developmental work on a similar topic 16 , 17 . Two participants were excluded from the reactivation analysis specifically, and as such, our overall sample size was slightly larger than this minimum at 86 participants. Note that these age bands were arbitrarily defined and used only to ensure even sampling across the age range, with greater representation among the narrower bands for the developing groups (7–17) relative to the adult group; however, all analyses reported here treat age as a continuous variable. In addition to participant-level exclusions, we excluded study runs that were (1) high motion, defined as more than one-third of the volumes with fast motion (‘ Motion-related participant-level and run-level exclusions ’), (2) incomplete or (3) associated with poor subsequent memory, defined as test performance for the direct pairs (AB, BC and non-overlapping) not reliably above chance (binomial test; minimum 13 correct trials of 24 total). Most participants contributed all four study runs (79.07% of participants; mean, 3.76 runs; range, 2–4; 95% CI, (3.65, 3.86)) and all three localizer runs (89.41% of participants; mean, 2.85 runs; range, 1–3; 95% CI, (2.74, 2.95)). Participant information, including how many runs (out of 7 total) were contributed by each person, are displayed in Supplementary Fig. 1a . The time between sessions ranged from 0 to 76 days (mean = 15.78, median = 13.5; Supplementary Fig. 1b ). For all analyses, a participant’s reported age is their age in years and months (converted to decimals) on the MRI session date. Experiment overview Data collection and analysis were not performed blind to the conditions of the experiments. The experiment unfolded across two sessions that usually took place on separate days (Supplementary Fig. 1b ). The primary purpose of the first session was to determine whether the participant would continue to the MRI session on the basis of eligibility and interest. During the first session, all participants (with their parents, if minors) were first exposed to the mock MRI scanner. Audio recordings of scanner noises were played over speakers while the participants lay supine in the mock scanner bore. The participants or their parents also provided information on demographics, socio-economic status and pubertal stage (Petersen Development Scale 88 ; participant-completed, only for ages 8–17 yr; these measures were for exploratory purposes only and are not considered further). The participants were screened for the presence of psychiatric symptoms using the Child Behavior Checklist 89 (parent-completed) for minors or Symptom Checklist 90-Revised 90 for adults. The participants also completed the Wechsler Abbreviated Scale of Intelligence, Second Edition 91 as a measure of IQ. In addition, the participants completed a stimulus rating task, enabling us to custom-select the faces and scenes that were familiar for each participant (‘ Memory task ’, ‘ Stimuli ’ section). Finally, the participants practiced both the memory and repeat detection tasks that would be performed in the MRI scanner on day 2. The practice tasks included different stimuli from the main experiment. Memory task Stimuli The memory task stimuli were familiar faces and scenes (20 per category) from popular children’s movies, as well as 160 common objects. The faces, scenes and objects were organized into 32 ABC triads (that is, groups of three stimuli: A, B and C) and 32 non-overlapping pairs. Of the 32 ABC triads, 16 were object–object–face (A O B O C F ), and 16 were object–object–scene (A O B O C S ), for which the face and scene stimuli were always in the C item position. All 32 non-overlapping pairs comprised two objects. Four triads and pairs were created for practice stimuli (four per condition). During study (‘ Pair study phase ’), the triads were presented to the participants as ‘overlapping’ AB and BC pairs, which are related by virtue of the identical B item (always an object) in an associative inference task 4 , 5 , 6 , 7 , 8 , 12 , 16 , 62 , 63 , 92 , 93 . The participants later inferred the relationship between A and C. This task is similar to others used to measure integration more frequently in the developmental literature, in which even younger children learn overlapping facts 9 , 14 , 18 , 20 , 94 , 95 , 96 , 97 , 98 , 99 , 100 , 101 or inequalities 102 , 103 , 104 , 105 to derive knowledge. We chose the associative inference task because participants can learn many arbitrary pairings of stimuli, thereby affording more trials; also, the straightforward nature of the content allows us to detect the retrieval of a single, held-out item (C) during learning. Custom selection of scenes and faces Our goal was to quantify the degree to which reactivation of a previously associated (C) content type (face or scene) is reflected in the neural patterns engaged during AB encoding, when visual presentation is held constant (always two objects). We anticipated that reactivation might be more likely for highly familiar stimuli 106 , 107 , 108 , 109 , 110 ; thus, we attempted to both maximize the familiarity of the C items for each person and equate it across conditions (scene versus face triads) and ages. To that end, faces and scenes were selected custom for each participant from a larger set according to their responses on a separate familiarity rating task completed during the behavioural screening session (day 1). In the familiarity rating task, the participants were shown up to 225 images (126 faces and 99 scenes) in a random order one at a time on a computer screen. For each image, the participants indicated how familiar they were with the picture using the following options: not at all (coded as 1), a little bit (2) or very (3). The participants made their responses verbally, and the experimenter input their choice into the computer. For pictures rated as very familiar, the participants were also asked to name the character or describe the scene. The experimenter scored these responses during the task as either correct or incorrect and input their accuracy (1 or 0) into the computer. From these ratings, stimuli were selected to maximize familiarity for each person, with the additional constraint that only one image from a given movie or show could be selected for a particular participant. For example, while participants might view multiple characters and scenes from the movie Frozen during the stimulus rating phase, only one image from Frozen would appear in the final task. Familiarity ratings were automatically calculated during the task such that the task ended as soon as a participant achieved maximum familiarity for a full stimulus set (that is, 20 faces and 20 scenes each from a unique movie were all rated a 3). Thus, the majority of participants did not make familiarity ratings for all 225 stimuli in our set. The average familiarity ratings for faces and scenes selected for the memory task are shown in Supplementary Fig. 2 . Selection of objects A single set of 160 objects was used for all participants. We made this choice for two reasons. First, from a logistical perspective, having participants rate familiarity for such a large set of objects would have made our behavioural (day 1) session prohibitively long. Second, more critically, as our goal was to measure face and scene (not object) reactivation during encoding, differences in familiarity among the object stimuli would not bias our results towards any particular outcome. Thus, in place of custom selection, we chose 160 objects that would probably be familiar to participants spanning our age range, taking into account published normative data on age of acquisition 111 for the objects’ names. In particular, we reasoned that if an object name was learned early in life, a photograph of that object would probably also be familiar to a child around the same age (or younger). The objects selected for our final set had ages of acquisition ranging from 2.5 to 14.67 (mean = 5.53, median = 5.42) years. The assignment of objects to conditions was determined randomly for each participant. Pair study phase There were four study–test cycles that each contained a unique set of pairs. During the scanned pair study phase (Fig. 1a ), the participants saw AB, BC and non-overlapping pairs on the screen (3.5 s stimulus presentation, 0.5 s inter-stimulus interval (ISI)) and were encouraged to imagine the two items interacting to aid their memory. No response was required during the pair study trials. There were a total of four triads per condition plus eight control pairs per run, yielding a total of eight AB (object–object; four related to a face, four related to a scene), eight BC (four object–face, four object–scene) and eight non-overlapping (object–object) pairs. The pairs were blocked by type (AB F , AB S , BC F , BC S , non-overlapping 1 and non-overlapping 2 , for which the non-overlapping pairs were arbitrarily split into two ‘conditions’ to match the triad block structure). The four pair-encoding trials within each block were jittered by interspersing them with a variable number of baseline task trials (1.5 s stimulus, 0.5 s ISI; range, 0–2; mean, 1 baseline trial between pair-encoding trials), during which the participants indicated with a button press the location at which a dot appeared on the screen (left, middle or right box). This jitter with baseline trials meant that the delay between pair-encoding trials (that is, from the offset of one pair to the onset of the next pair) ranged from 0.5 s (for zero intervening baseline task trials) to 4.5 s (for two intervening baseline task trials). The block durations were held constant at 24 s, and there was no additional interval between blocks. This mixed fMRI design enabled us to both extract single-trial estimates and analyse our data as a traditional blocked design. Each pair was presented three times across the run. This repetition gave the participants multiple opportunities to learn each pair, thus ensuring adequate memory, and it allowed for the possibility that neural signatures of differentiation 4 , 58 , 59 , 66 , 112 or integration would evolve—or only appear—across repeated experiences 6 , 63 . Repetitions were distributed across thirds of the run such that every pair was presented once before being shown a second time, and twice before being shown a third. The order of pair-encoding blocks was further constrained such that (1) two blocks of the same general type (AB, BC or non-overlapping) always occurred back-to-back, with the specific order shuffled across repetitions and runs within participants; (2) BC blocks occurred last within the repetition; and (3) AB and non-overlapping blocks occurred first within a repetition equally often for each participant. This final constraint was implemented to ensure that AB and non-overlapping blocks did not differ in their average delay from BC blocks, when faces and scenes were presented, as systematicity in this regard could have influenced our comparison of AB versus non-overlapping blocks. Test phase After each pair study phase, the participants completed a self-paced inference and memory test for the immediately preceding pairs (Fig. 2 depicts performance). The test was not scanned. We first tested the participants on their ability to make inference judgements for all eight ABC triads prior to testing any of the direct associations. This ordering was chosen to prevent further direct pair learning during the test that might influence inference behaviour. The participants first completed a category-level, two-alternative forced-choice judgement for all triads (Fig. 1b , left), in which they were presented with the A item (always an object) and asked to indicate whether the C item indirectly related through association with a common B was a face (character) or a scene. After completing all category-level inference trials, the participants then identified the specific face or scene indirect (C) associate for every A object in a three-alternative forced-choice test (Fig. 1b , right). Again, the A item (object) served as the cue, and C items served as the options. We included the category-level inference judgement to assess whether participants could recall some information about the indirectly related item when the correct answer was not currently present. Hereafter, we consider correct inference trials to be those for which participants got both the category-level and item-level judgements correct. This strategy has the benefit of reducing the likelihood that a correctly guessed item-level inference test trial will be treated as correct. Following the inference test, we tested the participants on their memory for the directly encoded pairs (AB, BC and non-overlapping; Fig. 1c ) in the same manner as the item-level inference test. A items were cues for all AB test trials, and B items were cues for all BC test trials. For all item-level inference and memory test trials, foils (incorrect options) were always other studied items of the same condition, position (A, B or C for overlapping pairs) and study run, to prevent the participants from using information other than the specific associative relationships to make their decisions. Note that because the foils were same-condition and same-position, the foils were always matched in stimulus type (face, object or scene) to the correct answer. Motivational interlude After completing each test, the participants viewed an animation of their avatar (chosen at the beginning of the experiment) climbing a staircase (Fig. 1d ). The distance the character moved was proportional to the participants’ accuracy on the direct pairs (AB, BC and non-overlapping) in the immediately preceding test. The staircase had three goal levels (represented by stars), and the participants were informed before beginning the experiment that they would receive a bonus payment in the amount of the highest star goal they had reached: US$5, US$10 or US$15. Our intention was to motivate the participants and keep them engaged with the task; as such, the threshold to reach the first goal was set low enough that all participants received some amount of bonus payment. The participants needed to achieve accuracies of 30%, 59% and 88% over the course of the whole experiment to earn a US$5, US$10 or US$15 bonus, respectively. After viewing the animation, the participants continued on to complete another study phase with new pairs until they completed all four cycles. Visual localizer (repeat detection) task The participants also performed a separate 1-back repeat detection task with faces, scenes, objects and scrambled objects. The fMRI data acquired during this task were used to train our MVPA classifier to decode viewing of different stimulus types. The visual localizer task always took place after all four study–test cycles of the memory task were completed and thus did not interfere with the memory task data. Stimuli The repeat detection task stimuli were 72 familiar faces, 72 scenes, 72 intact common objects and 72 scrambled common objects. The face and scene stimuli were drawn from the same set as those used in the main memory task, but for a given participant were different from those selected for the memory experiment. Task design The participants viewed the stimuli on the screen one at a time for 1.5 s with a 0.5 s ISI. The participants indicated with a button press when a stimulus was identical to (that is, an exact repeat of) the immediately preceding picture. The stimuli were blocked by type (six presentations per block, for a total block duration of 12 s), and there was exactly one repeat per block. There were four blocks of each stimulus type per run, as well as five baseline blocks of the same duration. During the baseline blocks, the participants performed the same baseline task as during encoding, in which they indicated the location of a dot in an array of three boxes (1.5 s stimulus, 0.5 s ISI). The participants completed up to three runs of the visual localizer task. Behavioural responses were collected purely to ensure that the participants were paying attention to the stimuli (Supplementary Fig. 3 ) and are otherwise not considered in our analyses. MR data acquisition Imaging data were acquired on a 3.0T Siemens Skyra MRI system. Functional data were collected in 75 oblique axial slices using an EPI sequence, oriented approximately 20° off the AC–PC axis (TR, 2,000 ms; TE, 30 ms; flip angle, 73°; 128 × 128 × 75 matrix; 1.7 mm isotropic voxels; multiband acceleration factor, 3; GRAPPA factor, 2). Between one and three field maps were collected (TR, 589 ms; TE, 5 ms/7.46 ms; flip angle, 5°; matrix size, 128 × 128 × 60; 1.5 × 1.5 × 2 mm voxels) for each participant to correct for magnetic field distortions. Field maps were planned (1) before the first study run, (2) before the first visual localizer run and (3) any time a participant came out of the scanner for a break. Four participants had only one field map acquired due to technical difficulty and/or operator error. Two or three oblique coronal T2-weighted structural images were acquired perpendicular to the main axis of the HPC and in approximately the same orientation as one another (TR, 13,150 ms; TE, 82 ms; 384 × 60 × 384 matrix; 0.4 × 0.4 mm in-plane resolution; 1.5 mm through-plane resolution; 60 slices; no gap); these images were not incorporated into the analysis for the present manuscript. A T1-weighted three-dimensional MPRAGE volume (256 × 256 × 192 matrix, 1 mm isotropic voxels) was also collected for automated segmentation using Freesurfer 113 and spatial normalization to the MNI template brain using Advanced Normalization Tools (ANTS) 114 . fMRI preprocessing The data were preprocessed and analysed using FMRI Expert Analysis Tool (FEAT) Version 6.00, part of FMRIB’s Software Library (FSL) Version 5.0.9 ( ), and ANTS 114 . Motion correction was applied to each functional run using MCFLIRT, and then non-brain structures were removed using BET, both part of FSL. All functional runs were then registered to the middle functional ‘reference’ run (in most cases, the third study run) by applying affine transformations calculated in ANTS. Anatomical images (mean coronal, MPRAGE) were then registered to the functional reference run after field-map-based unwarping of the functional data (implemented in FEAT as part of GLM analysis; see below) as follows. Each participant’s MPRAGE was directly registered to their functional data using ANTS affine transformations. Non-brain structures were removed from the anatomical images using a mask derived from Freesurfer output. The result of the registration process was that all data (functional and structural, including Freesurfer parcellations) were coregistered in each participant’s native functional space. All analyses were carried out in this native space except group-level GLMs. Pre-statistics processing In preparation for both univariate (GLM) and multivariate (MVPA) analyses, the following pre-statistics processing was applied: field-map-based EPI unwarping using PRELUDE+FUGUE, spatial smoothing using a Gaussian kernel with a full width at half maximum of 4 mm, grand-mean intensity normalization of the entire four-dimensional dataset by a single multiplicative factor, and highpass temporal filtering (Gaussian-weighted least-squares straight line fitting, with σ = 50 s). For most participants, the first field map was used to unwarp all study scans, and the second was used to unwarp all visual localizer scans. However, in many cases, participants took breaks between scans during which they were taken out of and then put back into the scanner to yield several mini scanning sessions. For these participants, the field map from the same mini-session was selected because it would most closely match the functional run in question in terms of the participant’s physical positioning in the scanner. In other words, we always chose the field map that would best reflect magnetic field inhomogeneities for the particular head position in a given functional run. Separate field maps were collected for all but four participants (‘ MR data acquisition ’) for the visual localizer and study runs to ensure a similar quality of correction for both phases of the experiment, which would serve as the training and test data for the MVPA classifier, respectively. Motion-related participant-level and run-level exclusions Realignment parameters from MCFLIRT were used to compute framewise displacement (FD) for each fMRI volume. For each participant and run, we then defined the number of ‘bad’ volumes as those exceeding an FD threshold of 0.5 mm, plus one volume before and two after each high-motion volume. We used these numbers to exclude runs for which more than one-third of the total run time was corrupted by motion. As noted in the ‘ Participants ’ section, we required that participants have at least two study runs that met this criterion to be included in any analyses, and at least one visual localizer run to be included in the multivariate analyses; however, the majority of participants contributed all runs for both tasks (Supplementary Fig. 1a ). ROI definition Content-sensitive ventral visual stream regions were defined anatomically for each participant by summing entorhinal cortex, fusiform gyrus, inferior temporal cortex and parahippocampal gyrus regions identified by Freesurfer. The resulting VTC region was used to mask functional data for MVPA. We also defined regions in MNI template space for small-volume correction of univariate analyses. Medial PFC was delineated by hand on the 1 mm MNI template, restricted to those regions in the “medial prefrontal network” described in previous work 115 . We used the Harvard–Oxford atlas to define both IFG and HPC ROIs. Univariate fMRI analysis Estimation of condition-level activation The task data were interrogated for regions that showed differential engagement during encoding of overlapping (AB) as compared with non-overlapping object–object pairs. The data were modelled using a GLM implemented in FEAT Version 6.00. Because we anticipated large gains in memory 47 , 52 , 53 , 77 , 78 , 116 , 117 , 118 , 119 , 120 and were interested specifically in developmental differences in the mechanisms engaged during successful overlapping versus non-overlapping encoding, we limited our analysis to those trials that were later remembered (that is, correct on the corresponding direct pair test). Individual trials (pair presentations and baseline trials) were modelled as 3.5 s events and convolved with the canonical (double-gamma) haemodynamic response function (HRF). The trials were split according to condition (AB F , AB S , BC F and BC S ; non-overlapping trials were split into two groups in a parallel fashion as the overlapping pairs) and repetition (one, two or three), yielding a total of 18 regressors of interest. Subsequently, incorrect trials were collapsed into a single regressor of no interest. The baseline task was also modelled in a separate regressor. Temporal derivatives were included for all task regressors. Motion parameters calculated during the motion correction step and their temporal derivatives were added as additional confound regressors. FD and DVARS, two measures of framewise data quality, were also added to the model as regressors of no interest 93 , 121 . Temporal filtering was then applied to the model. After modelling functional data within each run, we combined the resulting statistics images across study runs for each participant using fixed effects. As the data were already coregistered across runs, no additional registration or spatial normalization was necessary. Overlap-sensitive regions were defined as those that responded more on repetitions two and three (that is, after overlap had been introduced) for AB versus non-overlapping encoding (AB > non-overlapping) and vice versa (non-overlapping > AB), irrespective of the associated C item’s content type (that is, collapsed across AB F and AB S trials). We reasoned that any region that differentially responded to these AB and non-overlapping pairs must be involved in detecting or resolving overlap, and we were thus interested in both directions of this contrast. Contrast images for each participant were then warped to the 2 mm isotropic MNI template using ANTS and combined across participants using permutation tests (one-sample t -test; 1,000 iterations) implemented in FSL’s randomise 122 . As we wanted our later assessment of age-related differences in sensitivity to overlap (‘ Assessing effects of age ’) to be independent of region definition, age was not incorporated into the group analysis. The resulting group statistical maps were thresholded at a voxelwise P < 0.005 and submitted to cluster correction as follows. Smoothness was estimated using the residuals (warped to MNI template space) from every study run for each participant using AFNI’s 3dFWHMx utility. We used the spatial AutoCorrelation Function estimation method (-acf flag), which no longer assumes a Gaussian noise distribution and generally results in a larger (more conservative) estimate of smoothness relative to prior releases of this tool, thus reducing the likelihood of a type I error 123 . We then used these run-level values to compute the average smoothness parameters across all encoding runs within participants, and then finally across participants to yield a group-level mean smoothness estimate. This entire analysis was done separately for each ROI (grey matter, HPC, IFG and medial PFC) within which cluster correction was performed. The minimum cluster extents at a significance threshold of P < 0.05 were determined for each ROI using 3dClustSim. The minimum cluster sizes were determined to be 10 voxels for HPC, 17 voxels for IFG, 27 voxels for medial PFC and 71 voxels for grey matter. All clusters exceeding these criteria within the three a priori anatomical regions 2 , 4 , 5 , 6 , 7 , 8 , 12 , 93 and/or at the whole-brain grey matter level are reported here. Assessing effects of age All overlap-sensitive regions were identified for showing a main effect of overlapping versus non-overlapping pairs, irrespective of age. To determine whether there were in fact effects of age present within these clusters, we extracted contrast estimates (COPEs) for each participant and condition (AB and non-overlapping). We used linear models (‘ Statistical analyses ’) to assess whether the activation difference observed between encoding of overlapping and non-overlapping pairs was modulated by age. Note that because the functional regions were identified for showing either AB > non-overlapping or the reverse, the effects of condition are trivial; we were specifically interested in whether there were significant effects of age and/or interactions between age and condition. We used a model comparison approach to ask whether age explained any additional variability in activation beyond condition. Predictions from the best-fitting model and statistics for significant regions are provided in Supplementary Fig. 5 and Supplementary Table 1 , respectively. Estimation of trial-level neural patterns In addition to condition-level univariate analyses, we extracted neural patterns for individual trials. These patterns were used as inputs to our trialwise classification analysis (see below). Trial-level neural patterns were generated under the assumptions of the GLM using a modified LS-S approach 124 . Statistics images associated with each encoding trial were estimated for each repetition and participant using custom Python routines. Pair presentations were modelled as 3.5 s events and convolved with the canonical (double gamma) HRF. Motion parameters calculated during the motion correction step and their temporal derivatives were added as additional confound regressors. As in the other univariate models, FD and DVARS were also added to the model as regressors of no interest 94 , 122 . Temporal filtering was applied to the model. This process resulted in one statistic image for each of the 32 AB pairs, 32 BC pairs and 32 non-overlapping pairs for each of three repetitions, for a total of up to 288 images per participant (those participants contributing fewer runs had correspondingly fewer images). Multivariate fMRI classification analysis Blockwise reactivation analysis Our main classification analysis was carried out on the preprocessed time-series data in the native space of each participant. We first asked whether our classifier could discriminate between face and scene viewing during visual stimulus presentation (visual localizer task). We then applied our trained classifier to assess reactivation of related face or scene memories (C item content type) during study of overlapping (AB) object–object pairs. Decoding visually presented content We assessed whether the classifier could discriminate between the viewing of faces and scenes on the basis of VTC activation patterns from the visual localizer (repeat detection) task using a within-participant cross-validation approach. Specifically, we trained a pattern classifier (sparse multinomial logistic regression implemented in PyMVPA; λ = 0.1, the package default) to differentiate face from scene viewing on the basis of activation patterns from a subset of localizer runs (Fig. 3a ). We trained on all six volumes (12 s) of data in each face and scene block, with volume labels shifted by 6 s to account for haemodynamic lag. The classifier was then tested on patterns from the held-out run (one ‘fold’). This approach was repeated until all runs had been held out once. Cross-validation was performed on detrended and z -scored data within anatomically defined VTC; no further feature selection was performed. Accuracy was computed by comparing the classifier-predicted to actual stimulus type (face or scene) for each fMRI volume; an average accuracy was then calculated across all volumes and all folds to yield a single decoding accuracy score per participant (Fig. 3b ). Five participants (four children ages 8–9 and one adult age 19) were excluded from this cross-validation analysis because they contributed fewer than two fMRI runs of the localizer task, making it impossible to perform cross-validation across runs on these participants (total N = 81). Decoding internally generated content (reactivation) Our next analysis was designed to determine whether there are developmental differences in the tendency to reactivate related memories during encoding. To assess this, we trained our classifier (sparse multinomial logistic regression, λ = 0.1 as above) on all localizer task runs that met our inclusion criteria for each participant. The classifier was then applied to all study task volumes (Fig. 3c ). As above, this analysis was performed on detrended and z -scored data within VTC, with no other feature selection applied. We then computed a ‘reactivation index’ over the classifier evidence (probabilities) that summarized, for each participant, the degree to which their neural patterns reflected reinstatement of the related more than the unrelated type of content. The reactivation index was defined as face minus scene evidence for AB F trials plus scene minus face evidence for AB S trials (that is, the interaction term). Reactivation indices above zero thus indicated that classifier evidence was dependent on trial type. As control analyses, we also computed the same score for BC trials (‘perception index’) and non-overlapping trials (‘control index’), which should yield decoding of the perceived BC stimulus type and chance-level decoding of nonsense non-overlapping input, respectively. Two participants (eight- to nine-year-old children) were excluded from this analysis. One child had no localizer task data, and therefore this within-participant analysis was impossible; the other was an outlier in cross-validation accuracy for decoding of perception in VTC and was thus excluded (Fig. 3b , open circle; total N = 84). Note that for this analysis only, we decided to exclude this outlier participant, for whom classification performance was not reliably above chance. The reason for this exclusion was thus not so much that this person was an outlier per se, but rather that their results of applying a classifier that cannot discriminate among conditions in the training set will be uninterpretable when applied to a separate task (here, pair encoding). This group of 84 participants were included for all subsequent reactivation-related analyses that follow. Trialwise reactivation analysis Having established developmental differences in reactivation at the block level, we next quantified reactivation on a trial-by-trial basis. This analysis was performed to ask whether there are developmental differences in the behavioural and neural consequences of reactivation within participants. We used the same classifier trained to discriminate face from scene viewing as we did for the blockwise decoding analysis. However, instead of applying the classifier to each volume in the study runs, we applied it to the trial-level neural patterns (statistics images that reflected each repetition of each pair). This yielded an estimate of the degree to which each specific pair reflected reinstatement of its related type of content for each repetition. We then computed the log odds of the classifier output corresponding to the condition of interest as our trialwise measure of reactivation. For example, log[prob F /(1 − prob F )] would reflect face evidence for face-related trials. This transformation has been used in previous work 125 to correct for non-normality in the raw classifier output, which we also observed here. As our goal was to assess developmental differences in the neural mechanisms involved during successful memory formation rather than differences in memory ability per se, we restricted our analyses to correctly remembered trials only. Relating initial reactivation to subsequent engagement To assess whether initial reactivation modulates subsequent engagement, we asked whether the trialwise reactivation measures described above (‘ Trialwise reactivation analysis ’) from repetition two were related to activation on repetition three. Trialwise reactivation scores were mean-centred within participants and included as a trial-by-trial parametric modulator for all repetition-three AB trials. These GLMs were otherwise identical to the main models, except the addition of this parametric regressor. As in the previous analysis, only trials for which the corresponding direct memory test was correct were included. Note that because our reactivation measures were derived from repetition two to assess activation differences during the subsequent repetition three, the measures are coming from different time points and represent independent data, and the only relationship is through the pair (content) itself. Statistics images were then combined across runs within participants as above for the main models. At the group level, we were interested in effects that were consistent across the group as well as those that varied with age. We thus ran one main effects model disregarding age (mirroring our general approach in the main analyses) and a separate model that included mean-centred age as a parametric regressor. Both analyses were run using FSL’s randomise, as above. Cluster correction was performed using the same method as for the main models. Within identified clusters, we extracted the participant-level contrast estimates (COPEs in FEAT) associated with the parametric regressor for visualization of the effects (Fig. 5 ). Statistical analyses Model specifications As this study is a cross-sectional developmental study, age was an across-subjects factor; all other measures were repeated within subjects. Statistical analyses (except those carried out in FSL) were performed using R 126 . We primarily used (generalized) linear mixed effects models implemented in the lme4 package 127 to model individual trials, except when we had only one observation per participant (in which case we used linear models; stats::lm). For models assessing within-participant relationships (and optionally interactions with age), the predictors were scaled and centred within participants to remove subject-specific effects. Factors were effect coded to allow for the interpretation of lower-order terms as main effects in the presence of interactions. Repetition was typically treated as a factor so as not to require consistently increasing or decreasing reactivation across repetitions; one exception to this was for the analyses of motion ( Supplementary Results , ‘Effect of increased motion over repetitions is not significantly modulated by age’ section), in which we did expect consistent increases across repetitions. In addition, because the developmental trajectories in question are potentially nonlinear 31 , 40 , we opted to model age with a basis spline function. This approach uses a linear combination of basis functions, thereby allowing us to remain agnostic as to the particular shape of the relationship 128 . Basis splines have the advantage of being fit locally (that is, separately at specific parts of the age range). Such local fitting means that basis splines are less affected by values at either extreme end of our age range (that is, the youngest and oldest participants in our sample) compared with polynomials, which are fit globally 129 . In all analyses, participants were treated as random effects. Model comparison and statistical reporting For all analyses, we took a model comparison approach in which a model including age was compared with a base model in which age was not considered. The R package stats::anova was used to perform model comparison, and the age model was said to significantly improve on the base model at a threshold of P < 0.05 (two-tailed; uncorrected). We then report the statistics for the better-fitting model, either the base (in cases where adding age to the model did not significantly improve the fit) or the one additionally incorporating age (in cases where adding age did significantly improve the fit). We assessed statistical significance of each of our fixed effects including interaction terms in the better-fitting model using a Wald chi-squared test (type III SS for models including interaction terms, type II otherwise) for linear mixed effects models (lme4::lmer and lme4::glmer) and F -test for linear models (stats::lm; all Wald chi-squared tests were implemented in R using car::Anova) 130 . We used ggeffects::ggemmeans 131 in R to visualize the predicted responses and compute CIs at various ages (all data figures). Outlier exclusions We did not incorporate outlier exclusion into our primary analysis, with the single exception of removing one participant for whom we could not decode perceived stimulus type because this precluded the application of the trained classifier to the main memory task (Fig. 3b , open circle). However, to ensure that our other findings were not disproportionately influenced by outliers, we verified that the results were similar after removing statistical outliers, or data points with a standardized residual greater than 2.5 (in R, LMERConvenienceFunctions::romr.fnc with the default settings; the results excluding outliers are reported throughout the main text). For the analysis shown in Fig. 4c , all trials associated with the repetitions that were identified as outliers from the corresponding memory reactivation analysis were excluded (linear mixed effects model shown in Fig. 4a,b ; a total of five repetitions were excluded, one each from five participants), as identifying outlier observations with a binomial linking function is not straightforward. Reporting Summary Further information on research design is available in the Nature Research Reporting Summary linked to this article. Data availability The data that support the findings of this study are available on the Open Science Framework ( ) 132 . Code availability The custom code that supports the findings of this study is available on the Open Science Framework ( ) 132 .

A new study of brain activity patterns in people doing a memory task finds that the way we make inferences—finding hidden connections between different experiences—changes dramatically as we age. The study's findings might one day lead to personalized learning strategies based on a person's cognitive and brain development. The researchers found that whereas adults build integrated memories with inferences already baked in, children and adolescents create separate memories that they later compare to make inferences on the fly. "How adults structure knowledge is not necessarily optimal for children, because adult strategies might require brain machinery that is not fully mature in children," said Alison Preston, professor of neuroscience and psychology and senior author of the study published today in the journal Nature Human Behaviour. She co-led the study with first author Margaret Schlichting, formerly a doctoral student in Preston's lab and currently assistant professor of psychology at the University of Toronto. To understand the distinction between how adults and children make inferences, imagine visiting a day care center. In the morning, you see a child arriving with one adult, but in the afternoon that child leaves with a different adult. You might infer that the two grown-ups are the child's parents and are a couple, and your second memory would include both the second person you saw and information from your earlier experience in order to make an inference about how the two adults—whom you didn't actually see together—might relate to each other. This new study finds that a child who has the same experiences isn't likely to make the same kind of inference that an adult would during the second experience. The two memories are less connected. If you ask your child to infer who that child's parents are, your child can still do it; he or she just has to retrieve the two distinct memories and then reason about how each adult might be related. The neural machinery of children and adults differs, and the strategy that children use may be optimal for the way their brains are wired before key memory systems in the hippocampus and prefrontal cortex fully mature, the researchers believe. That difference could keep children from recalling past memories during new learning and limit their ability to connect events. "In the absence of a mature memory system, the best thing a child can do is lay down accurate, non-overlapping memory traces," Preston said. "From those accurate memory traces, children can later bring them to mind to promote inferences about their connections." Hannah Roome (right) and Nicole Varga prepare a study participant for an fMRI brain scan. Credit: Vivian Abagiu/University of Texas at Austin The researchers asked 87 subjects, ages 7 to 30, to look at pairs of images while lying in an fMRI (functional magnetic resonance imaging) scanner, which measures brain activity by detecting small changes in blood flow with images that, as in the day care example above, provide opportunities to infer relationships between objects that had not appeared together. The researchers found that the strategy adolescents used for making inferences was different from both that of young children and adults. Going back to the example of the parents at the day care, when an adolescent stores a memory of the second grown-up with the child, the adolescent suppresses the earlier memory involving the first one. Each memory becomes even more distinct than with younger children, and there are even fewer automatic inferences about how the two adults relate. "Teenagers may have learning strategies that are tuned to explore the world more so than exploiting what they already know," Preston said. This and other lessons from the study could inform strategies for improving teaching and learning at various ages. "From a brain maturation perspective, different people are going to be at different places," Preston said, "and we can devise learning strategies that take advantage of the neural machinery that an individual has at hand, no matter if they are 7 years old or 70 years old." The study's other authors are Katharine Guarino at Loyola University Chicago and Hannah Roome of UT Austin. This work was supported by the National Institutes of Health and by the Canada Foundation for Innovation. Preston holds the Dr. A. Wilson Nolle and Sir Raghunath P. Mahendroo Professorship in Neuroscience.

10.1038/s41562-021-01206-5

Earth

Improving poor soil with burned up biomass

Ogura T, Date Y, Masukujane M, Coetzee T, Akashi K, Kikuchi J (2016) Improvement of physical, chemical, and biological properties of aridisol from Botswana by the incorporation of torrefied biomass. Scientific Reports. DOI: 10.1038/srep28011 Journal information: Scientific Reports

http://dx.doi.org/10.1038/srep28011

https://phys.org/news/2016-06-poor-soil-biomass.html

Abstract Effective use of agricultural residual biomass may be beneficial for both local and global ecosystems. Recently, biochar has received attention as a soil enhancer and its effects on plant growth and soil microbiota have been investigated. However, there is little information on how the physical, chemical and biological properties of soil amended with biochar are affected. In this study, we evaluated the effects of the incorporation of torrefied plant biomass on physical and structural properties, elemental profiles, initial plant growth and metabolic and microbial dynamics in aridisol from Botswana. Hemicellulose in the biomass was degraded while cellulose and lignin were not, owing to the relatively low-temperature treatment in the torrefaction preparation. Water retentivity and mineral availability for plants were improved in soils with torrefied biomass. Furthermore, fertilization with 3% and 5% of torrefied biomass enhanced initial plant growth and elemental uptake. Although the metabolic and microbial dynamics of the control soil were dominantly associated with a C1 metabolism, those of the 3% and 5% torrefied biomass soils were dominantly associated with an organic acid metabolism. Torrefied biomass was shown to be an effective soil amendment by enhancing water retentivity, structural stability and plant growth and controlling soil metabolites and microbiota. Introduction In African dryland landscapes, improving nutrient-poor soils is important for increasing agricultural productivity, particularly because a significant population growth is expected in this region over the next 100 years. In the Republic of Botswana in southern Africa, Jatropha curcas L. has received attention as a biomass resource 1 , 2 although has exhibited unsatisfactory growth due to the arid climate, chilling injury and oligotrophic soil conditions (aridisols) 3 , 4 . Therefore, methods of soil amendment are expected to promote its agricultural production in nonfarming lands. In dryland ecosystems, such as arid African landscapes, termites, which build termite mounds, play a key role in soil amelioration 5 . Their effects may be artificially achieved through soil amendment using charcoal-like soil enhancers 6 , 7 . Charcoal has a porous structure and harbors soil microbes 8 that play roles in soil enrichment. Activated charcoal has been reported to increase nutrients, reduce nutrient leaching, enhance nutrient uptake and increase crop production 9 , 10 . Recently, biochar, which is made from post-harvest biomass residues, has been studied for its use to amend soils in various African countries 11 , 12 . Torrefied biomass, which is a kind of biochar made at low temperature under anaerobic conditions, is made by the torrefaction of plant biomass derived from grasses and/or woods. Torrefied biomass revealed isothermal pyrolyzed biomass at relatively low temperature ranges of 200 °C–300 °C 13 , 14 . The treatment evaporates the internal water from the biomass with an economic use of energy 15 ; therefore, this type of biochar exploits this resource of carbon-rich material. Watanabe et al. characterized torrefied Jatropha biomass components and they suggest that detoxification of phorbol ester by thermal degradation renders it suitable as a soil amendment 16 . The beneficial effects of biochar on plant growth and soil microbiota have also been investigated 12 , 13 , 17 , 18 . For example, Anders et al. 19 reported that biochar enhanced the positive correlation between nutrients and microbiota more in nutrient-poor soils than in nutrient-rich soils. Fox et al. 20 also reported that soil amended with biochar enhanced the microbially mediated nutrient mobilization of S and P resulting in an improvement in plant growth. However, the relationship between the metabolites and microbiota has not been evaluated in soil amended with biochar. A soil study focused on the humic substance or microbial properties in forest soils using nuclear magnetic resonance (NMR) 21 , 22 . However, a comprehensive approach based on physical, chemical and biological viewpoints was greatly anticipated to evaluate the effects of biochar on soil amendment. We had evaluated various soil properties, such as plant degrading abilities, using some analytical strategies and compared NMR and other meta-analytical methods 23 , 24 . In this study, we focused on torrefied biomass and evaluated the soil amendment effects of this biomass in an aridisol from Botswana. We evaluated water retentivity, chemical components, effect on plant growth and metabolic and microbial variations in the soil ( Fig. 1 ). Figure 1 Schematic representation of this study. Effects of soil amendment with torrefied biomass were evaluated using three steps: (1) characterization of Jatropha curcas pyrolysis profiles, (2) evaluation of soil physical and chemical properties and (3) evaluation of plant growth ability using J. curcas seedlings. Full size image Results and Discussion Torrefication profile of J. curcas To prepare the torrefied biomass, the thermal degradation profile of J. curcas was characterized by thermogravimetric (TG)-differential thermal analysis (DTA), attenuated total reflectance (ATR)-Fourier transform infrared (FTIR), grain size distribution and 1 H- 13 C heteronuclear single quantum coherence (HSQC) NMR spectra ( Figs S1–S4 ). Dehydration occurred at approximately 70 °C and degradation of the hemicellulose components at approximately 160 °C–280 °C ( Fig. S1 ). The torrefied biomass bonds, revealed by the vibration of the benzene C–H bond, the syringyl C–O vibration and cellulose and hemicellulose C–H deformations that were known to cleave under a fast pyrolysis treatment 25 , were broken between 240 °C and 250 °C ( Fig. S2; Table S1 ). The grain size of torrefied biomass was over 1,400, 710, 355 and 106 μm or smaller and the ratios were 11.40%, 25.30%, 28.45%, 23.06% and 11.79%, respectively ( Fig. S3 ). The chemical component comparison of raw and 240 °C torrefied biomass showed that some signals such as adipate, d -fructose and l -glutamine disappeared, but many saccharide signals such as maltodextrin remained ( Fig. S4; Table S2 ). Because denaturation of the supramolecular structures with the remaining main components such as cellulose at 240 °C may be able to provide adsorption of water and nutrient elements during incorporation into soil for plant growth, the torrefication temperature in these experiments was set at 240 °C. In addition, major toxic compounds (the profiles were already characterized in a previous study 16 ) were under the detection limit in the NMR spectra of biomass torrefied at 240 °C, suggesting very low concentrations of these toxins in the biomass. Therefore, we presumed that the soil community should be less affected by toxins when biomass torrefied at 240 °C was used for soil amendment experiments. Moreover, most bacteria that existed in the biomass were eliminated because the torrefaction process has the effect of dry sterilization. Thus, we considered that further analyses using soil ecosystems would be unaffected by external bacteria that existed in the biomass. Influence of torrefied biomass on soil The physical properties of soils with or without torrefied biomass were evaluated by measuring the water-holding capability, soil compression stress and relaxation time ( Figs 2 and S5 ). The water contents of the control and 1%, 3% and 5% torrefied biomass soils were significantly different at 30.0%, 33.3%, 34.1% and 35.7%/volume, respectively ( Fig. 2A ). A similar trend in water contents was observed in soils with raw biomass ( Fig. S6 ). Furthermore, the maximum mechanical stress values of the soils were 0.56, 1.39, 0.99 and 1.49 N, respectively ( Fig. 2B ). The 1% and 5% torrefied biomass soils were significantly different from the control at depths of 17.8 and 5.6 mm, respectively. The water T 2 relaxation times of these soils for bound water with biomass in the soils were 76.99, 25.05, 20.16 and 11.27 ms, respectively and significant differences between control and soils with torrefied biomass were observed ( Fig. 2C ). The T 2 relaxation times of free water were 222.78, 62.02, 46.44 and 19.13 ms, respectively ( Fig. 2D ). The abundance ratios of binding vs. free water in the control and 1%, 3% and 5% torrefied biomass soils were 63:37, 81:19, 66:34 and 79:21, respectively. The higher compression stress values and shorter T 2 relaxation times in the torrefied biomass soils compared with the control soil suggested that torrefied biomass soils facilitated soil structural stability, resulting in higher water contents and retention capabilities. These results indicated that the physical character of soil was greatly changed by the addition of torrefied biomass and that soil water retention was improved. Although a similar effect of water retention capability was observed in soils mixed with raw biomass, the chemical properties were different between raw and torrefied biomass ( Figs S1 , S2 and S4 ). The torrefied biomass retained the cellulose component by degradation of the supramolecular structure known as lignocellulose. Therefore, one of the advantages of utilizing torrefied biomass compared with raw biomass is the easy access and exchange to energy by microbiota. Figure 2 Physicochemical characterization of soils used for Jatropha cultivation with or without torrefied biomass. The moisture contents of saturated absorption ( A ) and compression mechanical stress under wet soil conditions ( B ). T 2 relaxation time of binding ( C ) and free water contents ( D ). The error bars show the standard error of the mean and the p value for comparison of the control with each sample calculated using Welch’s t test. * p < 0.05, † p < 0.01 and ‡ p < 0.005. Full size image Soil elements with or without torrefied biomass were characterized by water extraction followed by HNO 3 extraction using inductively coupled plasma-optical emission spectrometry (ICP-OES; Fig. 3 ). Principal component analysis (PCA) profiles showed similar trends between water and HNO 3 extractions, meaning that the elemental profiles were different between control soils and those with added torrefied biomass ( Fig. 3A,B ). Some elements such as K, P and S increased in the torrefied biomass soils compared with the control because these elements were derived from the torrefied biomass ( Fig. 3C ). In addition, the dissolution rates of some elements such as K, Na and P in water compared with those in HNO 3 were highly increased in the torrefied biomass soils compared with the control ( Fig. 3D ). Since water-soluble elements are readily accessible to plants, the elemental availability (especially of K, Na and P) to plants was improved by the addition of torrefied biomass to the soils. However, many elements were largely dissolved in HNO 3 , suggesting that in nature these elements are trapped by supramolecular structures in the soil. Figure 3 Evaluation of elemental components in soils. Soil elemental profiles were evaluated from extractions of water ( A ) and HNO 3 ( B ) using PCA score plots and the relative abundance of soil total elements compared the control with torrefied biomass ( C ) and extraction ratios compared water with total extracted elements ( D ). The error bars show the standard error of the mean and the p value for comparison of the control with each sample calculated using Welch’s t test. ∗ p < 0.05 and † p < 0.01. Full size image Soil maturation The metabolic dynamics of microbiota during soil maturation were evaluated using 1 H-NMR spectra ( Fig. S7 ). Metabolic profiles from day 0 to 21 in the soils with torrefied biomass were varied, but each profile at day 28 was similar to that of the control. This result indicated that organic components such as polysaccharides were digested during the period from day 0 to 21 and that the available organic components were finally lost by day 28 (as in the control). Moreover, the metabolic profiles of soil with fishmeal were largely varied; degradation of creatine and trimethylamine N -oxide (TMAO) was accompanied by the production of acetate, methylamine, dimethylamine and trimethylamine from day 0 to 13 ( Fig. S8 ). Creatine and TMAO are the most abundant components in fish water-soluble fractions 26 , 27 . This result indicated that the soil microbiota metabolized and utilized these fish components and produced some metabolites such as methylamine, dimethylamine and trimethylamine, which are derivatives of TMAO. However, a lot of nutrients remained in the fishmeal soil compared with the torrefied biomass, indicating that an excess of nutrients in the fishmeal soil prevented seedling and plant growth. Therefore, only the soils with torrefied biomass were used for further analyses. Effect of torrefied biomass on the initial growth stage of J. curcas To evaluate the effect of torrefied biomass on Jatropha growth, germinated J. curcas were transplanted to matured soils and grown for 4 weeks. The plant heights of the control and of the 1%, 3% and 5% torrefied biomass treatments were 10.50 and 9.23, 8.75 and 8.95 cm, respectively; the stem diameters were 3.04, 3.76, 4.02 and 4.58 mm, respectively ( Fig. 4A,B ); and the root lengths were significantly greater by 65.81, 73.75, 93.85 and 91.31 mm, respectively ( Fig. 4C ). The weights of the roots, stems and leaves also tended to increase with increasing torrefied biomass treatments ( Fig. 4D ). The uptake of soil elements by plants in the initial growth stages was characterized by ICP-OES ( Fig. 4E ). K and Na were present in higher concentrations, but Si, Mn and Ba were lower in plants grown in torrefied biomass-amended soils than in the control. Although Na is known to inhibit plant growth, the low concentration in the soils should not affect plant growth and less than 0.04% was incorporated into the plants. In addition, it is known that excessive Mn and Ba inhibit plant growth 28 , 29 . Moreover, the carboxyl functional group in organic compounds and polysaccharides such as hemicellulose in plants and alginic acid in algae is also known to be heavy metal adsorbents 30 , 31 , 32 . The result shows that torrefied biomass has the capacity to provide beneficial minerals such as K and to inhibit toxic element absorption. Thus, torrefied biomass can be utilized as a soil conditioner for soil amendment. Figure 4 Evaluation of the effect of torrefied biomass on the initial plant growth stage. The effect of torrefied biomass on the plant phenotype during the initial growth stage was evaluated by plant height ( A ), stem diameter ( B ), root length ( C ), whole weight ( D ) and elemental content ratios of torrefied biomass to the control ( E ). The error bars show the standard error of the mean and the p value for comparison of the control with each sample, calculated using Welch’s t test. * p < 0.05 and ‡ p < 0.005. Full size image Metabolic and microbial dynamics in soil during initial plant growth Metabolic soil dynamics during initial plant growth were evaluated using 1 H-NMR spectra ( Fig. 5A,B ) in combination with 2D-J spectra for annotation ( Fig. S9; Table S3 ). The metabolic profiles were clustered based on the differences between the control and torrefied biomass soils. In the 3% and 5% torrefied biomass soils, the profiles were PC1 positive with the factors being l -valine, lactate, acetate and succinate. Organic acids produced by anaerobic microbes due to cellulosic degradation are known to be phosphate-solubilizing and plant growth-promoting substances 33 , 34 . Thus, the torrefied biomass was considered to effect soil fertilization and enhance plant growth. Formate decreased from week 0 to 1 and methanol and butyrate showed a similar trend increasing in weeks 1 and 3 and decreasing in week 4 in the control ( Fig. S10A–C ). In contrast, acetate, lactate and succinate showed similar decreases from week 0 to 4 and l -valine increased from week 0 to 2 with 5% torrefied biomass ( Fig. S10D–G ). Figure 5 Soil metabolic and microbial profiles during plant growth. Metabolic ( A , B ) and microbial profiles ( C , D ) of soil during initial plant growth were evaluated using PCA score plots ( A , C ) and loading plots ( B , D ). Symbols in the score plot represent the control (blue), 1% (light orange), 3% (orange) and 5% torrefied biomass (brown) and weeks 0 (circle), 1 (triangle), 2 (diamond), 3 (square) and 4 (bar). Full size image Microbial profiles during plant growth were analyzed with a MiSeq sequencer ( Fig. 5C,D ). The control contributed to PC1 and 3% and 5% torrefied biomass soils contributed to PC2. In the control, the most predominant microbe was Methylotenera sp. In contrast, some microbes including Opitutus sp. and Devosia sp. were associated with the 3% and 5% torrefied biomass soil. Methylobacterium sp. and Methylotenera sp., which exist on plant leaves and are associated with methanol consumption and production 35 , 36 , were highly abundant and Methylotenera sp. greatly increased from week 0 to 1 in the control ( Fig. S11A,B ). Based on methanol variations, these microbes were inferred to be dominantly associated with a C1 metabolism in the control. Bacillus sp., Devosia sp. and Opitutus sp. were highly abundant and Devosia sp. and Opitutus sp. showed similar trends during plant growth in 3% and 5% torrefied biomass soils ( Fig. S11C–E ). Devosia sp. and Opitutus sp. are known to use lactate as a carbon source 37 , 38 and the time-course variations were associated with acetate, lactate and succinate dynamics. Thus, these microbes were inferred to be associated with the metabolism of these organic acids and considered to be key players in torrefied biomass adjusted soil environments for promoting plant growth. In the future, the effects of soil amendments using torrefied biomass should be evaluated by field experiments. In conclusion, the effects of soil amended with torrefied biomass were evaluated with respect to their physical properties, initial plant growth and metabolic and microbial dynamic soil profiles ( Fig. 6 ). Torrefied biomass improved the physical and structural properties of soil such as water retentivity and structural stability. Soil amended with 3% and 5% torrefied biomass enhanced the initial growth of J. curcas in the form of increased stem diameter, root length and element uptake ability. Although the metabolic and microbial dynamics of the control were associated with a C1 metabolism, those of the 3% and 5% torrefied biomass samples were associated with an organic acid metabolism. These results indicate that torrefied biomass is effective as a soil amendment by increasing water retentivity and structural stability, enhancing plant growth and controlling soil metabolites and microbiota. Figure 6 Schematic representation of the effects of incorporating torrefied biomass in soil on initial plant growth. Although the metabolic and microbial dynamics of the control were associated with a C1 metabolism (left), those of the 3–5% torrefied biomass samples were associated with an organic acid metabolism (right). This is attributed to the fact that torrefied biomass can improve physical and structural soil properties such as water retentivity and structural stability (right). Therefore, a soil amended with 3–5% of torrefied biomass can enhance the initial growth of Jatropha curcas in the form of increased stem diameter, root length and element uptake ability. Full size image Methods Sample preparation and experimental design The overall experimental design to evaluate the effects of soil amendments based on physical, chemical and biological characteristics is illustrated in Fig. 1 . For the torrefaction analysis, stem and leaf mixtures of J. curcas were milled with a food cutter, divided into 50-g samples and wrapped in aluminum foil. The samples were torrefied at 200 °C, 220 °C, 230 °C, 240 °C, 250 °C and 300 °C for 10 min under 5-L/min N 2 in an electric furnace (FO410; Yamato Scientific Co., Ltd., Tokyo, Japan). For growth experiments using J. curcas , a soil sample was collected from a Jatropha agricultural field in Gaborone, Botswana, in 2014. The soil was separated into four parts each weighing 3 kg into which 0, 30, 90 and 150 g of biomass torrefied at 240 °C [control, 1%, 3% and 5% (weight/weight), respectively] and 150 g of fishmeal [5% (weight/weight)] was incorporated. To stabilize the metabolic activities and microbial variations in soils, they were incubated in a chamber at 25 °C and 10% moisture for 1 month and sampled twice a week. Seeds of J. curcas IP2P accession 39 were germinated in 0.8 wt% agar gel with no nutrients as described in a previous study 40 . The germinated seeds after 10 days of growth were transplanted into the matured soils using quadruplicate experiments for the control and 1%, 3% and 5% of torrefied biomass. Jatropha growth experiments with and without torrefied biomass were conducted over 4 weeks in a chamber at 25 °C and 50% moisture. Samples of 50 g were taken from the soils once a week for 4 weeks during growth. Characterization of torrefied biomass A TG analysis was performed with an EXSTAR TG/DTA 6300 (SII Nanotechnology Inc., Tokyo, Japan) instrument following a previous study 23 . ATR FTIR was performed on a Nicolet 6700 FTIR (Thermo Fisher Scientific Inc., Waltham, MA, USA) instrument with a KBr disk following previous studies 23 , 41 . Grain size distribution was performed with a vibratory sieve shaker (Fritsch Japan Co., Ltd., Kanagawa, Japan) instrument and the percentage was calculated for each weight. Soil characterization For the analysis of soil compression stress, 9 g of dried soil samples was analyzed with an EZ-LX autograph (Shimazu, Kyoto, Japan) and TRAPEZIUM 2 software (Shimazu) at 5 mm/min to a depth of 20 mm and 25 N of stress using a lunge test jig. Soil samples with an addition of 2.5 ml ultrapure water were also measured under wet conditions using the same method. To measure the water content in soils, soil with or without 1%, 3%, or 5% torrefied biomass and raw biomass were divided into 600 g fractions and placed in plant pots into which 200 ml of ultrapure water was added. Soil water contents were measured with an ML3-theta probe soil moisture sensor (Delta-T Devices Ltd., Cambridge, England). Elemental analysis For the ICP-OES analysis, 50 mg samples of Botswana soils with or without torrefied biomass and J. curcas in the initial growth stage were extracted with ultrapure water and then with HNO 3 (6.9% v/v) following previous studies 42 , 30 and analyzed using an ICP-OES instrument (SPS5510; SII Nanotechnology Inc., Tokyo, Japan). One- and two-dimensional NMR analyses Samples of maturing phase soils with Jatropha growth (40 g) with an addition of sterile water were homogenized with a sonicator for 30 min and heated at 55 °C for 5 min in a Thermomixer comfort (Eppendorf AG, Hamburg, Germany). After centrifugation, the supernatants were collected and dried in a centrifugal evaporator (CVE-3000, Tokyo Rikakikai Co. Ltd., Tokyo, Japan). The dried samples were dissolved in 1 ml of D 2 O/KPi (100 mM, pH 7.0) and transferred into a 5 mm NMR tube. The torrefied biomass and soil after Jatropha growth were analyzed using 1 H- 13 C HSQC to identify the components. NMR spectra were acquired at 25 °C using an AVANCE II 700 MHz Bruker Biospin (Rheinstetten, Germany) instrument equipped with an inverse (with proton coils nearest to the sample) 5 mm 1 H/ 13 C/ 15 N cryoprobe. The peak of sodium 2,2-dimethyl-2-silapentane-5-sulfonate (DSS) was used as the internal reference (calibrated at δ C 140, δ H 0 ppm). NMR spectra were acquired from 11.704 to −2.296 ppm in F2 ( 1 H) using 2048 data points for an acquisition time of 104 ms, recycling delay of 2 s and 150−10 ppm in F1 ( 13 C) using 256 data points of 48 scans. All one-dimensional Watergate and two-dimensional (2D) J-resolved (2D-J) spectra were acquired with the same NMR instrument to evaluate metabolic profiles of soil microbiota. Watergate spectra were measured from 14 to −3 ppm at 25 °C using 32 k data points. 2D-J spectra were acquired from 11.7568 to −2.2458 ppm in F2 ( 1 H) using 16 k data points and from 20.0027 to −19.9973 Hz in F1 ( J coupling) using 16 data points of eight scans. HSQC and 2D-J spectra were further analyzed for annotation of chemical components using SpinAssign ( ) 43 , 44 , Biological Magnetic Resonance Bank ( ) 45 and Birmingham Metabolite Library ( ) 46 . Relaxation time analyses of the water content in soils were measured by solid-state NMR using a Bruker DRX-500 spectrometer operating at 500.13 MHz for 1 H equipped with the Bruker 4 mm double-tuned MAS probe. For the NMR measurements, approximately 80 mg of a sample and 100 μl of sterilized water were placed in a ZrO 2 rotor (outer diameter 4 mm) with a Kel-F cap. The magic angle (54.7°) pulse length for protons was set to 1.8 μs. The measurement program used 2D Carr Purcell Meiboom Gill and sampling of the decay/recovery curves was obtained at 2–80 ms. Metasequencing Microbial DNA extraction was performed using the PowerSoil ™ DNA Isolation Kit (Mo Bio Laboratories Inc., Carlsbad, CA, USA) according to the manufacturer’s instructions. The polymerase chain reaction (PCR) protocol used for metasequencing was described previously 47 , 48 . Sequencing was performed on a MiSeq sequencer (Illumina, San Diego, CA, USA) according to the manufacturer’s instructions. The data were analyzed using QIIME software ( ) 49 . Each sample was separated by a MiSeq barcode-attached 27F mod-534R primer and chimera check using Usearch software ( ) 50 . The resulting operational taxonomic unit data defined over 97% similarity were assigned to sequences using the Ribosomal Database Project ( ) classifier 51 . Statistical analysis All 1 H-NMR data were processed using Topspin 3.1 software (Bruker Biospin) and raw data were exported as text files. Exported data were processed over a range of 11 to −1 ppm with approximately 27.5 k data points for 1 H-NMR and binning using R 3.0.1 software ( ). The dataset was normalized using the sum of the DSS integral regions and analyzed by PCA using R software as previously described 26 , 33 , 52 . Additional Information How to cite this article : Ogura, T. et al. Improvement of physical, chemical and biological properties of aridisol from Botswana by the incorporation of torrefied biomass. Sci. Rep. 6 , 28011; doi: 10.1038/srep28011 (2016).

Researchers at the RIKEN Center for Sustainable Resource Science in Japan have shown that torrefied biomass can improve the quality of poor soil found in arid regions. Published in Scientific Reports, the study showed that adding torrefied biomass to poor soil from Botswana increased water retention in the soil as well as —the amount of plant growth. When high temperatures and the absence of oxygen are used to bring about the decomposition of biomass residue from agricultural products such as grains, the result is a charcoal-rich substance called biochar. Torrefied biomass—sometimes called bio-coal—is a type of biochar made at relatively lower temperatures that has recently received attention as a pretreatment method for biomass utilization. In order to characterize the biological properties of soil treated with biochar, the team incorporated torrefied plant residual biomass from the biodiesel crop Jatropha curcas into aridisol, a type of soil found in arid regions such as Botswana, and compared several soil properties with samples that had not been treated. Explained team leader Jun Kikuchi, "Jatropha is a potential biomass resource for dryland African landscapes, but the poor climate and soil conditions have limited its production. Our study shows that treating the poor soil with torrefied biomass improves a variety of factors that ultimately lead to greater plant growth." An important quality of good soil is its ability to retain water. Tests showed that water retention increased with the percentage of torrefied biomass, with 5% biomass yielding a soil that contained about 5% more water than the control soil. A good soil also remains structurally sound deeper in the ground where pressure from above is higher. Soil treated with 5% torrefied biomass showed significantly higher levels of compression stress than the control soil, and significantly shorter relaxation time—the time needed for it to relax back into its normal shape after being compressed. After finding that the torrefied biomass retained more water, the team tested the chemical properties of the soils. They found that levels of potassium, phosphorous, and sulfur were higher in the soil treated with torrefied biomass, as was the availability of potassium, sodium, and phosphorous—three elements regularly take up from the soil by plants. When they tested how well plants grew in the different soils, they found that plants grown in the torrefied biomass had thicker stems, much longer roots, and were heavier that those grown in the untreated soil. The plants grown in torrefied biomass also took up more potassium than controls and less manganese, an element known to inhibit plant growth. Other important features of soil are its metabolic and microbial components. Some compounds produced by the degradation and break down of cellulose are known to promote plant growth. The researchers found that levels of these organic acids, such as lactate and acetate, were higher in the treated soil, again supporting the idea that torrefied biomass can enhance soil fertilization. The treated soil also showed higher levels of Devosia sp. And Opitutus sp., bacteria that use lactate as a carbon source. This indicated that the soil metabolites available in the treated soil allowed for a different microbial environment that presumably acted to enhance plant growth. "Our next step," says Kikuchi, "is to elucidate the complicated reactions between symbiotic microbiota and plants for effective growth in nutrient-poor environments."

10.1038/srep28011

Nano

Using bottlebrush-shaped nanoparticles, researchers can identify and deliver synergistic combinations of cancer drugs

Irene Ghobrial, Molecular bottlebrush prodrugs as mono- and triplex combination therapies for multiple myeloma, Nature Nanotechnology (2023). DOI: 10.1038/s41565-022-01310-1. www.nature.com/articles/s41565-022-01310-1 Journal information: Nature Nanotechnology

https://dx.doi.org/10.1038/s41565-022-01310-1

https://phys.org/news/2023-01-bottlebrush-shaped-nanoparticles-synergistic-combinations-cancer.html

Abstract Cancer therapies often have narrow therapeutic indexes and involve potentially suboptimal combinations due to the dissimilar physical properties of drug molecules. Nanomedicine platforms could address these challenges, but it remains unclear whether synergistic free-drug ratios translate to nanocarriers and whether nanocarriers with multiple drugs outperform mixtures of single-drug nanocarriers at the same dose. Here we report a bottlebrush prodrug (BPD) platform designed to answer these questions in the context of multiple myeloma therapy. We show that proteasome inhibitor (bortezomib)-based BPD monotherapy slows tumour progression in vivo and that mixtures of bortezomib, pomalidomide and dexamethasone BPDs exhibit in vitro synergistic, additive or antagonistic patterns distinct from their corresponding free-drug counterparts. BPDs carrying a statistical mixture of three drugs in a synergistic ratio outperform the free-drug combination at the same ratio as well as a mixture of single-drug BPDs in the same ratio. Our results address unanswered questions in the field of nanomedicine, offering design principles for combination nanomedicines and strategies for improving current front-line monotherapies and combination therapies for multiple myeloma. Main Controlling the tissue exposure of drugs remains the most persistent challenge of modern cancer therapies and the holy grail of drug delivery 1 , 2 , 3 , 4 . By exploiting features such as size, shape, composition and release kinetics, nanocarriers can enhance the therapeutic indexes of drugs by increasing their exposure in diseased sites and/or avoiding major sites of toxicity 1 , 3 , 4 . The development of nanomedicine combination therapies represents a frontier of modern cancer treatment 5 , 6 , 7 , 8 , 9 , 10 . Although recent pioneering advancements in cancer biology have greatly improved the ability to identify and predict synthetic lethalities of drug combinations, the clinical translation of such combinations suffers from fundamental barriers 5 , 6 , 7 , 8 , 9 . For instance, due to the distinct physical properties of dissimilar drugs, combinations of those drugs that are synergistic in vitro may not accumulate in target tissues/cells in vivo 6 , 7 , 8 , 9 , 10 . Due to this disconnect, many clinical combination therapies are empirically derived based on the maximum tolerated dose (MTD) of each component drug rather than rational synergies 6 , 7 , 8 . Combination therapies present an exciting nanomedicine opportunity wherein multiple drugs that are pharmacologically different may be delivered to the same tissue/cell in precise ratios to empower their synergistic mechanisms. For example, Vyxeos (CPX-351), a clinically successful liposomal formulation of 5:1 cytarabine:daunorubicin, maintained a synergistic drug ratio (from 5:1 to 9:1) in the blood compartment over 24 h post-injection, whereas the free drugs exhibited a 1,923:1 ratio 15 min post-injection 11 , 12 . Although strategies for incorporating mixtures of structurally dissimilar drugs through encapsulation, chemical conjugation and/or self-assembly have been extensively studied, nanocarriers that simultaneously achieve controlled drug ratios, multidrug release kinetics and/or sequential release for two or more drugs remain rare 5 , 6 , 7 , 8 , 9 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 . Moreover, due to differences in cell uptake pathways, rates of cellular internalization and/or drug release kinetics, multidrug nanocarriers could exhibit synergistic ratios that are distinct from their free-drug counterparts, necessitating the identification of optimal ratios in the nanocarrier context. Given that most nanocarriers rely on supramolecular interactions between the drugs, vehicle and/or surfactant(s), which depend on the physical properties of the drugs, exchanging one drug for another may result in changes to the physical properties of the final nanocarrier. It is, thus, difficult to make multiple single-drug nanocarriers and multidrug nanocarriers with varying drug ratios but otherwise identical properties. Hence, in combination nanocarriers employed to date, the synergistic drug ratios exemplified for the free drugs are typically translated directly to the nanocarriers, without considering the possibility that these ratios may no longer be optimal 11 , 12 . Moreover, it remains unknown whether multidrug nanocarriers have fundamental advantages over mixtures of single-drug nanocarriers. Here we introduce a nanocarrier design that allows us to address these questions in the context of the second-most common haematologic malignancy in the United States—multiple myeloma (MM)—which remains incurable in most patients 21 . Our approach leverages ‘bottlebrush prodrugs’ (BPDs) comprising the clinically relevant three-drug combination of a proteasome inhibitor (PI) bortezomib (Btz), an immunomodulatory drug pomalidomide (Pom) and a corticosteroid dexamethasone (Dex). This drug combination is able to overcome resistance to the front line and standard-of-care regimen of lenalidomide (Len)/Btz/Dex as Pom allows for higher target-binding affinity compared to Len (refs. 22 , 23 , 24 , 25 ). In spite of the empirical derivation of this combination in the clinic, it offers prolonged progression-free survival in Len-refractory patients (17.8 versus 9.5 months) as well as in non-Len-refractory patients (22.0 versus 12.0 months); moreover, it improves the overall survival rates in both settings (85.90% versus 50.80% and 95.37% versus 60.00%, respectively) 26 . Nevertheless, the combination suffers from substantial drawbacks that primarily arise from off-tissue toxicities, poor stability and the development of Btz resistance. Although several examples of nanoparticle-based Btz formulations have been reported as monotherapies 27 , 28 , 29 , 30 , 31 , so far they have shown only minor improvements over free Btz in terms of efficacy 32 , 33 . By contrast, combination nanomedicines for MM are exceptionally rare, and nanocarriers incorporating the clinical combination of Btz, Pom and Dex have not been reported 32 , 33 , 34 , 35 , 36 , 37 , 38 . Moreover, no examples of more than two drug combination therapies with systematically optimized synergistic ratios have been demonstrated in any disease context. Here we show that (1) synergies between free drugs identified in vitro do not necessarily translate to BPDs and (2) BPDs bearing a statistical mixture of drugs in a synergistic ratio are more effective than a mixture of three different physically equivalent single-drug BPDs administered at the same ratio. The latter finding is explained using Monte Carlo simulations. BPD design and synthesis Our BPD manufacturing involves the synthesis of macromonomer prodrugs of Btz, Pom and Dex. For Btz, racemic 1,2-tertiary diol azide linker 1 was synthesized from tetraethylene glycol and 2,3-dimethyl-2-butene ( Supplementary Information provides the complete synthesis details). Azido-boronic ester Btz-N 3 was formed from 1 and Btz in 70% yield and was subsequently coupled to alkyne 2 through copper-catalysed azide–alkyne cycloaddition ‘click’ chemistry, affording Btz-M (Fig. 1a ) 39 . Following a similar workflow but with different linkers tailored to the inherent functionality of each active pharmaceutical ingredient (API), Pom-M and Dex-M were prepared (Fig. 1a ). The structures of each macromonomer and its precursors were characterized by 1 H and 13 C nuclear magnetic resonance spectroscopy and by either high-resolution mass spectrometry or matrix-assisted laser desorption ionization time-of-flight mass spectrometry where appropriate (Supplementary Figs. 1 – 17 ). Fig. 1: Synthesis and characterization of BPDs. a , Chemical structures of prodrug macromonomers used in this work. These macromonomers, or mixtures thereof, were subjected to ROMP via exposure to Grubbs third-generation bis-pyridyl complex to produce the corresponding BPDs. Schematic of multidrug BPD is provided (not drawn to scale). Maroon, blue and white spheres denote different drugs randomly arranged along the BPD backbone; green denotes cleavable linkers that activate to release the drugs; purple denotes the BPD backbone; blue strands denote poly(ethylene glycol) (PEG), which shrouds the drugs and backbone, providing similar physical properties for BPDs regardless of API identity. b , Size exclusion chromatography traces of BPDs. The minor peak at ~16 min elution time corresponds to residual macromonomers. c , D h of BPDs as determined by dynamic light scattering. The inset shows the cryogenic electronic microscopy image of three-drug BPD Syn (scale bar, 50 nm). d , e , Free drugs ( d ) and one-drug BPDs ( e ) were evaluated in MM.1S and KMS11 cell lines ( BBP refers to a drug-free PEGylated bottlebrush polymer). Cell viability ( n = 3 biologically independent samples) was evaluated by the CellTiter Glo assay at 48 h after incubation with varying concentrations. Data are presented as mean ± standard error of the mean (s.e.m.). Full size image Btz-M , Pom-M and Dex-M were polymerized by ring-opening metathesis polymerization to afford single-drug BPDs of Btz-BPD , Pom-BPD and Dex-BPD , respectively, with number-average degrees of polymerization of 10 (Fig. 1a ). Multidrug BPDs with varying ratios of Btz:Pom:Dex were synthesized by the copolymerization of these macromonomers in various feed ratios (Supplementary Figs. 18 – 20 and Supplementary Table 1 ). A drug-free control polymer ( BBP ) was synthesized for comparison 40 . For in vivo studies, a cyanine5.5 (Cy5.5) dye was incorporated into each BPD (ref. 41 ). Gel permeation chromatography (Fig. 1b ) and dynamic light scattering (Fig. 1c ) revealed efficient macromonomer-to-BPD conversions and hydrodynamic diameters ( D h ) of ~10–15 nm, respectively. All the samples, regardless of payload composition (that is, monodrug, multidrug or no drug), displayed consistent sizes (Supplementary Table 1 ). Cryogenic electron microscopy revealed ellipsoidal structures with dimensions of ~10 nm and average aspect ratios of 1.1 (Fig. 1c inset, Supplementary Table 1 and Supplementary Fig. 21 ). The release of Btz from Btz-BPD in pH 7.4 phosphate-buffered saline (PBS) was much slower (Supplementary Fig. 22 ) than from Btz-M (Supplementary Fig. 17 ), suggesting that the BPD architecture stabilizes the boronic ester linker from rapid hydrolysis. Nevertheless, exposure to glucose and adenosine triphosphate as well as acidic buffer—established triggers for boronic ester cleavage in the tumour microenvironment 42 , 43 , 44 —led to enhanced Btz release (for example, 25.9 ± 2.2% in 1 h at pH 4.0 or 34.6 ± 2.5% in 1 h at 100 mM glucose; Supplementary Fig. 23 ). We note that alkyl boronic esters are typically unstable in water at neutral pH; the placement of a boronic ester along the relatively hydrophobic BPD backbone shields it from immediate hydrolysis 42 , 43 , 44 , 45 , 46 . In vitro and in vivo characterization of single-drug BPDs The potency of each single-drug BPD was examined in vitro, using cell viability assays (Cell TiterGlo, Promega) performed after 48 h incubation and with two different MM cell lines (MM.1S and KMS11). In MM.1S cells, Btz-BPD was ~5-fold less potent than free Btz (half-maximal inhibitory concentration (IC 50 ) = 13.1 ± 0.9 nM versus 2.8 ± 0.4 nM, respectively; Fig. 1d,e ), which could be attributable to differences in cell uptake (transmembrane diffusion versus cellular endocytosis for free Btz and Btz-BPD , respectively) or the slowed release of Btz from Btz-BPD . Dex-BPD was similarly less potent than free Dex (IC 50 = 70.9 ± 1.9 nM versus 17.0 ± 2.8 nM, respectively; Fig. 1d,e ). Pom-BPD displayed a similar IC 50 value compared with free Pom (IC 50 = 354.2 ± 4.9 nM versus 308.6 ± 3.5 nM, respectively; Fig. 1d,e ). BBP was not toxic at any dose level, suggesting that the observed toxicities for the BPDs are due to release of the active APIs. As Btz-associated toxicities remain a hurdle in clinical MM therapy, we first assessed Btz-BPD as monotherapy in vivo. Gross toxicity was assessed in healthy BALB/c mice ( n = 5 animals per group) for free Btz (at 0.75, 1.00 and 1.25 mg kg –1 doses administered twice a week for four weeks via subcutaneous (s.c.) injection) and Btz-BPD (at 5.00, 10.00 and 18.75 mg kg –1 doses administered twice a week for four weeks via intravenous (i.v.) injection) (Fig. 2a ). For Btz-BPD , the 5.00, 10.00 and 18.75 mg kg –1 groups correspond to 0.47, 0.95 and 1.78 mg kg –1 of Btz, respectively. The drug-free polymer was not examined here as it was previously shown to be well tolerated at doses up to 2 g kg –1 (ref. 41 ). Moreover, since the s.c. administration of Btz is used in the clinic and displays improved safety with similar efficacy compared with i.v. administration, we use it here for the fairest possible comparison to Btz-BPD (refs. 46 , 47 ). For Btz, the 0.75 mg kg –1 dose was observed to be safe, which is consistent with previous reports (Fig. 2a ) 27 . Higher doses induced toxicities as reflected by decreased survival rates and dramatic losses in body weight. By contrast, Btz-BPD was tolerated at all doses with no evidence of mortality or substantial weight loss (Fig. 2a,b ). Toxicology studies were performed in BALB/c mice (twice a week over a two-week period; four injections per mouse) using the same test compounds. Metabolic profiles (Fig. 2c ), complete blood counts (Fig. 2d ) and white blood cell differential counts (Fig. 2e ) were obtained 13 days after the last injected dose of either Btz (0.75 mg kg –1 via s.c. injection) or Btz-BPD (18.75 mg kg –1 via i.v. injection). Animals in the Btz-BPD group did not display any changes with respect to the aforementioned parameters (two-tailed Student’s t -test; P > 0.05). The safety of Pom-BPD was evaluated by following similar protocols in the CRBN I391V mouse model known to be sensitive to immunomodulatory drug toxicity 48 (Supplementary Figs. 24 – 26 ). We did not test the MTD of Dex-BPD alone due to its low in vitro toxicity and the role of Dex as a mitigator of toxicity in clinical therapy. Fig. 2: Safety assessments of Btz-BPD . Healthy BALB/c mice were administered either PBS, Btz or Btz-BPD twice a week for four weeks. a , Kaplan–Meier survival curves for mice treated with each agent ( n = 5 mice per group). b , Body weight measurements of BALB/c mice administered Btz-BPD (i.v.) at various doses ( n = 5 mice per group). Data are presented as mean ± s.e.m. c – e , Basic metabolic profiles ( c ), complete blood counts ( d ) and white blood cell differential counts ( e ) for healthy BALB/c mice ( n = 3 mice per group) that were administered each treatment (twice per week for two weeks) followed by two weeks of rest (that is, no injection) before blood draw and analysis. Full size image Next, the accumulation of Btz-BPD in s.c. MM tumours (KMS11) was evaluated. Fluorescence microscopy revealed substantial intratumoural accumulation within 1 h of administration (Fig. 3a ). Additional s.c. KMS11 tumour-bearing mice ( n = 5 per group) were treated with either PBS, Btz (0.75 mg kg –1 via s.c. injection), Btz-BPD at a mass-equivalent dose of Btz (0.75 mg kg –1 via i.v. injection, or ‘low dose’) or Btz-BPD at its highest-tested dose level (18.75 mg kg –1 via i.v. injection, or ‘high dose’). We note that low dose corresponds to 0.071 mg kg –1 of Btz—more than tenfold lower than the free-drug dose. Groups of mice were treated twice a week for four weeks (Fig. 3b–d ); tumour volumes and body weights were monitored. The study endpoint was reached when a tumour measured >2 cm in the longest axis or the animal experienced >20% body weight loss. Fig. 3: Btz-BPD provides substantial therapeutic enhancements over Btz in s.c. and aggressive orthotopic models of MM. a , Evaluation of tumour accumulation and penetration of Cy5.5-labelled Btz-BPD at 1 h post-administration (i.v.) as assessed by fluorescence microscopy of the harvested tumour after animal euthanasia (scale bar, 200 µm); a representative micrograph is shown, and similar results were acquired in three independent biological samples. For efficacy evaluation, KMS11 s.c. tumour-bearing mice were administered PBS, Btz (s.c.) or Btz-BPD (i.v.), starting when their tumours reached 5 mm in the largest axis. b , c , Spider plots of tumour growth ( b ) and average tumour size (± s.e.m.) ( c ) over the course of the study ( n = 5 mice per group). A statistical analysis was performed using a two-tailed t -test between the Btz and Btz-BPD groups. P = 0.0025, Btz-BPD (18.75 mg kg –1 ) versus Btz (0.75 mg kg –1 ); P = 0.0325, Btz-BPD (0.75 mg kg –1 ) versus Btz (0.75 mg kg –1 ). d , Kaplan–Meier survival curves, revealing marked enhancements in therapeutic outcomes for animals treated with Btz-BPD versus Btz at equivalent doses and further improvement at increased Btz-BPD dose. The arrow indicates the last administered dose. A statistical analysis was performed using a log-rank test, P < 0.0002. e , Bioluminescence imaging of MM.1S LUC+/GFP+ cells after i.v. dissemination and as a function of time (day 0 versus day 20) after administration of PBS (control), Btz (0.75 mg kg –1 ) or Btz-BPD (18.75 mg kg –1 ). f , g , Individual spider plots ( f ) and average tumour size (± s.e.m.) ( g ) over the course of the study ( n = 5 mice per group). A statistical analysis was performed using a two-tailed t -test between the Btz-BPD and Btz groups. P = 0.0002, Btz-BPD (18.75 mg kg –1 ) versus Btz-BPD (0.75 mg kg –1 ); P = 0.0525, Btz-BPD (0.75 mg kg –1 ) versus Btz (0.75 mg kg –1 ). h , Kaplan–Meier survival curves confirm substantial enhancements in the therapeutic outcomes for animals treated at a high dose of Btz-BPD (18.75 mg kg –1 ) compared with those treated at the MTD of Btz (0.75 mg kg –1 ). The arrow indicates the last administered dose. A statistical analysis was performed using a log-rank test, P = 0.0002. For statistical tests, n.s. denotes non-significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001. Full size image Btz displayed modest activity in this aggressive MM model compared with the control (a mean survival time of 42 ± 6 days versus 22 ± 5 days for the control group) (Fig. 3b–d ). Btz-BPD outperformed free Btz at a low dose (a mean survival time of 61 ± 9 days versus 42 ± 6 days). Moreover, high dose further prolonged survival (a mean survival time of 84 ± 13 days, P < 0.0002, compared with Btz and other groups). The enhanced activity of Btz-BPD is attributable to its tumour accumulation and Btz release (Fig. 3a ) 42 , 43 , 44 , 45 , 46 . Next, we evaluated Btz-BPD in a more challenging, orthotopic model of MM, which primarily develops in the bone marrow compartment. Tumours were induced via the i.v. injection of luciferase-expressing MM.1S Luc+/GFP+ cells; tumour progression was quantified by bioluminescence imaging (Fig. 3e ). Mice were removed from the study when they exhibited hind limb paralysis or a loss of >20% body weight. Mice ( n = 5 per group) were treated with the same doses described above at four different time points (that is, day 1, 5, 8 and 12 after study initiation). Statistically significant efficacy ( P = 0.0525) was not observed for Btz and Btz-BPD when administered at the low dose (Fig. 3e–g ). On the other hand, the high dose of Btz-BPD offered marked improvements in tumour growth inhibition and survival (with a mean duration of 108 ± 11 days compared with 24 ± 4 days for the control group, for instance, P = 0.0002). Complete responses were observed in 40% of the animals (2 out of 5), whereas no complete responses were seen for either free Btz or PBS (Fig. 3f ). Thus, Btz-BPD is more effective as a single-agent PI therapy than Btz in this model. Lastly, we note that Pom-BPD gave similar trends as monotherapy (Supplementary Fig. 27 ). The serum distribution and PI activity of Btz and Btz-BPD were assessed to explain their differences in efficacy and MTD. Proteasomes are present in micromolar concentrations in red blood cells (RBCs); the binding of PIs to RBC proteasomes limits bioavailability and contributes to haematologic toxicity 49 , 50 . Stable boronic ester prodrugs like Btz-BPD may overcome this limitation. To test this hypothesis, Btz-BPD was incubated in human blood for various times. The plasma and cell fractions were separated, and the amount of Btz-BPD in each fraction was quantified (Supplementary Table 2 ). The concentration of Btz-BPD in plasma was 7-fold to 10-fold greater than in RBCs at all time points, which represents a >100-fold reversal compared to reported values for free PIs 49 , 50 . Next, the PI activity was assessed. The IC 50 values (concentrations of PI at which the proteasome is 50% active) for Btz and Btz-BPD were 11.83 and 80.50 nM, respectively (Supplementary Fig. 28 ). Thus, even when directly exposed to its target, Btz-BPD is relatively stable, which would shift its exposure away from RBCs and thereby improve bioavailability in vivo. In vitro characterization of combination nanomedicines Next, we investigated the potential synergies among Btz, Pom and Dex as free drugs and Btz-BPD , Pom-BPD and Dex-BPD as single-drug BPDs in vitro using a full-factorial design approach in MM.1S and KMS11 cell lines (Supplementary Fig. 29 ). Synergistic, additive or antagonistic relationships were determined using the Chou–Talalay method (Supplementary Figs. 29 and 30 ). Notably, free drugs and BPDs displayed distinct combination indexes (CIs) (Supplementary Fig. 30 ), suggesting that the direct translation of free-drug ratios to nanocarriers would be detrimental in this system. Additionally, the addition of Dex gives improvements in the efficacies of both Btz and Pom, and cell death is mostly driven by the concentration of Btz (Supplementary Fig. 30 ). CI maps of the three free drugs and of the three single-agent BPDs were generated using the Loewe additivity method (Fig. 4a ), holding the concentration of Dex constant (2 nM for free drugs and 20 nM for Dex-BPD ). Leveraging the Bliss independence model that predicts the toxicity of additive drug combinations, we identified a Btz:Pom:Dex ratio of 0.20:9.46:0.34 as synergistic (it is more toxic than the Bliss model prediction) and a ratio of 0.02:9.98:0.01 as antagonistic (as it is less toxic than the Bliss model prediction) for single-drug BPDs. Two new three-drug BPDs, namely, Syn and Ant , were synthesized bearing these average synergistic and antagonistic drug ratios, respectively, by the copolymerization of Btz-MM , Pom-MM and Dex-MM (Fig. 1a–d and Supplementary Table 1 ). Syn and Ant were incubated with four different MM cell lines (MM.1S, KMS11, U266 and KMS18); Syn exhibited greater toxicity and Ant showed lower toxicity than the Bliss model prediction (Fig. 4b ), which confirms the synergistic and antagonistic nature of these three-drug BPDs, respectively. Fig. 4: Three-drug BPD CI studies. a , CI maps obtained by using the Chou–Talalay method with a fixed dose of Dex (2 nM) and by varying the concentrations of Btz and Pom used in the free-drug combinations (top); a fixed dose of Dex-BPD (20 nM) and varying doses of Btz-BPD and Pom-BPD were employed for BPD combinations (bottom) after 48 h of treatment. b , Ratio validation using viability assays of three-drug BPDs at Syn and Ant were performed in four MM cell lines (KMS11, MM.1S, U266 and KMS18), confirming that the selected ratios were synergistic and antagonistic, respectively, compared with the corresponding Bliss model (for additive drug activity). Full size image In vivo evaluation combination nanomedicines We propose that three-drug BPDs should outperform mixtures of single-drug BPDs at the same synergistic ratio in vivo. To rationalize this proposal, the variance from the target drug ratio as a function of the number of BPD molecules for three-drug BPDs (assuming random copolymerization) and mixtures of one-drug BPDs was modelled (Supplementary Fig. 31 ). For small BPD sample sizes (<10,000 BPD molecules), the statistical mixture is more likely to reflect the target ratio. For example, if one randomly selects 1,000 BPD molecules, the sample will be ~80% reflective of the target ratio for the three-drug BPD and only ~20% reflective of the target ratio for the one-drug BPD mixture (Supplementary Fig. 31 ). Other reports have suggested that cells take up ~10 2 nanoparticles per vesicle regardless of the nanoparticle dose 51 , 52 , which could amplify this effect. Using the same MM models as above ( n = 5), Syn was administered at two doses: 5.30 mg kg –1 ( Syn LD , 0.01 mg kg –1 Btz, 0.38 mg kg –1 Pom, 0.02 mg kg –1 Dex) and 25.00 mg kg –1 ( Syn HD , 0.05 mg kg –1 Btz, 1.64 mg kg –1 Pom, 0.08 mg kg –1 Dex); Ant was administered at 50.00 mg kg –1 (0.01 mg kg –1 Btz, 3.48 mg kg –1 Pom and 0.01 mg kg –1 Dex); a mixture of single-drug BPDs ( 1D-BPD ) was administered at 5.30 mg kg –1 (0.10 mg kg –1 Btz-BPD , 5.00 mg kg –1 Pom-BPD and 0.20 mg kg –1 Dex-BPD ), corresponding to the same dose as Syn LD . Free drug ( FD ) was administered in a total mass that matched the mass of Syn LD (0.1 mg kg –1 Btz, 5.0 mg kg –1 Pom and 0.2 mg kg –1 Dex) (Fig. 5 ). Lower doses of PI were used for these combination therapy studies compared with the monotherapy studies above to more easily differentiate between the study groups. Bioluminescence imaging for the MM.1S model was done on day 0 (study initiation) and day 20—a known cutoff date for control mice in this model 27 . Fig. 5: Improved therapeutic efficacy of synergistic three-drug BPD in MM mouse models ( n = 5 mice per treatment group). a , Tumour fold-changes for various synergistic drug ratio delivery methods in the s.c. KMS11 mouse model. FD , free-drug combination; 1D-BPD , mixture of one-drug-loaded BPDs (synergistic ratio); Syn LD , low-dose Syn (synergistic ratio, three-drug BPD). Data are presented as mean ± s.e.m. Statistical analysis was performed using a two-tailed t -test to compare the different groups at fixed time points. 1D-BPD versus FD , P = 0.045; 1D-BPD versus Syn LD , P = 0.007. b , Tumour fold changes comparing the therapeutic outcomes for three-drug BPDs carrying synergistic versus antagonistic drug ratios in the s.c. KMS11 mouse model. Ant , antagonistic ratio; Syn LD , low-dose Syn (synergistic ratio); Syn HD , high-dose Syn (synergistic ratio). Data are presented as mean ± s.e.m. A statistical analysis was performed using a two-tailed t -test to compare the different groups at fixed time points. Ant versus Syn LD , P = 0.0075; Syn LD versus Syn HD , P = 0.045. c , d , Associated Kaplan–Meier curves comparing the therapeutic outcomes for various synergistic drug ratio delivery methods in the s.c. KMS11 mouse model ( c ) and for synergistic versus antagonistic three-drug BPDs in the same model ( d ). Statistical analysis was performed using a log-rank test, with P = 0.045 ( c ) and P = 0.0325 ( d ). e , Bioluminescence imaging of orthotopic MM.1S mouse models at day 0 and day 20 during treatment. f , g , Kaplan–Meier curves comparing the therapeutic outcomes for various synergistic drug ratio delivery methods in orthotopic MM.1S GFP + /LUC + mouse model ( f ) and for the synergistic versus antagonistic three-drug BPDs in the same model ( g ). A statistical analysis was performed using a log-rank test, with P = 0.0025 ( f ) and P = 0.025 ( g ). For statistical tests, ns denotes non-significant; *, P < 0.05; **, P < 0.01; ***, P < 0.001. The arrow indicates the last administered dose. We note that a – d and f and g display different study groups within the same experiment and sharing the same controls and Syn LD group, and these panels have been separated for visualization purposes. The mean survival times were as follows: FD (47 ± 6 days for KMS11 model and 41 ± 9 days for MM.1S model), 1D-BPD (53 ± 4 days for KMS11 model and 48 ± 4 days for MM.1S model), Syn LD (61 ± 9 days for KMS11 model and 53 ± 14 days for MM.1S model), Syn HD (unavailable for KMS11 model as >50% of the mice survived until the end of the study and 62 ± 8 days for MM.1S model) and Ant (52 ± 6 days for KMS11 model and 46 ± 5 days for MM.1S model). Full size image In support of our modelling, Syn LD outperformed 1D-BPD and FD (Fig. 5 ) in slowing tumour progression (Fig. 5a,e ) and increasing survival (Fig. 5c,f ). Moreover, Syn HD provided further enhancements in efficacy compared with Syn LD (Fig. 5b,d,g ) while still using less active API than FD . On the other hand, Ant displayed inferior efficacy compared with Syn LD despite having the same Btz dose and a tenfold higher dose of Pom (Fig. 5d,g ), suggesting that the synergistic ratio is preferred over a ‘more is better’ approach 6 , 7 , 8 . Interestingly, Ant outperformed FD despite having a smaller amount of drug, which may be due to the improved delivery of drugs to tumour cells via the BPD. Conclusions We report a nanomedicine strategy that offers a promising new PI-based treatment for MM and potentially other cancers and enables the rapid translation of three-drug synergies determined in vitro to in vivo 6 , 7 , 8 . First, PI-based monotherapy ( Btz-BPD ) is introduced that offers markedly improved efficacy compared with the standard PI Btz and displays no detectable toxicities in two in vivo models of MM. Then, by manufacturing single-drug BPDs of Btz, Pom and Dex, we observe that BPDs display synergistic, additive or antagonistic patterns, respectively, distinct from their corresponding free-drug counterparts, showing that synergies should be measured in the nanocarrier context. Finally, three-drug BPDs are shown to outperform a mixture of three single-drug BPDs and free drugs in vivo, which is quantitatively modelled. Overall, this work offers potentially translatable therapies for MM and offers new mechanistic insights into optimizing and manufacturing combination nanomedicines in other disease contexts. This approach also raises regulatory questions that will be important as the field of combination therapeutics moves forward. For example, could nanocarriers bearing a statistical mixture of drugs generally classify as single entities for regulatory purposes? If so, such an approach may be advantageous compared with mixtures of nanocarriers wherein each nanocarrier would need independent evaluation. Additionally, although it was shown here that synergy identified among Btz, Dex and Pom holds in four different cell lines, it is uncertain that this ratio would be optimal for all MM cell lines and patients given the heterogeneity of MM. A future clinical workflow could involve (1) biopsy to isolate a patient’s cancer cells; (2) CI screening to determine if synergy is maintained in those cells or if an alternative synergistic ratio exists; (3) if synergy is maintained, existing BPDs could be administered, and if not, BPDs with patient-specific ratios could be generated on demand. The latter would be facilitated if components of BPD combination therapies, such as prodrug macromonomers or single-drug BPDs, could undergo translational steps as one package 53 , 54 . Altogether, these questions and directions for the field of combination nanomedicine are fascinating. Methods Representative procedure for the synthesis of combination BPD with Pom:Btz:Dex ratio of 9.5:0.2:0.3 (Syn) To a vial containing a stir bar, Pom-M (34.3 mg, 8.7 μmol, 9.5 eq.) was added. To another three separate vials, a solution of Btz-M (20 mg ml –1 in tetrahydrofuran (THF)), a solution of Dex-M (20 mg ml –1 in THF) and a solution of third-generation Grubbs catalyst ( G3-Cat , 0.02 M in THF) were freshly prepared. THF (38.7 μl) was then added to the vial containing Pom-M , followed by the addition of the Btz-M (37.9 μl, 0.19 μmol, 0.2 eq.) and Dex-M (61.7 μl, 0.31 μmol, 0.3 eq.) solution. To the macromonomer mixture, G3-Cat solution (46.1 μl, 0.92 μmol, 1.0 eq.) was added, affording the desired total degree of polymerization of 10, a Pom:Btz:Dex ratio of 9.5:0.2:0.3 and a total macromonomer concentration of 0.05 M. The reaction mixture was allowed to stir for 3 h at room temperature. To quench the polymerization, a drop of ethyl vinyl ether was added. The reaction mixture was transferred to 8-kDa-molecular-weight cutoff dialysis tubing in 3 ml nanopure water; the solution was then dialysed against H 2 O (500 ml ×3; solvent exchange every 6 h). The dialysed solution of Syn was then concentrated as desired via centrifugation with a filter tube. Alternatively, Syn could be obtained by lyophilization, or precipitation from diethyl ether. Cell lines MM.1S (CRL-2974, ATCC) and U266 (TIB-196, ATCC) cells were obtained from ATCC (Manassas). KMS11 (JCRB1179, JCRB) and KMS18 (CVCL-A637, JCRB) cells were obtained from the JCRB Cell Bank. All the cell lines were cultured in Roswell Park Memorial Institute 1640 medium (Thermo Fisher Scientific) supplemented with 10% foetal bovine serum (VWR), 1% penicillin/streptomycin (Thermo Fisher Scientific) and 1% glutamine (Thermo Fisher Scientific). MM.1S Luc+/GFP+ cells were generated by retroviral transduction and authenticated by short tandem repeat DNA profiling. All the cell lines were confirmed to be mycoplasma free using the MycoAlert Mycoplasma kit (Lonza). The cell lines were housed in 37 °C incubators under 5% CO 2 . Animal usage All the experiments involving animals were reviewed and approved by the Dana-Farber Cancer Institute’s Committee for Animal Care. The maximum tumour size/burden permitted by the committee was not exceeded in these studies. For the free-drug comparison, Btz injection was administered via s.c. injection (as i.v. toxicity otherwise governed this route); Dex and Pom were administered via i.v. injection. All the BPDs were administered via i.v. injection. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Data availability All data supporting the findings of this study are available within the Article and its Supplementary Information and can also be obtained from the corresponding authors upon reasonable request. Change history 13 March 2023 A Correction to this paper has been published:

Treating cancer with combinations of drugs can be more effective than using a single drug. However, figuring out the optimal combination of drugs, and making sure that all of the drugs reach the right place, can be challenging. To help address those challenges, MIT chemists have designed a bottlebrush-shaped nanoparticle that can be loaded with multiple drugs, in ratios that can be easily controlled. Using these particles, the researchers were able to calculate and then deliver the optimal ratio of three cancer drugs used to treat multiple myeloma. "There's a lot of interest in finding synergistic combination therapies for cancer, meaning that they leverage some underlying mechanism of the cancer cell that allows them to kill more effectively, but oftentimes we don't know what that right ratio will be," says Jeremiah Johnson, an MIT professor of chemistry and one of the senior authors of the study. In a study of mice, the researchers showed that nanoparticles carrying three drugs in the synergistic ratio they identified shrank tumors much more than when the three drugs were given at the same ratio but untethered to a particle. This nanoparticle platform could potentially be deployed to deliver drug combinations against a variety of cancers, the researchers say. Irene Ghobrial, a professor of medicine at Harvard Medical School and Dana-Farber Cancer Institute, and P. Peter Ghoroghchian, president of Ceptur Therapeutics and a former MIT Koch Institute Clinical Investigator, are also senior authors of the paper, which appears today in Nature Nanotechnology. Alexandre Detappe, an assistant professor at the Strasbourg Europe Cancer Institute, and Hung Nguyen Ph.D. '19 are the paper's lead authors. Controlled ratio Using nanoparticles to deliver cancer drugs allows the drugs to accumulate at the tumor site and reduces toxic side effects because the particles protect the drugs from being released prematurely. However, only a handful of nanoparticle drug formulations have received FDA approval to treat cancer, and only one of these particles carries more than one drug. For several years, Johnson's lab has been working on polymer nanoparticles designed to carry multiple drugs. In the new study, the research team focused on a bottlebrush-shaped particle. To make the particles, drug molecules are inactivated by binding to polymer building blocks and then mixed together in a specific ratio for polymerization. This forms chains that extend from a central backbone, giving the molecule a bottlebrush-like structure with inactivated drugs—prodrugs—along the bottlebrush backbone. Cleavage of the linker that holds the drug to the backbone release the active agent. "If we want to make a bottlebrush that has two drugs or three drugs or any number of drugs in it, we simply need to synthesize those different drug conjugated monomers, mix them together, and polymerize them. The resulting bottlebrushes have exactly the same size and shape as the bottlebrush that only has one drug, but now they have a distribution of two, three, or however many drugs you want within them," Johnson says. In this study, the researchers first tested particles carrying just one drug: bortezomib, which is used to treat multiple myeloma, a cancer that affects a type of B cells known as plasma cells. Bortezomib is a proteasome inhibitor, a type of drug that prevents cancer cells from breaking down the excess proteins they produce. Accumulation of these proteins eventually causes the tumor cells to die. When bortezomib is given on its own, the drug tends accumulate in red blood cells, which have high proteasome concentrations. However, when the researchers gave their bottlebrush prodrug version of the drug to mice, they found that the particles accumulated primarily in plasma cells because the bottlebrush structure protects the drug from being released right away, allowing it to circulate long enough to reach its target. Synergistic combinations Using the bottlebrush particles, the researchers were also able to analyze many different drug combinations to evaluate which were the most effective. Currently, researchers test potential drug combinations by exposing cancer cells in a lab dish to different concentrations of multiple drugs, but those results often don't translate to patients because each drug is distributed and absorbed differently inside the human body. "If you inject three drugs into the body, the likelihood that the correct ratio of those drugs will arrive at the cancer cell at the same time can be very low. The drugs have different properties that cause them to go to different places, and that hinders the translation of these identified synergistic drug ratios quite immensely," Johnson says. However, delivering all three drugs together in one particle could potentially overcome that obstacle and make it easier to deliver synergistic ratios. Because of the ease of creating bottlebrush particles with varying concentrations of drugs, the researchers were able to compare particles carrying different ratios of bortezomib and two other drugs used to treat multiple myeloma: an immunostimulatory drug called pomalidomide, and dexamethasone, an anti-inflammatory drug. Exposing these particles to cancer cells in a lab dish revealed combinations that were synergistic, but these combinations were different from the synergistic ratios that had been identified using drugs not bound to the bottlebrush. "What that tells us is that whenever you are trying to develop a synergistic drug combination that you ultimately plan to administer in a nanoparticle, you should measure synergy in the context of the nanoparticle," Johnson says. "If you measure it for the drugs alone, and then try to make a nanoparticle with that ratio, you can't guarantee it will be as effective." New combinations In tests in two mouse models of multiple myeloma, the researchers found that three-drug bottlebrushes with a synergistic ratio significantly inhibited tumor growth compared to the free drugs given at the same ratio and to mixtures of three different single-drug bottlebrushes. They also discovered that their bortezomib-only bottlebrushes were very effective at slowing tumor growth when given in higher doses. Although it is approved for blood cancers such as multiple myeloma, bortezomib has never been approved for solid tumors due to its limited therapeutic window and bioavailability. "We were happy to see that the bortezomib bottlebrush prodrug on its own was an excellent drug, displaying improved efficacy and safety compared to bortezomib, and that has led us to pursue trying to bring this molecule to the clinic as a next-generation proteasome inhibitor," Johnson says. "It has completely different properties than bortezomib and gives you the ability to have a wider therapeutic index to treat cancers that bortezomib has not been used in before." Johnson, Nguyen, and Yivan Jiang Ph.D. '19 have founded a company called Window Therapeutics, which is working on further developing these particles for testing in clinical trials. The company also hopes to explore other drug combinations that could be used against other types of cancer. Johnson's lab is also working on using these particles to deliver therapeutic antibodies along with drugs, as well as combining them with larger particles that could deliver messenger RNA along with drug molecules. "The versatility of this platform gives us endless opportunities to create new combinations," he says.

10.1038/s41565-022-01310-1

Biology

The cement for coral reefs

Sebastian Teichert et al, A possible link between coral reef success, crustose coralline algae and the evolution of herbivory, Scientific Reports (2020). DOI: 10.1038/s41598-020-73900-9 Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-020-73900-9

https://phys.org/news/2020-11-cement-coral-reefs.html

Abstract Crustose coralline red algae (CCA) play a key role in the consolidation of many modern tropical coral reefs. It is unclear, however, if their function as reef consolidators was equally pronounced in the geological past. Using a comprehensive database on ancient reefs, we show a strong correlation between the presence of CCA and the formation of true coral reefs throughout the last 150 Ma. We investigated if repeated breakdowns in the potential capacity of CCA to spur reef development were associated with sea level, ocean temperature, CO 2 concentration, CCA species diversity, and/or the evolution of major herbivore groups. Model results show that the correlation between the occurrence of CCA and the development of true coral reefs increased with CCA diversity and cooler ocean temperatures while the diversification of herbivores had a transient negative effect. The evolution of novel herbivore groups compromised the interaction between CCA and true reef growth at least three times in the investigated time interval. These crises have been overcome by morphological adaptations of CCA. Introduction Coral reefs support the biologically most diverse marine ecosystems and have done so over substantial parts of earth history, starting in the Late Triassic, when scleractinian corals became prolific reef builders 1 . Mitigating the threats to modern coral reef ecosystems will thus benefit from a better understanding of the underlying causes in the rise and fall of ancient coral reefs 2 . Reefs, broadly defined as laterally confined limestone “structures built by the growth or metabolic activity of sessile benthic aquatic organisms” 3 comprise a large array of constructional styles and biota, and grew in a variety of environments. The concept of ‘true reefs’ in ancient reefs derives from the constructional style and environment of recent tropical, shallow-water coral reefs. True reefs with a syndepositional relief and a rigid framework constructed by skeletal organisms are known since Cambrian times 4 but occur alongside other reef types such as reef mounds, mud mounds and biostromes. There is a significant increase of both the relative and absolute abundance of true reefs over the Phanerozoic 5 . This trend may be caused by intrinsic (biological) factors as there are few significant correlations with earth system parameters such as temperature, sea level, or oceans chemistry 5 . Although hypothesized to play a major role 6 , the relative importance of biotic interactions in reef evolution is still poorly known. The evolution of crustose coralline red algae (Subclass Corallinophycidae Le Gall and Saunders 7 , hereafter referred to as ‘CCA’) may underline such interactions in an exemplary fashion. CCA play a key role in the construction of many modern coral reefs 8 , 9 , 10 in several regards: CCA are not only primary producers and contributors of calcareous sediment but also often act as consolidators and binders of coral reefs. The algae can form a distinct ridge providing a surf-resistant reef crest and bind loose sediment 11 . Even though there are examples of true reefs that grow successful without this CCA ridge under wave-exposure 12 , 13 , CCA are still considered as ‘the glue that holds coral reefs together’ 9 in many cases. Another key factor for the success of modern coral reefs are grazing organisms such as echinoids and parrot fish, because they remove fleshy algae from the reef surface 14 , 15 . The evolution of grazers also led to an increased feeding pressure on CCA, followed by adaptive strategies of the CCA 16 . It seems plausible that this development also impacted a potential capacity of CCA to facilitate coral reef formation in the geological past. However, this hypothesis has not been assessed quantitatively until now. There are hints for a long-term positive interaction between corals and CCA over geological timescales 17 . However, a finer temporal and taxonomic resolution and the inclusion of environmental parameters may explain if and why a potential importance of CCA for reef development varies over time. Using data from the PaleoReefsDatabase (PARED), a comprehensive compendium of geological and paleontological data of Phanerozoic reef sites as described previously 18 , 19 , 20 , we evaluate the role of CCA in coral reef development at the level of geological stages from the Early Cretaceous to the Pleistocene while considering additional factors such as ocean temperature and chemistry as well as interactions with relevant groups of grazing organisms. We test three hypotheses: (1) The probability of true reef formation is correlated with the presence of CCA reef cementers; (2) the capacity of CCA to reinforce coral reefs is linked to various oceanographic and ecological parameters; (3) the diversification of grazing organisms led to transient crises in the capacity of CCA to support true reef formation. Additionally, we highlight alternative explanations for the correlation between CCA occurrences and the formation of true reefs. This includes the impact of herbivore radiation on corals rather than on CCA directly and ecological niches for CCA provided by increased reef growth. Results The general hypothesis that the occurrence of true reefs is strongly correlated with CCA as secondary reef builders is supported by the linear regression model (Fig. 1 ). Regression residuals are not auto-correlated suggesting that no further data treatment is required. (Supporting Information Fig. S1 ). Figure 1 Role of CCA as secondary reef builders. Linear regression model between proportions of true reefs and proportions of coral reefs with CCA as secondary reef builders indicating a significant correlation between CCA cementation and true reef development. Full size image The temporal patterns of the investigated variables, proportions of true CCA-reefs, sea level, ocean temperature, CO 2 concentration, species diversity of CCA, and the origin and diversification patterns of the grazer clades echinoids and parrot fish are visualized in Fig. 2 . Four transient crises are evident, one between the Turonian and the Campanian, one in the Paleocene (Selandian–Thanetian), one in the Miocene (Serravallian), and the youngest one in the Pliocene (Zanclean–Piacenzian). Figure 2 Temporal patterns of true CCA-reef formation and potentially influencing parameters. Patterns of investigated variables, representing proportions of true CCA-reefs retrieved from PARED, relative sea level based on ocean volume change, ocean temperature derived from oxygen isotope data, CO 2 concentration relative to the current level, rarefied species diversity of CCA, echinoid evolution expressed as mean number of character changes per lineage per million years, and parrot fish origination and diversification expressed as lineage-through-time plot. Stars indicate the four transient crises in the CCA’s abundance within true coral reefs. Full size image All environmental variables except grazers are correlated with each other (Table 1 ). The high values of correlation between potential explanatory variables hinder a quantification of independent effects in the subsequent analyses. The only exceptions are the origination and diversification patterns of grazers. Table 1 Collinearities between environmental parameters. Full size table Stepwise model selection in a multivariate GLM selected the variables ocean temperature, grazers, and species diversity of CCA for the final model (Table 2 ) with reasonable goodness of fit measures (Table 3 , Fig. S2 ). The evaluation of residuals shows that residuals are not significantly auto-correlated (Supporting Information Fig. S3 ). However, the high correlation among explanatory variables (Table 1 ) makes it impossible to separate the effects of CCA diversity and ocean temperature with confidence . Nevertheless, visualization of the GLM shows that species diversity of CCA is positively correlated with the presence of true CCA-reefs while higher ocean temperatures and the origination and diversification of grazers have negative effects (Fig. 3 ). Table 2 Results of the GLM. Full size table Table 3 Goodness of fit measures for the GLM. Full size table Figure 3 Visualization of the GLM. CCA species diversity is positively correlated with the presence of true coral reefs that have CCA as secondary reef builders while origins and higher ocean temperatures have a negative effect. Full size image Discussion The role of CCA as reef consolidators We found a significant correlation between the proportion of reefs that contain CCA as secondary reef builders and the proportion of true reefs over the last 150 million years. Coral reefs can benefit from CCA in various ways. Relating to the reef ridge, the stony pavement made up by the algae protects the ridge from onrushing waves and also consolidates the reef flats behind the ridges 11 . With reference to the whole reef, CCA reinforce the structure created by corals, fill cracks, bind together much of the sand, dead corals and debris, and thereby create a stable substrate and reduce reef erosion 22 . Larval settlement, metamorphosis, and recruitment of several coral species is strictly determined by chemosensory recognition of specific signal molecules uniquely available in specific CCA 23 . However, it has to be considered that there are modern reefs that cope with wavy, high-energy environments without the aid of CCA, as for example the Alacran reef in Mexico 12 . CCA are not the only possibility to add rigidity to a reef. Submarine lithification can be more important than CCA in creating calcite precipitates, especially when environmental and ecological conditions are unfavourable for the growth of CCA, e.g. because of the lack of light. Submarine lithification in the form of Mg-calcite precipitates exists in many forms, including cemented micritic crusts and infillings of cracks. Additionally, their respective carbonate sources may be abiotic 24 or originate from a great variety of organisms, including reef fish 25 . Therefore, they do also play an important role for the structural integrity of coral reefs 24 . CCA abundance may benefit from reef growth in terms of ecological niches provided, additionally increasing the positive correlation. We thus suggest that the significant correlation between the proportion of reefs reinforced by CCA as secondary reef builders and the proportion of true reefs can be interpreted as a mutual benefit. On the one hand, the presence of CCA can add stability to coral reefs, especially when the reef ridge is exposed to heavy wave action. On the other hand, sufficient reef growth can be a prerequisite for a larger abundance of CCA. A shift towards one side in this mutual dependence is subject to the particular features of each reef, as for example if CCA rather benefit from the shelter of crevices in reefs with high grazing pressure or if corals rather benefit from the presence of CCA at sites of intense wave exposure. The physicochemical parameters ocean temperature, sea level, and RCO 2 CCA occur worldwide from the tropics 10 to polar latitudes 26 and temperature is one of the primary determinants in their geographical distribution, and the boundaries of their biogeographical regions are associated with isotherms 27 . Therefore, the identification of ocean temperature as an important driver of CCA reefs is reasonable. Aguirre, et al. 28 reported that throughout the history of CCA, species richness broadly correlates with global mean palaeotemperature. However, only the diversity of the order Sporolithales varies positively with temperature, whereas the diversity of the order Corallinales varies negatively with temperature. Accordingly, the warm-water Sporolithales were most species-rich during the warm Cretaceous, but they declined and were rapidly replaced by the Corallinales as Cenozoic temperatures declined. In recent environments, members of the Sporolithales are confined to greater water depths while in euphotic reefs, they do not play a role as reef stabilizers 28 and occupy only cryptic habitats sensu Kobluk 29 , i.e. cavities that serve as well-protected habitats and are not subject to the full spectrum of environmental and biotic controls that exist on the reef surface 28 . The wave-pounded intertidal algal ridges are built predominantly by Porolithon onkodes (Heydrich) Foslie 1909, P. gardineri (Foslie) Foslie 1909, P. craspedium (Foslie) Foslie 1909, and Lithophyllum kotschyanum Unger 1858 in the Indo-Pacific. In the Atlantic, the main reef reinforcers are Porolithon onkodes (Heydrich) Foslie 1909 and Lithophyllum congestum (Foslie) Foslie 1900. All these species belong to the ‘cool’-water adapted Corallinales. Thus, the increasing capacity of CCA to stabilize coral reefs is in line with the general trend of decreasing ocean temperatures. A change in sea level does not impact the capacity of CCA to reinforce coral reefs, likely because sea level changes measured on the level of geological stages have no effect on reef formation 5 . On shorter time scales, sea level is expected to influence the formation of coral reefs, but probably not the CCA’s reef enforcing capacity. We conclude this because the environmental tolerances of CCA in terms of sea level fluctuation are much wider than those of reef corals. Most CCA species appear uniquely tolerant of aerial exposure 10 . Additionally, many CCA are very well adapted to changes in salinity and especially to low photon irradiances 30 . The environmental tolerances of reef corals are narrower 31 , 32 . Considering our assumption that there is a mutual relationship between the presence of CCA and the growth of true reefs, another reason might be that one of the most important genera in modern coral reefs, Acropora Oken, 1815, is well adapted to cope with rapid sea-level changes. First observed as an important reef builder in the Oligocene 33 , Acropora has become a dominant reef builder from the Pleistocene until today, when sea-level fluctuations increased in rate and magnitude 34 . Indeed, there is a temporal overlap between the first decline in the fraction of CCA reefs—between the Turonian and the Campanian—and a maximum in sea level. Despite this, sea level is not selected as a relevant explanatory variable for the fraction of CCA reefs by the GLM because the relationship between the fraction of CCA reefs and sea level varies inconsistently throughout entire time series of the analysed 150 million years. While the decline in the fraction of CCA reefs may additionally be linked to an increase in temperature before and a significant drop in CCA diversity during the period with a low fraction of CCA reefs, data of the analysis are not suitable to conclusively identify the driver for this particular CCA crisis. For the entire time series, RCO 2 , was identified by the model as a minor driver, which may be explained by the fact that an increase of atmospheric p CO 2 has only little to no impact on mean ocean surface pH on timescales exceeding 10,000 years 35 . A plausible reason is that slow rates of CO 2 release lead to a different balance of carbonate chemistry changes and a smaller seawater CaCO 3 saturation response. This is because the alkalinity released by rock weathering on land must ultimately be balanced by the preservation and burial of CaCO 3 in marine sediments. The burial is controlled by the CaCO 3 saturation state of the ocean and therefore, the saturation is ultimately regulated by weathering on long time scales, and not by atmospheric p CO 2 . The effect of weathering on atmospheric p CO 2 is much weaker than the effect of weathering on ocean pH. The much stronger effect of weathering on ocean pH allows pH and CaCO 3 saturation to be almost decoupled for slowly increasing atmospheric p CO 2 35 . The influence of CCA species diversity The quantification of CCA species diversity in the geological past is associated to a number of challenges. While for recent CCA the extensive use of molecular phylogenetic methods resolved the four orders (Corallinales, Hapalidiales, Sporolithales, and Rhodogorgonales) currently recognized in the subclass Corallinophycidae as monophyletic lineages 36 , 37 , we have to rely on morphological characters since molecular methods are not available for the identification of fossil CCA. Because CCA show a pronounced phenotypic plasticity depending on environmental factors, their taxonomic identification depends on morphological characters like conceptacles (i.e. spore chambers) and the arrangement of cells in different areas of the thallus, features often not adequately preserved in fossil CCA. This has led to a great number of fossil CCA taxa that have been described on the basis of only a few anatomical characters of doubtful taxonomic value 38 . The inclusion of such taxa precludes fully reliable diversity estimations. To circumvent such problems, we used rarefied species data reviewed by experts on fossil CCA taxonomy 28 . Our results show that high CCA diversity is linked to a higher abundance of CCA in true coral reefs. This might seem to contrast with the fact that in modern reefs, the wave-pounded intertidal algal ridges are built predominantly by only a few species while the ones making up the majority of diversity have a cryptic, hidden mode of life protected from full or direct exposure to major physical environmental factors and therefore do not contribute significantly to reef stabilization. However, if several CCA species were contributing to the same ecosystem function, a higher species diversity may have buffered reef systems from losing all species associated with the key function of supporting reef development 39 . As discussed in detail in the next section, the abundance of CCA in true reefs was transiently reduced four times since the Early Cretaceous. Except for the earliest crisis, this was likely caused by the origin and diversification of echinoids and parrot fish, prominent groups of bioeroding organisms that denude CCA. However, the CCA-coral reef system successfully recovered all times. We argue that this was supported by functional redundancy of CCA, because a diverse group of abundant species with a wider range of responses can help absorb disturbances 39 . This redundancy of responses to events among species within a functional group—the reef cementers—is an important component of resilience and the maintenance of ecosystem services. The amount of CCA biomass is critical in terms of the cementing capacity. Multi-species community models 40 have shown that with consecutive native species’ extinctions at high diversity levels, species extinction usually only leads to a slight decrease in the total biomass of the native community. However, when starting from a lower initial diversity, a few consecutive species extinctions cause a relatively large biomass loss that ultimately leads to collapse. It should also be stressed that sometimes single species are responsible for the functioning of an ecosystem (i.e., keystone species), even if the ecosystem features a generally high biodiversity. Therefore, such ecosystems will decline if this key species is removed 41 . Experiments with plants in rangelands 42 showed that functional diversity maintains ecosystem functioning. At heavily grazed sites, some species dominant in the ungrazed communities were lost or substantially reduced. In four out of five cases, the minor species that replaced these lost ones were their functional analogues. Accordingly, we suggest that formerly less dominant but functionally analogous grazing-tolerant species increased in abundance and contributed to the maintenance of ecosystem functions. CCA species removed or reduced in biomass by grazing pressure can be replaced in terms of their ecosystem service, i.e. reef cementation, by other CCA that are better adapted to grazing. This implies that in recent coral reef environments, areas with high CCA diversity—potentially including species occupying cryptic habitats—are more resilient against disturbance. Because the skeletal mineralogies of CCA vary considerably among species 43 , this resilience possibly applies also to future ocean acidification. The evolution of herbivory and transient reef crises The data reveal four crises in the abundance of CCA within true reefs, during the Cretaceous (Turonian–Campanian), the Paleocene (Selandian–Thanetian), the Miocene (Serravallian), and the Pliocene (Zanclean–Piacenzian). The reason that the timing of the Paleocene crisis differs from the known Paleocene–Eocene crisis 20 might be that our study focuses on the number of true reefs, while the Paleocene-Eocene crisis is expressed by a change in cumulative metazoan reef volume. Except for the first one, all crises observed here occurred synchronous with pronounced evolutionary events in clades of grazing organisms. Cementing and binding is the main function of CCA in the facilitation of true coral reefs. The decline in CCA abundance during the Selandian–Thanetian corresponds with a marked increase in the rate of morphological evolution in echinoids (Fig. 2 ). This includes major shifts in lifestyle and the evolution of new subclades in this group 44 , with a net trend towards improved mobility and feeding ability also on CCA 16 . Regarding the Serravallian and Zanclean–Piacenzian crises, echinoids appear to play a very minor role as their evolutionary rates constantly decreased over time 44 . However, another important clade of coralline grazers, the parrot fishes (Scarinae Rafinesque, 1810) may have become major players 45 . Although reef-grazing fish have existed for nearly 400 Ma, specialized detritivores feeding on macroalgae have only been known since the Miocene 46 . This is also in line with the radiation of acroporid corals since the mid Miocene 47 , whose branched morphologies create interstitial niches for parrot fish but also for cryptic CCA species. The parrot fishes (Scarinae) first appeared in the Serravallian 45 , which may have caused the third crisis in CCA reef cementing capacity. The lineage diversification of Scarinae was most pronounced during the Zanclean-Piacenzian, which we deem responsible for the third crisis. The abundance of CCA in true coral reefs recovered relatively fast after all crises probably due to morphological adaptations developed within the CCA. Experiments have shown that echinoids are able to graze tissues to depths averaging 88 µm 16 , which is critical for CCA with thin crust morphologies. The resulting decline of thin crust morphologies led to the occupation of niches by branching CCA 16 . The twig-like morphologies of branching CCA prevent echinoids from denuding CCA thallus and confine this process to the tips of the branches. CCA are able to transfer nutrients within their thallus 16 . Therefore, these superficial grazing wounds can be rapidly healed if sufficient nutrient reservoirs are present in other, ungrazed parts of the algae. Meristems and conceptacles engulfed in the thallus may be another adaptation pertinent to the relatively low impact of echinoid grazing, as this is a plausible strategy to protect the reproductive and growth structures of the CCA. The more intense grazing pressure exerted by the parrot fishes, which bite CCA to an average depth of 288 µm 16 and are able to eat the tips of branched CCA 48 may have resulted in a greater abundance of CCA with very thick crusts. Thick-crust CCA possess larger nutrient reservoirs making them capable to recover also from grazing exerted by parrot fishes. All these adaptations and their development are congruent with the origination and diversification of the grazer clades as already outlined in other studies 16 , 49 , 50 . Today and potentially already during the geological history, CCA did not only successfully adapt to various grazer clades but even required the grazing pressure to stay free of epiphytes 49 . Here we show for the first time that the process of grazer evolution may also have affected the potential capacity of the CCA to reinforce coral reefs for three times during the geological past. Future implications for the capacity of CCA to reinforce coral reefs As it concerns some of the most important biodiversity hot spots on our planet 2 , the potential future impact of the ongoing global change on the capacity of CCA to reinforce coral reefs should become a focal point of reef research. Despite the implementation of numerous mesocosm and aquaria experiments 51 , 52 , 53 , long-term data in the magnitude of months on CCA responses to modified environmental parameters are still sparse. Also, the change from ambient to modified parameters (e.g. p CO 2 , temperature) happens much faster than at natural rates. The impact of elevated p CO 2 on CCA depends on the rate of change. While fast rates are critical, slow p CO 2 increase may even result in increased net calcification at moderately elevated p CO 2 levels 54 . However, this comes at the cost of structural integrity of the CCA skeleton which, in turn, makes the CCA likely more susceptible to bioerosion. Bioerosion by echinoids and parrot fishes is beneficial to CCA at the present state, as it removes fast growing fleshy algae and other epiphytes 49 , but nothing is known about the future of this interaction when the integrity of the CCA skeletons is altered. Additionally, it has been shown that elevated p CO 2 levels accelerate sponge reef bioerosion 55 , 56 , 57 . Therefore, a combination of increased bioerosion rates affecting corals and CCA might lead to strongly deteriorated conditions for coral reef formation. As outlined above, a greater CCA diversity might also increase their resilience against ocean acidification because of the great variety in skeletal mineralogies. Regarding elevated temperatures, the outcome for CCA is unpredictable. Depending on the examined species, elevated temperatures affect CCA primary production in different ways: some species show no or negligible response 30 , some change their skeletal chemistry in terms of dolomite concentration 58 , and others respond with strongly impaired germination success 59 or declining skeletal densities 60 . Due to the positive influence of cooler temperatures on CCA’s abundance in true reefs detected in our study, elevated temperatures will likely have a negative outcome but also here, the rate of change might be similarly important as the magnitude. To estimate the future of CCA’s potential to facilitate coral reef growth in the face of global change, we encourage long term experiments—preferably in near-natural mesocosm studies—including the main reef stabilizing CCA species. Materials and methods While CCA have existed at least since the Silurian 61 , pre-Cretaceous occurrences are scarce. We downloaded data on all reefs occurring from the earliest Cretaceous (Berriasian stage, 145.5 Ma) to the Late Pleistocene (0.01 Ma) and having corals as the main reef builders from the PaleoReefs Database (PARED, ) in May 2018. The 736 Cretaceous to Pleistocene coral reef sites were grouped in 33 geological stages (following the International Commission on Stratigraphy 62 ) representing 145 million years of Earth history with information on reef diversity and environmental parameters. The data in PARED contain information on the constructional style of each reef, distinguishing between true reefs, reef mounds, mud mounds or banks, and biostromes. True reefs are those where skeletal organisms in growth position form a dense, rigid framework. Reef mounds share abundant skeletal organisms with no evidence for a rigid skeletal framework. Mud mounds or banks predominately consist of carbonate mud, often of microbial origin. Biostromes consist of skeletal organisms but there is no syndepositional relief 19 . We focus on coral reefs in the constructional style of true reefs, because we hypothesize that their rigidity and three-dimensionality depend on the CCA’s cementing capacity. The data in PARED also contain information on the secondary reef builders and we only asserted a significant functioning as reef cementers to the CCA when they were listed as secondary reef builders within the particular coral reef. For each geological stage, we consequently tabulated the absolute numbers of four reef types: All reefs: all reefs in this study, regardless of their constructional style and if they have CCA as secondary reef builders. True reefs: reefs with the constructional style of true reefs, but regardless if they have CCA as secondary reef builders or not. CCA-reefs: reefs that have CCA as secondary reef builders, but regardless of their constructional style. True CCA-reefs: reefs both features apply for: the constructional style of true reefs and CCA as secondary reef builders. From these data, we calculated the proportions (mean ± one standard error) of true reefs, CCA-reefs, and true CCA-reefs for each geological stage by dividing the number of reefs of each type by the number of all reefs (Table 4 ). The standard error was calculated using the equation $$SE= \sqrt{\frac{p\left(reef \,type\right)*(1-p\left(reef \,type\right))}{number \,of \,reefs}}$$ where p(reef type) is the calculated proportion of true reefs, CCA-reefs, and true CCA-reefs, respectively. Table 4 Absolute numbers and proportions of reef types retrieved from the PaleoReefs Database (PARED). Full size table Our analyses aim to identify the environmental parameters, which are crucial for the capacity of CCA to facilitate the formation of true reefs. As potential explanatory variables, we explor sea level, ocean temperature, CO 2 concentration, species diversity of CCA, and the origin and diversification patterns of the grazer clades echinoids and parrot fish. Information on sea level represents the relative sea level based on ocean volume change as mean value per geological stage 63 . Temperature data derive from the oxygen isotope dataset published by Veizer and Prokoph 64 . CO 2 concentration for focal time periods (mean values per stage) is expressed relative to the current level (RCO 2 multiproxy model; Berner and Kothavala 65 ). Data on CCA diversity represent rarefied species diversities per stage 28 . Data on echinoids are from Hopkins and Smith 44 and represent rates of morphological evolution, which are measured as the mean number of character changes per lineage per million years. To quantify the diversification rates of herbivorous parrot fish, we used lineage-through-time (LTT) estimates in a dated phylogeny (Choat et al. 45 , Fig. 1 ). This phylogeny was obtained by analyses of three loci (16S, control region, S7I1) and comprises 16 species of the genus Chlorurus Swainson, 1839 (excavating feeding mode), 45 species of Scarus Forsskål, 1775 (scraping feeding mode), and two species of Hipposcarus Smith, 1956 (scraping feeding mode), all belonging to the family Labridae Cuvier, 1816. Testing the hypothesis that novel herbivore characteristics influence the capacity of CCA to facilitate the formation of true reefs, we distinguished background intervals from intervals in which distinct increases in character change or lineage origination occurred (parameter ‘grazers’ with values 1/0). All environmental parameters are compiled in Table 5 . Table 5 Environmental parameters. Full size table We performed all statistical tests in R version 3.5.2 66 and removed all stages containing missing values in any parameters (Berriasian to Barremian, Selandian, and Gelasian) before analyses. We tested the general hypothesis that the occurrence of true reefs is strongly linked to CCA as secondary reef builders using a linear regression model between the proportions of true reefs and the proportions of CCA-reefs and tested the regression residuals for autocorrelation. Prior to further analysis, we tested all explanatory variables for collinearity. We then assessed the influence of the environmental parameters on the coral-coralline interaction by implementing multivariate analyses (generalized linear models, GLM). As the dependent variables represent proportions (percentage of true CCA-reefs relative to all reefs), we implemented GLMs with a binomial error distribution (logit-link function). The binomial error distribution for percentage values has the advantage to account for the fact that a particular percentage value is more accurate if it is based on a larger number of observations (here the number of reefs). We used stepwise model selection based on a version of the Akaike information criterion (AIC, which is used in statistics), that has a correction for small sample sizes (AICc) to estimate the relevance of different environmental parameters. We quantified model performance using different goodness of fit measures (Nagelkerke’s pseudo-R 2 , Mc Fadden’s pseudo-R 2 , maximum likelihood pseudo-R 2 ), tested model residuals for autocorrelation and visualised the model predictions. Data availability The reef data supporting the results are freely available in the PaleoReefs Database (PARED).

Coral reefs are hotspots of biodiversity. As they can withstand heavy storms, they offer many species a safe home, and at the same time, they protect densely populated coastal regions as they level out storm-driven waves. However, how can these reefs that are made up of often very fragile coral be so stable? A team of researchers from Friedrich-Alexander Universität Erlangen-Nürnberg (FAU) and the University of Bayreuth have now discovered that a very specific type of 'cement' is responsible for this—by forming a hard calcareous skeleton, coralline red algae stabilize the reefs, and have been doing so for at least 150 million years. The wide variety of life they support is immediately apparent on images of tropical coral reefs. Their three-dimensional scaffolding provides a habitat for a large number of species. However, the skeletons of the coral are often so fragile that they would not be able to withstand heavy storms by themselves. Even if scientists have long suspected that coralline red algae provide support to reefs with their calcareous skeletons, this is the first time that this link has been proven. Coralline red algae have been supporting coral reefs for at least 150 million years The researchers from FAU and the University of Bayreuth were able to prove this supporting function by analyzing more than 700 fossilized reefs from 150 million years of the Earth's history. "The coralline red algae form a calcareous skeleton and cement the coral reefs together," explains Dr. Sebastian Teichert from the Chair of Palaeoenvironmental Research at FAU. "However, several crises over the course of millions of years have limited their capacity to do so." Successful adaptive strategies against plant grazers These crises include the evolution of plant grazing marine animals such as sea urchins and parrot fishes who have repeatedly decimated populations of coralline red algae over the course of time. The algae, however, developed defense mechanisms such as special growth forms in order to defend themselves against their attackers. "The algae have adapted so well that they now even benefit from these plant grazers," says Teichert. "They rid the coralline red algae of damaging growth such as green algae, allowing it to grow unhindered." This means coralline red algae are more successful at supporting coral reefs today than ever before in the Earth's history. The extent to which climate change affects the supporting role of coralline red algae is not yet known. Any deterioration to their living conditions would not only affect the coral and other inhabitants of reefs, but also humans, as coral reefs level out storm-driven waves and make an important contribution to coastal protection. They also provide a nursery habitat for several fish and shellfish which are an important source of food.

10.1038/s41598-020-73900-9

Earth

Human-engineered changes on Mississippi River increased extreme floods

Climatic control of Mississippi River flood hazard amplified by river engineering, Nature (2018). nature.com/articles/doi:10.1038/nature26145 Journal information: Nature

http://nature.com/articles/doi:10.1038/nature26145

https://phys.org/news/2018-04-human-engineered-mississippi-river-extreme.html

Abstract Over the past century, many of the world’s major rivers have been modified for the purposes of flood mitigation, power generation and commercial navigation 1 . Engineering modifications to the Mississippi River system have altered the river’s sediment levels and channel morphology 2 , but the influence of these modifications on flood hazard is debated 3 , 4 , 5 . Detecting and attributing changes in river discharge is challenging because instrumental streamflow records are often too short to evaluate the range of natural hydrological variability before the establishment of flood mitigation infrastructure. Here we show that multi-decadal trends of flood hazard on the lower Mississippi River are strongly modulated by dynamical modes of climate variability, particularly the El Niño–Southern Oscillation and the Atlantic Multidecadal Oscillation, but that the artificial channelization (confinement to a straightened channel) has greatly amplified flood magnitudes over the past century. Our results, based on a multi-proxy reconstruction of flood frequency and magnitude spanning the past 500 years, reveal that the magnitude of the 100-year flood (a flood with a 1 per cent chance of being exceeded in any year) has increased by 20 per cent over those five centuries, with about 75 per cent of this increase attributed to river engineering. We conclude that the interaction of human alterations to the Mississippi River system with dynamical modes of climate variability has elevated the current flood hazard to levels that are unprecedented within the past five centuries. Main Flooding of the lower Mississippi River in the spring of 2011 was among the largest discharge events since systematic measurements began in the late nineteenth century, and it caused US$3.2 billion in agricultural losses and damages to infrastructure 6 . This and other recent flood events on the Mississippi River—including those in 2016 and 2017—have repeatedly, although controversially, been attributed to an aggressive campaign of river engineering designed and implemented over the past 150 years 3 , 4 , 5 . Federally mandated efforts to reduce the impacts of flooding began in the late nineteenth century and initially relied almost exclusively on the use of artificial levees, but this strategy was revised in the wake of a particularly devastating flood in the spring of 1927 that overwhelmed the levee system 7 . The current flood management system—the Mississippi River & Tributaries Project (MR&T)—includes a series of spillways that can be opened to relieve pressure on an enlarged levee system, as well as an artificially shortened and straightened main channel that is held in place by concrete retaining walls (revetments) and isolated from most of its natural floodplain 2 , 6 , 7 . Although these modifications are credited with protecting communities and croplands within the floodplain from inundation, artificial channelization has altered the relationship between discharge and river stage 3 , 4 and accelerated the rate of land loss in the Mississippi River delta 8 , necessitating additional investments in flood mitigation infrastructure and coastal restoration 9 . Although fluvial processes are sensitive to flood mitigation infrastructure, climate variability can also shape the dynamics of continental drainage networks, particularly over decadal to centennial timescales that are difficult to detect using short observational records 10 , 11 . Precipitation and soil water storage over the Mississippi River basin are influenced by climate variability driven by sea-surface-temperature anomalies in both the Pacific and Atlantic Oceans 12 , 13 . Yet establishing the natural controls on discharge extremes of the lower Mississippi has proved challenging because gauging-station measurements record a limited range of variability, particularly before major investments were made in river engineering. As a result, analyses of historical streamflow records disagree over the role that dynamical modes of climate variability play in modulating the discharge 12 , 14 , 15 . To plan flood mitigation and other infrastructure projects, it is critical to understand the climate controls on the discharge of the lower Mississippi River, but the short length of the instrumental record limits our ability to evaluate the range of natural hydrological variability from observational data alone. Recent advances in palaeoflood hydrology could extend the instrumental record back in time to diagnose the controls on the discharge of large alluvial rivers such as the lower Mississippi. Traditional approaches in palaeoflood hydrology, which include the use of slackwater deposits as flood event indices 16 , are of limited use on the low-relief landscapes that characterize the Mississippi River alluvial plain. One new approach uses the sedimentary archives held in floodplain lakes, which act as sediment traps during overbank floods, to develop continuous, quantitative and event-scale records of past flood frequency and magnitude 17 , 18 . Parallel work in dendrochronology demonstrates that when trees are inundated by floodwaters they exhibit anatomical anomalies in that year’s growth ring such that they provide a precise chronology of flood events that occurred during the growing season 19 . Together, these methodological advances provide an opportunity to evaluate interannual to multi-decadal scale trends in flood frequency and magnitude on a large alluvial river such as the lower Mississippi, before and during the era of river engineering. Here we analyse records of individual overbank flood events derived from sedimentary and tree-ring archives from the lower Mississippi River’s floodplain ( Fig. 1 ). We collected sediment cores from the in-filling thalwegs of three oxbow lakes, Lake Mary (MRY), False River Lake (FLR) and Lake Saint John (STJ), that formed by neck cut-offs of the lower Mississippi River in ad 1776, ad 1722 and roughly ad 1500, respectively 20 ( Extended Data Figs 1 , 2 , 3 ). In these sedimentary archives, we identified individual flood events by using grain-size analysis, bulk geochemistry (from X-ray fluorescence scanning, XRF) and radiography; developed age–depth models constrained by multiple independent chronological controls ( Extended Data Figs 4 , 5 , 6 ); and estimated flood magnitudes from a linear model that relates the coarse grain-size component to the discharge of historical flood events 18 ( Extended Data Fig. 7 ; see Methods for details). We also include tree-ring records from the floodplain of the lower Mississippi, collected and described by ref. 21 ; each tree-ring series was examined for anatomical evidence of flood injury to produce a record of overbank flood events that extends back to the late seventeenth century 21 . A composite time series for flood frequency describing the number of flood events in a moving 31-year window derived from sedimentary and tree-ring archives ( Fig. 2b ) is highly correlated with instrumental flood frequency ( r = 0.90, t = 19.12, effective degrees of freedom ν eff = 3.77, p < 0.001) for the interval of overlap, while reconstructed flood magnitudes ( Fig. 2c ) track trends observed in gauging-station measurements (see Supplementary Information for additional validation), indicating that the palaeoflood archives provide robust reconstructions of hydrological extremes on the lower Mississippi River beyond the period of instrumental record. Figure 1: The lower Mississippi River and the Mississippi River basin in North America. River engineering modifications (artificial cut-offs and levees) that contribute to channelization, the locations of palaeoflood records (FLR; MRY; STJ; and Big Oak Tree, BOT) and river gauging stations on the lower Mississippi used in this study (Memphis, Helena, Arkansas City, Vicksburg and Baton Rouge) are shown. Shaded relief shows relative topographic highs (dark shades) and lows (light shades) using the National Elevation Dataset 25 . PowerPoint slide Full size image Figure 2: Instrumental and reconstructed flood frequencies and magnitudes of the lower Mississippi River. a , Human impacts to the lower Mississippi River (MR&T refers to a major river engineering initiative): timing and intensity of agricultural land use 26 and river engineering. b , Flood frequencies (number of flood events in a 31-year moving window) derived from palaeoflood records, including mean and bootstrapped 2 σ confidence intervals of all palaeoflood archives, and the instrumental frequency of all floods attaining major flood stage (>1.5 m above flood stage) at the Mississippi River gauging station at Baton Rouge (station number 07374000). c , Flood magnitudes derived from the sedimentary palaeoflood records, with 1 σ uncertainties, and instrumental flood magnitudes for the Mississippi River gauging station at Vicksburg (station number 07289000). PowerPoint slide Full size image Our multi-proxy palaeoflood dataset extends the record of extremes in the discharge of the lower Mississippi River back to the early sixteenth century and demonstrates that both the frequency and magnitude of flooding have increased over the past 150 years as land use and river engineering efforts have intensified ( Fig. 2 ). Flood frequencies and magnitudes exhibit multi-decadal oscillations that increase in amplitude around the beginning of the twentieth century such that the highest rates of overbank flooding and the largest discharge events of the past 500 years have occurred within the past century. The amplification of flood magnitudes that has occurred over the past 150 years corresponds in time with the intensification of anthropogenic modifications to the lower Mississippi River and its basin, particularly the artificial channelization of the river with levees, revetments and cut-offs in the late nineteenth and early twentieth centuries 2 , 7 . Yet the continued presence of multi-decadal oscillations in flood frequency and magnitude throughout the entire period of record indicates that anthropogenic modifications to the Mississippi River system are acting in concert with other factors to alter flood hazard through time. To evaluate the role of climate variability on flood hazard, we examined the relationships between flood frequency, the El Niño–Southern Oscillation (ENSO) and the Atlantic Multidecadal Oscillation (AMO), to find that sea-surface temperature anomalies in both the Pacific and Atlantic Oceans exert a strong influence on the occurrence of lower Mississippi River floods ( Fig. 3 ). Over the past five centuries, correlations between composite flood frequency and the frequency of El Niño events ( r = 0.73) and the AMO index ( r = −0.39) derived from instrumental and palaeoclimate data sets are significant ( p < 0.001; see Methods for details). The strength and direction of these relationships support the hypothesis that discharge extremes on the lower Mississippi River arise through the interaction of ENSO, which influences antecedent soil moisture, with the AMO, which controls the flux of moisture from the Gulf of Mexico inland 12 , 15 . Extreme precipitation events over the Mississippi River basin are associated with a stronger and more westerly position of the North Atlantic Subtropical High that is characteristic of the negative phase of the AMO 12 , 13 , and these heavy precipitation events are more likely to generate discharge extremes if they fall on the saturated soils that tend to be left in the wake of El Niño events 15 . Figure 3: Lower Mississippi River flood frequency and its relation to dominant modes of climate variability. a , AMO derived from instrumental 27 and palaeoclimate 28 datasets. b , Frequency of El Niño events (the warm phase of the ENSO) in a 31-year moving window derived from instrumental 27 and palaeoclimate 28 , 29 , 30 , 31 data sets (mean with 2 σ bootstrapped confidence interval). SST, sea surface temperature. c , Frequency of lower Mississippi River floods derived from palaeoflood data (mean with bootstrapped 2 σ confidence interval). d , Correlation field of monthly precipitation 32 with the AMO 27 ( ad 1901–2014) smoothed with a common 121-month filter. e , Correlation field of monthly Palmer Drought Severity Index 33 with the Niño 3.4 index 27 ( ad 1948–2011). Correlation fields are interpolated to a common 2° × 2° grid, and individual points with significant correlations at the P < 0.05 level are marked with a hollow circle. PowerPoint slide Full size image Despite the strong influence of climatic variability on lower Mississippi River flood occurrence, the amplification of flood magnitudes that we observe over the past 150 years is primarily the result of human modifications to the river and its basin ( Fig. 4 ). The magnitude of the 100-year flood ( Q 100 ; a flood with a 1% chance of exceedance in any year) estimated from gauging-station measurements ( ad 1897–2015) is (20 ± 7)% larger than Q 100 for the period before major human impacts to the river and its basin ( ad 1500–1800), as estimated from the palaeoflood data (see Methods for details). To identify the influence of human activities on this observed increase in Q 100 , we use a linear model that relates peak discharge to the AMO index over the period before major human impacts to the river, ad 1500–1800 ( R 2 = 0.35, degrees of freedom ν = 18, p < 0.01) and use this model to predict flood magnitudes over the entire period of record. This ‘climate-only’ regression predicts that, in the absence of human modifications to the land surface, Q 100 would have increased by only (5 ± 6)% over the same period, accounting for only about 25% of the observed increase in Q 100 and implying that the remainder (about 75%) of this elevated flood hazard is the result of human modifications to the river and its basin. Figure 4: Attribution of the observed increase in flood magnitudes over the past five centuries. a , Composite peak discharges from palaeoflood archives and the instrumental record from Vicksburg. The red line indicates observed trends in the largest flood of the century in a moving window; the blue line indicates trends under ‘climate-only’ conditions, estimated from a statistical model (see text for details). Both lines are shown with 1 σ confidence intervals. Instrumental peak discharge estimates are reported without uncertainty and are therefore plotted without confidence intervals. b , Comparison of the 100-year flood observed during the baseline period ( ad 1500–1800, before major human modifications to the Mississippi River and its basin; grey boxplot) with that estimated using a statistical model under ‘climate-only’ conditions (blue boxplot) and observed (red boxplot) during the modern period of instrumental record ( ad 1897–2015). Boxplots show mean (centre line) and 1 σ confidence intervals (box top and bottom) for Q 100 estimates. PowerPoint slide Full size image The timing and nature of the amplification of flood magnitudes at the onset of the twentieth century strongly imply that it reflects the transformation of a freely meandering alluvial river to an artificially confined channel, because the confinement of flood flows to a levee-defined floodway can speed up the downstream propagation of a flood wave and increase peak discharge for a given flood 22 . The establishment of widespread agricultural activity in the Mississippi River basin occurred in the nineteenth century, before the divergence of the observed and ‘climate-only’ flood magnitudes, indicating a secondary and possibly lagged influence of agricultural expansion 23 on flood magnitudes relative to that of river engineering. In short, this analysis identifies artificial channelization of the lower Mississippi River, and its effects on the river’s gradient, channel area and flow velocity 2 , 7 , as having significantly increased the discharge of a given flood event relative to pre-engineering conditions. Our main finding—that river engineering has elevated flood hazard on the lower Mississippi to levels that are unprecedented within the past five centuries—adds to a growing list of externalized costs associated with conventional flood mitigation and navigation projects, including a reduction in a river’s ability to convey flood flows 3 , 4 , the acceleration of coastal land loss 8 and hypoxia 24 . Despite the societal benefits that these major infrastructure projects convey 6 , the costs associated with maintaining current levels of flood protection and navigability will continue to grow at the expense of communities and industries situated in the river’s floodplain and its delta. For those interested in improving seasonal and longer-term forecasts of flood hazard or management strategies that reconnect the river with its floodplain, the Mississippi River’s discharge of freshwater—and by extension the flux of sediment, nutrients and pollutants—to its outlet should be viewed as highly sensitive both to anthropogenic modifications to the basin and to variability of the global climate system. Methods Instrumental streamflow data We obtained daily stage data for Mississippi River gauges at Vicksburg (station number 07289000) and Baton Rouge (07374000) from the United States Army Corps of Engineers (USACE) and the United States Geological Survey (USGS). Discharges for the Vicksburg, Memphis (07032000), Helena (07047970), Arkansas City (07146500) and Baton Rouge gauges were compiled from multiple sources. For the early instrumental record (pre-1927), peak discharges and measured discharges were compiled from historical documents 34 , 35 . In the few cases in which annual peak discharges were not recorded during this period, we used the measured discharges to create rating curves from which to determine the peak discharge for the annual peak stage. Discharge data after ad 1927 were acquired either from the USACE or from the USGS. The discharge record at Vicksburg is the longest and most continuous of the available discharge records, and its peak annual discharge is highly correlated ( r > 0.86, p < 0.01) with that of other lower Mississippi River gauging stations in the study area (see Supplementary Information ) and was thus used to reconstruct flood magnitudes from the sedimentary archives. Sedimentary archives We collected sediment cores from the infilling thalwegs of MRY, FLR and STJ with a rod-driven vibracore system in July 2012 and March 2016 ( Extended Data Figs 1 , 2 , 3 ). For each core, we collected a replicate drive using a 7.5-cm-diameter polycarbonate piston corer to ensure recovery of an intact sediment/water interface. The targeted lakes were selected because the lateral position of the active channel near the lake’s arm has remained relatively stable from the time of cut-off to the mid-twentieth century 20 . We cannot eliminate the possibility that minor lateral and/or vertical channel migration has occurred near these lakes since the time of cut-off, but we reduce the influence of this potential bias on our analysis by (i) using a low-pass filter on the grain-size data (see below) and (ii) validating the resulting flood frequency and magnitude data sets against the instrumental record (see Supplementary Information ). At FLR and STJ, mainline levees of the MR&T have inhibited the deposition of fluvial sediment in the lake during overbank floods after about ad 1950 and 1937, respectively; MRY is not protected by artificial levees and it continues to be inundated during overbank floods. Oxbow lakes can continue to exchange water and sediment with the main channel when the river is below flood stage 36 to create high rates of fine-grained ‘background sedimentation’ that differs in texture and composition from the coarser material that is mobilized during high-magnitude flood events. Cores were collected along an arm of the oxbow lakes at locations proximal to the ‘plug’ that separates the active channel from the lake to maximize the contrast between background and flood event sediments. Core locations at each site were targeted based on bathymetric surveys before core collection. Cores were transported back to the Woods Hole Oceanographic Institution (WHOI) where they were split, described and photographed. Archived core halves were subjected to high-resolution XRF (4,000 μm resolution) and radiography (200 μm resolution) in an ITRAX core scanner housed at WHOI. For grain-size analysis, sediment sub-samples at continuous 1-cm intervals were dispersed in water using a vortex mixer before 5 s sonication and analysis in a Beckman Coulter LS 13 320 laser diffraction particle-size analyser; randomly selected replicate samples showed a <1% volume difference in any detector. Complex, multi-modal grain-size distributions were modelled as mixtures of discrete, simple distributions and decomposed using end-member calculations into four representative populations, or end-members (EMs), that were considered geologically meaningful, using the EMMAgeo package run in RStudio. The score of each sample on the coarsest end-members (EM1), representing deposition of bedload during overbank floods 18 , was normalized with a low-pass (41-cm) moving minimum filter to remove long-term trends in sediment composition caused by local geomorphic processes. We then identified potential flood deposits as normalized EM1 scores that exceeded a high-pass (11-cm) moving mean with a 0.1 EM1 score threshold, and we verified identified peaks against the XRF and radiography ( Extended Data Figs 4 , 5 , 6 ). To estimate flood magnitudes from the sediment records, we used the method of ref. 18 and developed linear models that describe the normalized EM1 scores as a function of historical flood event discharge at the Mississippi River gauging station at Vicksburg. Using this, we assigned each flood deposit to a historical flood event approximating ‘major flood stage’ as defined by the USGS at a nearby gauging station, in stratigraphic order, and within the 2 σ age estimate for the deposit ( Extended Data Fig. 7 ). The requirement for flood deposits to be assigned to historical floods in stratigraphic order eliminated ambiguity in cases in which more than one historical flood fell within a deposit’s 2 σ age estimate. There were no cases for which a flood deposit could not be assigned to a historical flood within the period of instrumental observations ( ad 1897–2015), but there were three cases at FLR ( ad 1944, 1929 and 1920) and two cases at STJ ( ad 1920 and 1913) for which a major historic flood did not leave an identifiable flood deposit. These ‘missing’ flood deposits are rare and occurred during periods of high flood frequency, and they may reflect reduced sediment availability 37 during these events. The sedimentary record reconstructs peak annual discharge at the Vicksburg gauge, not at individual site locations. We developed age–depth models using Bacon v.2.2 38 , a Bayesian age–depth modelling program, informed by multiple independent dating techniques (see Supplementary Information ), including: (i) 137 Cs and 210 Pb activity in desiccated and powdered bulk sediment samples in a Canberra GL2020RS well detector for low-energy germanium gamma radiation, for which we used the constant rate of supply model 39 to estimate the age of a sampled depth; (ii) radiocarbon ( 14 C) dating via accelerator mass spectrometry of a terrestrial plant macrofossil at the National Ocean Sciences Accelerator Mass Spectrometers facility at WHOI, calibrated using the IntCal13 curve embedded in Bacon; (iii) optically stimulated luminescence (OSL) dating with the fast component of silt-sized quartz 40 using a Risø DA-15 B/C luminescence reader at the University of Liverpool, UK; (iv) core tops as the date of collection and, when appropriate, the age of lake formation 20 as the core bottom. Sedimentation rate priors were increased to near-instantaneous rates through thick (>20 cm) flood deposits 17 . Tree-ring records Tree-ring samples from 33 living and 2 dead oak ( Quercus lyrata and Q . macrocarpa ) trees were collected from Big Oak Tree State Park (BOT) in southeast Missouri 21 . One to four core samples were extracted from each tree at or below breast height (about 1.4 m) using a 5-mm-diameter Swedish increment borer. Cross-sections from dead trees were collected as close to the base of the tree as possible. All samples were absolutely cross-dated using the skeleton-plot method of dendrochronology. Tree-ring widths were measured on a stage micrometer to a nominal resolution of 0.001 mm. We crosschecked the accuracy of our visual dating using the computer program COFECHA. We visually determined flood-ring years by examining each tree-ring series for any evidence of flood injury consistent with the anomalous anatomical features caused by flooding as described by previous flood-ring studies 19 . Additional characteristics used in our identification included ‘jumbled ranks’ or ‘additional ranks’ of early wood vessels or zones of ‘extended earlywood’ and disorganized flame parenchyma as well as ‘offset’ early wood ranks 19 . We used the same criteria as ref. 21 to identify flood events (that is, a year in which more than 10% of sampled trees exhibited signs of flood injury) as this threshold encompasses all historic floods that attained major flood stage and occurred during the growing season 21 . Historical climate and palaeoclimate data Historical (late nineteenth century to present) indices of ENSO and AMO 27 were extended back to the sixteenth century with annual palaeoclimate reconstructions of ENSO 28 , 29 , 30 , 31 and AMO 41 . To compare the ENSO series, we identified El Niño events in the historical Niño 3.4 index as periods of five consecutive overlapping 3-month windows at or above +0.5 °C, and as years with anomalies of more than +0.5 °C in the palaeoclimate series. We then derived El Niño event frequencies using a 31-year moving window on each record, and we computed the mean of the historical and all palaeoclimate El Niño frequencies and bootstrapped 2 σ confidence intervals using the boot function in RStudio. For the composite AMO series, we used the detrended historical AMO index 27 back to ad 1871, and then transitioned to a palaeoclimate AMO reconstruction 41 to ad 1572. We sampled this composite AMO index at the median age probability of the 20 palaeofloods that occurred between ad 1500–1800, and used these data to develop a linear model (using the lm function in RStudio) that relates peak discharge from the AMO index; the El Niño frequency timeseries was not a significant predictor of flood magnitudes, presumably because Pacific sea-surface temperatures do not control the inland flux of Gulf of Mexico moisture that triggers high-magnitude discharge events 15 , so only the AMO index was used to statistically estimate flood magnitudes under ‘climate-only’ conditions. The AMO is detrended to remove recent warming of North Atlantic sea surface temperatures, so the ‘climate-only’ estimates of Q 100 do not consider the potential effects of recent greenhouse warming on flood magnitudes—although we note that the inverse relationship between AMO and Mississippi River flood magnitudes implies that warming of North Atlantic sea-surface temperatures would act to suppress flood magnitudes. When evaluating the significance of Pearson correlations between climate and hydrological time-series that exhibited high degrees of serial autocorrelation, we estimated the effective degrees of freedom with the following relation 42 : where N is the number of independent samples, and φ x and φ y are the lag-1 autocorrelation coefficients of time series x and y respectively. Flood hazard attribution The magnitude of Q 100 was estimated both empirically and through statistical modelling. The sedimentary palaeoflood archives record major flood events over periods greater than 100 years, and are suitable for estimating recurrence intervals empirically through the relation: where t r is the recurrence interval (the inverse of t r is the probability that the event magnitude will be exceeded in any one year), n is the number of years in the window being considered, and m is the number of recorded occurrences of the event being considered. The same approach was used to estimate Q 100 in the statistically modelled ‘climate-only’ peak annual discharges derived from palaeoclimate and historical climate records. The instrumental record at the Vicksburg gauge provides a measurement for peak annual discharge in every year, but is relatively short, so the modern Q 100 was estimated statistically by fitting a log Pearson type III distribution to the data set following standard protocols outlined by the United States Interagency Advisory Committee of Water Data 43 for instrumental hydrological data sets. We compared the observed Q 100 baseline ( ad 1500–1800) with the observed and ‘climate-only’ Q 100 estimates for the modern period ( ad 1897–2015) and attributed the proportion of the observed change that was not explained by the ‘climate-only’ estimates to human alterations to the river channel and basin. The modern Q 100 estimated empirically from sedimentary records and the modern Q 100 estimated by fitting a generalized extreme value distribution to the instrumental data both fall within the 1 σ confidence intervals of the modern Q 100 estimated by fitting a log Pearson type III to the instrumental record (see Supplementary Information ), indicating that our findings are robust to different estimations of flood hazard. Data availability statement The datasets generated by this study are available as Supplementary Data . Code availability The R code used to produce the figures in this paper is available from the corresponding author on reasonable request.

A new study has revealed for the first time the last 500-year flood history of the Mississippi River. It shows a dramatic rise in the size and frequency of extreme floods in the past century—mostly due to projects to straighten, channelize, and bound the river with artificial levees. The new research, led by scientists at Woods Hole Oceanographic Institution (WHOI), also uncovered a clear pattern over the centuries linking flooding on the Mississippi with natural fluctuations of Pacific and Atlantic Ocean water temperatures. This newly recovered long-term record provides a historical context that spotlights how more recent river engineering has intensified flooding to unprecedented levels. "The floods that we've had over the last century are bigger than anything we've seen in the last 500 years," said Sam Muñoz, a former postdoctoral scholar at WHOI and the lead author of the study, published April 5, 2018, in the journal Nature. The research shows that in the past 150 years, the magnitude of the 100-year flood—a flood that has a 1 percent chance of occurring in any given year—has increased by about 20 percent. The research team found that about three-quarters of that elevated flood hazard can be attributed to human modifications of the river and its basin. "There's been a longstanding question about the extent to which all the changes we have made to the Mississippi River—one of the most engineered rivers in the world—have altered the probability of really large floods," said Muñoz, now an assistant professor at Northeastern University. To answer that question, the scientists used a technique that WHOI paleoclimatologist Jeff Donnelly pioneered in the coastal ocean to track the history of hurricanes: extracting long cores of bottom sediment from lakes and marshes. "It's sort of the equivalent of sticking a straw in a milkshake, putting your thumb over the top, and pulling it out," said Donnelly, a co-lead on the study. In this case, the "straw" was a 30-foot-long aluminum tube deployed from a small pontoon boat, and the "milkshake" was the mud, sand, and silt at the bottom of three oxbow lakes adjacent to the Mississippi River. During large floods, faster moving water from the river channel stirs up larger-grained sediments and flows into the usually disconnected lakes, carrying sediments and debris along with it. The material from the river is caught in the lakes and eventually sinks. It forms a layer on the bottom that is subsequently buried over time. The layers are telltale clues of past floods. The deeper the cores, the further back in time the scientists can reach. The grain size in the layers provides clues to the size of the floods. The larger the flood, the more energy is generated by the river water, and the larger the grains that get deposited in the lakes, said WHOI geoscientist Liviu Giosan, another lead member of the research team. By analyzing grain size and flood size for known flood events—for example, the Great Mississippi River flood of 1927—Muñoz could estimate the sizes of previously unknown floods represented in the sediment cores. To find out when those floods had occurred, the team used isotopes of lead, cesium, and carbon to date the coarse sediment layers. Zhixiong Shen of Coastal Carolina University used a technique called optically-stimulated luminescence—which determines a material's age by analyzing when it was last exposed to sunlight. Matthew Therrell of the University of Alabama used annual tree rings to reconstruct a detailed record of more recent regional flooding. Combining these methods, the team traced the history of floods back over 500 years—about 350 years further back in time than the oldest written flood records. Next, they compared what they found to records of naturally oscillating climate cycles that affect sea surface temperatures in the Atlantic and Pacific, such as the El Niño-Southern Oscillation (ENSO). They found that the Mississippi's flood cycles corresponded with ocean and climate cycles. In particular, El Niño events bring more storms and rainfall to central North America, which saturates the ground around the Mississippi. One phase of the Atlantic Ocean oscillation brings extreme rainfall over the Mississippi basin. When the two coincide, flooding is more likely. "We're able for the first time to really parse out how the natural variability of the climate system influences flooding, and then how people have modified that," Muñoz said. The sediment data also showed that the natural rhythm of flooding caused by ocean changes was greatly amplified by major federally-funded river engineering projects that began after 1928 to facilitate commercial navigation on the river and to protect communities and cropland from floods. The societal benefits of river engineering should be weighed against the risks that more large floods pose to agriculture, infrastructure, and communities, the scientists said. In addition, big floods sweep more pollutants and fertilizers into the Gulf of Mexico, causing oxygen-depleted "dead zones." Giosan says restoring more natural flood patterns to the river and allowing sediments to flow onto the floodplain during floods would help rebuild the drowning Mississippi delta, whose planned restoration is projected to cost tens of billions of dollars. According to the team, the next steps will be to dig deeper into the river sediments to extend the flood record even further back in time and to apply this new method to understand what drives flooding on other major rivers systems around the world.

nature.com/articles/doi:10.1038/nature26145

Medicine

Genomic analysis underscores need for precision therapies that target pediatric cancer

Pan-cancer genome and transcriptome analyses of 1,699 paediatric leukaemias and solid tumours, Nature (2018). DOI: 10.1038/nature25795 Journal information: Nature

http://dx.doi.org/10.1038/nature25795

https://medicalxpress.com/news/2018-02-genomic-analysis-underscores-precision-therapies.html

Abstract Analysis of molecular aberrations across multiple cancer types, known as pan-cancer analysis, identifies commonalities and differences in key biological processes that are dysregulated in cancer cells from diverse lineages. Pan-cancer analyses have been performed for adult 1 , 2 , 3 , 4 but not paediatric cancers, which commonly occur in developing mesodermic rather than adult epithelial tissues 5 . Here we present a pan-cancer study of somatic alterations, including single nucleotide variants, small insertions or deletions, structural variations, copy number alterations, gene fusions and internal tandem duplications in 1,699 paediatric leukaemias and solid tumours across six histotypes, with whole-genome, whole-exome and transcriptome sequencing data processed under a uniform analytical framework. We report 142 driver genes in paediatric cancers, of which only 45% match those found in adult pan-cancer studies; copy number alterations and structural variants constituted the majority (62%) of events. Eleven genome-wide mutational signatures were identified, including one attributed to ultraviolet-light exposure in eight aneuploid leukaemias. Transcription of the mutant allele was detectable for 34% of protein-coding mutations, and 20% exhibited allele-specific expression. These data provide a comprehensive genomic architecture for paediatric cancers and emphasize the need for paediatric cancer-specific development of precision therapies. Main Paired tumour and normal samples from 1,699 patients with paediatric cancers enrolled in Children’s Oncology Group clinical trials were analysed, including 689 B-lineage acute lymphoblastic leukaemias (B-ALL), 267 T-lineage ALLs (T-ALL), 210 acute myeloid leukaemias (AML), 316 neuroblastomas (NBL), 128 Wilms tumours and 89 osteosarcomas ( Extended Data Fig. 1a–c ). All tumour specimens were obtained at initial diagnosis, and 98.5% of patients were 20 years of age or younger (see Methods, Extended Data Fig. 1d ). The median somatic mutation rate ranged from 0.17 per million bases (Mb) in AML and Wilms tumours to 0.79 in osteosarcomas ( Fig. 1a, b ), lower than the 1–10 per Mb found in common adult cancers 6 . Genome-wide analysis (see Methods) identified 11 mutational signatures (T-1 through T-11; Fig. 1c–e and Supplementary Table 1a–c ). Signatures T-1 through T-9 corresponded to known COSMIC signatures 7 , whereas T-10 and T-11 were novel but enriched in mutations with a low (<0.3) mutant allele fraction (MAF). Figure 1: Somatic mutation rate and signature. Sample size of each histotype is shown in parentheses. Mutation rate using non-coding SNVs from WGS ( a ) and coding SNVs from WGS and WES ( b ). Red line, median. a and b are scaled to the total number of samples with WGS ( n = 651), WGS or WES ( n = 1,639), respectively. c , Mutational signatures identified from WGS and T-ALL WES data and their contribution in each histotype. d , Mutation spectrum of representative samples in each histotype. Hypermutators (three s.d. above mean rate of corresponding histotype) are labelled with an asterisk. e , Mean and s.d. of MAF of each signature in each histotype. PowerPoint slide Full size image Signatures T-1 and T-4 (clock-like endogenous mutational processes) were present in all samples and contributed to large proportions of all mutations in T-ALL (97%), AML (63%), B-ALL (36%), and Wilms tumours (28%). T-2 and T-7 (APOBEC (apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like)) were highly enriched in B-ALLs with ETV6-RUNX1 fusions (15-fold and 9-fold enrichment for T-2 and T-7, respectively; Supplementary Table 1e ). T-3 (homologous recombination deficiency) was present in many childhood cancers, including osteosarcomas (18 of 19), NBLs (59 of 137), Wilms tumours (28 of 81), and B-ALL (47 of 218). T-8 (8-oxoguanine DNA damage) was present in a small proportion (4.5–12%) of AML, B-ALL, osteosarcoma, and Wilms tumour samples. T-8 was also present in many (36%) NBL samples and was associated with age at diagnosis ( Supplementary Table 1d ). T-9 (DNA repair deficiency) was present in two B-ALLs, including one (sample PARJSR) with a somatic MSH6 frameshift mutation. T-2, T-3, T-5, T-7, T-8, and T-9 were enriched among the 39 samples with elevated mutation rates in each histotype ( Fig. 1d ). The T-5 ultraviolet-light (UV)-exposure signature was unexpectedly present in eight B-ALL samples ( Extended Data Fig. 2a–c ). Although its mutation rate in B-ALL, ranging from 0.06 to 0.72 per Mb, was 100-fold lower than the average rate in adult (15.8 per Mb) 8 and paediatric (14.4 per Mb) 9 skin cancer, T-5 exhibited other features associated with UV-related DNA damage. Specifically, CC>TT dinucleotide mutations were enriched 110-fold in these eight B-ALL samples when compared with other samples ( P = 1.07 × 10 −7 ), which is consistent with pyrimidine dimer formation. Moreover, transcriptional strand bias in T-5 indicated that photodimer formation contributed to cytosine damage. The validity of T-5 was further confirmed by analysis of the mutation clonality, cross-platform concordance, genomic distribution and mutation spectrum of each sample (see Methods, Extended Data Fig. 2d–i ), indicating that UV exposure or other mutational processes 10 , 11 may contribute to paediatric leukemogenesis. Notably, all T-5 B-ALLs had aneuploid genomes ( P = 3 × 10 −5 ; two-sided binomial test; cohort frequency 24%) without any oncogenic fusions. By analysing the enrichment 12 , 13 of somatic alterations within each histotype or the pan-cancer cohort (see Methods), we identified 142 significantly mutated driver genes ( Fig. 2a , Supplementary Table 2 , Extended Data Fig. 3a ). Somatic alterations in CDKN2A , which were predominantly deletions, occurred at the highest frequency, affecting 207 of 267 (78%) T-ALLs, 91 of 218 (42%) B-ALLs and 2 of 19 (11%) osteosarcomas ( Extended Data Fig. 3b ). More than half (73) of the driver genes were specific to a single histotype, such as TAL1 for T-ALL and ALK for NBL ( Extended Data Fig. 3c ). Genes that were mutated in both leukaemias and the three solid tumour histotypes accounted for only 17% of driver genes ( Extended Data Fig. 3e ), of which some genes had various types of somatic alteration. For example, STAG2 , a known driver gene for Ewing’s sarcoma 14 and adult AML 15 , exhibited five different types of somatic alteration (single nucleotide variants (SNVs), small insertions or deletions (indels), copy number alterations (CNAs), structural variants and internal tandem duplications (ITDs)) across five histotypes ( Extended Data Fig. 4a–d ). Nine STAG2 variants were predicted to cause protein truncation, including four predicted by aberrant transcripts in RNA sequencing (RNA-seq). Notably, 78 of 142 driver genes ( Supplementary Table 2 ) were not found in adult pan-cancer studies 1 , 2 , 3 , 4 , and 43 ( Fig. 2a and Extended Data Fig. 3a ) were not found in the Cancer Gene Census (v81) 16 . Thirty-seven were absent from both sources, although mutations in cancer have been reported for 29 of these genes, such as NIPBL 17 , 18 , 19 and LEMD3 20 ( Extended Data Fig. 4p, q ). Nearly half (40–50%) of point mutations in leukaemia and NBL driver genes had low MAFs (<0.3), indicative of subclonal mutations contributing to tumorigenesis ( Extended Data Fig. 3f ). Figure 2: Candidate driver genes in paediatric cancer. a , Top 100 recurrently mutated genes: case count for each histotype is shown in the same colour as the legend. Asterisk indicates gene not reported in prior adult pan-cancer analyses. b , Statistically significant pairwise relationships ( P < 0.05; two-sided Fisher’s exact test) for co-occurrence (red) or exclusivity (blue) in each histotype. Gene pairs with Q < 0.05 are coloured dark red (co-occurring) or dark blue (exclusive) to account for false discovery rate. Significance detected only in WGS + WES samples is marked with an asterisk. Shown in parentheses are number of mutated samples. PowerPoint slide Full size image Three hundred and four gene-pairs exhibited statistically significant ( P < 0.05, two-sided Fisher’s exact test; Fig. 2b , Supplementary Table 3 ) co-occurrence (for example, USP7 and TAL1 in T-ALL 21 ) or mutual exclusivity (for example, MYCN and ATRX in NBL 22 ). The analysis also unveiled novel co-occurrences (for example, ETV6 and IKZF1 in AML and CREBBP and EP300 in B-ALL) and mutual exclusivities (for example, SHANK2 and MYCN in NBL and PAX5 and TP53 in B-ALL). Because of reduced power for detecting low-frequency drivers 2 (detection limits were 1% for the entire cohort and 3% for individual histotypes with more than 200 samples; Extended Data Fig. 5 and Methods), we performed subnetwork analyses 3 and variant pathogenicity classification 23 (see Methods), identifying 184 variants in 82 additional genes ( Supplementary Table 4 and Extended Data Fig. 4e, f ). A notable example is the MAP3K4 G1366R mutation, which was found in one T-ALL, two B-ALLs, and one Wilms tumour. MAP3K4 is a member of the MAPK family 24 and structural modelling indicates that the G1366R mutation is likely to cause disruption of normal inhibitory domain binding and kinase dynamics 24 ( Extended Data Fig. 4l, m ). Several genes in which structural variants were found ( PDGFRA , CDK4 , YAP1 , UBTF ) are listed in Extended Data Fig. 4 . While the percentage of tumours with point mutations in driver genes was highly consistent between whole-genome sequencing (WGS) and whole-exome sequencing (WES) ( Fig. 3a ), WGS makes it possible to detect CNAs and structural variants, which are frequently driver events for paediatric cancers. For example, 72% of NBL tumours analysed by WGS had at least one driver variant compared to 26% of those analysed by WES ( Fig. 3a and Extended Data Fig. 4j, k ). Furthermore, integrative analyses of CNAs and structural variants with WGS data revealed chromothripsis (that is, massive rearrangements caused by a single catastrophic event) in 11% of all samples (13 in osteosarcomas, 15 in Wilms tumours, 22 in NBL, 14 in B-ALL, and 6 in AML; Extended Data Fig. 1f ). We next performed pathway analyses (see Methods) on 654 samples analysed by WGS and 264 T-ALL samples analysed by both WES and single nucleotide polymorphism (SNP) arrays, totaling 682 leukaemias and 236 solid tumours. Figure 3: Biological processes with somatic alterations in paediatric cancer. a , Percentage of tumours with at least one driver alteration are shown for each histotype. WGS-analysed tumours may have point mutations (light grey), CNAs or structural variations (SV) (dark grey), or both (black). For T-ALL, CNAs were derived from SNP array. b , Percentage of tumours within each histotype that have somatic alterations in 21 biological pathways; histotype ordering is as in a . The coloured portion of each pathway indicates the percentage of variants in genes that are absent in three TCGA pan-cancer studies. c , Mutation occurrence by histotype in RAS, tyrosine kinase, and PI3K pathways. PowerPoint slide Full size image The 21 biological pathways that were disrupted by driver alterations were either common (for example, cell cycle and epigenetic regulation) or histotype-specific (for example, JAK–STAT, Wnt/β-catenin, and NOTCH signalling) ( Fig. 3b ). More importantly, the genes that were mutated in each pathway differed between histotypes. One example is signalling pathways such as RAS, JAK–STAT and PI3K ( Fig. 3c ). For genes in these pathways, somatic alterations in solid tumours primarily occurred in ALK , NF1 , and PTEN , whereas nearly all mutations in FLT3 , PIK3CA , PIK3R1 , and RAS were found in leukaemias. Although many biological processes are dysregulated in both paediatric and adult cancers 1 , 2 , 4 , the affected genes may be either paediatric-specific (for example, transcription factors and JAK–STAT pathway genes) or common to both (for example, cell cycle genes and epigenetic modifiers). Notably, two novel KRAS isoforms were detected in 70% of leukaemias but rarely in solid tumours ( Extended Data Fig. 6 ). Evaluation of mutant allele expression makes it possible to assess the effects on the gene product and to detect potential epigenetic regulation that may cause allelic imbalance. Here we present this analysis on 6,959 coding mutations with matching WGS and RNA-seq data. RNA-seq expression clusters confirmed the tissue of origin of each histotype ( Extended Data Fig. 7 ). Mutant alleles were expressed for 34% of these mutations, which is consistent with previous reports 25 , 26 , 27 . The expression of mutant alleles is generally associated with corresponding DNA MAF and the expression levels of host genes ( Fig. 4a ); however, exceptions can be found due to X-inactivation, imprinting, nonsense-mediated decay or complex structural re-arrangements ( Extended Data Fig. 8a ). Figure 4: Mutant allele expression. a , Percentage of expressed mutations (red) categorized by DNA MAF ( x axis) and expression level ( y axis). Circle size is proportional to mutation counts. b , Detection of ASE in expressed mutations by comparing DNA and RNA MAF in 443 samples (solid colours, statistically significant (two-sided Fisher’s exact test Q < 0.01 and effect size >0.2); grey, not significant). c , Confirming ASE for WT1 D447N (red arrow in b ) by single-cell sequencing. Presence of subclonal 11p LOH leads to two possible outcomes: the mutant allele is in either non-LOH subclone (top) or LOH subclone (bottom): the former suggests ASE and the latter rejects ASE due to homozygosity. No-ex: WT1 not expressed. PowerPoint slide Full size image Allele-specific expression (ASE) was evaluated for 2,477 somatic point mutations with sufficient read-depth in DNA and RNA-seq (see Methods). Of 486 candidate ASE mutations ( Supplementary Table 5 ), 279 had no detectable expression of the mutant allele, and a comparable DNA MAF distribution was found for truncating and non-truncating mutations ( P = 0.5, two-sided Wilcoxon rank-sum test, Extended Data Fig. 8b ). Of the remaining 207 candidate ASE mutations, 76% of truncating mutations exhibited suppression of the mutant allele ( P = 7 × 10 −5 ; two-sided binomial test), while 87% of hotspot mutations showed the opposite trend of elevated expression ( P = 6 × 10 −5 ; two-sided binomial test; Fig. 4b , Extended Data Fig. 8c ). Excluding hotspot mutations resulted in equal distribution of suppression versus elevation (66 versus 55) for the remaining 121 non-truncating ASE mutations ( P = 0.4; two-sided binomial test). Subclonal loss-of-heterozygosity (LOH) in tumours is a confounding factor for ASE analysis. For example, significant allelic imbalance between tumour DNA and RNA MAF of WT1 D447N in an AML that also harboured a subclonal 11p copy-neutral LOH ( Fig. 4c ) could be attributed to ASE or WT1 expression of a subclone with a double-hit of D447N mutation and 11p LOH. To address this, we performed single-cell DNA sequencing on 63 germline variants on 11p and the somatic point mutations. We confirmed ASE by establishing that WT1 D447N and 11p LOH occurred in separate subclones ( Fig. 4c and Extended Data Fig. 9a, b ). The resulting genotype data projected that one WT1 allele was silenced in a common ancestor and the other was lost in the three descendant subclones by 11p LOH, acquisition of the WT1 D447N mutation, or focal deletion. Two additional AMLs with WT1 D447N also exhibited ASE ( Extended Data Fig. 9c ), implying that loss of WT1 expression by epigenetic silencing or mutations in cis -regulatory elements is not rare in AML. Similarly, single-cell sequencing of an ALL sample confirmed ASE of a JAK2 hotspot mutation ( Extended Data Fig. 9d ). The somatic variants used for this study are available at the National Cancer Institute TARGET Data Matrix and our ProteinPaint 28 portal, which provides an interactive heat map viewer for exploring mutations, genes, and pathways across the six histotypes ( Extended Data Fig. 10 ). The portal also hosts the somatic variants analysed by the companion paediatric pan-cancer study of 961 tumours from 24 histotypes, including 559 central nervous system tumours 29 . We anticipate that these complementary pan-cancer datasets will be an important resource for investigations of functional validation and implementation of clinical genomics for paediatric cancers. Methods Patient samples Specimens were obtained through collaborations with the Children’s Oncology Group (COG) and the Therapeutically Applicable Research to Generate Effective Treatments (TARGET) project. Institutional review boards from the following institutions were responsible for oversight: Ann & Robert H. Lurie Children’s Hospital, Fred Hutchinson Cancer Research Center, National Cancer Institute, St Jude’s Children’s Research Hospital, The Children’s Hospital of Philadelphia, The University of New Mexico, Texas Children’s Hospital, and The Hospital for Sick Children. In our cohort, osteosarcoma has a higher percentage of older patients because the age of onset has a bimodal distribution: the first peak occurs among adolescents and young adults, and the second (associated with Paget disease and with a different underlying biology 30 ) occurs among the elderly. We used an age cutoff of 40 years, which is typical for COG-conducted osteosarcoma trials 31 . Informed consent was obtained from all subjects. Genomic datasets WGS, WES, and RNA-seq data were downloaded from dbGaP with study identifier phs000218 (including phs000463, phs000464, phs000465, phs000467, phs000471, and phs000468). Among the 1,699 cases analysed, 45 B-ALLs 32 , 33 , 197 AMLs 34 , 264 T-ALLs 21 , 240 NBLs 35 and 115 Wilms tumours 36 have been included in published studies of individual histotypes. No statistical methods were used to predetermine sample size. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment. WGS data analysis WGS data were generated with Complete Genomics Inc. (CGI) technology with an average genome-wide coverage of 50× using 31- to 35-bp mate-paired reads, which was powered for detecting mutations in 94% of mappable exonic regions 37 , 38 . Read pairs were mapped to hg19/GRCh37, and somatic SNVs, indels, and structural variants were analysed by comparing paired tumour and normal genomes using the CGI Cancer Sequencing service pipeline version 2 38 , 39 . For each case, we downloaded CGI-generated WGS files for somatic SNVs, indels, structural variants, and CNAs from the TARGET Data Matrix as the starting point for our analysis. Filtering of point mutations Putative somatic point mutations including SNVs and indels were extracted from Mutation Annotation Format files and run through a filter to remove false-positive calls. First, germline variants were filtered by using: (1) NLHBI Exome Sequencing Project ( ); (2) dbSNP (build 132); (3) St Jude/Washington University Paediatric Cancer Genome Project (PCGP); and (4) germline variants present in five or more TARGET CGI WGS cases in each cohort. Second, a variant was removed unless it met the following criteria: (1) at least three reads supported the mutant allele in the tumour; (2) the mutant read count in the tumour was significantly higher than normal ( P < 0.01 by two-sided Fisher’s exact test); and (3) the normal MAF was below 0.05. Finally, a BLAT search 40 was run on the mutant allele with 20-bp flanking to verify unique mapping. A ‘rescue’ pipeline was implemented to avoid over-filtering, by using the customized AnnoVar annotation and pathogenicity identification tool Medal_Ceremony 23 (M.N.E. et al ., unpublished). Pathogenic variants were rescued and further curated with ProteinPaint 28 . This filtering has reduced the original 51 million SNVs and 38 million indels from the CGI files to a set of 711,490 SNVs and 57,700 indels. Of these, 9,397 SNVs and 1,000 indels are in protein coding regions. A comparison with gnomAD database (version r2.0.1; ) indicated that 1.1% of our detected SNVs overlap with SNPs with population frequency greater than 0.1%. Verification of somatic point mutations after filtering is presented in Supplementary Note 1 . Filtering of structural variation CGI structural variants were filtered to remove germline rearrangements, including those found in the Database of Genomic Variants, dbSNP, PCGP, recurrent germline rearrangements from CGI Mutation Annotation Format files, low-confidence somatic calls (>90% reference similarity to the assembled sequence) and those with both structural variant breakpoints falling into gap regions (hg19). Each structural variant was required to have an assembled contig length of at least 10 bp on each breakpoint. CNAs in each tumour were integrated into the structural variant analysis by matching breakpoints within a 5-kb window to rescue rearrangements with CNA support by manual curation. A comparison of CGI structural variants with the known oncogenic re-arrangement in AML and B-ALL is presented in Supplementary Note 2 . Copy number alterations We adapted the CONSERTING algorithm 41 to detect CNAs from CGI WGS data. In brief, germline single nucleotide polymorphisms (SNPs) reported by CGI in Mutation Annotation Format files were extracted, and paralogous variants identified from 625 germline WGS cases generated by PCGP were removed. A coverage profile was constructed using the mean of SNP read counts within a sliding window of 100 bp, and the differences between tumour and normal samples were used as inputs for CONSERTING. To detect LOH, we used SNPs with variant allele fraction (VAF) in normal sample within an interval of (0.4, 0.6) and >15× coverage in tumour and normal samples. Allelic imbalance (AI; |Tumour_VAF-0.5|) was used to detect LOH. Regions with concomitant copy number changes (|log ratio|>0.2) or LOH (AI >0.1) were subjected to manual review. Finally, regions less than 2 Mb were considered focal and included in the GRIN 12 analysis to determine the significance of the somatic alterations. A comparison of our CNA detection with clinical information is provided in Supplementary Note 3 . For osteosarcomas, manual reviews of candidate genes affected by CNA were prioritized for the following three groups owing to the high number of rearrangements caused by chromothripsis in this histotype 42 : (1) gene expression change matched the CNA status; (2) genes with recurrent loss and gain; and (3) published osteosarcoma driver genes 42 . This resulted in the discovery of 13 focal CNAs affecting CCNE1 , CDKN2A , RB1 , PTEN , TUSC7 , and YAP1 in addition to TP53 . WES data analysis Of the 1,131 tumour-normal WES pairs, all but 23 osteosarcoma pairs exhibited the expected binomial distribution of B-allele fraction for germline SNPs. The 23 outlier samples were therefore used neither for the discovery of driver genes nor for calculating mutation rate in coding regions ( Fig. 1b ). They were included only for determining driver mutation prevalence. Somatic SNVs and indels were detected by the Bambino 43 program, followed by postprocessing and manual curation as previously described 44 , 45 . To address 8-oxo-G artefacts 35 , we implemented the D-ToxoG filtering algorithm 46 . Somatic mutation rate The median mutation rate of 651 CGI WGS samples ( Fig. 1a ) was calculated from tier3 non-coding SNVs 47 . This analysis did not include the T-ALL cohort as only three T-ALLs were analysed by the CGI platform. Mutations in coding regions were based on coding SNVs from 1,639 samples analysed by WGS or WES ( Fig. 1b ). Among these, 120 samples were analysed by both WGS and WES, and the union of coding SNVs from WGS and WES were used. Twenty-three osteosarcoma WES samples were excluded from coding mutation analysis owing to quality issues described in ‘WES data analysis’. For osteosarcomas, the mutation rate in coding regions (0.53 per Mb) is lower than in non-coding regions (0.79 per Mb). Nineteen osteosarcoma samples were analysed by both CGI and WES. For these samples, the mutation rate in coding regions derived from either CGI or WES was 0.54 per Mb while the mutation rate in the non-coding regions was 0.79 per Mb, indicating a potential contribution of kataegis 42 in the elevated mutation rate in non-coding regions. Within each histotype, hypermutators were defined as having mutation rates 3 s.d. above the mean (trimming 5% outliers). Mutational signature analysis Mutational catalogues were generated for each sample by using a 96-bin classification ( Supplementary Table 1b ). These were examined for all samples with our previously established methodology 48 to decipher mutational signatures and to quantify their activities in individual samples. The correlation between age of diagnosis and mutational signature activities was computed by using robust regression 49 . We also compared the cosine similarity between original and reconstructed samples and found that samples with more than 100 mutations had cosine similarities greater than 0.85, whereas samples with less than 100 mutations mostly (93.5%) had cosine similarities less than 0.85. To calculate the average MAF values for each signature ( Fig. 1e ), each of the 96 mutation types was assigned to the signature with the highest probability (the same result was obtained if we required the highest probability to be higher than the second (by Δ = 0.05, 0.1, and 0.2; data not shown). This assignment was also used for Extended Data Fig. 2e–i . The two novel signatures, T-10 and T-11, were enriched in low MAF mutations. T-11 was the only signature that was significantly correlated ( r 2 = 0.9) with the presence of multi-nucleotide variations composed of co-occurring SNVs separated by 3 or 4 bp which were not verified by Illumina WES. Therefore, it is likely to be associated with platform-specific sequencing artefacts. For the eight B-ALL cases identified with mutation signatures of UV-light exposure, only 0.96% of the somatic SNVs overlap with SNPs that have population allele frequencies (AFs) >0.1% in the gnomAD database (version r2.0.1; ). The overlap is only 0.22% if using AF >1%. The overlap rate is comparable to the 1.1% observed for non-UV somatic SNVs across the entire cohort (0.27% match if using AF >1%). For each of these eight B-ALL cases, UV- and non-UV-mutations were stratified according to the ploidy of their genomic locations ( Extended Data Fig. 2e–g ; cluster centres estimated using R package mclust). Inter-mutational distances were plotted for comparison of genomic distribution of UV- versus non-UV mutations. Chromosomal ploidy and tumour purity were obtained from TARGET clinical files and prior publications 50 . By adjusting for ploidy and corresponding tumour purity, we calculated expected MAFs for clonal mutations as follows: denoting the tumour purity as π , the expected MAF for clonal mutations was π /(2 − π ) in the 1-copy loss region, π /2 in the diploid region, and π /(2 + π ) in the 1-copy gain (wild-type allele) region. Age-specific incidence rates for childhood ALL reported by the Surveillance, Epidemiology, and End Results (SEER) program show that the rate of incidence in African American children is half of that in white children ( Extended Data Fig. 2h ). While none of our eight patients is African American according to clinical information and genomic imputation, we were not able to test the significance of this observation as 6.6% of the children enrolled in the COG ALL trial are African American. Chromothripsis analysis To detect chromothripsis, we first assessed whether the distribution of structural variant breakpoints in each tumour departed from the null hypothesis of random distribution using Bartlett’s goodness-of-fit test 42 . The distribution of structural variant types (deletion, tandem duplication, head-to-head and tail-to-tail rearrangements) was also evaluated using a goodness-of-fit test for chromosomes with a minimum of five structural variants. Chromosomes with P < 0.01 for Bartlett’s test and with P > 0.01 for the structural variant type test were further reviewed for oscillation between restricted CNA states. Discovery of candidate driver genes For the 654 CGI samples, we ran GRIN 12 analysis with all somatic variants (structural variants, CNAs, SNVs and indels) for both individual histotypes and a combined pan-cancer cohort. Similarly, we combined coding SNVs and indels identified in both WGS and WES for MutSigCV 13 . Putative genes with Q < 0.01 by GRIN or MutSigCV were subjected to additional curation to determine their driver status. Only one candidate gene was included in this analysis for somatic alterations affecting multiple genes such as fusion pairs ( Supplementary Note 4 ). We discovered 142 candidate driver genes by this approach ( Supplementary Table 2 ). Of these, 133 were significant by GRIN analysis (87 genes common to both GRIN and MutSigCV) and nine were significant only by MutSigCV. HotNet2 analysis We applied HotNet2 3 to somatic mutations using interaction data obtained from the HINT, HI2014, and KEGG databases. We reviewed all predicted sub-networks and identified the cohesin complex with three additional genes ( STAG1 , PDS5A and PDS5B ; Extended Data Fig. 4e, f ). Pathway analysis Biological pathways for candidate driver genes were assigned using public pathway databases (KEGG and version 2.0 of the NCI RAS Pathway, ), literature reviews, and biological networks produced by HotNet2. For each pathway in each histotype, a tumour was counted if any genes of that pathway were mutated. The percentage of variants in genes unique to paediatric cancers was calculated by excluding genes reported in the three TCGA pan-cancer studies 1 , 2 , 4 . Mutual exclusivity and co-occurrence of mutations We tested mutual exclusivity and co-occurrence of mutations for the 142 driver genes. For each histotype, we performed this analysis in two separate sample sets: (1) samples with WGS (T-ALL with WES and SNP6), and (2) WGS and WES together (only SNVs and indels considered to avoid detection bias due to platform differences for CNVs and structural variants). For a gene pair A and B (mutated in five or more samples), we performed two-sided Fisher’s exact test according to their mutation status. The R package qvalue 51 was used to control for multiple testing. Although the co-occurrence test is well-powered for most gene pairs, we recognize that the mutual exclusivity test is not powered for most gene pairs, and pairs with P < 0.05 were reported even if Q > 0.05 ( Supplementary Table 3 ). Saturation analysis To study the effect of sample size on detecting driver genes, we performed down-sampling analysis in the pan-cancer cohort and in each histotype 2 , for GRIN and MutSigCV separately. For each combination, we repeated the statistical analysis for a series of subsets of cases from 1 to the total number of samples. The number of genes (of the 142 driver genes) with false discovery rate less than 0.01 were counted for the corresponding subset. Analysis for individual histotypes was limited to those with at least 200 samples (osteosarcomas and Wilms tumours excluded). Somatic variant pathogenicity analysis We implemented a somatic mutation classifier Medal_Ceremony 23 (M.N.E. et al ., unpublished) to identify additional driver variants in genes that did not pass the statistical testing. Pathogenic variants include (1) hotspot SNVs and indel mutations for known cancer genes in any cancer type; (2) pathogenic mutations in ClinVar; (3) truncation mutations in known tumour suppressor genes that were expressed in the cancer histotype; and (4) known recurrent gene fusions, focal deletions, truncations, and amplifications that affect key pathways of any cancer type and that were simultaneously corroborated by an aberrant expression profile. We identified 184 variants in another 82 genes ( Supplementary Table 4 ). BRAF was the most frequently mutated, with nine variants. We also reviewed novel hotspot mutations detected in three or more samples. After removing low-confidence mutations and those without expression, one hotspot was found ( MAP3K4 G1366R, n = 4). Recurrent internal tandem duplication (ITD) was also reviewed for evidence in both DNA and RNA, yielding the discovery of UBTF -ITD in AML. Tumour purity assessment We used regions with copy number loss or copy neutral LOH as well as SNVs (coding and noncoding) from diploid regions to estimate tumour purity. For regions with LOH, a previously described method was used 42 . For SNVs, an unsupervised clustering analysis was performed with the R package mclust. Tumour purity was defined to be two times the highest cluster centre that was <0.5. The maximal CNA and SNV purity was used. We compared our estimates with blast counts for 197 AML and 9 B-ALL samples. Of the 135 tumours with blast count >70% (value ‘many’ in clinical file was mapped as >70%), we identified 127 (94%) with purities >70% (seven of the other eight tumours had purities >50%). An additional 40 tumours were estimated with purities >70%, although their blast count was below 70%. Thirty-one tumours were classified as low purity (<70%) by both our analysis and blast count. KRAS isoform analysis We investigated alternative splicing in KRAS ( Extended Data Fig. 6 ), as differential oncogenic activity of mutant alleles expressed in KRAS 4a or 4b isoforms has been reported previously 52 . We detected splice junction reads connecting exon 3 to one of the two novel acceptor sites in the last intron (53 bp apart). This aberrant splicing is predicted to create two novel isoforms, each incorporating one of the two novel exons (40 bp and 93 bp, respectively) located 2.2 kb downstream of exon 4A ( Extended Data Fig. 6b ). These novel isoforms would form truncated KRAS proteins (154/150 amino acid), each retaining the GTPase domain but losing the hypervariable region that is critical for targeting KRAS to the plasma membrane 53 . One of the two novel isoforms (novel isoform 2) was detected in myeloid cells from three healthy donors (data not shown). Protein products of KRAS isoforms in AML cells were analysed by western blot ( Supplementary Notes 5 , 6 ). RNA-seq data analysis RNA-seq data were mapped with StrongArm 23 , and rearrangements identified with CICERO 23 , followed by manual review. We performed RNA-seq clustering to confirm the tissue of origin and analysed immune infiltration using ESTIMATE 54 and CIBERSORT 55 ( Extended Data Fig. 7 , Supplementary Notes 7 , 8 ). Allele-specific expression (ASE) in RNA-seq CGI and WES allele counts were combined whenever possible. Point mutations were required to have DNA and RNA coverage ≥20×. Variants with |RNA_MAF – DNA_MAF|>0.2 and a false discovery rate of <0.01 (calculated with R package qvalue 51 on two-sided Fisher’s exact test P ) were considered to show ASE. Within-sample analysis was performed to distinguish ASE from potential artefacts caused by normal-in-tumour contamination ( Extended Data Fig. 8d , Supplementary Note 9 , Supplementary Table 5 ). Single-cell targeted re-sequencing One cryopreserved vial each from patients PAPWIU and PABLDZ was thawed using the ThawSTAR system (MedCision) and then diluted in RPMI supplemented with 1% BSA. The cells were then washed five times with C1 DNA-seq wash buffer according to the manufacturer’s instructions (Fluidigm), counted and viability estimated using the LUNA-FL system (Logos Biosystems), then diluted to 300 cells per μl and loaded in a small C1 DNA-seq chip according to the manufacturer’s instructions (Fluidigm, except the suspension buffer to cell ratio was changed from 4:6 to 6:4). The cells also underwent an on-chip LIVE/DEAD viability stain (Thermo Fisher). Each capture site was imaged using a Leica inverted microscope and phase contrast images, as well as fluorescent images with GFP and Y3 filters, were acquired to determine the number of cells captured and the viability of each. The cells then underwent lysis, neutralization, and MDA WGA according to the manufacturer’s instructions (Fluidigm) using the GenomePhiv2 MDA kit (GE Life Sciences). One C1 chip was run per patient. Selected variants and germline SNPs then underwent microfluidic PCR-based targeted resequencing in the bulk sample or genomes amplified from the single cells using the Access Array System as previously described 56 . Target-specific assays were designed using primer3plus ( ) and employed oligos purchased from Integrated DNA Technologies; multiplexing was performed according to guidelines in the Access Array manual (Fluidigm). All samples were loaded with the Access Array loader and underwent PCR cycling in an FC1 system, followed by sample-specific barcoding using standard PCR, all according to the manufacturer’s instructions (Fluidigm). Amplicons were run on the MiSeq using v2 chemistry with 2 × 150-bp paired-end reads (Illumina), using custom sequencing primers, according to the Access Array manual (Fluidigm). Single-cell sequencing data analysis Mapped BAM files for each of the 96 single-cell assays were genotyped for all designed markers. Assays with two captured cells (6 assays for both cases) or assays with fewer than 50% of designed markers with coverage 10× or greater, were dropped, resulting in 48 assays for case PAPWIU ( Supplementary Table 6 ) and 64 assays for case PAPEWB ( Supplementary Table 7 ). The assays were called tumour cells if they had one or more somatic markers with MAFs greater than 0.05. Germline markers with MAFs greater than 0.05 were called positive. The R package pheatmap was used to visualize the single-cell data using hierarchical clustering with ‘binary’ distance and ‘complete’ agglomeration method. Code availability Custom codes are available from the authors upon request. Data availability The somatic variants used for this study are available at the National Cancer Institute TARGET Data Matrix ( ) and our ProteinPaint 28 portal ( ), which also hosts variant data generated by Gröbner et al . 29 ( ).

Researchers have determined that children and adults with cancer usually have different mutated genes driving their disease, which suggests they would likely benefit from different therapies. The finding, from a collaborative study led by St. Jude Children's Research Hospital, underscores the need to develop precision medicines for pediatric cancer. The research, along with a companion study led by scientists at the Hopp Children's Cancer Center at NCT Heidelberg (KiTZ) and the German Cancer Research Center, Heidelberg, appears as an advance online publication today in the scientific journal Nature. The findings from the back-to-back papers provide the most comprehensive analysis yet of the genomic landscape of childhood cancer across multiple subtypes. The results will help guide clinical and laboratory research to improve understanding, diagnosis and treatment of pediatric cancer patients worldwide, including the estimated 15,780 children and adolescents diagnosed annually in the U.S. Rather than focusing on a single tumor type, the St. Jude—led study evaluated data generated by members of the Therapeutic Applicable Research to Generate Effective Treatments (TARGET) program. The study was funded in part by the National Cancer Institute of the National Institutes of Health. The scientists evaluated six cancer subtypes using three different, next-generation sequencing approaches. The researchers compared somatic genetic mutations—genomic changes found in cancer cells but not in normal cells—and their impact on key biological processes in tumor and normal tissues of 1,699 pediatric cancer patients. The approach, called pan-cancer analysis, revealed that only 45 percent of the mutated genes "driving" cancer in children are the same as the genes "driving" cancer in adults. "This shows for the first time that pediatric and adult cancers frequently arise from different genes with different mutations," said corresponding author Jinghui Zhang, Ph.D., chair of the St. Jude Department of Computational Biology. The first authors are Xiaotu Ma, Ph.D., and Yu Liu, Ph.D., both members of her laboratory. "The results really bring home the message that pediatric cancer patients are not small adults and their disease should not be treated as if that were the case," Zhang said. "Better treatments and more sensitive diagnostic tests require understanding the biology driving pediatric cancer. These results provide a better roadmap for researchers working in the laboratory and the clinic." Treatment advances have pushed cure rates for pediatric cancer to better than 80 percent, but cancer remains the leading cause of death by disease in U.S. children ages 1 to 19. Treatments often involve major life-long side effects, including the risk of second cancers. The Genomic Architecture of Pediatric Cancers The analysis included patients enrolled in the clinical trials of the Children's Oncology Group, a clinical research cooperative. The patients had acute lymphoblastic leukemia, both B and T-cell ALL; acute myeloid leukemia; the bone cancer osteosarcoma; the kidney cancer called Wilms tumor; and neuroblastoma, a tumor of the sympathetic nervous system. Unlike a previously published adult pan-cancer analysis, which focused primarily on coding sequence mutations, the St. Jude team analyzed variations in DNA copies and chromosomal rearrangements. The data were from whole genome sequencing of the complete DNA of patients' tumor and normal tissue. Transcriptome sequencing was also performed. That provided a snapshot of gene expression in cells. The analysis also included the first comprehensive evaluation of quantitative and qualitative expression of mutant versus wild-type alleles in multiple cancer subtypes. The sequencing data was provided by the National Cancer Institute's TARGET initiative. TARGET was launched to understand the genetic changes that underlie pediatric cancer. The intention of the program was to use the data by the research community to improve the precision oncology treatment of pediatric cancers. For this study, Zhang and her colleagues processed the sequencing data using a uniform analytic framework. "Looking across tumor types helps to identify really important mutations in pediatric cancer and understand the pathways involved," she said. Genomic Alterations Revealed For example, the analysis identified unexpected patterns of mutations, or mutational signatures, in eight of the 689 patients with B-ALL. The patients' DNA had a mutational signature consistent with exposure to ultraviolet light, which was previously observed exclusively in skin cancers. The patients also shared another genomic abnormality—too many or too few chromosomes. "This finding suggests that exposure to ultraviolet light may be a previously unrecognized environmental factor that increases the risk for developing leukemia in some children," said co-author Ludmil Alexandrov, Ph.D., of the University of California, San Diego. In addition, the analysis revealed that rather than point mutations in DNA, a majority (62 percent) of the mutations driving pediatric cancer were copy-number alterations and structural variations. Copy-number alterations leave patients with too many or too few copies of particular genes. Structural variations involve gene rearrangements. "This shows that as genomics moves into the clinic, diagnostic testing for pediatric patients must include copy-number changes and structural variations," Zhang said. Data Being Shared Zhang and several other St. Jude scientists are also co-authors of a study led by investigators at the German Cancer Research Center. That study included whole genome and whole exome data from 914 pediatric cancer patients with blood, central nervous system and solid tumors. The sequencing data came from the St. Jude Children's Research Hospital—Washington University Pediatric Cancer Genome Project and other sources. The Pediatric Cancer Genome Project, which began in 2010, sequenced the genomes of more than 800 children and adolescents with some of the least-understood and most difficult-to-treat cancers. Researchers worldwide can explore somatic mutation data used for both pediatric pan-cancer studies through the St. Jude interactive ProteinPaint portal. The information, including the genes, mutations and cellular pathways underlying a variety of pediatric cancers, is an unmatched resource for pediatric cancer research going forward. The TARGET initiative somatic variants used in the analysis are available at the NCI TARTGET Data Matrix, and the sequencing data will be available at the NCI's Genomic Data Commons.

10.1038/nature25795

Medicine

Researchers discover possible drug targets for common non-Hodgkin's lymphoma

Bojie Dai, X. Frank Zhao, Krystyna Mazan-Mamczarz, Patrick Hagner, Sharon Corl, El Mustapha Bahassi, Song Lu, Peter J. Stambrook, Paul Shapiro and Ronald B. Gartenhaus. "Functional and molecular interactions between ERK and CHK2 in diffuse large B-cell lymphoma." Published online in Nature Communications on July 19, 2011. dx.doi.org/10.1038/ncomms1404

http://dx.doi.org/10.1038/ncomms1404

https://medicalxpress.com/news/2011-07-drug-common-non-hodgkin-lymphoma.html

Abstract Distinct oncogenic signalling cascades have been associated with non-Hodgkin lymphoma. ERK1/2 signalling elicits both transcriptional and post-transcriptional effects through phosphorylation of numerous substrates. Here we report a novel molecular relationship between ERK1/2 and CHK2, a protein kinase that is a key mediator of the DNA damage checkpoint that responds to DNA double-strand breaks. Our studies are the first to demonstrate the co-localization and overexpression of ERK1 /2 and CHK2 in diffuse large B-cell lymphoma (DLBCL). The physical interaction between ERK and CHK2 was highly dependent on phosphorylated Thr 68 of CHK2. Concurrent administration of an ERK inhibitor enhances the antitumour activity of CHK2 inhibition in both a human DLBCL xenograft model as well as primary human DLBCL cells. Our data suggest a functional interaction between ERK and CHK2 and support the potential combined therapeutic targeting of ERK and CHK2 in human DLBCL. Introduction Recent evidence implicates DNA repair and the genome integrity checkpoints as major barriers whose deregulation is largely responsible for the enhanced genetic instability of cancer cells 1 . The CHK2 kinase is a key mediator of the DNA damage checkpoint that responds to DNA double-strand breaks (DSBs). DNA DSBs induce the activation of ATM (ataxia telangiectasia mutated), which subsequently phosphorylates CHK2 on threonine 68 (ref. 2 ). Phosphorylation of CHK2 induces CHK2 dimerization, which is required for CHK2 activity. Once dimerized, CHK2 is further activated by autophosphorylation in trans at residues 383, 387, 546 and 516 (ref. 2 ). CHK2 phosphorylates a range of proteins involved in cell cycle control and apoptosis including cdc25A, cdc25C, Mdmx, p53, BRCA1, PML, E2F1, and phosphatase 2A (ref. 2 ). CHK2 also mediates stabilization of the FoxM1 transcription factor to stimulate expression of DNA repair genes 3 . Cells derived from CHK2-deficient mice exhibit defects in their ability to delay entry into S phase, sustain a G2 cell cycle arrest, and undergo apoptosis in response to DNA damage 4 . It is newly reported that CHK2, independent of p53 and DNA damage, is required for proper progression of mitosis, and for the maintenance of chromosomal stability in human somatic cells 5 . CHK2 can also protect genome integrity by promoting apoptosis through interacting with a number of other substrates. Inhibition of CHK2 by transfection of a dominant-negative CHK2 mutant or a chemical inhibitor, debromohymenialdesine, stabilizes centrosomes, maintains high cyclin B1 levels, and allows for a prolonged activation of Cdk1 (ref. 6 ). Under these conditions, multinuclear HeLa syncytia do not arrest at the G2/M boundary and rather enter mitosis and subsequently die during the metaphase of the cell cycle 6 . Therefore, inhibition of CHK2 can sensitize proliferating cells to chemotherapy-induced apoptosis. The first indication of the role of CHK2 in cancer came from a study that reported the presence of germline mutations in CHK2 in families with Li-Fraumeni Syndrome 7 . It has been subsequently shown that defects of CHK2 occur in subsets of diverse sporadic malignancies and predispose to several types of hereditary carcinomas 8 . However, there is increasing evidence that CHK2 is a cancer susceptibility gene, but not a tumour suppressor gene in the classical sense 9 , 10 . CHK2 is aberrantly and constitutively activated in invasive urinary bladder carcinomas, and the putative proapoptotic checkpoint signalling can be disabled by inactivation of CHK2 and/or p53 tumour suppressors in subsets of these tumours 8 . More than 50% of CHK2 was phosphorylated at Thr68 in surgically resected lung and breast tumour specimens from otherwise untreated patients 11 . Targeting of CHK2 with small interfering RNA prevents survivin release from the mitochondria and enhanced apoptosis following induction of DNA damage by ionizing radiation (IR) or doxorubicin and inhibits the growth of resistant in vivo tumours. Expression of a dominant negative CHK2 potentiates cytotoxicity in HCT116 colon carcinoma cells to doxorubicin 9 . These findings suggested that activated CHK2 manifests both tumour-suppressor functions as well as the capacity to promote tumour and cell survival 9 , 12 . Relatively little is known about the contribution of CHK2 or the efficacy of CHK2 inhibitor in diffuse large B-cell lymphoma (DLBCL). The extracellular signal-regulated kinases 1 and 2 (ERK1/2) regulate cell proliferation and survival. Deregulation of ERK signalling is associated with genomic instability and cancer 13 . In earlier work, we showed that ERK inhibition induced the apoptosis of human DLBCL cells and has marked antitumour activity in human DLBCL xenograft models, implying the potential role of targeting ERK in the therapy of DLBCL 14 . It was shown that disruption of ERK1/2 activation by pharmacologic MEK1/2 inhibitors results in a dramatic increase in apoptosis of hematopoietic malignant cells 15 , 16 . Considering that there is little data on CHK2 and ERK signalling in lymphoid malignancies, in this study, we explored the molecular mechanisms underlying the functional interaction between ERK2 and CHK2. Furthermore, we investigated the potential therapeutic efficacy of combining ERK and CHK2 inhibitors in a pre-clinical model of DLBCL. We report here that the interaction between ERK and CHK2 was highly dependent on phosphorylated Thr 68 of CHK2 and that concurrent administration of an ERK inhibitor enhances the antitumour activity of CHK2 inhibition in human DLBCL. Results ERK1/2 and CHK2 are highly expressed in human DLBCL To examine the expression of both ERK1/2 and CHK2 in human DLBCL, we performed immunohistochemistry analysis to evaluate the ERK1/2 and CHK2 protein levels on 50 primary DLBCL specimens and 24 reactive lymph nodes. The expression of ERK1/2 and CHK2 was examined in the adjacent slides from the same sample. The DLBCL was further classified into germinal centre B-cell (GCB) and activated B-cell (ABC) subtypes, according to well-accepted immunohistochemistry-based algorithms 17 , 18 . Based on the criteria defined in the Methods, the specimens were divided into three staining groups: negative, low and high. As summarized in Table 1 , ERK1/2 levels are markedly elevated in 98% of all the DLBCL and 100% of the GCB subtype DLBCL samples as compared with the normal GCB cells of the reactive lymph nodes (21%). CHK2 is also elevated in 100% of the GCB DLBCL samples as compared with the normal GCB cells of reactive lymph nodes (46%). Interestingly, our correlation analysis revealed that ERK1/2 expression levels in human DLBCL samples are strongly correlated with CHK2 protein levels (Pearson correlation coefficient ρ =0.737, P <0.0001). Representative sections of human DLBCL arrays are shown in Figure 1a . These data indicate that increased coexpression of ERK1/2 and CHK2 may be associated with the development of DLBCL. These data were further confirmed by western blot analysis in normal B cells and DLBCL cells ( Fig. 1b ). Together, ERK1/2 and CHK2 protein levels appear to be elevated in the vast majority of DLBCL, supporting their value as potential therapeutic targets. Table 1 Immunohistological analysis of ERK1/2 and CHK2 expression in DLBCL and normal GCB cells. Full size table Figure 1: ERK1/2 and CHK2 protein levels are elevated in human DLBCLs. ( a ) Representative fields of reactive germinal centre B-cell centroblasts and DLBCL tissue arrays immunohistochemically stained with anti-CD20, anti-CHK2 and anti-ERK1/2 antibodies as indicated. Low (original magnification, X20) and high (original magnification, X400) power views of CD20, CHK2, and ERK1/2 expression in reactive germinal centre B-cell centroblasts and DLBCL cells (immunoperoxidase stains). H&E staining was used for morphologic examination. ( b ) The correlation between ERK1/2 and CHK2 expression was shown in normal B cells and DLBCL cells. Five cases of normal B cells from different donors and seven DLBCL cells were lysed and subjected to the western blot analysis with the indicated antibodies. 1–5, Normal B cells from different donors. 6, OCI-LY3; 7,SUDHL4; 8, SUDHL6; 9, Farage; 10, Pfeiffer; 11, Toledo; 12, Karpars-422. Full size image Inhibition of ERK and CHK2 increases DLBCL cells apoptosis It has been reported that inhibition of CHK2 expression attenuated DNA damage-induced cell cycle checkpoints and enhanced apoptotic activity in HEK293 cells 10 , 19 . We asked whether inhibition of CHK2 would induce apoptosis in DLBCL cells. The CHK2 Inhibitor II shows 1,000-fold greater selectivity for the CHK2 serine/threonine kinase than for the Cdk1/B and CK1 kinases and was first discovered to be a potent, selective small molecule showing radioprotection towards human T cells 20 . CHK1 and CHK2 are structurally unrelated yet functionally belong to the checkpoint kinase family 1 . Therefore, we compared CHK1 phosphorylation with CHK2 phosphorylation after CHK2 Inhibitor II treatment. Phosphorylation of CHK2 at Thr68 by ATM functions as a surrogate marker for CHK2 activation 11 , 21 . We found that different doses of CHK2 inhibitor II specifically inhibit CHK2 phosphorylation at Thr68 at different time course, but not CHK1 phosphorylation ( Fig. 2a,b ). These data indicated that the CHK2 inhibitor II is specific for the inhibition of CHK2 activity. Exposure of SUDHL4, SUDHL6 and Farage cells to 5 μM CHK2 inhibitor II resulted in the apoptosis of less than 30% cells ( Fig. 2c ). We previously demonstrated that less than 30% DLBCL cells underwent apoptosis when exposed to 20 μM of a novel ERK inhibitor 14 . Because DLBCL has two major subtypes (ABC versus GC), we subsequently examined the efficacy of combining CHK2 inhibitor II with ERK inhibitor in both ABC (for example, OCI-LY3) and GC-type (Farage, SUDHL4 and SUDHL6) cells. Treatment with combination of CHK2 inhibitor II and ERK inhibitor resulted in substantially more apoptosis compared with treatment of either drug alone ( Fig. 2c ). Similar result was obtained in OCI-LY3 cells ( Supplementary Fig. S1 ). DNA DSBs are demonstrated by examining the phosphorylation of histone H2AX on Ser139 residues (designated γH2AX) 15 . Exposure to either the ERK inhibitor or CHK2 inhibitor II alone induced a modest increase in H2AX phosphorylation, whereas co-administration of both inhibitors dramatically increased H2AX phosphorylation. Apoptosis by the combination of these two inhibitors was also demonstrated by examining PARP cleavage ( Fig. 2d ). Figure 2: Inhibition of ERK and CHK2 increases cell apoptosis. ( a ) The specificity of CHK2 inhibitor II. Farage and SUDHL6 cells were treated with different doses of CHK2 inhibitor II in RPMI 1640 complete medium for 4 h. The levels of pCHK1, CHK1, pCHK2 and CHK2 in the total cell lysates were monitored by immunoblotting with the indicated antibodies. ( b ) Farage and SUDHL6 cells were treated with 5 μM CHK2 inhibitor II in RPMI 1640 complete medium for different time course. The levels of pCHK1, CHK1, pCHK2 and CHK2 were examined by immunoblotting. ( c ) SUDHL4, SUDHL6 or Farage cells were exposed to 5 μM CHK2 inhibitor II or 20 μM ERK inhibitor alone or in combination for 72 h after which the percentage of apoptotic cells was determined by Annexin V analysis as described in Methods. Data show means and s.d. from at least three separate experiments. P values were calculated using a paired Student t -test. **Indicates values significantly greater ( P <0.01) than treatment with either agent alone. ( d ) Farage or SUDHL6 cells were treated as described in c , after which whole cells were lysed and subjected to western blot analysis to monitor expression of γH2AX and cleavage of PARP. CF, cleavage fragments. Full size image ERK and CHK2 physically interact with each other in DLBCLs To better understand the relationship between ERK and CHK2 in vivo , we conducted double-immunofluorescence analysis of ERK1/2 and CHK2 in human primary DLBCLs to detect their co-localization. As shown in Figure 3a , CHK2 was localized to both the cytoplasmic and nuclear factions of DLBCLs, and ERK was predominately localized in the cytoplasm of DLBCLs. In addition, endogenous ERK and CHK2 were predominately co-localized in the cytoplasm of human DLBCLs. We next examined whether ERK physically associated with CHK2. We were able to detect the reciprocal co-immunoprecipitation of endogenous ERK1/2 with CHK2 in Farage, SUDHL4 and SUDHL6 cells ( Fig. 3b ). Consistent with their co-localization, the endogenous ERK and CHK2 proteins co-immunoprecipitated predominately from the cytoplasm fraction of SUDHL6, Farage and OCI-LY3 cells ( Fig. 3c ; Supplementary Fig. S2a ). To further confirm this interaction and to demonstrate a direct physical interaction, in vitro glutathione S -transferase (GST) fusion protein pull-down assays were performed. We showed that the GST–ERK2 fusion protein was able to pull down CHK2 ( Fig. 3d ). Taken together, these data establish the physical interaction between ERK1/2 and CHK2. Figure 3: ERK1/2 and CHK2 physically interact with each other. ( a ) Co-localization of ERK1/2 and CHK2 in human lymphoma tissues. Double immunofluorescent staining for CHK2 (green) and ERK (red) at X600 magnification indicates that there is strong overlapping in human lymphoma tissues. DAPI, 4′,6-diamidino-2-phenylindole; Merge, merged image of CHK2, ERK1/2, and DAPI. ( b ) ERK1/2 associates with CHK2 in cells. Farage, SUDHL4 and SUDHL6 cells were subjected to immunoprecipitation using either anti–ERK1/2 or anti-CHK2 antibody, followed by immunoblotting with indicated antibodies. TCL, total cell lysate. F, Farage; S4, SUDHL4; S6, SUDHL6. ( c ) Equal amounts of cytoplasmic (Cy) and nuclear fraction (Nu) were immunoprecipitated with anti-CHK2 antibody followed by immunoblotting with anti-ERK1/2 or anti-CHK2 antibody. Tubulin and HDAC1 were used as markers of the cytoplasmic and nuclear fractions respectively. ( d ) ERK2 associates with CHK2 in vitro . GST-fused ERK2 was immobilized onto Glutathione S -transferase beads and then incubated with Farage or SUDHL6 cell lysates. Immunoblotting was followed using anti-CHK2 antibody. Full size image Thr 68 of CHK2 mediates the interaction between ERK and CHK2 Because CHK2 inhibitor II modestly induced the apoptosis of DLBCL cells, we asked whether CHK2 inhibition could affect the interaction between ERK and CHK2. Different doses of CHK2 inhibitor II dramatically diminished the physical interaction between ERK and CHK2 ( Fig. 4a ). Because the CHK2 inhibitor II specifically inhibits CHK2 phosphorylation at Thr68, this indicated that CHK2 phosphorylation at Thr68 is involved in the physical interaction between ERK and CHK2. We examined the CHK2 Thr68 to Ala (T68A) mutant that was established previously 22 . Due to high levels of CHK2 protein in DLBCL cells and the technical difficulty in transfecting SUDHL4 and SUDHL6 cells using liposomal or electroporation methods, we expressed CHK2 wild type (WT) and its mutant T68A in both HEK293 and HeLa cells. Wild-type CHK2 co-immunoprecipitated with ERK2, but the mutant CHK2 T68A did not ( Fig. 4b ). These results are further corroborated by GST pull-down experiments. Figure 4c shows that the GST–ERK2 protein can pull down CHK2 WT but not the T68A mutant. Mimicking T68 phosphorylation with a negatively charged amino-acid Asp (T68D) promoted ERK interaction with CHK2 ( Fig. 4d ). In addition, IR, which enhances CHK2 T68 phosphorylation, also increases their interaction ( Fig. 4e ). Together, these data demonstrated that Thr68 phosphorylation is required for the interaction between ERK and CHK2. Figure 4: CHK2 Thr68 is required for the physical interaction between ERK1/2 and CHK2. ( a ) Farage and SUDHL6 cells were treated with CHK2 inhibitor II in RPMI 1640 complete medium for 4 h. The cell lysates were subjected to immunoprecipitation with anti–ERK1/2 or anti-CHK2 antibody followed by immunoblotting with anti-CHK2 or anti-ERK1/2 antibody. The expression of pCHK2, CHK2 or ERK1/2 was monitored by immunoblotting of the total cell lysates. Actin was used as a loading control. ( b ) HEK 293 and HeLa cells were transiently transfected with the expression plasmids as indicated. At 24 h post-transfection, immunoprecipitation was performed using anti-ERK1/2 antibody followed by immunoblotting with anti-Flag or anti-ERK1/2 antibody. ( c ) HEK 293 and HeLa cells were transiently transfected with the Flag (MYC)-tagged CHK2 expression plasmids as indicated. At 24 h post-transfection, Cell lysates were incubated with GST–ERK2 bound to glutathione-Sepharose 4B beads at 4 °C for 1 h. The beads were washed intensively, and the bound proteins were detected by immunoblotting with an anti-Myc antibody. The GST fusion proteins were detected with anti-GST. ( d ) HeLa cells were transiently transfected with GFP–CHK2–WT or GFP–CHK2–T68D expression plasmids as indicated. At 24 h post-transfection, immunoprecipitation was performed using anti-ERK1/2 antibody followed by immunoblotting with anti-CHK2 or anti-ERK1/2 antibody. The expression of pCHK2,pERK1/2,CHK2, ERK1/2 or GFP was monitored by immunoblotting of the total cell lysates. Actin was used as a loading control. 1,2,3,4, four different clones. ( e ) Farage and SUDHL6 cells were treated or untreated with 1 Gy of IR) and collected 30 min or 4 h later for immunoprecipitation and western blotting analysis. Full size image CHK2 negatively regulates ERK activity The interaction of CHK2 and ERK suggests that they may regulate each other and ERK may be involved in CHK2-mediated DNA damage responses. This is also supported by our observation that ERK inhibitor potentiates CHK2 inhibitor-induced cell apoptosis. However, it is still unclear how these two proteins regulate each other and why co-treatment with ERK and CHK2 inhibitors increased cell apoptosis. To begin to address these questions, we next assessed whether inhibition of CHK2 regulates ERK. The CHK2 inhibitor II decreases CHK2 phosphorylation at Thr68 but increases ERK phosphorylation in SUDHL6, Farage and OCI-LY3cells ( Fig. 5a ; Supplementary Fig. S2b ). Furthermore, when CHK2 was genetically modified to inhibit Thr68 phosphorylation, a similar result was observed. Figure 5b showed that mutant T68A activated ERK but CHK2 WT failed to do so. In contrast, the phosphorylation-site mimic mutant T68D decreased ERK phosphorylation compared with CHK2 WT ( Fig. 4d ). These data indicated that ERK activity is regulated by CHK2 and the disruption of the physical interaction between ERK and CHK2 is necessary for the activation of ERK. Knockdown of CHK2 by its specific siRNA decreased CHK2 protein level and its phosphorylation; however, it also enhanced ERK phosphorylation ( Fig. 5c ). We next examined how ERK phosphorylation was affected when CHK2 phosphorylation was increased. It has been previously reported that etoposide is a potent activator of the DNA damage-signalling pathway 23 , 24 . Figure 5d shows that when CHK2 phosphorylation was increased after etoposide exposure, ERK phosphorylation was decreased. Knocking down of CHK2 attenuates the regulation of ERK by etoposide ( Fig. 5e ). In addition, while CHK2 phosphorylation at Thr68 was enhanced by IR, ERK phosphorylation was also decreased ( Fig. 4e ). Figure 5: Inhibition of CHK2 activity increases ERK1/2 phosphorylation. ( a ) SUDHL6 and Farage cells were treated with CHK2 inhibitor II in RPMI 1640 complete medium for 4 h. The levels of pCHK2, CHK2, pERK1/2 or ERK1/2 in the total cell lysates were monitored by immunoblotting with the indicated antibodies. (M, media; D, DMSO). ( b ) HeLa and HEK 293 cells were transiently transfected with the expression plasmids as indicated. At 24 h post-transfection, cell lysates were subjected to immunoblotting with anti-pERK1/2, anti-ERK1/2, anti-pCHK2, anti-CHK2 or anti-Myc antibody. ( c ) Knocking down of CHK2 increases ERK1/2 phosphorylation. Farage cells were transfected with control siRNA or CHK2 siRNA. At 72 h after transfection, the expression of pCHK2, CHK2, pERK1/2, ERK1/2, or Actin was monitored by immunoblotting. ( d ) Farage and SUDHL6 cells were treated with 5 μM Etoposide in RPMI 1640 complete medium for 15 or 30 min. Cell lysates were subjected to immunoblotting with anti-pERK1/2, anti-ERK1/2, anti-pCHK2 or anti-CHK2 antibody. ( e ) Knocking down of CHK2 attenuates the regulation of pERK by etoposide. Farage cells were transfected with control siRNA or CHK2 siRNA. At 48 h after transfection, Farage cells were treated with 5 μM Etoposide in RPMI 1640 complete medium for 30 min. Cell lysates were subjected to immunblotting with anti-pERK1/2, anti-ERK1/2, anti-pCHK2 or anti-CHK2 antibody. M, media; D, DMSO. ( f ) CHK2 directly inhibits ERK2-mediated phosphorylation through a kinase-independent mechanism. Purified wild-type (WT) or inactive (KD) GST–CHK2 and MBP were incubated in the absence or presence of ERK2, and protein phosphorylation was determined. Top panels: autoradiography of CHK2 and ERK2 autophosphorylation and phosphorylation of MBP. CHK2 does not inhibit ERK2 autophosphorylation but does inhibit ERK2-mediated phosphorylation of MBP. Bottom panels: Coomassie stain of CHK2, ERK2, and MBP in the kinase assays. GST was used as a control. Full size image We next investigated whether CHK2 directly inhibits ERK activity in vitro . GST–CHK2–KD is defective for kinase activity because of substitution of the catalytically essential Asp347 residue with Ala 25 . This mutation prevents autophosphorylation of CHK2 kinase activity 25 . In Figure 5f , in vitro kinase assays showed ERK2 and CHK2 WT can autophosphorylate and phosphorylate the generic kinase substrate, myelin basic protein (MBP). Interestingly, CHK2–KD was able to inhibit ERK2's ability to phosphorylate MBP but had no apparent effect on ERK2 autophosphorylation, indicating that ERK2 catalytic activity was intact. These findings indicate that CHK2 can directly inhibit ERK2's ability to phosphorylate downstream substrates by an activity-independent mechanism. Together, these findings establish that CHK2 negatively regulates ERK and suggest that inhibition of CHK2 is most effective in killing DLBCL cells if ERK functions are also inhibited. ERK inhibition enhances the activity of CHK2 inhibitors in vivo Next, we examined whether CHK2 inhibition cooperated with ERK inhibition in the treatment of DLBCL tumours in vivo . We established SUDHL6 DLBCL xenografts in 28 severe combined immunodeficient (SCID) mice (7 mice per group). Once the tumour reached the size of 60–163 mm 3 , we treated the mice every other day intraperitoneally with either vehicle, ERK inhibitor (5 mg kg −1 ), CHK2 inhibitor II (1 mg kg −1 ), or both ERK inhibitor and CHK2 inhibitor II for 20 days. At these doses, no lethal toxicity, significant weight loss or any gross abnormalities were observed among treated animals (data not shown). No evidence of tissue damage was detected by microscopic examination of mouse organs ( Supplementary Fig. S3 ). Both 5 mg kg −1 ERK inhibitor and 1 mg kg −1 CHK2 inhibitor II modestly inhibited tumour growth but combined treatment with ERK inhibitor and CHK2 inhibitor II resulted in a statistically significant suppression of tumour growth ( Fig. 6a,b ). Moreover, western blot analysis revealed that combined treatment with the ERK inhibitor and CHK2 inhibitors caused a striking increase in H2AX phosphorylation and PARP cleavage ( Fig. 6c ). This further confirmed that co-administration of the CHK2 inhibitor markedly potentiates DNA damage in DLBCL cells exposed to ERK inhibitor and this event precedes induction of extensive apoptosis. Immunohistochemical analysis of tumour sections showed that co-administration of the ERK and CHK2 inhibitors markedly reduced expression of the cell proliferation marker Ki67 whereas each inhibitor alone had slight impact ( Fig. 6d ). In addition, TdT-mediated dUTP nick end labelling (TUNEL) assay showed that the combined treatment with ERK inhibitor and CHK2 inhibitor II resulted in greater induction of apoptosis compared with either inhibitor alone ( Fig. 6d ). Figure 6: Cotreatment with ERK inhibitor and CHK2 inhibitor II results in decreased tumour growth in a xenograft model. ( a ) 28 mice bearing s.c. tumours of SUDHL6 cells were divided into 4 groups (vehicle versus ERK inhibitor versus CHK2 inhibitor II versus combination), with 7 mice each. Representative mice were photographed before death to visually demonstrate tumour growth. ( b ) The average tumour volume of each group ( n =7) with s.d. is shown as a function of time. P values were calculated using a Student t -test and statistical significance was determined when a P value less than 0.05 was attained. * P <0.05; ** P <0.01 versus single treatments. ( c ) At the moment of mice death, xenograft tumours were excised, and the pooled protein extracts randomly from three mice of each group was subjected to western blot analysis using the indicated antibodies. ( d ) Tumour tissue sections were subjected to immunochemistry for Ki67 expression (original magnification, X400) and TUNEL staining (original magnification, X600). Full size image Inhibition of ERK and CHK2 in human DLBCLs cells Because combined treatment of the CHK2 and ERK inhibitors is a candidate treatment strategy for human lymphomas, we wished to examine the effects of this combination on the apoptosis and viability of primary DLBCLs from human patients. Single cell suspensions were generated from two confirmed DLBCL lymph node biopsy specimens and exposed to either control (DMSO), ERK inhibitor, CHK2 inhibitor II or the combination of both inhibitors for 48 h. Co-treatment of the ERK and CHK2 inhibitors enhanced apoptosis of human primary DLBCL cells as compared with either inhibitor alone ( Fig. 7a ). Four cases of normal B cells are unresponsive to the inhibitor alone and the combination ( Supplementary Fig. S4a ). To further exclude the toxicity of the inhibitors in different cell phenotypes, immortalized untransformed fibroblasts (ID:GM04390, skin) were examined and insensitive to the combination of the inhibitors ( Supplementary Fig. S4b ). In addition, combined treatment with the ERK inhibitor and CHK2 inhibitor II dramatically reduced cells viability whereas the inhibitor administered individually had less effect ( Fig. 7b ). Therefore, these results are consistent with our DLBCL cell line and xenograft data and further support the concept that inhibition of ERK functions potentiates the antitumour activity of CHK2 inhibition in human DLBCLs. Figure 7: ERK and CHK2 inhibitors enhance apoptosis induction in primary human DLBCL cells. ( a ) Single-cell suspensions were obtained from lymph node biopsies of patients diagnosed with DLBCL and were treated with DMSO, ERK inhibitor and CHK2 inhibitor II alone or in combination for 48 h after which the percentage of apoptotic cells was determined by Annexin V analysis as described in Methods. Flow cytometric profile was showed in the left panel. Pt 1, patient one; Pt 2, patient two. ( b ) Single-cell suspensions were treated as described in the ( a ). After the treatment, the percentages of non-viable cells were determined by trypan blue dye uptake in a hemocytometer. Data represents mean and s.d. from three separate experiments. P values were calculated using a Student t -test. * P <0.05; ** P <0.01. Full size image Discussion Our study addresses the therapeutic targeting and functional interaction between ERK and CHK2 in the setting of DLBCL. Our immunohistochemical analysis on tissue microarrays showed that both ERK and CHK2 protein levels were elevated in the majority of human DLBCL samples as compared with the normal GCB cells of reactive lymph nodes. This supports their value as potential therapeutic targets. The potential utility of ERK targeted therapy is supported by our previous data that treatment with the ERK inhibitor has marked antitumour activity in human DLBCL xenograft models 14 . Previous reports using western blotting have shown that CHK2 protein expression level is high in most primary diffuse large B-cell lymphomas (41/43), and absent or very low in six samples of normal peripheral blood lymphocytes 26 , which is consistent with our results. However, their western blot data also showed that CHK2 protein levels in DLBCL samples was similar to that in 6 reactive lymphoid tissues 26 . This discrepancy may be explained by the presence of other cell types such as T cells, follicular dendritic cells and stromal cells in addition to reactive B cells in the reactive germinal centre that was examined by western blot analysis. Furthermore, we were able to demonstrate using IHC that CHK2 levels of GC subtype of DLBCL are also elevated as compared with the normal GC cells. The mechanism for elevated CHK2 protein levels in DLBCL is still poorly described. Further experiments are necessary to investigate CHK2 modification and the CHK2 signalling network in DLBCL. It has been previously reported that Raf/MEK/ERK signalling was linked to DNA damage response and DNA repair through an ATM-dependent process 13 . This indicates that ERK is related with CHK2 signalling. We for the first time demonstrate the physical interaction between ERK1/2 and CHK2 and their co-localization in primary DLBCL cells. In addition, the interaction between CHK2 and ERK is disrupted by both CHK2 inhibitor II and CHK2 mutation at Thr68. We also found that ERK protein level is strongly correlated with the level of CHK2 protein (Pearson correlation coefficient ρ =0.737). These data together suggest that the interaction between ERK and CHK2 is likely to have biological relevance in DLBCL. This is further supported by our finding that inhibition of CHK2 phosphorylation of T68 disrupted the interaction between these two proteins. Thus, the functional interaction between ERK and CHK2 requires CHK2 phosphorylation. However, the ability for CHK2 to inhibit ERK phosphorylation of downstream substrates is kinase-independent and likely involves steric hindrance of ERK accessibility to substrate proteins. Determination of the molecular contacts involved in CHK2 interactions with ERK2 will provide insight into how CHK2 negatively regulates ERK functions. We found that inhibition of CHK2 activity modestly induced apoptosis of SUDHL4, SUDHL6 and Farage cells. CHK2 inhibition also induced DNA breaks by itself, manifested by increased phosphorylation of histone H2AX at serine 139. Inhibition of CHK2 expression has also been found to attenuate DNA damage-induced cell cycle checkpoints and to enhance apoptotic activity in HEK293 cells 19 . Targeting of CHK2 with siRNA enhanced apoptosis following induction of DNA damage by IR or doxorubicin, and inhibited the growth of resistant in vivo tumours 9 , 12 . Thus, induction of DNA damage by CHK2 inhibition may contribute to its antitumour activity. It should be noted that the response of a tumour to CHK2 manipulation depends on a specific cellular context. An intriguing proposal put forth by Laurent Antoni and his colleagues is that CHK2 kinase is similar to the two sides of the same coin 10 . In the case of wild-type CHK2 tumours that possess other defects in checkpoint and repair pathways, short-term inhibition of CHK2 could have therapeutic value 10 . In contrast, in those tumours where CHK2 function is diminished or eliminated, inhibition of other proteins involved in checkpoint and repair pathways might be lethal 10 . Defining how CHK2 kinase elicits such distinct cellular outcomes as cell survival through DNA repair versus apoptosis or senescence, depending on the genetic background is a critical area of investigation. The Ras/MEK/ERK signalling cascade is key pathway implicated in hematopoietic malignant cells survival signalling. Under physiologic conditions, ERK activation stimulates transcription leading to cell cycle progression and proliferation by activated growth factor signalling. ERK1/2 signalling has been implicated in attenuation of DNA damage through positive regulation of DNA repair mechanism 15 . Although the CHK2 inhibitor induced DNA breaks and damage, an unexpected finding was that inhibition of CHK2 activity leads to activation of ERK, which may preserve cell survival and proliferation. This provides a plausible explanation for the modest induction of cell apoptosis by CHK2 inhibitor and suggests that activation of ERK signalling pathway may represent a compensatory response to CHK2 inhibition. In other words, functional CHK2 may be required for cells to survive the basal level of genomic instability known to be present in many malignancies. However, the resulting ERK activation in response to CHK2 inhibition provides a potent survival signal that attenuates the apoptotic drive. This also provides a molecular mechanism by which interruption of ERK functions potentiates the killing of DLBCL cells treated with the CHK2 inhibitor. Similarly, the CHK1 inhibitor, UCN-01, also triggers compensatory activation of the prosurvival Ras/MEK/ERK cascade in various tumour cell types, including human multiple myeloma cells 15 . Our data is consistent with the previous report that activation of ATM by radiation downregulates phospho-ERK1/2 and its downstream signalling in both cell culture and tumour models 27 . It is imperative to identify drugs that can improve the efficacy and reduce the toxicity of standard anti-lymphoma therapy 28 . We have demonstrated that a novel ERK2 inhibitor enhances the antitumour activity of CHK2 inhibition in a human DLBCL xenograft model and primary DLBCL cells from human patients. It is also worth noting that there was no evidence of toxicity in animals treated with either ERK or CHK2 inhibitors alone or in combination. Additionally, exposure of normal human B cells and fibroblasts to both ERK and CHK2 inhibitors revealed little cytotoxicity. The lack of apparent toxicity to normal tissues makes targeting both ERK and CHK2 signalling an attractive clinical approach. Moreover, making this potential combination clinically appealing is the fact that CHK2 activation occurs after exposure to commonly used chemotherapeutic agents in DLBCL therapy as well as IR, suggesting that incorporating this dual inhibitor strategy could increase the therapeutic efficacy of standard regimens without increasing side effects. In summary, we have shown a positive correlation between ERK and CHK2 expression in human DLBCL. ERK physically interacts with CHK2 and inhibition of CHK2 activates ERK, which may serve as a compensatory response to enhance cell survival. Thus, combined inhibition of ERK and CHK2 exerts a potent anti-tumour effect in a human DLBCL xenograft model and primary DLBCL cells from human patients and may provide a therapeutic advantage. Whether these findings can be extended to other types of lymphoma is unknown at present and warrants further investigation. Our study demonstrates the mechanistic basis for the therapeutic targeting of DLBCLs with inhibition of both ERK and CHK2 and provides a rationale for using this combinatorial therapy. Methods Cell culture and treatments DLBCL (SUDHL4, SUDHL6 and Farage) cells were grown in RPMI 1640 (Invitrogen) containing 10% fetal bovine serum and 1% Penicillin/Streptomycin. HeLa and HEK293 cells were grown in DMEM (Invitrogen) containing 10% fetal bovine serum. ERK inhibitor no. 76 (3-(2-aminoethyl)-5-((4-ethoxyphenyl) methylene)-2,4-thiazolidine- dione, HCl) was purchased from Calbiochem. CHK2 inhibitor II (2-(4-(4-chlorophenoxy) phenyl)-1H-benzimidazole-5-carboxamide hydrate) was from Sigma Aldrich. Myc (or Flag)-Wild-type CHK2 (WT), and Myc (or Flag) -T68A CHK2 constructs were gifts from Dr Sheau-Yann Shieh. GFP–WT CHK2, GFP–T68D CHK2, GST–WT CHK2 and GST–Kinase inactive CHK2(KD) constructs were kindly provided by Drs Bahassi and Stambrook. HeLa and HEK293 cell transfections were done with Lipofectamine 2000 (Invitrogen). Specific oligo siRNAs for CHK2, and the negative siRNA control were obtained from Santa Cruz. Transfection experiments were performed using Amaxa Nucleofector kit V (Amaxa), program A-020. Primary cells De-identified tissue samples were obtained from patients in accordance with the guidelines and approval of the University of Maryland Medical School Institutional Review Board and conform to the Declaration of Helsinki. Single-cell suspensions were obtained from lymph node biopsies by physical disruption of tissues (using scalpels and cell restrainers), followed by cell density ficoll-gradient separation. Cell number and viability were determined by trypan blue dye exclusion with at least 70–80% viable cells before treatment exposure. Primary cells were cultivated in medium containing 80% RPMI and 20% human serum supplemented with antibiotics, L -glutamine and HEPES for 48 hours +/− drug treatment. Cells were exposed to varying doses of the ERK and/or CHK2 inhibitors as indicated. Assessment of cell apoptosis and percentage non-viable cells Cells were seeded at equal density and then treated with the ERK and/or CHK2 inhibitors in complete RPMI 1640 medium. Forty-eight or seventy-two hours after drug treatment, cells were collected and apoptosis was analysed by flow cytometry using Annexin V staining kit (Southern Biotech). After designated treatments, cells were stained with Trypan Blue (Sigma-Aldrich). The numbers of non-viable cells were determined by counting the cells that showed trypan blue uptake using a hemocytometer. Statistical significance between experimental conditions was determined using the Student's t -test. Co-immunoprecipitation and immunoblotting After the treatment, the cells were lysed in the buffer (20 mM Tris–HCl pH 7.4, 150 mM NaCl, 1 mM EDTA, 1 mM EGTA, 1% Triton X-100, 1 mM Na 3 VO 4 , 1 mg ml −1 aprotinin, 1 mg ml −1 leupterin and 1mM PMSF). Insoluble material was removed by centrifugation, and antibodies were added to lysates for 1 h at 4 °C. Antibodies were collected with protein A or protein G-sepharose beads, and protein complexes were washed three times at 4 °C with the lysis buffer. Immunoblotting was performed using standard methods. Rabbit polyclonal anti–CHK2 (Santa Cruz), Anti-pCHK2, anti-pERK1/2, anti-ERK1/2 (Cell Signaling) and anti-PARP (Enzo) antibodies were used at 1:1,000, whereas anti-γH2AX (Millipore) and anti-actin (C2, Santa Cruz) were used at 1:5,000. Glutathione S -transferase fusion protein pull-down assays GST fusion proteins were expressed in Escherichia coli BL21(DE3) cells, and purified using glutathione-Sepharose beads. Briefly, the GST fusion proteins were coupled to glutathione beads at 4 °C for 1 h and then washed three times with the lysis buffer. The immobilized GST fusion proteins were incubated with the cell lysates for 1 h at 4 °C. The beads were washed with the lysis buffer four times and then the protein complexes were loaded in 10% SDS–PAGE, followed by immunoblotting. In vitro kinase assay ERK2 was expressed in bacteria and purified as previously described 29 . The GST-tagged Chk2 wild type and kinase inactive mutant (KD) was obtained from Drs Bahassi and Stambrook and expressed in bacteria 14 . ERK2 wild type (2 μg) and 0.5 μg MBP were incubated in the presence or absence of 1 μg GST–Chk2 wild type or inactive GST–Chk2 (KD) for 15 min at 300C in 50 mM Tris–HCl, 10 mM MgCl 2 , 1 mM EGTA, 2 mM DTT, 0.01% Brij 35 (pH 7.5) containing 20 μM cold ATP and 2 μCi γ- 32 P–ATP. Reactions were stopped with an equal volume of 2×SDS–PAGE sample buffer, and the proteins were resolved by SDS–PAGE, stained with Coomassie Blue, and phosphorylation of Chk2, ERK2 and MBP was determined by autoradiography or phosphoimager analysis. Tissue microarrays and protein detection Eight tissue microarrays were purchased from Biomax (#LM801). These arrays included a total of 10 reactive lymph nodes, 54 DLBCL and 6 other types of lymphoma. Four more DLBCL were excluded because of fall-off of the tissue sections during staining. Six of the 10 reactive lymph nodes were excluded in this analysis because of their lack of germinal centres. In addition, 20 reactive lymph node samples were provided by the UMGCC Pathology Biorepository and Research Core to serve as normal GCB controls. Human sample studies have been approved by the University of Maryland Medical School Institutional Review Board and conform to the Declaration of Helsinki. The paraffin sections were routinely stained with hematoxylin-eosin (H&E) reagents for morphologic examination. The Vectastain Elite ABC kit, purchased from Vector Laboratories, was used for immunohistochemical analysis.The paraffin sections were incubated overnight at 4 °C with the primary antibody. Anti-Ki67 and anti-CD20 antibodies (Dako) were used at 1:1,000. Anti-CD10,anti-BCL6 and anti-MUM1 antibodies (Abcam) were used at 1:20 and anti-FOXP1 antibody (Abcam) was used at 1:1,200. Anti-ERK1/2 antibody (Cell Signaling) was used at 1:300 and anti-CHK2 antibody (Santa Cruz) was used at 1:200. Slides were then treated with biotinylated secondary antibody and incubated with performed avidin–peroxidase complex. The sections were counterstained with hematoxylin, dehydrated, and mounted. Appropriate positive and negative controls were included in each assay. Immunohistochemical staining was evaluated independently by two hematopathologists (XFZ and SL with an inter-observer consistency >90%), according to our scaled criteria: 0, <25% neoplastic cells staining; 1+, 26–50% neoplastic cells staining; 2+, 50–75% neoplastic cells staining; 3+, >75% neoplastic cells staining. Based on the intensity of staining, the lymphomas were also divided into negative, weak, and strong staining patterns. Because the staining intensity may vary from assay to assay, our scaled criteria overrule the staining intensity, if only small number of lymphoma cells show strong staining. Occasional inter-observer inconsistencies were reconciled by re-reviewing together by the two hematopathologists to reach agreement. In combination, the lymphomas were roughly divided into negative (0 and/or negative), low (1+ and/or weak) and high (2+, 3+ and strong) expresser groups. The images were acquired with a Nikon eclipse 50i microscope (Nikon Instruments) and formatted with Adobe PhotoShop CS software. For fluorescence double staining, the section was treated as above and first stained with mouse monoclonal anti-ERK1/2 antibody (Cell Signaling, 1:300). After biotin blocking, the section was then stained with rabbit polyclonal anti-CHK2 antibody (Santa Cruz, 1:200). The sections were then washed with PBS and incubated with goat anti-rabbit Alexa Fluor 488 and goat anti-mouse Alexa Fluor 568 at a concentration of 1:200 each. The stained sections were mounted using Vectashield mounting medium (Vector). Control sections were similarly processed, except that the primary antisera were omitted, which yielded no staining. Images were captured using an Olympus FV1000 confocal microscope (Olympus) and processed with the Olympus Fluoview Version 1.7c software. In vivo tumour growth in the xenograft models All procedures involving animals were approved by the Institutional Animal Care and Use Committee of the University of Maryland. Female SCID Beige mice were housed in a pathogen-free environment under controlled conditions of light and humidity and received food and water ad libitum . SUDHL6 cells (1.2×10 6 ) were resuspended in 100 μl PBS and then mixed with an equal volume of Matrigel. The mixture was s.c. injected into the left and right dorsal flanks of 5- to 7-week-old female SCID mice. When the tumour reached the size of 60–163 mm 3 , the treatment was administered intraperitoneally every other day for a total of 20 days. ERK inhibitor no. 76 in DMSO was freshly prepared in PBS for a final dose of 5 mg kg −1 . CHK2 inhibitor II in alcohol was diluted in PBS for a final dose of 1 mg kg −1 . Injection of the vehicle alone was used as a control. Tumours were measured three times a week with calipers, and tumour volumes were calculated by the formula ½× r 1 2 × r 2 ( r 1 < r 2 ). At the moment of death, tumours were collected for further analysis. The significance of differences between treatment arms was determined using the Student's t -test. Additional information How to cite this article: Dai, B. et al . Functional and molecular interactions between ERK and CHK2 in diffuse large B-cell lymphoma. Nat. Commun. 2:402 doi: 10.1038/ncomms1404 (2011).

Researchers at the University of Maryland School of Medicine have discovered a novel interaction between two proteins involved in regulating cell growth that could provide possible new drug targets for treating diffuse large B-cell lymphoma, the most common type of non-Hodgkin's lymphoma. In a study published online in Nature Communications, the scientists report that they have found a complex molecular and functional relationship between ERK (extracellular signal-regulated kinase), a protein that helps to regulate cell proliferation and survival, and CHK2 (checkpoint kinase 2), a protein that is involved in the cellular DNA damage response. They also demonstrated, for the first time, elevated levels of both proteins in diffuse large B-cell lymphoma cells, compared to non-cancerous cells. Ronald B. Gartenhaus, M.D., associate professor of medicine at the University of Maryland School of Medicine and the senior author, says researchers found that CHK2 appears to regulate the activity of ERK, although the exact mechanism is not clear. "The two proteins physically interact, which was not known before, and we may be able to use this interaction for therapeutic advantage. We found that treating human B-cell lymphoma cells with both an ERK inhibitor inhibitor and a CHK2 inhibitor killed substantially more cancer cells than treating the cells with either drug alone," he says. "Based on our findings, we believe that a combination therapy targeting both ERK and CHK2 could offer a potential new approach to treating diffuse large B-cell lymphoma," says Dr. Gartenhaus, who is co-leader of the Program in Molecular and Structural Biology at the University of Maryland Marlene and Stewart Greenebaum Cancer Center. The drugs used to inhibit ERK and CHK2 caused the cancer cells to die through a process called apoptosis, or programmed cell death. Human cells normally self-destruct in a controlled manner, but cancer cells lose this ability and consequently grow uncontrollably. E. Albert Reece, M.D., Ph.D., M.B.A., vice president of medical affairs at the University of Maryland and dean of the University of Maryland School of Medicine, says, "This is a very important discovery that may ultimately benefit patients with diffuse large B-cell lymphoma, a common hematological malignancy that can be difficult to treat. These findings provide valuable new insight into the molecular make-up of this cancer that may lead to new targeted drug therapies." Lymphoma is a cancer that originates in the lymphocytes (a type of white blood cell) of the immune system. Diffuse large B-cell lymphoma is a fast-growing, aggressive form of non-Hodgkin's lymphoma. It accounts for about 30 to 35 percent of all non-Hodgkin's lymphomas, and about 25,000 new cases are diagnosed each year. Non-Hodgkin's lymphomas usually are treated with several types of chemotherapy – cyclophosphamide, doxorubicin, vincristine and prednisone (CHOP) – and a biological therapy, such as the monoclonal antibody rituximab (Rituxan). Radiation therapy may be used on occasion, and bone marrow or stem cell transplantation may also be a treatment option. Dr. Gartenhaus says researchers hope the study findings will help to develop new therapies that will be effective and well-tolerated by patients. "We believe it is important to identify drugs that can improve the efficacy and reduce the toxicity of standard anti-lymphoma therapy," he says. The new research showed that using compounds to inhibit ERK and CHK2 did not cause any significant damage to normal cells or tissue examined in the lab, according to Bojie Dai, Ph.D., a postdoctoral fellow at the University of Maryland School of Medicine and the lead author. "We hope that new therapies directed at these two proteins would have modest side effects because they would target only the lymphoma cells," Dr. Dai says. Dr. Gartenhaus says that researchers don't know yet whether the interaction between ERK and CHK2 occurs in other types of lymphoma.

dx.doi.org/10.1038/ncomms1404

Medicine

New method improves transplant safety in mice

Agnieszka Czechowicz et al. Selective hematopoietic stem cell ablation using CD117-antibody-drug-conjugates enables safe and effective transplantation with immunity preservation, Nature Communications (2019). DOI: 10.1038/s41467-018-08201-x Journal information: Nature Communications

http://dx.doi.org/10.1038/s41467-018-08201-x

https://medicalxpress.com/news/2019-02-method-transplant-safety-mice.html

Abstract Hematopoietic stem cell transplantation (HSCT) is a curative therapy for blood and immune diseases with potential for many settings beyond current standard-of-care. Broad HSCT application is currently precluded largely due to morbidity and mortality associated with genotoxic irradiation or chemotherapy conditioning. Here we show that a single dose of a CD117-antibody-drug-conjugate (CD117-ADC) to saporin leads to > 99% depletion of host HSCs, enabling rapid and efficient donor hematopoietic cell engraftment. Importantly, CD117-ADC selectively targets hematopoietic stem cells yet does not cause clinically significant side-effects. Blood counts and immune cell function are preserved following CD117-ADC treatment, with effective responses by recipients to both viral and fungal challenges. These results suggest that CD117-ADC-mediated HSCT pre-treatment could serve as a non-myeloablative conditioning strategy for the treatment of a wide range of non-malignant and malignant diseases, and might be especially suited to gene therapy and gene editing settings in which preservation of immunity is desired. Introduction Hematopoietic stem cell transplantation (HSCT) is a powerful treatment modality that enables replacement of host hematopoietic stem cells (HSCs) with HSCs from a healthy donor or genetically improved/corrected HSCs from the patient 1 . This procedure often results in life-long benefits and can curatively treat many malignant and non-malignant blood and immune diseases. Hence >1,000,000 patients have been transplanted in the last 60+ years for a wide range of blood and immune diseases, including leukemias, hemoglobinopathies, metabolic diseases, immunodeficiencies, and even HIV 2 . HSCT has also been demonstrated to be a beneficial treatment for autoimmune diseases 3 , and, with modern gene-modification techniques such as lentiviral transduction and ZFN, TALEN, or CRISPR/Cas9 gene editing, HSCT application may be expanded to an even wider range of diseases 4 . However, despite its broad curative potential, HSCT is currently mainly restricted to otherwise incurable malignant diseases and it is estimated that <25% of patients that could benefit from HSCT undergo transplantation 5 . This is largely due to undesirable morbidity/mortality from cytotoxic chemotherapy and irradiation-based conditioning currently necessary to enable donor HSC engraftment and the risks associated with graft versus host disease (GvHD). Due to their non-specific nature, classic conditioning regimens lead to both detrimental short-term and long-term complications including multi-organ damage, mucositis, need for frequent red blood cell and platelet transfusions, infertility, and secondary malignancies 6 , 7 . Additionally, these agents result in profound and prolonged immune ablation, which predisposes patients to serious and sometimes fatal opportunistic infections necessitating extended hospitalizations and exposure to toxic side effects of anti-infective agents 8 . Although much work has led to the development of reduced intensity conditioning (RIC) methods, which use lower dose combination chemotherapy with or without low dose irradiation, patients still experience many of these debilitating side effects 9 . Eliminating such harsh conditioning regimens would dramatically improve HSCT and expand its use, especially when combined with gene therapy or gene editing where the native hematopoietic system can be repaired without the need for allogeneic transplantation which carries GvHD and immune suppression risk. Traditionally, conditioning involves total body irradiation (TBI) and/or various chemotherapy prior to HSCT. These agents have been thought essential to make “space” in host bone marrow (BM) for donor HSC engraftment 10 , but they are non-specific and induce significant collateral damage. We previously demonstrated in immunodeficient mice that host HSC competition specifically limits donor HSC engraftment 11 , 12 . Subsequently, we showed that host HSCs in this model could be depleted using an antagonistic anti-murine CD117 monoclonal antibody (ACK2), resulting in an effective, safe, alternative single-agent conditioning approach enabling high donor HSC engraftment 11 . However, this naked antibody conditioning approach only functions as a stand-alone agent in certain disease models; such as immune deficiency 11 , 13 and Fanconi anemia 14 . In other settings, it has been found necessary to combine ACK2 with agents such as low-dose irradiation 15 or CD47 antagonism 13 to increase potency, making clinical translation of this approach challenging. We have recently shown that an alternative antibody-based approach to transplant conditioning is through use of CD45.1 or CD45.2 antibodies conjugated to the drug saporin 16 . Saporin is a ribosome-inactivating protein with potent cell-cycle-independent cytotoxic activity 17 . Unlike other toxins, it lacks a general cell entry domain and on its own is non-toxic. It can be targeted to specific cell types by coupling to antibodies directed to various cell-surface antigens and it is believed that upon receptor-mediated internalization, saporin is released intracellularly halting protein synthesis and inducing cell death 17 . As CD45 is present on most hematopoietic cells, including HSCs, we found CD45-antibody-drug-conjugates (CD45-ADCs) to be effective conditioning agents in various syngeneic immunocompetent mouse models 16 . However, as CD45 is also present on all lymphocytes, CD45-ADCs lead to profound lymphodepletion 16 and therefore likely will maintain opportunistic infection susceptibility. Therefore, if this approach is translated, it may be suited for HSCT contexts where immune depletion is required (e.g. allotransplant and autoimmune disease treatment), but is likely to be suboptimal for other applications. For settings such as autologous gene therapy, an improved solution in which HSCs can be specifically depleted while maintaining intact immunity would be optimal. Here, we show that CD117 antibody-drug-conjugates (CD117-ADCs) effectively and specifically deplete host HSCs in vivo with minimal toxicity. This allows safe and highly efficient transplantation of immunocompetent mice with whole bone marrow (WBM) or purified HSCs, whereas downstream effector cells are spared due to a lack of CD117 expression, thereby leading to the preservation of immunity. Results CD117-ADC potently depletes endogenous HSCs We created a single-agent antibody-drug-conjugate targeting the antigen CD117 (c-kit), which is a critical hematopoietic stem and progenitor cell receptor for the cytokine stem cell factor (SCF) 18 . Within the hematopoietic compartment, CD117 expression is largely restricted to HSCs and proximal multi-potent and oligopotent progenitors, though it is also expressed on some downstream hematopoietic effector cells such as mast cells and rare cells of other tissues 19 . The CD117-ADC was prepared by combining biotinylated anti-CD117 antibody (clone 2B8) with streptavidin–saporin (Fig. 1a ). In separate experiments, the CD117-ADC and naked CD117 antibody were administered by intravenous injection to immunocompetent, wild-type mice, and HSC depletion was assayed in BM 8 days after injection. HSCs (defined immunophenotypically as Lin−cKit+Sca1+CD150+CD48−) were quantified by flow cytometry of the BM of treated mice (Fig. 1b–d ), and functionally by transplantation of the BM into lethally irradiated recipients (Fig. 1e, f ). These studies revealed that 1.5 mg/kg of CD117-ADC (~12 µg streptavidin-saporin) optimally resulted in depletion of >99% of immunophenotypic and functional HSCs. Fig. 1 One-time, intravenous treatment with CD117-ADC potently depletes hematopoietic stem cells (HSCs). a Experimental outline for assessing the ability of antibody-drug-conjugates to deplete HSCs in immunocompetent, wild-type C57BL/6 mice. HSCs and progenitors were assessed through flow cytometric (FACS) phenotypic analysis and colony-forming cell-forming assays, and in vivo reconstitution potential was assessed by competitive transplantation assays. b Representative FACS plots of bone marrow from untreated control and CD117-ADC-treated animals (2B8 clone bound to saporin toxin). c , d Dose-dependent phenotypic depletion of HSCs (Lin−cKit+Sca1+CD150+CD48−) by CD117-ADC and lack thereof by naked CD117 mAb as assessed 8 days after IV administration, displaying decreased HSC frequency ( c ) and total HSC number ( d ). Non-treated mice served as controls. e , f In vivo depletion of HSCs affirmed by lack of long-term reconstitution activity of treated WBM as assessed by donor total blood chimerism ( e ) and donor granulocyte chimerism ( f ) in competitive transplantation assays. Statistics calculated using unpaired t test. Data represent mean ± SEM. ( n = 3–5 mice/group, assayed individually); all data points significant as indicated vs. untreated control (** P < 0.01, *** P < 0.001, **** P < 0.0001) Full size image As CD117 is restricted to approximately 5–10% of BM cells, overall BM cellularity remained unchanged with such treatment (Supplementary Fig 1a, b ). Surprisingly many CD117-expressing immunophenotypic progenitors remained present at the day 8 timepoint post CD117-ADC treatment (Supplementary Fig 1c ), although functional progenitor activity, as assessed by in vitro colony-forming activity, showed continued dose-dependent depletion at this timepoint (Supplementary Fig 1d ). Although all CD117-expressing hematopoietic progenitor sub-populations assayed were decreased at early timepoints after CD117-ADC treatment, depletion was to a lesser degree than after sub-lethal 5 Gy TBI and importantly these progenitors rapidly rebounded to near normal levels by day 6 post treatment (Supplementary Fig 1e–j , Supplementary Fig 2 ). CD117-ADC enables efficient donor-cell engraftment We next determined whether the HSC depletion achieved by CD117-ADC conditioning enabled successful engraftment of donor hematopoietic cells. Consistent with our previously published study investigating CD45-ADCs 16 , ten million congenic CD45.1 or syngeneic CD45.2–GFP WBM cells were used as donor cells for transplantation following CD117-ADC treatment (Fig. 2a ). In this setting, optimal engraftment was observed when BM transplantation was performed 8–9 days post CD117-ADC administration (Supplementary Fig 3a, b ). Under such conditions as early as 4 weeks after transplantation, we observed >98% donor myeloid chimerism within the peripheral blood for both donor cell types in mice conditioned with 1.5 mg/kg CD117-ADC but not the naked CD117 antibody, and subsequently over time with further hematopoietic turnover, all peripheral blood compartments became mostly donor-derived by 20 weeks post CD117-ADC conditioning and transplantation (Fig. 2b, c ; Supplementary Fig 3c–f , Supplementary Fig 4 ). Engraftment was found to correlate with CD117-ADC dose with significant donor chimerism already observed with 0.3 mg/kg CD117-ADC dose and maximum engraftment observed with 1.5 mg/kg CD117-ADC dose (Supplementary Fig 3g ). To confirm true HSC replacement, recipient mice were sacrificed and BM was assessed to reveal that post 1.5 mg/kg CD117-ADC conditioning and WBM transplantation >99% of HSCs were of donor origin (Fig. 2d ). Additionally, when BM from these primary recipients was transplanted into lethally irradiated secondary recipients, high levels of donor-derived multi-lineage engraftment were observed, confirming bona fide stem cell function (Fig. 2e ). Total donor chimerism post CD117-ADC-conditioning was similar to that achieved using CD45.2-ADC or conventional reduced-intensity 5 Gy TBI conditioning (Fig. 3a ); however, myeloid donor engraftment post CD117-ADC was achieved with faster kinetics as compared to conditioning with CD45.2-ADC or 5 Gy TBI (Fig. 3b ). Donor B-cell and T-cell turnover was delayed post CD117-ADC and CD45.2-ADC conditioning as compared with 5 Gy TBI; however ultimately higher donor lymphocyte chimerism was reached in the ADC conditioned groups which continued to increase with time (Fig. 3c, d ). Moreover, unlike with irradiation, neither CD117-ADC or CD45.2-ADC conditioning resulted in any neutropenia in the post-transplant period (Fig. 3e ). In contrast to CD45.2-ADC and 5 Gy TBI, CD117-ADC conditioning also uniquely avoided lymphopenia in this period likely due to the lack of CD117 expression on the lymphocytes (Fig. 3f, g ). Fig. 2 CD117-ADC conditioning durably and robustly enhances donor murine whole bone marrow (WBM) engraftment. a Experimental outline for assessing the ability of antibody-drug-conjugates to condition immunocompetent, wild-type C57BL/6 mice allowing for efficient engraftment of donor murine WBM. b CD117-ADC pre-treatment 8 days before infusion of 10 × 10 6 CD45.1+ donor whole bone marrow cells leads to robust turnover of recipient peripheral blood as assessed through flow cytometric (FACS) analysis unlike naked CD117 pre-treatment. Representative FACS plots of peripheral blood showing >80% total donor peripheral blood CD45.1+ cells in otherwise CD45.2+ host. c CD117-ADC conditioning with WBM transplantation results in rapid multi-lineage engraftment with kinetics paralleling lifespan of cell populations. d Almost complete donor HSC (Lin−cKit+Sca+CD150+CD48−) and HPC (Lin−cKit+Sca+) chimerism post CD117-ADC treatment and donor WBM transplantation confirmed by phenotypic bone marrow analysis of transplanted animal 20 weeks post transplantation. Representative FACS plots of bone marrow HSCs showing >99.9% donor HSC chimerism. e Donor HSC engraftment confirmed through secondary transplantation of WBM into lethally irradiated recipients. Data represent mean ± SEM ( n = 5 mice/group assayed individually). Statistics calculated using unpaired t test; all data points significant as indicated vs. untreated control (**** P < 0.0001) Full size image Fig. 3 CD117-ADC conditioning leads to comparable total multi-lineage engraftment as alternative conditioning regimens post WBM transplantation with improved engraftment kinetics. a CD117-ADC conditioning pre-WBM transplantation leads to comparable total peripheral blood chimerism as CD45.2-ADC and 5 Gy TBI conditioning. Robust multi-lineage donor engraftment with rapid donor granulocyte chimerism ( b ) and slower donor B-cell ( c ) and T-cell turnover ( d ). e – g No significant leukopenia observed post CD117-ADC treatment and WBM transplantation, as opposed to CD45.2-ADC and 5 Gy TBI conditioning with grossly normal absolute myeloid cells ( e ), absolute B-cells ( f ), and absolute T-cells ( g ). Data represent mean ± SEM ( n = 5 mice/group, assayed individually). Statistics calculated using unpaired t test; all data points significant as indicated vs. untreated control (** P < 0.01; *** P < 0.001; **** P < 0.0001) Full size image To determine if the efficacy profile of this CD117-ADC was unique to the non-antagonistic 2B8 clone, we tested two additional anti-CD117 clones, reported non-antagonistic 3C11 and antagonistic ACK2. Of the tested clones, 2B8 was most efficiently internalized by EML cells (a CD117+ hematopoietic progenitor line dependent on SCF) (Supplementary Fig 5a ), though all antibodies were internalized and cytotoxic to EML cells when conjugated to saporin (Supplementary Fig 5b ). All three CD117-ADCs were effective at depleting immunophenotypic and functional HSCs in vivo with no notable gross toxicity (Supplementary Fig 5c–e ); however, the 2B8-ADC was the most potent. Not surprisingly, when the three CD117-ADCs were used as conditioning agents prior to congenic CD45.1 WBM transplantation, donor engraftment efficacy roughly correlated with the extent of recipient HSC depletion (Supplementary Fig 5f–i ). All CD117-ADCs tested effectively depleted recipient HSCs and enhanced donor HSC engraftment to some degree, implicating the CD117-ADC conditioning concept is broadly applicable and not specific to one agent. However, as 2B8-saporin was the most effective CD117-ADC, it was carried forward for subsequent studies. Although WBM transplantation has wide clinical utility, donor T-cells contained in such grafts can cause GvHD. T-cell-depleted grafts reduce this complication; however, their engraftment has historically been more challenging with delayed immune reconstitution and increased infectious complications 20 . We tested whether CD117-ADC conditioning could overcome this limitation and enable engraftment of purified HSCs (Fig. 4a ), which is also relevant to gene therapy settings where optimally these isolated cells are exclusively transplanted to minimize the number of cells manipulated and thereby amount of vector required. As expected, we observed robust engraftment of purified congenic CD45.1 HSCs with CD117-ADC conditioning, which increased in a donor cell dose-dependent manner (Fig. 4b, c ) and led to >85% donor granulocyte chimerism at the highest HSC cell dose at late timepoints. Importantly, we also observed faster and more robust immune reconstitution using purified HSCs in CD117-ADC-conditioned animals than other reduced-intensity conditioning regimens (Fig. 4d, e , Supplementary Fig 6a–c ), obviating the historic concern of increased infection susceptibility associated with T-cell-depleted grafts 20 . Stem cell engraftment was further confirmed by assessing BM HSC chimerism that matched long-term peripheral blood granulocyte chimerism (Fig. 4f ), and through serial transplantation into lethally irradiated secondary recipients (Fig. 4g ). As human HSCs have been hypothesized to utilize similar microenvironments to mouse HSCs, we also tested whether the CD117-ADC enabled the engraftment of human HSCs in mice to give rise to irradiation-free xenografts, which can be used to model human hematopoiesis in vivo. Indeed, when we transplanted 25,000 human CD34+ cord blood cells into CD117-ADC-conditioned adult immunocompromised NSG recipients, we observed robust total peripheral blood donor human chimerism with multi-lineage engraftment that was similar to irradiation-conditioned controls (Fig. 4h–j ). Fig. 4 CD117-ADC conditioning effectively enhances purified mouse and human HSC engraftment, with increased HSC cell dose resulting in increased multi-lineage reconstitution. a Experimental outline for assessing the ability of antibody-drug-conjugates to condition immunocompetent, wild-type C57BL/6 mice allowing for efficient engraftment of donor murine HSC. b , c CD117-ADC pre-treatment 8 days before infusion of FACS-purified CD45.1+ donor HSCs leads to robust total donor peripheral blood ( b ) and donor granulocyte chimerism ( c ). Rapid donor B-cell ( d ) and T-cell ( e ) reconstitution post transplantation of purified HSCs into CD117-ADC-conditioned animals as compared to 5 Gy TBI controls. f Similarly enhanced donor HSPC (Lin−cKit+Sca+) and HSC (Lin−cKit+Sca+CD150+CD48−) chimerism post CD117-ADC treatment and transplantation of 2000 purified HSCs confirmed by phenotypic bone marrow analysis of transplanted animal 20 weeks post transplantation. g Donor HSC engraftment confirmed through secondary transplantation of WBM into lethally irradiated recipients. h Experimental outline for assessing the ability of antibody-drug-conjugates to condition immunocompromised, NSG mice for efficient engraftment of donor human CD34+ cord blood cells. i Single treatment of CD117-ADC effectively enhances donor human cord blood HSPC engraftment in NSG mice, enabling creation of irradiation-free xenografts with multi-lineage reconstitution with total donor chimerism nearing similar levels as 2 Gy TBI. j Multi-lineage human chimerism observed in all xenografted mice, with B-cell engraftment predominant regardless of the conditioning method. Data represent mean ± SEM ( n = 3–5 mice/group, assayed individually). Statistics calculated using unpaired t test; all data points significant as indicated vs. untreated control (* P < 0.05; ** P < 0.01; *** P < 0.001; **** P < 0.0001) Full size image CD117-ADC spares mature cells and preserves immunity As CD117-ADC, CD45.2-ADC, and 5 Gy TBI yielded equivalent levels of total peripheral blood chimerism post WBM transplantation (Fig. 3a ), we assayed the toxicities of each in treated mice that did not undergo transplantation. As CD117 is expressed on erythroid progenitors, previous antagonistic anti-CD117 antibody conditioning approaches were found to cause anemia and require red blood cell transfusion 11 , 13 . However, interestingly, we observed no significant anemia with either CD117-ADC or CD45.2-ADC treatment (Fig. 5a ). Additionally, despite CD117 expression on megakaryocyte progenitors, CD117-ADC-treated recipients experienced only a minor decrease in platelet counts (Fig. 5b ). Neither CD117-ADC nor CD45.2-ADC was myeloablative, with both resulting in an increase in myeloid cells (Fig. 5c ); however, CD117-ADC uniquely spared lymphoid cells and did not cause profound lymphopenia seen with other methods (Fig. 5d, e ). Fig. 5 CD117-ADC is uniquely non-ablative to the peripheral blood with no clinically significant cytopenias, and results in preservation of phenotypic and functional immunity. a , b Despite HSC depletion, only very minor and not clinically significant decreases in hemoglobin ( a ) and platelet counts ( b ) were observed 8 days post CD117-ADC treatment. c Unlike TBI, ADC treatments did not lead to neutropenia, and myeloid counts were increased 8 days post CD117-ADC treatment at the time of transplantation. Unlike CD45.2-ADC, peripheral B-cell ( d ) and T-cell counts ( e ) remained largely intact 8 days post administration of CD117-ADC. f T-cell numbers not only remained intact post CD117-ADC treatment, but animals were uniquely able to mount viral immune responses to LCMV infection post CD117-ADC as indicated in the experimental outline with results showing present LCMV-specific activated CD8 T-cells. g Additionally, post CD117-ADC treatment LCMV-specific activated CD8 T-cells generated from prior LCMV infection uniquely remained, as indicated in the experimental outline with results showing similar numbers of LCMV-specific activated CD8 T-cells CD117-ADC similar to post no conditioning. h Animals were also able to mount functional immune response to Candida challenge post CD117-ADC treatment, as indicated in the experimental outline with animals showing persistent neutrophils, control of Candida load, and increased survival compared to other treatment groups. Data represent mean ± SEM ( n = 3–5 mice/group, assayed individually). Statistics calculated using unpaired t test; all data points significant as indicated vs. untreated control (* P < 0.05; ** P < 0.01; *** P < 0.001; **** P < 0.0001) Full size image As viral infections are a significant cause of morbidity and mortality with traditional hematopoietic cell transplantation, treated recipients were challenged with a murine T-cell dependent virus, lymphocytic choriomeningitis virus (LCMV), and assessed for the ability to generate viral-specific responses. Only post CD117-ADC treatment were LCMV-specific T cell responses robustly observed (Fig. 5f ). Additionally, viral-specific immunity was preserved after this treatment, and many LCMV-specific T cells were observed in recipients that had previously been challenged with LCMV and then treated with CD117-ADC (Fig. 5g ). Fungal immunity was also preserved following CD117-ADC treatment, as after intravenous challenge with the human fungal pathogen, Candida albicans , CD117-ADC recipients could mount effective neutrophil responses as evidenced by decreased kidney fungal organ loads with increased survival (Fig. 5h ). In contrast, busulfan and irradiation conditioning resulted in profound neutropenia, high C. albicans organ loads, and early mortality (Fig. 5h ). Even at the highest 1.5 mg/kg CD117-ADC dose, this treatment was well-tolerated and non-myeloablative; moreover, treatment permitted long-term survival (>6 months) without requiring hematopoietic cell transplantation or transfusions unlike alternative treatments such as irradiation conditioning or ACK2 with CD47 blockade 13 . As CD117 is also expressed on a variety of other rare cell types throughout the body such as melanocytes and germ cells, toxicity assessment including full necropsy was performed after CD117-ADC treatment. Unlike post alternative conditioning strategies such as TBI or ACK2, recipients appeared healthy with full coat color and remained fertile with vigor (Supplementary Fig 7a ). At the highest 1.5 mg/kg optimal treatment dose, CD117-ADC conditioning did result in mild liver toxicity with transiently elevated transaminases and rare apoptotic hepatocytes (Supplementary Fig 7b ), and at increased doses animal mortality was intermittently observed. However, transaminitis resolved without intervention and no other toxicity was documented at this optimal dosing with subsequent resolution of all histologic liver changes. Upon review by an experienced pathologist, all other organs examined had normal histology post treatment (Supplementary Fig 7c ). Collectively, these results indicate that CD117-ADC is less detrimental systemically than conventional conditioning, while achieving efficient HSC depletion enabling rapid, high, sustained donor BM/HSC engraftment and preservation of immunity allowing for protection post viral and fungal challenge. Furthermore, we found that alternative HSC antigens, including CD27, CD90, CD110, and CD184 even in non-optimized settings, could be similarly targeted to generate further restricted conditioning agents implying broad applicability of such an HSC-ADC approach (Supplementary Fig 8a, b ). CD117-ADC depletes human HSCs Although the CD117-ADC generated with the 2B8 clone enabled powerful, proof-of-concept studies and serves as an important tool, this agent only targets mouse CD117. We tested whether similar CD117-ADCs could potently deplete human HSC by first creating a human CD117-ADC using the biotinylated anti-human CD117 (clone 104D2) with streptavidin–saporin linkage. Parallel studies showed that this human CD117-ADC inhibited human HSC growth in vitro (Fig. 6a ). This human CD117-ADC was then administered to humanized xenograft mice, and in vivo human HSC depletion was assayed by measuring human myeloid cells in the peripheral blood which are human HSC and progenitor derived (Fig. 6b ). Human HSC activity as represented by the inability to produce human myeloid cells was found to be diminished 440-fold by this human CD117-ADC (Fig. 6c ). These results imply that similar human CD117-ADCs can deplete human HSCs, suggestive of potential clinical application of such a strategy in conditioning prior to hematopoietic cell transplantation in patients. Given these results, we are optimistic of the rapid advancement and development of similar clinical-grade CD117-ADCs. Fig. 6 Human CD117-ADC inhibits human HSCs in vitro and robustly depletes human HSCs in vivo. Human CD117-ADCs were prepared similarly by conjugating biotinylated anti-human CD117 (104D2 clone) with streptavidin–saporin toxin. a Human CD117-ADC inhibited human cord blood growth in vitro, as compared to unconjugated antibody controls. b Experimental outline for assessing the ability of anti-CD117 antibodies and antibody-drug-conjugates to deplete human HSCs in xenograft mice. c Only antibody conjugated to saporin is effective at eliminating human myeloid cells. Data represent mean ± SEM ( n = 3–5 samples/group, assayed individually). Statistics calculated using unpaired t test; all data points significant as indicated vs. unconjugated antibody-treated control animals (* P < 0.05; ** P < 0.01; **** P < 0.0001) Full size image Discussion Historically, non-selective chemotherapy, either alone or with irradiation, has been necessary for effective hematopoietic cell transplantation 10 . These classical cytotoxic interventions drive the morbidity and mortality of HSCT and lead to many of the complexities of the procedure 7 . Even at reduced doses, which are described as “non-myeloablative”, some myeloid ablation with non-specific organ toxicity occurs 9 . We previously demonstrated that host HSC depletion is a critical requirement to enabling donor HSC engraftment and long-term hematolymphoid turnover, and can be achieved using specific HSC-targeting strategies 11 . This study uniquely highlights that long-term donor BM or HSC engraftment can be achieved in a setting with exclusive HSC depletion. Moreover, donor engraftment occurs with mostly intact host hematopoietic progenitor and myeloid cells, highlighting successful transplantation in a true non-marrow and non-myeloablative context. These studies question the common synonymous use of the term “myeloablation” with “HSC ablation”, as here we display strong evidence that the two can be separated. Based upon this work, we further advocate for the development of antibody-based conditioning strategies that selectively deplete host HSCs. We primarily targeted CD117 in these studies but given our results, presumably any HSC-specific antigen could be targeted, though the strategy we employed herein also likely requires antigen/antibody internalization to effectively deliver the drug. Antibody-based methods targeting HSCs are advancing to patients, with the most advanced antagonistic anti-CD117 naked antibody derived from our prior work already in clinical trials in severe combined immunodeficiency patients which is showing encouraging clinical results 21 . However, although such approaches are very promising for certain patient groups, all prior reported methods have significant limitations. To address these challenges, we show here that a single-agent CD117-ADC is a superior conditioning agent which safely and effectively enables robust BM/HSC transplantation in an immunocompetent setting while uniquely preserving immunity. This approach enables transplantation while avoiding clinically significant collateral damage across tissues. In particular, no anemia, neutropenia, lymphopenia, coat color discoloration, or infertility was observed. We foresee that the preservation of immunity would minimize additional severe infectious co-morbidities associated with conventional conditioning, making HSCT safe and potentially achievable as a simple outpatient procedure. Furthermore, the preservation of past immunity makes this approach additionally uniquely attractive. These features would be especially useful in settings of autologous gene therapy or gene editing where immune depletion is not required and presents no added benefit as the host generally does not reject its own gene-modified cells. As current conditioning methods are a major limitation for gene therapy transplantation, successful translation of our approach to humans could dramatically enhance the utility of gene therapy and gene editing across disease areas ranging from sickle cell anemia, beta thalassemia, congenital immunodeficiencies to HIV and potentially even broadly enable HSCs to be used as a vector for any type of protein production. Although transient immune suppression would be necessary in the allogeneic setting to prevent immune-mediated donor graft rejection, we predict that CD117-ADC conditioning may still have broad adoption in this setting with advantages over classic conditioning. In addition to decreased toxicity and decreased tissue inflammation, CD117-ADC conditioning enables efficient transplantation of purified, T-cell-depleted grafts which likely would result in significantly decreased GvHD 20 . This approach could also enable more rapid lymphoid recovery which is a major concern with manipulated grafts and further enable shorter immune suppression with decreased infectious complications. Additionally, such CD117-ADC conditioning may lead to deeper remissions in malignant settings, especially for patients with leukemias expressing CD117 and/or when combined with conventional therapies. Moreover, in complementary studies with combined CD117-ADC and transient immune suppression, such conditioning enables unique fully HLA-mismatched transplantation and concurrent immune tolerance likely due to robust donor HSC engraftment (Li et al., co-published) enabling transplantation with many more diverse donors. Our data suggest that a single-agent CD117-ADC may be the conditioning agent of choice for both mouse and human. Although the CD117-ADCs to saporin through biotin–streptavidin linkage developed here are proof-of-concept agents, with proper development and optimization, equivalent anti-human CD117-ADCs may be extremely powerful and dramatically expand the application of HSCT. As saporin toxin is known to have some hepatotoxicity 22 , which we similarly observed in our studies, alternative drug-conjugates may be preferred and utilized in similar fashions to further improve the safety profile of such a conditioning approach 23 . However, saporin may also be a viable ADC, given that it has been previously effectively used in clinical trials and in this type of conditioning setting would only need to be used as a one-time administration. Additionally, although some transient hepatotoxicity was seen with this CD117-ADC at the maximum 1.5 mg/kg dose, this is still likely a much-improved toxicity profile over current standard-of-care chemotherapy and/or irradiation conditioning. Furthermore, we have shown that a fivefold decreased dosing of this CD117-ADC can also generate therapeutically meaningful levels of low mixed chimerism. Such dosing may be ultimately preferred for many diseases which would be expected to have minimal or negligent hepatotoxicity. Interestingly, although CD117 is expressed on other rare cells throughout the body, such as germ cells and melanocytes, no toxicity was documented in these tissues after this CD117-ADC administration unlike after alternative conditioning treatments. Moreover, post this treatment, many CD117-expressing hematopoietic progenitors were also surprisingly spared in the BM despite robust HSC depletion, further implicating that not all CD117-expressing cells are killed by this CD117-ADC treatment, which likely further contributes to its improved toxicity profile. We anticipate such agents to not only be effective at enabling robust HSCT, but also be gentler to other tissues and the hematopoietic microenvironment itself especially with further optimization. As such, this may permit revisiting basic aspects of hematopoiesis since much of our scientific understanding of both human and murine hematopoiesis is based upon transplantation into irradiation or chemotherapy-conditioned hosts. The significant tissue damage associated with classic methods may have introduced artifacts which potentially confounded interpretations of prior studies. Therefore, in addition to its clinical potential for treatment of patients, CD117-ADC conditioning in experimental transplantation may also be a useful tool that provides new scientific insights. Given these findings, we anticipate CD117-ADC conditioning to be broadly used in both experimental and clinical transplantation, possibly expanding the use of this powerful curative procedure across disease settings. Methods Mice About 8–10-week-old female C57BL/6J (CD45.2+) mice were purchased from Charles River or Jackson Laboratories and 7-week-old female NOD.Cg-Prkdc scid Il2rg tm1Wjl /SzJ mice were purchased from Jackson Laboratory and used as recipients in all studies. Newborn NOD.Cg-Prkdc scid Il2rg tm1Wjl /SzJ pups were used to generate xenografts for human CD117-ADC studies, and recipients were used 14 weeks post xenograft generation. About 10–12-week-old female B6.SJL -Ptprc a Pepc b /BoyJ (B6.SJL, CD45.1+) mice were purchased from Jackson Laboratories and used as donors for most C57BL/6J (CD45.2+) mice experiments. In additional experiments, 10–12-week-old female B6 GFP+ (GFP + ) mice were purchased from Jackson Laboratories and used as donors. All animal experiments were performed in compliance with the Institutional Animal Care and Use Committee and approved by Harvard Medical Area Standing Committee or Stanford University Administrative Panel on Laboratory Animal Care. Antibody-drug-conjugate generation and administration Biotinylated anti-CD117 (Biolegend; clones 2B8 [Lot # B170777, B205356], 3C11, ACK2, 104D2), biotinylated anti-CD45.2 (Biolegend; clone 104), biotinylated anti-CD27 (Biolegend; clone LG3A10), biotinylated anti-CD90.2 (Biolegend; clone 30-H12), biotinylated anti-CD110 (Clontech; clone AMM2), and biotinylated anti-CD184 (Biolegend; clone L276F12) were combined with streptavidin-SAP conjugate (Lot # 94-31; Advanced Targeting Systems) in 1:1 molar ratio and then diluted in PBS to desired concentration. CD117-ADCs were generally administered at 1.5 mg/kg (~12 µg of streptavidin-saporin) per recipient, unless otherwise indicated. CD45.2-ADCs were generally administered at 3 mg/kg (~24 µg of streptavidin-saporin) per recipient, as previously optimized. Additional HSC antigen ADCs were administered as follows: CD27-ADCs at 1.5 mg/kg, CD90-ADCs at 3 mg/kg, CD110-ADCs at 3 mg/kg, and CD184-ADCs at 3 mg/kg. Alternative reagent lots were utilized in subsequent studies with slightly altered properties requiring further optimization. Antibody-drug-conjugates were administered in 300 µL via retro-orbital intravenous injection. Naked anti-CD117 antibody (BioXCell; clone 2B8) was administered in separate experiments at 500 µg (~25 mg/kg) per recipient as a control. Murine BM and HSC transplantation BM cells were harvested from donor CD45.1 or GFP mice. WBM cell numbers were determined by counting via CBC (complete blood count) using VetScan HM5 (Abaxis); 10 × 10 6 WBM cells or 500–5000 purified HSCs were transplanted via retro-orbital intravenous infusion 8 days post antibody-drug-conjugate injection or 1 day post 5 Gy TBI into C57BL/6J (CD45.2+) immunocompetent mice as previously described 16 . Purified mouse HSCs used in transplantation studies were obtained by double sorting via fluorescence-activated cell sorting (FACS) Lin−cKIT+Sca1+CD150+CD34− cells on a BD FACS Aria II, as previously described 11 . Competitive transplants were performed by transplanting 1×10 6 WBM cells from treated C57BL/6J (CD45.2+) animals and 1×10 6 WBM cells from untreated B6.SJL (CD45.1+) animals into 9 Gy, lethally irradiated, B6.SJL (CD45.1+) wild-type recipients as previously described 16 . Secondary transplants were performed by transplanting 2×10 6 WBM cells from transplanted animals into 9 Gy, lethally irradiated, C57BL/6J (CD45.2+) wild-type recipients as previously described 16 . Human cord blood transplantation Single donor-derived cord blood CD34+ cells were obtained from AllCells (Alameda, CA) under local IRB-approved protocol with written informed consent. Cells were counted via hemocytometer and suspended to desired dilution of 2.5 × 10 4 cells/200 µL. Each animal received 2.5 × 10 4 cells via retro-orbital intravenous infusion 8 days post antibody-drug-conjugate injection or 1 day post 2 Gy TBI into adult NSG immunodeficient mice. To test for in vivo efficacy of human CD117-ADCs, xenografts were generated by transplanting 500 FACS-sorted human HSCs (Lin−CD34+CD38−CD90+) via facial vein injection into newborn NSG immunodeficient mice post 1 Gy TBI as per previously described methods 24 . Peripheral blood analysis Peripheral blood was sampled from treated animals at 8 days post antibody-drug-conjugate administration to assess for peripheral blood effects. Transplanted mice were assessed via peripheral tail vein bleeds at 2, 4, 8, 12, 16 weeks or later as indicated post donor cell transplantation. White blood cells, hemoglobin, platelets, absolute lymphocyte count, and absolute neutrophil count were determined by a VetScan HM5 (Abaxis) or a Hemavet 950 (Drew Scientific). In addition, the blood samples were red blood cell-lysed and fixed using BD Staining buffer, and subsequently stained for CD3 (T-cells), B220 (B-cells), Gr1 (Granulocytes), CD45.1 (donor), and CD45.2 (host) to assess chimerism by flow cytometry. Xenografts were assessed via peripheral tail vein bleeds at indicated timepoints. Blood samples were similarly red blood cell-lysed and stained for hCD45 (pan-blood), hCD13 (myeloid), hCD3 (T-cell), and hCD19 (B-cell) to assess human chimerism by flow cytometry. Data analysis was done on FlowJo 10 software (version 10.2; Treestar, Ashland, OR) using previously published methods 11 . BM analysis BM cells were harvested post sacrifice and single cell suspensions were obtained by crushing lower extremity bones. Cells were filtered and stained for appropriate markers to quantify HSC and progenitors and to determine donor chimerism. Samples were run on flow cytometry using various panels to assess cellular makeup including: CD45.1, CD45.2, CD117, CD34, CD48, CD41, Flk2, CD105, CD150, Sca1, FcGammaRa, and IL7Ra (all purchased from Biolegend). Data analysis was done on FlowJo 10 software (version 10.2; Treestar, Ashland, OR) using previously published subset definitions 25 . Colony forming cell (CFC) assays were performed by plating 25,000/mL WBM cells in M3434 methyl cellulose media (Stem Cell Technologies) in 3 cm dishes. After 7 days of incubation, the total number of colonies formed were counted using a light microscope. In vitro cell death assays In vitro cell death experiments were performed using the EML (ATCC CRL-11691) cell line cultured in IMDM media in the presence of 200 ng/mL murine stem cell factor (mSCF1, R&D Systems). Cells were plated in 96-well plates with 5000 cells/well in 100 μL volume cell culture media containing various concentrations of antibody-drug-conjugate. Three independent experiments were performed with three technical replicates within each experiment. After 72 h, cell viability was determined using the CellTiter MTS assay (Promega). PBS-treated and 10 μM staurosporin-treated cells (Sigma) were used as live and dead controls, respectively. Additionally, in vitro human HSC cell death experiments were performed using FACS-sorted primary cord blood HSCs. Twenty-five human cord blood HSCs (Lin−CD34+CD38−CD90+CD45RA−) were sorted using previously described methods 25 into 96-well plates with StemSpan media supplemented with 100 ng/mL of human cytokines SCF, TPO, FLT3, and IL3 (R&D Systems) as well as 10 μg/mL of biotinylated antibodies or saporin conjugates. Cells were quantified by counting under light microscopy 2 days and 6 days post plating. Measurement of antibody internalization EML cells (200,000/mL) were plated into 96-well plates in complete media (IMDM, 10% FBS, 200 ng/mL mSCF) and 20 nM of a 1:1 mixture of biotinylated antibodies with streptavidin-AF488 (Life Technologies) with n = 6 technical replicates. After 4 h of incubation, the cells were washed twice and resuspended in PBS containing 2% FBS. Samples were split into two and one sample was incubated with 0.25 mg/mL polyclonal anti-AF488 quenching antibody (clone A-11094, Life Technologies). AF488 signal in samples with and without quenching antibody was quantitated by flow cytometry. Unstained and time zero stained controls (stained on ice) were used to determine the quenching efficiency on the AF488-quenching antibody and calculate internalization frequency of the test samples. Toxicity assessment C57BL/6 mice were sacrificed at various timepoints (2 days or 1 year after treatment), submerged individually in 300 mL containers of Bouin’s solution (Sigma) or 10% Formalin solution (VWR) and delivered to the Harvard Medical Area Core Specialized Histopathology Services or Stanford Department of Comparative Medicine’s Animal Histology Services. Necropsy with resulting H&E histology was performed using standard methods. Pathology was assessed by a trained and certified veterinary or hematopathologist. Liver function testing was also assessed in additional C57BL/6 mice (2 or 4 days after treatment) via Charles River Clinical Pathology Services. Immune memory and response with LCMV LCMV Armstrong was a gift from Dr. John Wherry (University of Pennsylvania, Philadelphia, PA) and was propagated according to the standard protocol 26 . LCMV (2 × 10 5 pfu) was intraperitoneally injected into 6–12-week-old C57BL/6 mice. LCMV-specific CD8 T-cells were stained by PE-conjugated H2DbGP33-KAVYNFATM tetramers, which was kindly provided by the NIH Tetramer Facility (Atlanta, GA). Irradiation post LCMV infection was not possible due to facility policies. Systemic challenge with C. albicans C. albicans , wild-type stain SC5314 (ATCC MYA-2876), was grown overnight from frozen stocks in yeast extract, peptone, and dextrose (YPD) medium (BD Biosciences) with 100 μg/mL ampicillin (Sigma) in an orbital shaker at 30 °C. After pelleting and washing with cold PBS, yeasts were counted using a hemocytometer and cell density adjusted in PBS to 150,000 yeasts per 200 μL. C57BL/6 mice (non-conditioned control, 8 days post 1.5 mg/kg intravenous CD117-ADC, 4 days post 20 mg/kg intraperitoneal busulfan, or 2 days post 5 Gy TBI) were injected via lateral tail vein with 150,000 yeasts, and the animals were monitored daily. Moribund mice were euthanized humanely. For kidney Candida organ loads, mice were sacrificed, kidneys harvested, and homogenized using a hand-held tissue homogenizer for 30 s. Organ homogenates were diluted in PBS and plated on YPD solid agar. Candida colonies were counted at 48 h. Fungal burden was expressed as CFU per gram of kidney. Statistics Animal groups were typically n of five for all experiments, and experiments were performed in duplicate unless otherwise noted. All statistics were calculated using unpaired t tests using two-sided analysis except Kaplan–Meier data, which were analyzed by log-rank (Mantel–Cox) test. Alphanumeric coding was used to blind pathology samples and CFC counting. Reporting Summary Further information on experimental design is available in the Nature Research Reporting Summary linked to this article. Data availability The authors declare that all the data supporting the finding of this study are available with the paper and its supplementary information files.

Blood stem cell transplants—also known as bone-marrow transplants—can cure many blood, immune, autoimmune, and metabolic disorders, from leukemia to sickle-cell disease. But to make sure the healthy blood cells "take," doctors first have to deplete the patient's own, defective blood stem cells using intensive chemotherapy or whole-body radiation. This wipes out the patient's immune system, raising the risk of infection and often causing serious side effects, including anemia, infertility, secondary cancers, organ damage, and even death. "We know that stem cell transplants can cure dozens of blood disorders, including exciting progress with treating autoimmunity," said Professor David Scadden, co-director of the Harvard Stem Cell Institute and director of the Center for Regenerative Medicine at Massachusetts General Hospital. "Yet they are rarely used beyond treating blood cancers because of the extreme rigors the patients must endure." New research led by Scadden and his colleagues at Harvard University has demonstrated in mice an effective alternative to this toxic pretreatment. Published in Nature Communications, the findings pave the way for clinical studies that will show whether the method would work in humans, making blood stem cell transplants safer and available to more patients. "If the principles demonstrated in these studies translate to the clinic, they open the possibility of expanding stem cell transplantation to many more patients who we know would benefit if only it were made more tolerable," said Derrick Rossi, co-senior author of the paper. A promising antibody Previous research showed that a certain antibody blocks a receptor called CD117, which is carried mainly by blood-forming stem cells. When mice with genetically deficient immune systems were treated with the antibody, it selectively targeted those cells to die, making chemotherapy and radiation unnecessary. On its own, the antibody wasn't enough. Clinical trials with a human version of this antibody are now underway in patients with a rare disease called severe combined immunodeficiency. Building on this knowledge, the scientists identified an antibody that targets CD117, but is internalized by cells. Realizing that this could allow them to deliver a toxin directly—and selectively—into stem cells that have CD117, they attached it to a drug called saporin. Saporin, which has already been used in cancer patients, inhibits ribosomes, the protein-building structure in all cells. The team hoped that the combination would effectively kill blood-forming stem cells, and only those cells, by stopping their ability to make important proteins. It did. The Nature Communications study showed that a single dose of the antibody-drug combination eliminated more than 99 percent of blood-forming stem cells in mice. This allowed transplanted stem cells to take up residence in the recipient animals, effectively replacing their damaged blood and immune systems. The combination targeted the recipients' stem cells without harming other kinds of blood cells, and without causing clinically significant side effects. The animals' immune cell functions were preserved and responded effectively to pathogens. "We theorized it would be effective, but were both reassured and excited when it worked so well," said first author Agnieszka Czechowicz of Boston Children's Hospital, who was also deeply involved in previous CD117 studies. Future research will evaluate the safety and efficacy of the antibody-drug combination, and explore other promising combinations. Magenta Therapeutics of Cambridge has licensed this technology and is working toward developing and testing the approach in patients. The company presented preclinical data on anti-CD117 and other antibody-drug conjugates at the American Society of Hematology meeting in December 2018. "Collectively, these results are transformative for the field of transplant research. They open up the possibility of transplanting safely and effectively, without putting patients through toxic treatments," said Rahul Palchaudhuri of Harvard's Department of Stem Cell and Regenerative Biology, who was co-first author on the paper. "If this approach works in people, we can make a difference for patients suffering from many different diseases," said Scadden. "Patient conversations with their doctors would be more about benefit and cure and less about risk and suffering. That is our goal."

10.1038/s41467-018-08201-x

Biology

The gene that starts it all

Alberto de Iaco, Evarist Planet, Andrea Coluccio, Sonia Verp, Julien Duc and Didier Trono. DUX-family transcription factors regulate zygotic genome activation in placental mammals. Nature Genetics 01 May 2017. DOI: 10.1038/ng.3858 Journal information: Nature Genetics

http://dx.doi.org/10.1038/ng.3858

https://phys.org/news/2017-05-gene.html

Abstract In animal embryos, transcription is mostly silent for several cell divisions, until the release of the first major wave of embryonic transcripts through so-called zygotic genome activation (ZGA) 1 . Maternally provided ZGA-triggering factors have been identified in Drosophila melanogaster and Danio rerio 2 , 3 , but their mammalian homologs are still undefined. Here, we provide evidence that the DUX family of transcription factors 4 , 5 is essential to this process in mice and potentially in humans. First, human DUX4 and mouse Dux are both expressed before ZGA in their respective species. Second, both orthologous proteins bind the promoters of ZGA-associated genes and activate their transcription. Third, Dux knockout in mouse embryonic stem cells (mESCs) prevents the cells from cycling through a 2-cell-like state. Finally, zygotic depletion of Dux leads to impaired early embryonic development and defective ZGA. We conclude that DUX-family proteins are key inducers of zygotic genome activation in placental mammals. Main Dux genes encode double-homeodomain proteins and are conserved throughout placental mammals 4 , 5 . Human DUX4 , the intronless product of an ancestral DUXC , is nested within the D4Z4 macrosatellite repeat of chromosome 4 as an array of 10 to 100 units (ref. 6 ). DUX4 , DUXC , and Dux genes from other placental mammals display the same repetitive structure: DUX4 from primates and Afrotheria and DUXC from cows and other Laurasiatheria localize at telomeric or pericentromeric regions, and mouse Du x tandem repeats lie adjacent to a mouse-specific chromosomal fusion point that resembles a subtelomeric structure 4 , 5 . Overexpression-inducing mutations in DUX4 are associated with facioscapulohumeral muscular dystrophy (FSHD), the third most common muscular dystrophy 7 , 8 , and forced DUX4 production in human primary myoblasts leads to upregulation of genes active during early embryonic development 9 . On the basis of this premise, we analyzed publicly available RNA-seq data sets corresponding to this period, focusing on DUX4 and the 100 genes most upregulated in DUX4 -overexpressing muscle cells 10 , 11 ( Fig. 1a and Supplementary Table 1 ). DUX4 RNA was detected from the oocyte to 4-cell (4C) stages, whereas transcripts from putative DUX4 targets emerged on average at the 2-cell (2C) stage and peaked at the 8-cell (8C) stage, as previously defined for human ZGA 12 . Transcripts upregulated in DUX4 -overexpressing muscle cells 11 were also enriched at the 8C stage ( Supplementary Fig. 1a,b ), and after clustering of genes according to their patterns of early embryonic expression ( Fig. 1b ), we delineated (i) 1,517 genes whose transcript expression was already detected in oocytes, plateaued up to 4C and abruptly decreased afterward (maternal gene cluster); (ii) 94 genes and 124 genes whose expression started at 2C and peaked at 4C and 8C, respectively, before rapidly decreasing, a pattern consistent with that of early-ZGA genes (2C–4C and 2C–8C gene clusters); and (iii) 1,352 genes expressed only from 4C, peaking at 8C, and then progressively decreasing, as expected for late-ZGA genes (4C–8C gene cluster). Only the two early-ZGA clusters (2C–4C and 2C–8C) were highly enriched in genes upregulated in DUX4 -overexpressing myoblasts ( Fig. 1c and Supplementary Fig. 1c ). Figure 1: DUX4 promotes transcription of genes expressed during early ZGA. ( a ) Comparative expression during early human embryonic development of DUX4 (red) and the top 100 genes upregulated after DUX4 overexpression in human primary myoblasts (blue; solid line, average; dashed lines, 95% confidence interval around the mean). Oo, oocyte; Zy, zygote; Mo, morula; Bl, blastocyst. ( b ) Cluster of genes differentially expressed during early embryonic development were selected from the previously identified subsets of genes ( Supplementary Fig. 1a ) on the basis of high expression at 4C (top) and 8C (bottom). Blue solid and dotted lines delineate the mean and 95% confidence interval, respectively. ( c ) Expression of genes from each cluster illustrated in b , after ectopic expression of DUX4 in human primary myoblasts. Lower parts of the panels depict the fold change in expression of genes within these clusters, all randomly distributed along the y axes, with kernel density plotted in the upper part. Full size image It is technically challenging to reliably analyze chromatin from the very low number of cells that make up an early embryo, but chromatin immunoprecipitation (ChIP)–seq data obtained in DUX4-overexpressing human embryonic stem cells (hESCs) ( Fig. 2a,b and Supplementary Fig. 2 ) and myoblasts 9 ( Supplementary Fig. 3 ) showed a marked enrichment of the transcription factor around the annotated transcription start site (TSS) regions of early-ZGA genes (2C–4C and 2C–8C clusters) but not zygotic (maternal) and late-ZGA (4C–8C) genes. Interestingly, several genes were bound not on their annotated TSSs, but on neighboring sequences, and their transcription was found to start near this DUX4-binding site ( Supplementary Fig. 4 ). DUX4 has been demonstrated to drive expression of many of its target genes from alternative promoters 11 . After examining publicly available single-cell RNA-seq data quantifying the far 5′ ends of transcripts (TFEs) in early human development 13 , we found that the TFEs of 24 out of 31 early-ZGA genes overlapped with DUX4-binding sites ( Fig. 2c,d and Supplementary Fig. 3c,d ). DUX4 was also recruited to several groups of transposable elements (TEs), notably endogenous retroviruses such as HERVL, MER11B, and MER11C, whose expression increased at ZGA ( Supplementary Fig. 2b,c ). Furthermore, DUX4 overexpression in hESCs led to early-ZGA-gene induction, as previously observed in myoblasts 11 ( Fig. 2e ). Figure 2: DUX4 binds TSSs of genes expressed during early ZGA and activates their expression in hESCs. ( a ) Average coverage normalized for sequencing depth of the ChIP–seq signal of DUX4 (blue) overexpressed in hESCs in a window of 5 kb from the annotated TSSs of genes belonging to the 2C–4C and 2C–8C clusters from Figure 1b . Total input is represented in gray (line, average; shading, s.e.m.). ( b ) Fraction of genes belonging to each cluster from Figure 1b with a DUX4 peak within 5 kb of their annotated TSSs. Two-sided Fisher's exact test was performed to compare maternal versus 2C–4C and 2C–8C (*** P = 3.54 × 10 −61 and *** P = 2.23 × 10 −13 , respectively). ( c ) Average coverage of ChIP–seq signal of DUX4 (blue) overexpressed in hESCs within 5 kb of TFEs of transcripts specifically upregulated at oocyte–4C and 4C–8C transitions. Total input is represented in gray (line, average; shading, s.e.m.). ( d ) Fraction of TFEs from oocyte–4C ( n = 32) and 4C–8C ( n = 128) transitions with a DUX4 peak overlapping with the 5′ end. Two-sided Fisher's exact test was performed to compare 4C–8C versus oocyte–4C TFEs (*** P = 4.48 × 10 −17 ). ( e ) Comparative expression in hESCs of three genes activated at ZGA ( ZSCAN4 , MBD3L2 , and DUXA ), and two control housekeeping genes ( ACTB and TBP ) 24 h after transfection with plasmids expressing LacZ (green squares) or DUX4 (blue circles). Expression was normalized to that of ACTB. Horizontal lines represent the mean. *** P ≤ 0.001, two-sided unpaired t -test. Full size image Dux and DUX4 have largely conserved amino acid sequences, particularly within the two DNA-binding homeodomains and the C-terminal region, which has previously been described to be responsible for recruiting p300–CBP 14 ( Supplementary Fig. 5b ). The mouse Dux tandem repeat encodes two main transcripts: full-length Dux (or Duxf3 ) and a variant named Gm4981 , which lacks the first homeodomain ( Supplementary Fig. 5a ). Both Dux and Gm4981 are expressed in mouse embryos before ZGA-defining genes and TEs (for example, mouse ERVL or MERVL) at the middle 2C stage, thus indicating that their products are probably functional homologs of DUX4 (ref. 15 ) ( Fig. 3a ). To consolidate these results, we used mESCs, a small percentage of which, at any given time, display a 2C-like transcriptome in culture, with expression of ZGA genes, including some from the MERVL promoter 16 , 17 . After analyzing single-cell RNA-seq data from 2C-like mESCs 18 , we confirmed that Dux transcripts were markedly enriched, as were early-ZGA RNAs such as Zscan4 , Zfp352 , and Cml2 ( Fig. 3b and Supplementary Fig. 6 ). We used CRISPR–Cas9-mediated genome editing to delete the Dux -containing macrosatellite repeat in mESCs expressing a GFP reporter under control of a MERVL promoter, thus resulting in a complete absence of GFP + 2C-like cells and in the loss of a large fraction of 2C-like-cell-specific transcripts ( Fig. 3c,d and Supplementary Fig. 7 ). Overexpression of Dux but not DUX4 rescued the 2C-like state in the mESC knockout (KO) clones ( Fig. 3e,f and Supplementary Figs. 8 and 9 ), albeit not in all cells in which Dux was produced ( Fig. 3g ). Interestingly, both mouse Dux and human DUX4 were able to induce the transcription of ZGA genes in the human 293T cell line ( Supplementary Fig. 10 ). Figure 3: Dux is necessary for formation of 2C-like mESCs. ( a ) Comparative expression of the two alternative transcripts of Dux , Dux (pink) and Gm4981 (orange), with genes (blue) and TEs (MERVL; green) specifically expressed during mouse ZGA. Solid lines, average; dashed lines, 95% confidence interval around the mean. Zy, zygote; e, early; m, middle; l, late; Bl, blastocyst. ( b ) Single-cell RNA-seq comparison between mESCs sorted for expression of both Tomato and GFP reporters driven by MERVL and Zscan4 promoters, respectively (revelators of 2C-like cells), and the double-negative population. Average gene expression was quantified, and the fold change between positive and negative cells is plotted. Dots are randomly distributed along the y axes. The top plot represents the kernel-density estimate of the middle-2C stage (blue line) and the rest of the genes (gray line). The Dux macrosatellite repeat was deleted in mESCs carrying a MERVL-GFP reporter by CRISPR–Cas9-mediated excision. ( c ) Fraction of GFP + cells in wild type (WT) or Dux -deleted cells. ( d ) RNA-seq analysis of WT and Dux -KO mESC clones. The dot plot displays the average gene expression of three independent clones from each cell type. ( e ) GFP expression in Dux -KO (blue circles) and WT (green squares) mESC clones carrying an integrated MERVL-GFP reporter and transiently expressing LacZ , DUX4 , Dux , or Gm4981 transgenes. Data are shown in log 10 scale. ( f ) RNA-seq analysis of Dux -KO mESC clones transiently expressing Dux or control. The dot plot displays the average gene expression of two independent clones from each cell type. ( g ) Dux -KO mESCs carrying an integrated MERVL-GFP reporter and transiently expressing a hemagglutinin (HA)-tagged form of Dux were stained for HA, and immunofluorescence was detected by confocal microscopy. DAPI nuclear stain, blue; GFP, green; HA, red. Horizontal bars in c and e represent the mean. *** P ≤ 0.001, two-sided unpaired t -test. Full size image After depletion of the transcriptional repressor tripartite motif-containing protein 28 (TRIM28; also known as KAP1) from mESCs, expression of 2C-specific genes increased, as previously observed 17 , as did levels of Dux transcripts ( Fig. 4a and Supplementary Fig. 11b–d ). Remarkably, this phenotype was completely abrogated in Dux -depleted mESCs ( Fig. 4b,c and Supplementary Figs. 9 and 11a–d ). Correspondingly, we found that TRIM28 associated with the 5′ end of the Dux gene and that trimethylated histone H3 Lys9 (H3K9me3), a canonical marker of TRIM28-mediated repression, was enriched at the Dux locus and was lost after knockdown (KD) of the heterochromatin inducer ( Fig. 4d and Supplementary Fig. 11e,f ). Figure 4: TRIM28 regulates formation of 2C-like mESCs by repressing Dux expression. ( a ) RNA-seq analysis of WT and Trim28 -KO mESCs. Average gene expression was quantified, and the fold change between KO and WT cells is plotted. Dots are randomly distributed along the y axes. The top plot represents the kernel-density estimate of genes specifically expressed in 2C-like mESCs (green line) and the rest of the genes (gray line). ( b ) WT (blue circles) and Dux -KO (green squares) mESC clones carrying an integrated MERVL-GFP reporter were transduced with lentiviral vectors encoding short hairpin RNAs (shRNAs) targeting Trim28 or a control (ctrl). 4 d later, GFP expression was quantified. Horizontal lines represent the mean. *** P ≤ 0.001, two-sided unpaired t -test. ( c ) RNA-seq of Trim28 -depleted or control Dux -KO mESC clones. The dot plot represents the average gene expression of three independent KO clones transduced with lentiviral vectors encoding a control or a Trim28 -specific shRNA. ( d ) Average coverage of ChIP–seq signal of TRIM28 (top plot; blue lines; two independent experiments) and H3K9me3 (bottom plot; two replicates) in control (red lines) and Trim28 -KD mESCs (green lines) around the Dux gene. Total input is represented in gray. ChIP–seq reads were mapped on the genome before the analysis focused on a 500-bp window around the main Dux gene. H3K9me3 peaks over the Dux macrosatellite repeat were called in only the control-KD mESCs (SICER; false discovery rate, 0.05). Full size image Finally, we addressed the role of Dux during mouse early embryonic development. To do so, we injected zygotes with plasmids encoding the Cas9 nuclease and either the two single guide RNAs (sgRNAs) used to generate Dux -KO mESCs or a nontargeting sgRNA control. We then determined the RNA profiles of 2C embryos approximately 7 h after the first cell division or monitored their ex vivo development into blastocysts over 4 d ( Fig. 5a ). We found that Dux-depleted embryos presented a major differentiation defect, in which most failed to reach the morula/blastocyst stage and did not exhibit transcriptional changes typical of ZGA, such as the induction of MERVL, Zscan4 , and several other tested early-ZGA genes, and the decrease in the Mpo maternal transcript ( Fig. 5b,c and Supplementary Fig. 12 ). Figure 5: Dux is necessary for mouse early embryonic development. ( a ) Schematic of the Dux loss-of-function experiment in mouse preimplantation embryos. ( b , c ) Zygote pronuclei were first injected with plasmids encoding the Cas9 nuclease and sgRNAs targeting the flanking region of the Dux macrosatellite repeat or a nontargeting sgRNA, then were either monitored for their ability to differentiate ex vivo ( b ) or collected at 2C stage for mRNA quantification ( c ). ( b ) Average percentage of embryos reaching the morula/blastocyst stages (white) or failing to differentiate (black, delayed/dead embryos; gray, defective morula/blastocyst) 4 d after pronuclear injection. The plot represents an average from 3 independent experiments with 16–23 embryos per condition. Two-sided Fisher's exact test was performed to compare the embryonic stage of control and Dux KO ( P = 1.54 × 10 −10 ). ( c ) Comparative expression of Dux , early-ZGA genes ( Zscan4 , Sp110 , B020004J07Rik , Usp17la , Tdpoz4 , Eif1a , Tcstv3 , and Cml2 ), 2C-restricted TE (MERVL, whose long terminal repeat and internal regions were detected with MT2_mm and MERVL-int primers, respectively), a gene ( Mpo ) whose expression decreases at ZGA, two genes ( Actb and Zbed3 ) stably expressed during preimplantation embryonic development and a control TE (IAPEz) in 15 2C stage embryos (5 from each of 3 independent experiments) 15–24 h after pronuclear injection with plasmids expressing Cas9 and control or Dux -specific sgRNAs. Box limits, twenty-fifth and seventy-fifth percentiles; lines in the boxes, median. Whiskers are shown as implemented in the ggplot2 package of R. The upper whisker extends from the hinge to the largest value, no further than 1.5× the interquartile range (IQR) from the hinge. The lower whisker extends from the hinge to the smallest value, at most 1.5× the IQR of the hinge. Expression was normalized to that of Actb . * P ≤ 0.05; ** P ≤ 0.01; *** P ≤ 0.001, two-sided Wilcoxon rank-sum test. Full size image In sum, our data highlight DUX genes as key regulators of early embryonic development. The demonstrated ability of DUX4 to recruit the p300–CBP complex and to induce local chromatin relaxation 14 , as well as the mechanism of action of Zelda, a master inducer of ZGA in Drosophila 19 , 20 , suggests that DUX proteins may act as pioneer factors in transcriptional activation, by opening chromatin around the TSSs of early-ZGA genes, thereby facilitating access by other transcription factors. Nonetheless, the genomic recruitment of pioneer factors such as OCT4, NANOG, and KLF4 may be hindered if heterochromatin marks are overly abundant at their target loci 21 . Many mouse ZGA genes are expressed from the long terminal repeats of endogenous retroviruses, which in mESC cells, are typically enriched in repressive marks 17 . At any given time, these marks may be relieved in only a small percentage of mESCs in culture. What drives this fluctuation remains to be determined. Furthermore, what controls expression of DUX genes themselves remains to be defined, although the conserved genomic localization of all placental-mammal DUX orthologs close to telomeric and subcentromeric regions suggests that this genomic context, characterized by high levels of repression, might be of primary relevance 4 , 5 , 22 . DUX genes indeed appear to be expressed only during events associated with major chromatin relaxation, for instance in early embryos and after loss of repression of the D4Z4 macrosatellite repeat in myoblasts of people with FSHD 23 , 24 . Our data indicated that TRIM28 plays a major role in mouse Dux repression, but the mild increase in cells entering the 2C state after TRIM28 depletion ( ∼ 5% of mESCs) and the demonstrated ability of several other transcriptional modulators (for example, SETDB1, EHMT2, HP1, CHAF1A, CHAF1B, RYBP, and KDM1A) to prevent cycling of mESCs through this state indicate that control of the Dux macrosatellite repeat is most probably multifactorial 16 , 25 , 26 , 27 , 28 , 29 . Broad derepression of the human and mouse DUX -containing repeats may similarly occur immediately after fertilization in either species. Future investigations of the chromatin states of these loci in early embryos should shed light on the epigenetic changes responsible for this process and on the nature of their molecular mediators. Methods Cell culture. mESCs, either WT or KO for Trim28 (ref. 30 ), and E14 mESCs containing the MERVL regulatory sequence driving expression of a 3×turboGFP-PEST 16 were cultured in feeder-free conditions on 0.1% gelatin-coated tissue culture plates in 2i medium, an N2B27 base medium supplemented with the MEK inhibitor PD0325901 (1 μM), the GSK3β inhibitor CHIR99021 (3 μM), and LIF. E14 mESCs expressed the main markers of pluripotency (RNA-seq). H1 hESCs (WA01, WiCell) were maintained in mTesRI (StemCell Technologies) on hES-qualified Matrigel (BD Biosciences). 293T cells were maintained in DMEM supplemented with 10% FCS. All cells were regularly tested for the absence of mycoplasma contamination. Plasmids and lentiviral vectors. The MT2/gag sequence was amplified from the pGL3 plasmid 29 , and the human PGK promoter was amplified from pRRLSIN.cPPT.R1R2.PGK-GFP.WPRE 30 and was cloned upstream of luciferase in pGL4.20. Supplementary Table 2 shows the primers used to obtain truncations of the MT2/gag sequence. sgRNAs targeting sequences flanking the 5′ and 3′ ends of the Dux-containing macrosatellite repeat were cloned into px459 (version 2) through a standard protocol 31 . Supplementary Table 2 shows the primers used to clone the sgRNAs. The pLKO.1-puromycin shRNA vector was used for the Trim28 KD 30 . The pLKO.1 vector was further modified to express a blasticidin- S -deaminase drug-resistance cassette in place of the puromycin N -acetyltransferase. The resulting pLKO.1-blasticidin backbone was used to clone shRNAs against the mouse Dux transcript. The sequence of the primers used to clone the Dux shRNA is shown in Supplementary Table 2 . The Gm4981 cDNA was cloned from the genome of E14 mESCs, and codon-optimized DUX4 and Dux were synthesized (Invitrogen). Gm4981 , DUX4 , Dux , and LacZ cDNAs were cloned in the pAIB HIV-1-based transfer vector, which also confers blasticidin resistance, by using an In-Fusion HD Cloning Kit (Clontech) 31 . pMD2-G encodes the vesicular stomatitis virus G protein (VSV-G). The minimal HIV-1 packaging plasmid 8.9NdSB, carrying a double mutation in the capsid protein (P90A/A92E), was used to achieve higher transduction of the weakly permissive mESCs 32 . Production of lentiviral vectors, and transduction and transfection of mammalian cells. Lentiviral vectors were produced by transfection of 293T cells with polyethylenimine (PEI) (Sigma) 32 . To generate stable KD, mESCs were transduced with empty pLKO.1 vector or vectors for expression of shRNA targeting Trim28 or Dux transcripts 30 . Cells were selected with 1 μg/ml puromycin or 3 μg/ml blasticidin starting 1 d after transduction. hESCs expressing LacZ and DUX4 were generated by transfection of the corresponding AIB plasmids with TransIT-LT1 Transfection Reagent (Mirus Bio), and nucleofection (with an Amaxa P3 Primary Cell 4D-Nucleofector X Kit) was used to engineer mESCs expressing LacZ, DUX4, Dux, and Gm4981. Creation of Dux -KO mESC lines. E14 mESCs containing the MERVL regulatory sequence driving expression of a 3×turboGFP-PEST were cotransfected with px459 plasmids encoding Cas9, the appropriate sgRNAs and a puromycin-resistance cassette 33 , through nucleofection (Amaxa P3 Primary Cell 4D-Nucleofector X Kit). 24 h later, the cells were selected for 48 h with 1 μg/ml puromycin, single-cell cloned by serial dilution, and expanded, and their DNA was extracted to detect the presence of WT and/or KO alleles. Three WT and three homozygous Dux -KO clones were selected and used in this study. Luciferase assay. 293T cells or E14 mESCs were cotransfected with the various pGL4.20 derivatives, the Renilla plasmid and the pAIB transfer vector encoding LacZ, Dux, Gm4981, or DUX4, by using Lipofectamine 3000 (Invitrogen). Luciferase activity was quantified 24 h after transfection. Firefly luciferase activity was normalized to the activity of Renilla luciferase. Light emission was measured on a luminescence plate reader. Immunofluorescence assay. mESC clones expressing an HA-tagged DUX protein were fixed for 20 min with 4% paraformaldehyde, permeabilized for 5 min with 0.1% Triton X-100, and blocked for 30 min with 1% BSA in PBS. Cells were then incubated for 1 h with anti-HA.11 (Covance), anti-NANOG (61628, Active Motif), or anti-SOX2 (39824, Active Motif) antibodies diluted in PBS with 1% BSA. Validation information is available on the manufacturers' websites. After being washed three times, the cells were incubated with anti-mouse (for HA) or anti-rabbit (for NANOG and SOX2) Alexa Fluor 647–conjugated secondary antibodies for 1 h and washed again three times. Every step until this point was carried out with cells in suspension. Pelleted cells were then resuspended in VECTASHIELD Mounting Medium with DAPI (Vector Laboratories) and mounted on the coverslip. The slides were viewed with a Zeiss LSM700 confocal microscope. Fluorescence-activated cell sorting (FACS). FACS analysis was performed with a BD FACScan system. Trim28 -KD mESCs containing the MT2/gag-GFP reporter were subjected to FACS sorting with an AriaII instrument (BD Biosciences). Standard PCR, RT–PCR and RNA sequencing. For the genotyping of Dux WT and KO alleles, genomic DNA was extracted with a DNeasy Blood & Tissue Kit (Qiagen), and the specific PCR products were amplified with PCR Master Mix 2× (Thermo Scientific) combined with the appropriate primers (design in Supplementary Fig. 6a ; primer sequences in Supplementary Table 2 ). Total RNA from cell lines was isolated with a High Pure RNA Isolation Kit (Roche). cDNA was prepared with SuperScript II reverse transcriptase (Invitrogen). An Ambion Single Cell-to-CT Kit (Thermo Fisher) was used for RNA extraction, cDNA conversion, and mRNA preamplification of 2C-stage embryos. Primers listed in Supplementary Table 2 were used for SYBR Green qPCR (Applied Biosystems). Library preparation and 150-bp paired-end RNA-seq were performed with standard Illumina procedures for the NextSeq 500 platform ( GSE94325 ). ChIP and ChIP–seq. ChIP and library preparation were performed as previously described 30 . DUX4-HA ChIP was done with anti-HA.11 (Covance) antibody. Sequencing of TRIM28 and H3K9me3 ChIP was performed with the Illumina HiSeq 2500 platform in 100-bp-read runs. Sequencing of DUX4 ChIP was performed with an Illumina NextSeq 500 in 75-bp paired-end-read runs. RNA-seq data-set preprocessing. Single-cell RNA-seq of human and mouse early embryo development ( GSE36552 and GSE45719 respectively), single-cell RNA-seq of 2C-like cells ( E-MTAB-5058 ), DUX4 overexpression in human myoblasts ( GSE45883 ), and TRIM28-KO ( GSE74278 ) data sets were downloaded from different repositories (NCBI GEO and EBI ArrayExpress) 34 , 35 . Reads were mapped to the human genome (hg19) or mouse genome (mm9) with TopHat (v2.0.11) 36 in sensitive mode (the exact parameters were: tophat -g 1 --no-novel-juncs --no-novel-indels -G $gtf --transcriptome-index $ transcriptome --b2-sensitive -o $localdir $index $reads1 $reads2). Gene counts were generated with HTSeq-count. Normalization for sequencing depth and differential gene expression analysis was performed with Voom 37 , as implemented in the limma package of Bioconductor 38 . TEs overlapping exons were removed from the analysis. Counts per TE integrant (genomic loci) were generated with the multiBamCov tool from bedtools software 39 . Normalization for sequencing depth was performed with Voom, with the total number of reads on genes as the size factor. To compute the total number of reads per TE family, counts on all integrants of each family were summed. Analysis of single-cell expression data from human and mouse embryonic stages. For every embryonic stage, we identified the genes that had different expression levels compared with those at other stages 10 , by using a moderated F test (comparing the interest group against every other), as implemented in the limma package of Bioconductor. Genes were selected as being expressed in a specific stage if they had a significant P value (<0.05 after adjustment for multiple testing with the Benjamini and Hochberg method) and an average fold change greater than ten, with respect to the other embryonic stages. We additionally removed all genes exhibiting a 1.1-fold-higher expression in any of the embryonic stages compared with the stage analyzed ( Supplementary Fig. 1a ). Notably, through this approach, a gene may be marked as being expressed in more than one stage. Codes are available upon request. Correspondence between DUX4 overexpression and single-cell expression data from human embryonic stages. For every stage, we classified the genes in four patterns of expression by performing hierarchical clustering (with Pearson's correlation as the distance measure and the complete agglomeration method). Figure 1b shows the two most relevant patterns derived from the 4C and 8C stages. Expression of the genes identified with this method was then compared between DUX4- and GFP-overexpressing human myoblast cells. For a gene to be considered differentially expressed, a P value (after multiple testing correction with the Benjamini and Hochberg method) <0.05 and a fold change greater than two were imposed. A moderated t -test was used as the statistical test, as implemented in the limma package of Bioconductor. ChIP–seq data processing. The ChIP–seq data set of DUX4 overexpressed in human myoblasts ( GSE94325 ) was downloaded from GEO. Reads were mapped to the human genome assembly hg19 with Bowtie2 by using the sensitive-local mode 40 . SICER was used to call histone-mark peaks 41 . For the peaks that did not represent histone marks, we used MACS (with default parameters) for single-end data and used MACS2 (with the parameters macs2 callpeak -t $chipbam -c $tibam -f BAM -g $org -n $name -B -q 0.01 --format BAMPE) for paired-end data 42 . Both SICER peaks with an FDR >0.05 and MACS peaks with a score <50 were discarded. RSAT was used for motif discovery and to compute motif abundance 43 . To compute the percentage of bound TE integrants in each family, we used the bedtools suite. Coverage plots. ChIP–seq signals on features of interest were extracted from bigWig files beforehand and normalized for sequencing depth (in reads per hundred million). Each signal was then smoothed with a running window average of 75 bp for DUX4, 250 bp for Trim28, and 500 bp for H3K9me3. Finally, the mean and s.e.m. of the signals were computed and plotted for each set of features of interest. Scripts are available upon request. Pronuclear injection of mouse embryos. Pronuclear injection was performed according to the standard protocol of the Transgenic Core Facility of EPFL. In brief, five-week-old B6D2F1 mice were used as egg donors. Mice were injected with PMSG (10 IU) and then with HCG (10 IU) 48 h later. After females were mated with B6D2F1 males, zygotes were collected and kept in KSOM pregassed under 5% CO 2 at 37 °C. Embryos were then transferred to M2 medium and microinjected with 10 ng/μg of either a px459 plasmid containing a nontargeting sgRNA or a mix of the two plasmids used to obtain the KO in mESCs, in injection buffer (10 mM Tris-HCl, pH 7.5, 0.1 mM EDTA, pH 8, and 100 mM NaCl). After microinjection, embryos were cultured in KSOM at 37 °C in 5% CO 2 for 4 d. In each of three independent experiments, five embryos per condition were collected approximately 7 h after the first cell division (2C formation) for qPCR analysis, and the remaining embryos were differentiated. At day 4, all the fertilized embryos (16–23 per condition) were classified according to their developmental state. Randomization and blinded outcome assessment were not applied. All animal experiments were approved by the local veterinary office at EPFL and were carried out in accordance with the EU Directive (2010/63/EU) for the care and use of laboratory animals. No statistical method was used to predetermine sample size. Sample sizes and statistical tests. We used nonparametric statistical tests (two-sided Wilcoxon test), when we had sufficient sample sizes (low-cell-number qPCR). Otherwise, we used two-sided unpaired t -tests (standard qPCR and FACS). Fisher's exact test was used to test for differences in proportions in contingency tables. Data availability. RNA-seq and ChIP–seq data generated in this study have been deposited in the NCBI Gene Expression Omnibus (GEO) database under accession number GSE94325 . Additional information Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Accession codes Primary accessions Gene Expression Omnibus GSE94325 Referenced accessions ArrayExpress E-MTAB-5058 Gene Expression Omnibus GSE36552 GSE45719 GSE45883 GSE74278

The formation of a human embryo starts with the fertilization of the oocyte by the sperm cell. This yields the zygote, the primordial cell that carries one copy each of the maternal and paternal genomes. However, this genetic information starts being expressed only after the zygote divides a couple of times. But what triggers this process, called "zygotic genome activation", was unknown until now. EPFL scientists have just found that members of the DUX family of proteins are responsible for igniting the gene expression program of the nascent embryo. Published in Nature Genetics, this discovery is a milestone for developmental biology. Alberto de Iaco, a postdoc in the lab of Didier Trono at EPFL, drew upon a seemingly irrelevant study of patients suffering from a form of muscular dystrophy where mutations lead to the production in muscle cells of a protein called DUX4, which is normally detected only at the earliest stage of human embryonic development. De Iaco also found that when DUX4 is forcibly produced in muscle cells, it turns on a whole set of genes that are expressed during zygotic genome activation. This was what first suggested that DUX4 could be the key regulator of this seminal event. To confirm this, the researchers analyzed publicly available data to determine what components of the human genome are expressed during the first few days of embryonic development. They found that DUX4 is one of the very first genes expressed at this stage, releasing a high concentration of its protein product just before zygotic genome activation. In line with this lead, the scientists could show that the DUX4 protein binds to the regulatory region of genes that are induced during zygotic genome activation, stimulating their expression. Next, they looked at mouse embryonic stem cells, which contain the mouse version of the DUX4 gene (called simply DUX). When in culture, a small fraction of these cells exhibit a any given time the gene expression pattern of 2-cell stage embryos, before cycling back to the features of more advanced embryonic cells. But when the EPFL researchers deleted the DUX gene, this process stopped, the appearance of the 2-cell stage-like subpopulation was suppressed. The final piece of evidence came when the EPFL scientists removed the DUX gene from fertilized mouse oocytes using CRISPR/Cas9 genome editing. This prevented zygotic genome activation altogether, and precluded the growth of embryos beyond the first couple of cell divisions. The study points to DUX4, and by extension the DUX family of proteins, as the master regulator responsible for kick-starting genome expression at the earliest stage of embryonic life in humans, mouse and probably all placental mammals. "An old enigma is solved," says Didier Trono. "The study sheds light on what triggers the genetic program that ultimately makes us what we are. It can also help us understand certain cases of infertility and perhaps guide the development of new treatments for DUX-related muscle dystrophies". He and his team are now curious about what could unleash, in the first few hours of our embryonic life, the ephemeral yet so crucial production of this master regulator.

10.1038/ng.3858

Physics

Scientists produce 3-D chemical maps of single bacteria

Tiffany W. Victor et al. X-ray Fluorescence Nanotomography of Single Bacteria with a Sub-15 nm Beam, Scientific Reports (2018). DOI: 10.1038/s41598-018-31461-y Journal information: Scientific Reports

http://dx.doi.org/10.1038/s41598-018-31461-y

https://phys.org/news/2018-11-scientists-d-chemical-bacteria.html

Abstract X-ray Fluorescence (XRF) microscopy is a growing approach for imaging the trace element concentration, distribution, and speciation in biological cells at the nanoscale. Moreover, three-dimensional nanotomography provides the added advantage of imaging subcellular structure and chemical identity in three dimensions without the need for staining or sectioning of cells. To date, technical challenges in X-ray optics, sample preparation, and detection sensitivity have limited the use of XRF nanotomography in this area. Here, XRF nanotomography was used to image the elemental distribution in individual E . coli bacterial cells using a sub-15 nm beam at the Hard X-ray Nanoprobe beamline (HXN, 3-ID) at NSLS-II. These measurements were simultaneously combined with ptychography to image structural components of the cells. The cells were embedded in small (3–20 µm) sodium chloride crystals, which provided a non-aqueous matrix to retain the three-dimensional structure of the E . coli while collecting data at room temperature. Results showed a generally uniform distribution of calcium in the cells, but an inhomogeneous zinc distribution, most notably with concentrated regions of zinc at the polar ends of the cells. This work demonstrates that simultaneous two-dimensional ptychography and XRF nanotomography can be performed with a sub-15 nm beam size on unfrozen biological cells to co-localize elemental distribution and nanostructure simultaneously. Introduction Trace elements are essential for carrying out biochemical reactions and act as structural components in cells. The appreciation of the important roles that trace elements, especially metals, play in cell metabolism has resulted in an increase in the number of studies in this area 1 , 2 , 3 , 4 , 5 . However, to truly understand the functions of trace elements in cells and tissues, the distribution within a cell must be imaged and quantified in the native cellular environment, which has proven to be difficult. Hard XRF microscopy is a well-suited microanalytical technique for assessing the elemental distribution in a wide range of materials including whole cells and tissues 6 , 7 , 8 . In XRF microscopy, multiple elements can be mapped simultaneously, which provides precise elemental co-localization and speciation. High detection sensitivity down to sub-parts-per-million has been demonstrated in biological studies 9 , 10 . To understand the functions of trace elements in biological cells, the X-ray probe must resolve subcellular compartments on the sub-micron to nanometer scale such as mammalian nuclei, mitochondria, and cell membranes, which measure ~6 μm, ~1 μm and ~10 nm, respectively 11 . In prokaryotic cells such as E . coli (1 μm × 3 μm), the outer and inner cell membranes are each ~10–15 nm wide, while ribosomes are about 20 nm in size 12 . Nanoscale imaging with X-rays requires high-resolution nano-focusing optics like Kirkpatrick-Baez mirrors 13 , 14 , 15 and Fresnel zone plate optics 16 . A more sophisticated, higher-resolving, nano-focusing approach is to use Multilayer Laue Lenses (MLLs) 17 , 18 , 19 , incorporated into a specialized X-ray microscope located at a beamline with infrastructure for good vibration isolation and thermal stability 20 , 21 , 22 . It is well-accepted that X-ray imaging of living biological specimens can be a challenge due to radiation-induced damage, so cells are frequently imaged in the fixed/dried or frozen-hydrated state 23 . While fixed/dried cells are least subject to radiation damage from water hydrolysis and free radical production, the subcellular structure can be compromised, especially at the nanoscale. Frozen-hydrated cells represent the most natural state of the cell while reducing the effects of radiation damage; however, this method requires the cells to be flash frozen, e.g. with a plunge freezer, and imaged in a cryostage. Since optical configurations of nanotomography instruments with MLLs 21 have very short working distances and rotation constraints, integration of a cryostage is a significant technical challenge. In addition, since the trace elements are typically micro- to nanomolar in concentration, sensitive detectors with a large solid angle situated close to the sample are required for quantification, further increasing technical complexity. Whereas two-dimensional XRF microscopy is the most straightforward approach for imaging cells and tissues, these two-dimensional data can be ambiguous when imaging three-dimensional objects, e.g. when aiming to differentiate between elements located on the surface of a cell from those distributed throughout the cell. At present, only a few three-dimensional XRF tomography studies from whole biological specimens have been reported. Kim et al . have studied the role of the vacuolar iron uptake transporter in the uptake and distribution of iron in the Arabidopsis seed, which was performed at 12 μm spatial resolution 24 . More recently, de Jonge et al . performed three-dimensional XRF tomography of a whole diatom, Cyclotella meneghiniana , at 400 nm spatial resolution using 150 nm voxels over a 15 μm field of view 9 . This work elegantly showed the distribution of several species including silicon and zinc within the organelles including the siliceous frustule, a diatom’s cell wall. Most surprising was the detection of Fe and Mn rings in the cell wall that would not have been detectable using only two-dimensional projections. To date, XRF nanotomography has been limited by sub-micron spatial resolution and weak absorption contrast in biological cells, inhibiting its ability to distinguish the ultra-structures of cell organelles 25 . In contrast, coherent diffraction imaging techniques, such as ptychography, can be used to image the fine structural components in samples 26 but cannot determine the distribution of trace elements. However, by combining XRF nanotomography and ptychography, it is possible to obtain complementary data on elemental and structural composition, respectively, at the nanoscale from biological cells 8 , 27 . The results described here represent a combination of two-dimensional ptychography and XRF nanotomography performed on E . coli cells using a MLL microscope with a sub-15 nm focused beam. Methods Cell culture LB media (10 mL) was inoculated with a single colony of BL21 (DE3) E . coli cells and the culture grown at 37 °C with shaking to an OD of 0.6–0.8. Cells were harvested in 1 mL aliquots via centrifugation at 3,500 RPM, flash frozen in liquid nitrogen, and stored at −80 °C. Sample preparation, light microscopy and scanning electron microscopy A 20 μl solution containing 2 μM NaCl, 0.01 OD E . coli cells, and 0.025 OD 100 nm gold nanoballs (CytoDiagnostics, #GRF-100-20) was mixed thoroughly. The gold nanoballs were added as fiducial markers for the tomography reconstruction. One microliter of this solution was pipetted onto a silicon chip sample holder with Cr fiduciary grids (Fig. 1C,D ), also known as a “diving board” (substrate area 1.4 mm by 0.4 mm with a 10 μm thickness) (Norcada NCT4155P-III-Cr), and allowed to dry overnight. The diving board was then glued onto a stainless-steel insect pin (0.2 mm thick, 16.5 mm tall) (Fig. 1C ) and mounted on the sample mount (Fig. 1B ) for the X-ray microscope (Fig. 1A ). The diving board has 5 μm wide chromium-patterned gridlines for sample alignment in the X-ray microscope. Images of cells prepared on the diving boards were taken using both a visible light microscope (Nikon Eclipse LVDIA-N) and a scanning electron microscope (SEM, JEOL 300) for cell identification, correlation, and navigation during data acquisition at the beamline. Figure 1 ( A ) X-ray microscope at the Hard X-ray Nanoprobe (HXN, 3-ID) beamline at NSLS-II. ( B ) Sample mounted on the rotation stage for tomography, ( C ) E . coli bacteria deposited on a Si substrate (“diving board”) that was mounted to an insect pin. ( D ) Size comparison between the Si substrate and a US penny. Substrate is 1.4 × 0.4 mm. Full size image Synchrotron X-ray fluorescence microscopy and ptychography measurements The distribution of elements was mapped using nano-Mii (Nanoscale Multimodal Imaging Instrument) at beamline 3-ID (HXN) at the National Synchrotron Light Source II (Fig. 1A ). Beamline 3-ID provides a scanning X-ray microscope capable of multimodal imaging including absorption, fluorescence, differential phase contrast, and ptychography 28 . In this study, a 12 keV incident beam was focused to a sub-15 nm spot (14 nm horizontal and 12 nm vertical) using Multilayer Laue Lenses (MLLs). The coherent illumination was selected by a 50 × 30 µm secondary source aperture placed about 15 m in front of the nanofocusing optics. A three-element silicon drift detector (Vortex-ME3) was placed perpendicular to the incident beam to collect the fluorescence signals, and the transmitted scattering data for differential phase contrast or ptychography reconstructions were acquired by using a pixelated area detector (Merlin) with a pixel size of 55 µm at a distance of 0.5 m downstream the sample. Two-dimensional fluorescence/ptychography scans, via continuous fly-scan, were performed using a step size of 20 nm/pixel and a dwell time of 250 ms. For the tomography data set, the cells were scanned from -39 to 132 degrees at 3 degree intervals using a dwell time of 100 ms and a 20 nm step size. Each projection was ~14 min. After each rotation interval, a short coarse scan (~1 min) was used to center the sample in the field of view of the projection image, resulting in a total imaging time for a complete cell of approximately 15 hours, excluding overhead. The fitting of the X-ray fluorescence data was accomplished using PyXRF, an X-ray fluorescence analysis package developed at NSLS-II 29 . For the fitting process, the summed spectrum was first fitted using the non-linear least squares method to determine the global parameters such as the energy-calibration values, widths of the global peaks, and parameters related to the Compton and elastic peaks. Once the correct global parameters were obtained, the peak area of each element under the XRF spectrum of each single pixel was then fitted using a nonnegative least squares approach. TomoPy was used for the tomography reconstruction of the data sets 30 . The projection images were aligned by calculating the cross-correlation of a pair of images taken at successive projection angles. The aligned image stack was reconstructed using the maximum-likelihood method coded in the TomoPy package 30 , and 20 iterations were used to reconstruct each slice. To correct for any attenuation (due to the thickness of the diving board, since the XRF photons are detected through the diving board for some projections), the total intensity of the Zn fluorescence images was normalized to be constant for all the images taken at different projection angles. For the ptychography reconstruction, a 96 × 96 array was cropped from the recorded raw dataset, which gave a reconstruction pixel size of about 10 nm. The cropped data array was fed into 100 iterations of a difference map algorithm for ptychography reconstruction 31 . From the preliminary reconstruction result, we noticed that the measurement was not performed exactly at the focal plane (due to run-out error of the rotary stage). In order to achieve the best reconstruction image quality, the exact illumination function at the measurement plane was searched by running the reconstruction process using a series of propagated wavefronts as the initial probe guesses, and the one that gave the best contrast was selected. A series of illumination functions propagated to various distances from the focus were tested, and the one that gave the best reconstruction image quality was selected. To assist the reconstruction convergence, the allowable ranges for the amplitude and phase parts of the E . coli image were constrained to [0.95, 1.0] and [−0.1, 0.0], respectively. The last 20 iterations were averaged to give the final reconstructed image. Results E . coli cells were deposited on the diving board and examined under a visible light microscope in epi-brightfield mode followed by visualization using SEM. The SEM image (Fig. 2A ) displays E . coli cells (typically 1 μm × 3 μm in size) embedded in small sodium chloride crystals (3–20 μm) that formed on the silicon substrate, where the cells appear as dark (negative contrast) regions in the crystals. The inset in Fig. 2A shows a magnified view of a group of cells lined up side-by-side within a sodium chloride crystal. Figure 2 ( A ) SEM image showing E . coli bacteria embedded in NaCl crystals on the Si substrate with Cr-patterned fiducial grids. Cells appear as negative (dark) contrast regions in images. Scale bar is 20 μm. XRF images showing ( B ) Zn, ( C ) Ca, and ( D ) Cl distribution in the E . coli cells. ( E ) XRF image showing the co-localization of Zn, Ca and Cl in the E . coli cells. ( F ) Average XRF spectrum from the sample. Scale bar is 1 μm in ( B – E ). Full size image Two-dimensional elemental images (Fig. 2B–E ) were taken using XRF microscopy, mapping the distribution and co-localization of zinc, calcium, and chlorine. The chlorine image (Fig. 2D ) shows the location of the sodium chloride crystals and the dark regions indicate the areas occupied by the cells embedded in the crystal, which correlates well with the SEM image in Fig. 2A (inset). Most notably, zinc exhibits an inhomogeneous distribution (Fig. 2B ) whereas a more uniform distribution of calcium (Fig. 2C ) is observed throughout the cell. An XRF spectrum showing elements present in the sample is seen in Fig. 2F . It should be noted that peaks from Fe, Ni, and W are contributions from the microscope, the Cr signal arises from the bars on the substrate, and the Au signal comes from the nanoball fiducials. In Fig. 3A , the two-dimensional zinc distribution is shown in a series of two cells lined end-to-end. Again, the distribution of zinc is uneven within the cells and varies from cell to cell. The gold elemental map (Fig. 3B ) was used to localize the gold nanoballs, which were added to the cell solution as fiducial markers for the alignment and reconstruction of the tomography data. These gold nanoballs are also visible in the ptychography reconstructed phase image (Fig. 3C ) along with the clear outline of two cells that appear to be dividing. An overlay of the ptychography and XRF images can be seen in Fig. 3D . It should be noted that the edge enhanced contrast in the ptychography image is most likely due to the propagation uncertainty along the beam direction 32 . As the reconstruction plane in the ptychography method is determined by the illumination function 26 , the object transmission functions that propagate within the depth of focus range (±2 μm for MLLs) are allowed in the phase-retrieval process. For an object with extremely small contrast, which is the case for this study, the edge-enhanced artifact is more pronounced than the objects with higher contrast. Figure 3 XRF images showing the ( A ) Zn and ( B ) Au distribution in several E . coli cells lined end-to-end. ( C ) Ptychography reconstructed phase image of the cells from ( A ) showing the cell boundaries and the 100 nm Au nanoballs that were added to the cell solution as fiducial markers for alignment and tomography reconstruction. ( D ) Overlay of the ptychography and XRF images from ( A – C ). Scale bar is 1 μm. Full size image The zinc distribution was further investigated in a three-dimensional XRF nanotomography dataset of a single E . coli cell (Fig. 4 ). The tomography reconstruction in Fig. 4A as well as the 2D cut in Fig. 4C clearly show that zinc is inhomogeneously distributed throughout the cell. The 2D cut through the cell in Fig. 4B shows that Zn is heavily concentrated at one pole of the cell. Figure 4 Zinc XRF nanotomography profile of an E . coli cell. ( A ) Three-dimensional view of the cell. The green box indicates the cell orientation during the tomography data collection. ( B ) Zinc distribution in a slice of the cell through the plane outlined in the yellow box in ( A ) showing the dense distribution of Zn at the polar end of the cell. ( C ) Zinc distribution in a slice of the cell through the plane outlined in blue in ( A ) showing that zinc is unevenly distributed throughout the cell. Scale bar is 500 nm. Full size image Since bacteria do not have many well-defined subcellular structures, the spatial resolution of the two-dimensional XRF images was estimated using a power spectral analysis 27 , 33 . According to Deng et al ., for low-photon statistics images like XRF images, one can use the fact that Poisson fluctuations (‘shot noises’) are uncorrelated from pixel to pixel so that, at high spatial frequencies, one arrives at a noise floor in Fourier power analysis. Figure 5 shows the result of the power spectral analysis, where we found an estimated spatial resolution of approximately 36 nm (horiz) × 36 nm (vert) using the 2D zinc fluorescence image from the tomography projection at 15 ° . Figure 5 Estimate of the spatial resolution of the Zn XRF E . coli image using power spectral analysis. The power spectrum signal declines with the spatial frequency before reaching a noise floor. The cutoff point is at a half period of approximately 36 × 36 nm. The Zn XRF image is from a 2D projection using a 20 nm step size and 0.1 s dwell time at 15°. Full size image Another method to estimate the spatial resolution is by using the Fourier Ring Calculation approach which requires two independent datasets acquired under the same conditions 34 , 35 . Based on previously published work 22 , the beam size routinely produced by the MLLs is about 14 nm x 12 nm when performed with 2D imaging using a well-aligned sample. This results in a depth of focus of ±2 μm for the HXN MLL microscope. However, the rotation stage used for these experiments had a poor run-out error of ~5 μm. Thus, for some rotation angles, the sample moved away from its ideal rotation center. Although the sample was centered in the direction perpendicular to the beam, the run out along the direction parallel to the beam was not corrected. Considering these factors, the XRF image resolution in this study was larger than the published work by Yan, et al . 22 . Discussion In this study, X-ray fluorescence microscopy, nanotomography, and ptychography were performed on individual E . coli cells using a sub-15 nm beam size with a MLL microscope. The distributions of chlorine, calcium, and zinc were imaged, focusing on the sub-cellular distribution of zinc in the cell. E . coli cells have few internal structures; they possess a nucleoid, which is a highly condensed and membrane-free, irregular structure containing the genetic material of the bacterial cell 36 , 37 and a dense distribution of ribosomes. The tomography results showed an inhomogeneous zinc distribution in the cell, with a higher concentration at one pole. Most of the internal labile zinc is thought to be stored in the ribosomes 38 , 39 , 40 . Generally, ribosomes are distributed among and around the central nucleoid. This ribosome distribution leads to ribosome-rich regions especially at the polar end-caps of the cell and in a cylindrical shell surrounding the nucleoid 36 which would explain the higher amounts of zinc observed in these regions 39 , 40 , 41 . In rapidly growing E . coli , it is believed that the majority of the translation events are presumably carried out on mature, freely diffusing mRNAs within this localized ribosome-rich, “protein-factory” region using poly-ribosomes 42 . In bacteria, zinc can play a structural, regulatory, or catalytic role in proteins and is tightly controlled by dedicated systems for high-affinity zinc uptake and export systems involving the cell membranes 40 , 43 . One of the challenges of using a MLL microscope is the short working distance between the sample position and the order-sorting aperture (OSA), which was ~0.8 mm in this experiment. Because there is a short working space available to rotate the sample to collect XRF maps at different angles, it was necessary to have a sample and substrate small enough to prevent sample collision with the OSA. To mitigate this, several iterations of sample holders were tested including Mitegen’s Microcrystal Mounts TM , direct addition of the cells embedded in NaCl crystals to a tungsten pin, and different sizes of the HXN diving board Si-chip sample holder. In the end, we chose a Si-chip substrate that was less than 0.5 mm wide, which was small enough to safely rotate the sample but large enough to accommodate more than 100 cells for the experiment. A further limitation of nanoscale X-ray imaging, in particular for living biological specimens, is the inherent radiation damage that occurs with hard X-ray exposure, which can change the structure and chemistry of samples. Generally, samples are measured dried or under cryogenic conditions to mitigate chemical and structural changes when exposed to an intense X-ray beam. With the short working distance of the MLL microscope, sample transfer and data collection at cryogenic conditions presented a significant technical challenge. Instead, we chose to embed the bacteria in small (5–20 µm wide, 1–2 µm thick) sodium chloride crystals, which was a simple and straightforward approach to provide a non-aqueous matrix that retained the three-dimensional structure of individual cells. Furthermore, the XRF and SEM contrast provided by chlorine greatly enhanced the ability to locate individual cells on the sample substrate, also minimizing the radiation dose. While this method may not be applicable to all cells, it is likely to be most effective for many nonadherent cell types including bacteria, archaea, and algae that can easily be suspended in the salt solution prior to drying. The estimated radiation dose imparted on the sample for the duration of the tomography data collection was 1.3 × 10 9 Gray. No signs of cell shrinkage or compositional changes were observed. This dose was calculated assuming the main consistency of the cell material was of protein composition of H 48.6 C 32.9 N 8.9 O 8.9 S 0.6 with a condensed density of 1.35 g/cm 3 44 , 45 . This dosage was similar to those imparted on biological samples reported by others without seeing any significant mass loss 33 , 46 , albeit these studies were done under cryogenic conditions. Having the ability to assess the zinc distribution in E . coli – as is possible with XRF nanotomography – is essential for establishing the role of zinc in cellular processes and disease. Importantly in bacteria, infectivity can readily be altered by varying levels of zinc in the host system 47 . More broadly, zinc plays a vital role in the physiology of all organisms and its homeostasis is widely studied in areas such as aging, neurodegenerative diseases, cancer, the immune system, and energy metabolism. This work represents one example of how XRF nanotomography and ptychography can be used to image trace element distribution/concentration and structural morphology, respectively, in biological cells. It was made possible through the development of a synchrotron source with high brightness and coherence, nanofocusing optics, and exceptional beam and sample stability along with a new approach for sample preparation and image collection. Conclusions In summary, this study demonstrates that XRF nanotomography can be used to image the distribution of trace elements in intact biological cells using a sub-15 nm beam in three dimensions, and that simultaneous ptychography can be used for the co-localization of the trace elements with subcellular structures. An approach such as this presents new possibilities in understanding subcellular biochemistry in individual organelles and other subcellular compartments, which are usually analyzed at the organelle population level.

Scientists at the National Synchrotron Light Source II (NSLS-II)—a U.S. Department of Energy (DOE) Office of Science User Facility at DOE's Brookhaven National Laboratory—have used ultrabright x-rays to image single bacteria with higher spatial resolution than ever before. Their work, published in Scientific Reports, demonstrates an X-ray imaging technique, called X-ray fluorescence microscopy (XRF), as an effective approach to produce 3-D images of small biological samples. "For the very first time, we used nanoscale XRF to image bacteria down to the resolution of a cell membrane," said Lisa Miller, a scientist at NSLS-II and a co-author of the paper. "Imaging cells at the level of the membrane is critical for understanding the cell's role in various diseases and developing advanced medical treatments." The record-breaking resolution of the X-ray images was made possible by the advanced capabilities of the Hard X-ray Nanoprobe (HXN) beamline, an experimental station at NSLS-II with novel nanofocusing optics and exceptional stability. "HXN is the first XRF beamline to generate a 3-D image with this kind of resolution," Miller said. While other imaging techniques, such as electron microscopy, can image the structure of a cell membrane with very high resolution, these techniques are unable to provide chemical information on the cell. At HXN, the researchers could produce 3-D chemical maps of their samples, identifying where trace elements are found throughout the cell. "At HXN, we take an image of a sample at one angle, rotate the sample to the next angle, take another image, and so on," said Tiffany Victor, lead author of the study and a scientist at NSLS-II. "Each image shows the chemical profile of the sample at that orientation. Then, we can merge those profiles together to create a 3-D image." XRF images show the zinc (B), calcium (C), chlorine (D) distributions in the single bacteria. XRF image E shows all three elements in the cell. Image A shows bacteria embedded in sodium chloride crystals. Credit: Brookhaven National Laboratory Miller added, "Obtaining an XRF 3-D image is like comparing a regular X-ray you can get at the doctor's office to a CT scan." The images produced by HXN revealed that two trace elements, calcium and zinc, had unique spatial distributions in the bacterial cell. "We believe the zinc is associated with the ribosomes in the bacteria," Victor said. "Bacteria don't have a lot of cellular organelles, unlike a eukaryotic (complex) cell that has mitochondria, a nucleus, and many other organelles. So, it's not the most exciting sample to image, but it's a nice model system that demonstrates the imaging technique superbly." Yong Chu, who is the lead beamline scientist at HXN, says the imaging technique is also applicable to many other areas of research. "This 3-D chemical imaging or fluorescence nanotomography technique is gaining popularity in other scientific fields," Chu said. "For example, we can visualize how the internal structure of a battery is transforming while it is being charged and discharged." A 3-D view of the single bacteria produced through XRF. Credit: Brookhaven National Laboratory In addition to breaking the technical barriers on X-ray imaging resolution with this technique, the researchers developed a new method for imaging the bacteria at room temperature during the X-ray measurements. "Ideally, XRF imaging should be performed on frozen biological samples that are cryo-preserved to prevent radiation damage and to obtain a more physiologically relevant understanding of cellular processes," Victor said. "Because of the space constraints in HXN's sample chamber, we weren't able to study the sample using a cryostage. Instead, we embedded the cells in small sodium chloride crystals and imaged the cells at room temperature. The sodium chloride crystals maintained the rod-like shape of the cells, and they made the cells easier to locate, reducing the run time of our experiments." The researchers say that demonstrating the efficacy of the X-ray imaging technique, as well as the sample preparation method, was the first step in a larger project to image trace elements in other biological cells at the nanoscale. The team is particularly interested in copper's role in neuron death in Alzheimer's disease. "Trace elements like iron, copper, and zinc are nutritionally essential, but they can also play a role in disease," Miller said. "We're seeking to understand the subcellular location and function of metal-containing proteins in the disease process to help develop effective therapies."

10.1038/s41598-018-31461-y

Biology

Mutations that affect aging: More common than we thought?

"Deleterious mutations show increasing negative effects with age in Drosophila melanogaster", Martin I. Brengdahl, Christopher M. Kimber, Phoebe Elias, Josephine Thompson and Urban Friberg, (2020), BMC Biology, published online 30 September, DOI: 10.1186/s12915-020-00858-5 Journal information: BMC Biology

http://dx.doi.org/10.1186/s12915-020-00858-5

https://phys.org/news/2020-09-mutations-affect-aging-common-thought.html

Abstract Background In order for aging to evolve in response to a declining strength of selection with age, a genetic architecture that allows for mutations with age-specific effects on organismal performance is required. Our understanding of how selective effects of individual mutations are distributed across ages is however poor. Established evolutionary theories assume that mutations causing aging have negative late-life effects, coupled to either positive or neutral effects early in life. New theory now suggests evolution of aging may also result from deleterious mutations with increasing negative effects with age, a possibility that has not yet been empirically explored. Results To directly test how the effects of deleterious mutations are distributed across ages, we separately measure age-specific effects on fecundity for each of 20 mutations in Drosophila melanogaster . We find that deleterious mutations in general have a negative effect that increases with age and that the rate of increase depends on how deleterious a mutation is early in life. Conclusions Our findings suggest that aging does not exclusively depend on genetic variants assumed by the established evolutionary theories of aging. Instead, aging can result from deleterious mutations with negative effects that amplify with age. If increasing negative effect with age is a general property of deleterious mutations, the proportion of mutations with the capacity to contribute towards aging may be considerably larger than previously believed. Background Aging is the decline in physiological function with age, which results in a gradual decrease in survival and/or reproductive performance [ 1 , 2 , 3 , 4 ]. Aging affects many organisms and studies have shown that it has a genetic basis (reviewed in [ 5 ]). Theory suggests that aging evolves because the strength of selection declines with age [ 6 , 7 , 8 , 9 , 10 , 11 ], in combination with mutations with age-specific effects on performance. Two main evolutionary/genetic theories of aging have been proposed based on this idea. They both suggest that aging results from mutations with late acting deleterious effects, while they differ in that Mutation Accumulation (MA; [ 8 ]) assumes aging mutations to be neutral early in life whereas antagonistic pleiotropy (AP; [ 9 ]) assumes them to be beneficial. The AP and MA theories of aging have both been tested extensively, and evidence in favor of each has been found in the laboratory as well as in the wild (AP, e.g., [ 12 , 13 , 14 , 15 , 16 , 17 , 18 ], MA, e.g., [ 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 ]). The two theories make contrasting predictions about how early and late life performances are associated with one another. While AP predicts a negative association, the initial formulation of the MA theory, based on mutations with effects spanning very short age intervals, predicts no association [ 2 , 30 ]. These predictions stand in sharp contrast to the positive pleiotropy often observed between early and late life performance (e.g., [ 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 ]). Charlesworth [ 43 ] later extended the MA theory to also encompass mutations with effects spanning several (up to all) adult age classes, a modification which allows for positive pleiotropy under MA. Alternatively, aging can be reconciled with positive pleiotropy if variation in aging, at least within populations, is primarily caused by deleterious mutations with increasing negative effects with age. This possibility is also logically appealing, since from a biological perspective it is difficult to imagine a genetic architecture where mutations have deleterious effects exclusively confined to specific adult age classes. Aging through mutations with gradually increasing negative effects has rarely been considered (see discussions in [ 30 , 33 , 44 , 45 ]), but is predicted by recent theory [ 46 , 47 ]. The genetics of aging has primarily been studied using experimental evolution and quantitative genetics. These approaches have provided a general understanding of how genetic variants collectively contribute to aging, but provide limited information on the age-specific effects of individual mutations [ 48 ]. QTL studies and GWAS partly circumvent this limitation (e.g., [ 18 , 28 , 49 , 50 , 51 ]), but can generally only detect signals from mutations with large effects segregating at appreciable frequencies. To learn more about how the selective effect of individual deleterious mutations is distributed over age classes, we here measure the effect on female fecundity across adult life for 20 mutations in Drosophila melanogaster . From an evolutionary perspective, aging concerns both the elevation of mortality and the reduction in reproductive performance with age. Studying the latter however has several experimental advantages: as the same individuals can be measured at several time points, it provides a general measure of somatic condition, and experimental effort can be focused to key ages. Our results show that 16 of the 20 tested mutations have a negative effect on fecundity and that the negative effect increases with age for 14 of these 16. Increasing negative effects with age could hence be an inherent property of deleterious mutations, causing most deleterious mutations to contribute towards aging. Results To test if deleterious mutations have an increasing negative effect with advancing age, we independently introgressed 20 dominant mutations ( Bl[1] , Bsb[1] , bw[D] , Dfd[1] , Dr[1] , Frd[1] , Gl[1] , H[2] , Ki[1] , Kr[If-1] , L[rm] , Ly[1] , nw[B] , Pin[1] , Pri[1] , Pu[2] , Rap1[1] , Sb[1] , sna[Sco] , wg[Sp-1] ) by backcrossing for ≥ 11 generations into our outbred Drosophila melanogaster base population (Dahomey). To test for age-specific effects on reproductive performance, we then estimated the fecundity of 500–550 females of each mutation when expressed in the heterozygous state in young (day 5), middle-aged (day 19), and moderately old (day 33) individuals and compared this to the fecundity of 500–550 paired wildtype females (see the “ Methods ” section and Additional file 1 : Fig. S1 for details). Since our focus was on the effect of deleterious mutations on aging, we first asked which of the 20 mutations are deleterious with respect to fecundity. Analyzing fecundity data from young flies (5 days old) only, we find that 15 of the 20 mutations have a significantly deleterious effect, while one additional mutation had a significant deleterious effect when we also took fecundity at days 19 and 33 into account (Fig. 1 , Additional file 2 : Table S1). Fig. 1 Relative fecundity of females carrying each of the different mutations at ages 5, 19, and 33 days old (mean ± SE). Blue background: mutations that are deleterious and have an increasing negative effect on fecundity with age. Pink background: mutations that are deleterious but do not show an increasing negative effect on fecundity with age. Gray background: mutations for which no deleterious effect was detected. Mutations are ordered after their effect on fecundity at day 5, separately for each category/color Full size image By fitting a generalized linear mixed model (GLMM) with age as a covariate across the three age-specific measures of fecundity (as mutant versus wildtype egg counts at the vial level), we next tested if the deleterious mutations had an increasing negative effect on fecundity with age and find that this is the case for 14 of the 16 deleterious mutations (Fig. 1 , Additional file 2 : Table S2). As such, these 14 mutations increase the rate of reproductive aging. We also tested if we could detect an increasing negative effect when only comparing the effect on fecundity between two ages at a time (days 5 to 33, 5 to 19, and 19 to 33) using a similar model with age as a factor and find largely similar results (Additional file 2 : Table S3). Since several studies have suggested that both standing and mutational genetic variation show positive pleiotropy in their effects on life history traits between ages (e.g., [ 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 ]), we next tested if this was true for the 14 deleterious mutations that elevate the rate of aging studied here (using Kendall’s Tau which tests for an ordinal association between variables). Our results show a strong positive correlation in relative fecundity (ln[mut/wt]) between days 5 and 33, 5 and 19, as well as between 19 and 33 (all p ≤ 0.0001; Fig. 2 ). Similarly, strong positive correlations are also found when all 16 deleterious mutations are included (5 to 33: τ = 0.62, p = 0.0005; 5 to 19: τ = 0.62, p = 0.0005; 19 to 33: τ = 0.80, p < 0.0001). We also tested if relative fecundity at day 5 correlated to relative survival at day 33 and again find evidence for positive pleiotropy ( τ > 0.51 and p < 0.005, irrespectively of whether we use the 14 aging inducing or the 16 deleterious mutations; Additional file 2 : Fig. S2). Fig. 2 Correlations between relative fecundity (ln [mut/wt]) estimated at different ages. a Day 5 vs. day 33, b day 5 vs. day 19, and c day 19 vs. day 33, for the 14 deleterious mutations with increasing negative effects with age. Associations were tested with Kendall’s tau. Two-tailed p values are reported Full size image We were next interested in testing if the rate of aging induced by a mutation is associated with its degree of deleteriousness. To do this, we calculated the change in relative fecundity for each mutation between two ages at a time (old–young), and then tested if this difference was associated with the relative fecundity of the mutant at the first of the two ages using Kendall’s Tau. To be conservative, we first performed this test using all 16 deleterious mutations, regardless of whether they induced significant aging. We find no support for such an association when studying early-life aging (days 5 to 19: τ = − 0.10, p = 0.63), but when studying aging from early to late life (days 5 to 33: τ = 0.40, p = 0.03) and late-life aging (days 19 to 33: τ = 0.52, p = 0.005), we find that the deleteriousness of a mutation is associated with the rate of aging it induces. We find qualitatively similar results when analyzing only the 14 mutations that induced significant aging (results presented in above order: τ = − 0.12, p = 0.59; τ =0.47, p = 0.02; τ = 0.43, p = 0.04). If the rate of aging induced by a mutation at a particular age is positively associated with its deleteriousness at that age, this in turn suggests that the rate of aging a mutation induces should accelerate with age in a self-reinforcing process. We find strong support for this hypothesis as 10 of the 14 deleterious aging inducing mutations show significantly higher rates of aging between days 19 and 33 compared to between days 5 and 19 (Fig. 3 ). In addition, the other four mutations, as well as the two that are deleterious but do not induce aging when taking all three ages into account, all have (non-significant) point estimates (posterior means) pointing towards faster aging later in life (Fig. 3 ). Fig. 3 Difference between the rates of aging between the intervals 5 to 19 and 19 to 33 days of age (posterior mean, 95% credibility interval and posterior density distribution displayed). The dashed line indicates a constant rate of aging and values to the right indicate acceleration in the rate of aging with age. Deleterious mutations that increase the rate of aging are displayed in blue, while deleterious mutations that do not increase the rate of aging are displayed in pink Full size image Evolution of aging through changes in the relative frequency of mutations that show increasing negative effects with age and positive pleiotropy across ages does not necessarily follow from changes in how the strength of selection declines with age. For this to follow, the age specific variance in fitness induced by these mutations must increase with age [ 2 , 30 , 33 ]. Our finding that deleterious mutations cause the rate of aging to accelerate with age suggests that this is the case, which a direct test for changes in the genetic variance in relative fecundity among mutations with age also confirmed (Fig. 4 ). The genetic variance increased both between day 5 and 19 ( p mcmc < 0.007) and between days 19 and 33 ( p mcmc < 0.001), irrespective of whether all 16 deleterious mutations or only the 14 aging inducing ones are included. Fig. 4 Variance in relative fecundity across the 16 deleterious mutations at 5, 19, and 33 days of age (posterior mean, 95% credibility interval and posterior density distribution displayed) Full size image Discussion It is the declining strength of selection with age that creates the potential for evolution of aging, but for aging to actually evolve, mutations with age-specific effects are also required. The discussion of genetic variants with such properties has hitherto almost exclusively been limited to those assumed by the AP and MA theories of aging, while deleterious mutations with increasing negative effects with age have rarely been considered. Their shape in terms of age-specific effects is nevertheless conceptually very similar to those assumed by both the AP and MA theories: all three classes of mutations have effects that become worse with age, but differ in that they start out as either beneficial, neutral or deleterious in early life. In this study, we set out to test for age-specific effects of 20 presumably deleterious mutations. For 16 of these, we could detect a negative effect on fecundity and for 14 of these the negative effect increased with age. This result suggests that increasing negative effects with age could be a general property of deleterious mutations and that the pool of mutations contributing towards aging may be larger than previously thought. If deleterious mutations in general induce genetic variation in aging, this property is not enough for these mutations to respond to a steeper decline in the strength of selection with age and cause evolution of faster aging. For this to follow, it is required that their increasing negative effect with age depends on how deleterious they are and thus that the variance in age-specific fitness they cause increases with age [ 30 ]. Our results suggest that deleterious mutations, in general, have this property, which implies that aging could evolve through these mutations in relation to age-specific intensities of extrinsic mortality, similarly to what is predicted when aging evolves through MA and AP mutations [ 30 ]. While this study provides direct support for the existence of deleterious mutations with increasing negative effects, many previous studies have given indirect support for allelic variants with these properties. Developments of the MA theory of aging [ 19 , 52 ] predict that the genetic variance in fitness and inbreeding depression should increase with age and that “hybrid” vigor of individuals from crosses between different populations with small effective population sizes should be primarily confined to late ages. These predictions have been supported by a suite of studies (e.g., [ 19 , 20 , 21 , 22 , 23 , 24 , 25 , 52 , 53 , 54 ]). However, provided that deleterious mutations with increasing negative effects show increasing variance in effect size with age, and that they in general are at least partly recessive, the above findings are also compatible with aging through deleterious mutations with increasing negative effects. Additional findings that are compatible with aging through this class of mutations are extended lifespan in populations that have experienced intensified selection on early adult performance [ 35 , 40 , 42 ], strong correlations between mutations reducing early adult fitness and lifespan [ 33 , 41 ], and identification of mutations with increasing deleterious effects in a GWAS [ 50 ]. The few studies that have directly aimed at measuring age-specific effects of spontaneous mutations in D . melanogaster have, at least at face value, found results that differ from those reported here. Pletcher et al. [ 55 , 56 ] studied a set of mutation accumulation lines and found that the variance in mortality rate among lines was larger early compared to late in life. Similarly, Yampolsky et al. [ 57 ]) and Mack et al. [ 58 ], assaying a different set of mutation accumulation lines, found that the effect of spontaneous mutations is larger early compared to late in life for mortality and fecundity, respectively, at least at early generations of mutation accumulation. These findings are also corroborated by theory, extending Fisher’s geometric model for adaptation to different ages [ 59 ]. There are however several viable alternative explanations to these results, including ones that suggest that the true effect of spontaneous mutations actually increases with age. Of the alternatives, heterogeneity in frailty [ 60 , 61 ] among individuals within lines is the strongest candidate. The mutation accumulation lines studied by Pletcher et al. [ 55 , 56 ] were all derived from an inbred line. Inbred individuals are known to be particularly sensitive to environmental perturbations [ 52 , 62 ], which may have been further exacerbated since flies carried the mutation ebony. Variation between genetic lines could hence have been artificially reduced at late ages due to selective deaths of frailer individuals [ 63 , 64 ]. The diminishing effect of reproduction on mortality at late ages could also have contributed to the estimated reduction in variance at old ages [ 65 ]. The studies by Yamplosky et al. [ 57 ] and Mack et al. [ 58 ] accumulated mutations through the Middle Class Neighborhood design, which causes individuals within lines to vary in their mutational load in addition to environmentally induced heterogeneity. A smaller departure of mutation accumulation lines from control lines (in mortality rate and fecundity) at late compared to young ages could hence have followed from selective death of individuals more heavily laden with mutations [ 57 ]. Our study largely avoided the potential problem of heterogeneity, since we studied fecundity only up until an age most flies survive to. To be able to detect a signal of individual mutations in our experiments, we were constrained to focus on mutations with relatively large effect sizes. From this perspective, the mutations we assayed are not representative of spontaneous mutations, which in theory also could explain why our results differ from those of spontaneous mutations mentioned above. We cannot however see any reason why the age specificity of a mutation should depend on its effect size. This said, mutations with effect sizes comparable to those studied here would probably not contribute much to aging at the population level, since they should be rapidly purged from natural populations. The negative effects of a deleterious mutation could plausibly increase with age because the effect of the mutation itself induces an increase in the degradation of somatic quality (either early in or continuously across life), altering the context in which the mutation’s effects in later life are felt. This could be realized through several different potential mechanisms. One possibility is that deleterious mutations commonly interfere with the efficiency of biochemical pathways and elevate the production of harmful/toxic byproducts. If some of these byproducts gradually accumulate with age, they could have an increasing detrimental impact with age and cause aging [ 44 , 45 , 66 , 67 ]. Another possibility is that deleterious mutations reduce the robustness (i.e., quality/redundancy) of individuals. With imperfect repair systems, wear and tear would then result in faster failure of individuals carrying more deleterious mutations (for similar arguments see [ 68 , 69 ]). In line with this scenario, environmental stress during early development hastens reproductive aging [ 70 ]. Yet, another possibility is that the adverse effect of deleterious mutations on organismal function and capacity to acquire resources results in fewer resources available for somatic repair. Deleterious mutations would then cause aging similarly to how aging is generated according to the disposable soma theory (a physiological account of AP), which states that aging follows from insufficient resources being invested into somatic maintenance due to strong selection for resource allocation into current reproduction [ 71 , 72 ]. An extension of this argument is that individuals experiencing the negative effect from a deleterious mutation will age faster because they allocate relatively more resources into early reproduction in response to an elevated risk of death. Alternatively, it is possible that the negative effects of deleterious mutations do not increase with age per se, but that their impact is amplified with age because traditional aging mutations (MA or AP) cause the soma to be more sensitive to their effect at older ages [ 45 ]. This argument is similar to the one that the selective effect of deleterious mutations is larger in stressful environments, which has gained some empirical support [ 73 ] but does not seem to be generally true [ 74 , 75 ]. In any case, even if the increasing negative effect with age of deleterious mutations depends on other mutations initiating aging, deleterious mutations will contribute to aging in the presence of any AP or MA mutations. Conclusions The evolution of aging ultimately requires genetic variants with deleterious late-life acting effects. If these mutations primarily have beneficial or neutral effects early in life has been vividly debated, while mutations already expressing smaller deleterious effects early in life have only rarely been considered [ 30 , 33 , 44 , 45 , 46 ]. Our study does however demonstrate that deleterious mutations indeed can have negative effects that amplify with age and hence that increasing negative effect size could be a common property of deleterious mutations. Expressing a deleterious effect early in life exposes these mutations to negative selection to a much higher extent than those having a neutral effect early in life, keeping them at a considerably lower frequency, on a per locus basis, at mutation-selection-drift balance. The genome-wide influx of this type of aging mutation may however largely exceed those suggested by established aging theories, since logic suggests that most deleterious mutations should already manifest their harmful effect early in life, as most genes are expressed throughout life and show little expression dynamics during adulthood [ 76 , 77 ]. If future studies on a wider diversity of mutations corroborate our findings, the pool of mutations we know to be contributing to senescence will be substantially expanded. Methods Fly population and mutations We used flies from a laboratory-adapted population of Drosophila melanogaster known as Dahomey in our experiments, collected from the wild in 1970. Since then, it has been maintained as a large outbred population, housed in population cages with overlapping generations and under constant conditions (12 L ∶ 12D light cycle, 25 °C, 60% relative humidity, and a standard yeast/sugar-based food medium), which we maintained during our experiments. Due to its population size (not strictly controlled but in the thousands) and overlapping generations, the Dahomey population has become widely used in studies of aging and lifespan (e.g., [ 15 , 78 , 79 , 80 , 81 , 82 , 83 , 84 , 85 , 86 ]). Twenty autosomal mutations ( Bl[1] , Bsb[1] , bw[D] , Dfd[1] , Dr[1] , Frd[1] , Gl[1] , H[2] , Ki[1] , Kr[If-1] , L[rm] , Ly[1] , nw[B] , Pin[1] , Pri[1] , Pu[2] , Rap1[1] , Sb[1] , sna[Sco] and wg[Sp-1] ) were included in this study. They were chosen on the basis that they had a reliable dominant visible phenotype, to make it possible to efficiently backcross them into our base population, and based on that, several of them have successfully been used earlier to address the effect of deleterious mutations on various evolutionary processes [ 87 , 88 , 89 , 90 ]. The mutations were obtained from Bloomington Stock Center and introgressed separately into the Dahomey genetic background for at least eleven generations of backcrossing. During the first two generations, mutant males were crossed with virgin Dahomey females; thereafter, virgin mutant females were crossed with Dahomey males. From generation F4 and onwards, the backcrosses consistently involved at least 100 virgin females paired with 50 males per mutation line (Additional file 1 : Fig. S1A). Prior to use in our experiments, the mutation lines were reared at controlled densities (~ 180 eggs per vial) for two generations; eggs in each vial were laid by 15 Dahomey virgins mated with 15 males heterozygous for the mutant. To sample a range of Dahomey haplotypes, this was replicated across 10 vials per line for the first generation, and then increased to 30 vials for the generation producing experimental flies. With this design, no maternal effects should have influenced our results, since mutant and wildtype flies all shared the same mothers. Any paternal effects [ 91 ] should have been equally shared between mutant and wildtype flies as they, for each line, had the same fathers (Additional file 1 : Fig. S1B). Experimental design We measured age-specific fecundity in four experimental blocks, each including five different mutations. We housed flies in vials containing 33 mutant and 33 wildtype females along with 33 males marked with ebony (introgressed into the Dahomey background), using 20 to 22 vials for each mutation. Flies were transferred to new vials four times a week throughout the experiment. At three equally spaced ages (day 5, 19, and 33 of adult life), we randomly separated out 25 females of each type from every vial to lay eggs for 24 h; mutant and wildtype females were separately housed in small bottles with lids made of petri dishes filled with standard medium as egg-laying substrate. Petri dishes were replaced after ~ 18 h to simplify counting. In cases when fewer than 25 females of either type were alive in a vial, all available flies of the type with fewest survivors were selected to lay along with a matching number of randomly selected individuals of the other type. During the laying period, all excess females and the males from each vial were retained, and at the end of the laying period, all flies originating from an experimental vial were placed back together. Eggs laid on both petri dishes from each vial were counted and subsequently used to estimate the relative fecundity for each mutation at each time point. Ebony males were replaced with younger males (~ 4 days of adult age) at days 13 and 27 to standardize the influence of males across the different ages female fecundity was assayed at (Additional file 1 : Fig. S1C). Statistical analysis All statistical analysis was performed in the R statistical environment [ 92 ]. All generalized linear mixed-effect models (GLMMs) were fit using Bayesian Hamiltonian Markov chain Monte Carlo via the rstanarm package [ 93 ]. We used weakly informative, normally distributed priors for both the intercept (mean = 0, SD = 10) and the coefficients (mean = 0, SD = 2.5), with autoscaling for the coefficient priors. Each model was run with four chains of 4000 iterations each and the first 1000 discarded as warm-up, and chain convergence was evaluated using the Gelman-Rubin potential scale reduction factor. Generally speaking, we then used the posterior distributions of model coefficients to calculate relevant posterior distributions that allowed us to test hypotheses of interest. We report the equivalent to a two-tailed P value (p mcmc ) for these tests, calculated as two times the proportion of samples in a posterior distribution that was smaller/greater than the critical value (0), whichever was smaller. We tested whether mutations were deleterious using a GLMM with a Poisson error distribution and log link function, with the syntax: fecundity ~ type * mutation + (1|ID). This model was run twice, once where fecundity was calculated only for day 5, and once where fecundity was summed across days 5, 19, and 33. Type and mutation are fixed factors representing mutant or wildtype females for each mutation line respectively, their interaction is a fixed factor, and ID is a random factor representing the vial they were housed in. To test whether mutations cause aging, we used a model that directly compares the number of mutant eggs to wildtype eggs from females in each vial as a binomial response. This model took the form of a GLMM with a binomial error distribution and logit link function, with the syntax: (mutant fecundity|wildtype fecundity) ~ age * mutation + (1|ID), where age in days is modeled as a covariate, mutation and its interaction with age are fixed factors representing each mutation line, and ID is a random factor representing vial. We also fit a very similar model to the one described above where age was modeled as a fixed linear factor. This model enabled us to test directly for differences in aging during each discrete interval between fecundity measures. The posterior distributions of coefficients from this model were also used for two additional tests. We first tested whether the rate of aging accelerated with age by calculating a posterior distribution for the difference between the rate of aging in the day 19–33 interval and the rate in the day 5–19 interval and then testing whether or not this difference was different from 0. We then tested whether variance among mutations increased with age by calculating the posterior distribution of among-line variances at each of days 5, 19, and 33 and then testing whether the difference in variance between each pair of ages was different from 0. When testing whether there is positive pleiotropy in deleteriousness between ages, we first calculated the median relative fecundity (ln[mut/wt]) for each mutation across all replicate vials at each age (adding 1 to all egg counts to account for the occasional cases when no egg was laid). We then tested for a correlation (in the form of an ordinal association) between medians for each pair of ages using Kendall’s Tau. To test if there is a relationship between a mutation’s deleteriousness and the increase in rate of aging it induces, we first calculated the difference in median relative fecundity between each pair of ages for each mutation (ln[(mut t+1 /wt t+1 )/(mut t /wt t )] = ln[mut t+1 /wt t+1 ] − ln[mut t /wt t ]). We then tested for a positive correlation between this difference and the median relative fecundity at the first age using Kendall’s Tau. We note that this test is conservative, since random error in the estimate of the relative fecundity of the mutant at time t will contribute with the opposite effect on the difference in relative fecundity of the mutant between time points t and t +1, biasing the association towards negative values. Availability of data and materials The dataset generated and analyzed during the current study are available at the Dryad repository [ 94 ].

The number of mutations that can contribute to aging may be significantly higher than previously believed, according to new research on fruit flies. The study by scientists at Linköping University, Sweden, supports a new theory about the type of mutation that can lie behind aging. The results have been published in BMC Biology. We live, we age and we die. Many functions of our bodies deteriorate slowly but surely as we age, and eventually an organism dies. This thought may not be very encouraging, but most of us have probably accepted that this is the fate of all living creatures—death is part of life. However, those who study evolutionary biology find it far from clear why this is the case. "The evolution of aging is, in a manner of speaking, a paradox. Evolution causes continuous adaptation in organisms, but even so it has not resulted in them ceasing to age," says Urban Friberg, senior lecturer in the Department of Physics, Chemistry and Biology at Linköping University and leader of the study. Nearly 70 years ago, evolutionary biologists proposed two theories concerning two different types of mutation that contribute to aging. Both of these mutations have a detrimental effect as the organism becomes older—which leads to aging—while they are either advantageous or neutral early in life. Researchers have, however, not been able to determine which of the two types of mutation contributes most to aging, despite experimental studies. A new theory was proposed a few years ago suggesting that aging is caused by mutations with a detrimental effect early in life, and whose negative effects increase with age. Those who support this hypothesis believe that many of the mutations that arise have negative effects right from the start, compared with the normal variant of a gene. Martin Iinatti Brengdahl, PhD student at Linköping University, examining fruit flies in a microscope. Credit: Magnus Johansson/Linköping University The study now published describes experiments to test the theory of mutations that have a detrimental effect throughout life and contribute to aging. The authors used one of the most well-studied animals in the world, namely the fruit fly, or Drosophila melanogaster. They tested 20 different mutations that they had placed into the genetic material of the flies. For each individual mutation, they studied a group of flies with the mutation and a control group without it. Each mutation had a specific, visible effect, which made it easy to follow, such as a somewhat different appearance of the wings or a different shape of the eyes. As an organism ages, the probability that an individual dies increases, and its ability to reproduce falls. The researchers determined the fertility of the fruit flies and used it as a measure of aging. They counted the number of eggs laid by each female early in life, after two weeks, and finally after a further two weeks (which is a ripe old age for a fruit fly!). The researchers wanted to see whether the difference between flies with the mutations and the control group changed as they aged. The results support the theory they were testing. Most of the mutations had a negative effect on the fertility of the fruit flies early in life, and most of them also caused reproductive aging to occur more rapidly. "The results suggest that mutations that are detrimental early in life can also contribute to aging. Thus it may be that mutations that bring on aging are significantly more common than we previously believed," says Martin Iinatti Brengdahl, doctoral student in the Department of Physics, Chemistry and Biology and principal author of the study.

10.1186/s12915-020-00858-5

Computer

Air Learning: A gym environment to train deep reinforcement algorithms for aerial robot navigation

Air learning: a deep reinforcement learning gym for autonomous aerial robot visual navigation. Machine Learning(2021). DOI: 10.1007/s10994-021-06006-6.

http://dx.doi.org/10.1007/s10994-021-06006-6

https://techxplore.com/news/2021-08-air-gym-environment-deep-algorithms.html

Abstract We introduce Air Learning, an open-source simulator, and a gym environment for deep reinforcement learning research on resource-constrained aerial robots. Equipped with domain randomization, Air Learning exposes a UAV agent to a diverse set of challenging scenarios. We seed the toolset with point-to-point obstacle avoidance tasks in three different environments and Deep Q Networks (DQN) and Proximal Policy Optimization (PPO) trainers. Air Learning assesses the policies’ performance under various quality-of-flight (QoF) metrics, such as the energy consumed, endurance, and the average trajectory length, on resource-constrained embedded platforms like a Raspberry Pi. We find that the trajectories on an embedded Ras-Pi are vastly different from those predicted on a high-end desktop system, resulting in up to \(40\%\) longer trajectories in one of the environments. To understand the source of such discrepancies, we use Air Learning to artificially degrade high-end desktop performance to mimic what happens on a low-end embedded system. We then propose a mitigation technique that uses the hardware-in-the-loop to determine the latency distribution of running the policy on the target platform (onboard compute on aerial robot). A randomly sampled latency from the latency distribution is then added as an artificial delay within the training loop. Training the policy with artificial delays allows us to minimize the hardware gap (discrepancy in the flight time metric reduced from 37.73% to 0.5%). Thus, Air Learning with hardware-in-the-loop characterizes those differences and exposes how the onboard compute’s choice affects the aerial robot’s performance. We also conduct reliability studies to assess the effect of sensor failures on the learned policies. All put together, Air Learning enables a broad class of deep RL research on UAVs. The source code is available at: . Working on a manuscript? Avoid the common mistakes 1 Introduction Deep Reinforcement Learning (RL) has shown promising results in domains like sensorimotor control for cars (Bojarski et al., 2016 ), indoor robots (Chiang et al., 2019 ), as well as UAVs (Gandhi et al., 2017 ; Sadeghi & Levine, 2016 ). Deep RL’s ability to adapt and learn with minimum apriori knowledge makes them attractive for use in complex systems (Kretchmar, 2000 ). Unmanned Aerial Vehicles (UAVs) serve as a great platform for advancing state of the art for deep RL research. UAVs have practical applications, such as search and rescue (Waharte & Trigoni, 2010 ), package delivery (Faust et al., 2017 ; Goodchild & Toy, 2018 ), construction inspection (Peng et al., 2017 ). Compared to other robots such as self-driving cars, robotic arm, they are vastly cheap to prototype and build, which makes them truly scalable. Footnote 1 Also, UAVs have fairly diverse control requirements. Targeting low-level UAV control (e.g. attitude control) requires continuous control (e.g., angular velocities) whereas, targeting high-level tasks such as point-to-point navigation can use discrete control. Last but not least, at deployment time they must be a fully autonomous system, running onboard computationally- and energy-constrained computing hardware. But despite the promise of deep RL, there are several practical challenges in adopting reinforcement learning for the UAV navigation task as shown in Fig. 1 . Broadly, the problems can be grouped into four main categories: (1) environment simulator, (2) learning algorithms, (3) policy architecture, and (4) deployment on resource-constrained UAVs. To address these challenges, the boundaries between reinforcement learning algorithms, robotics control, and the underlying hardware must soften. The figure illustrates the cross-layer, and interdisciplinary nature of the field, spanning from environment modeling to the underlying system. Each layer, in isolation, has a complex design space that needs to be explored for optimization. In addition, there are interactions across the layers that are also important to consider (e.g., policy size on a power-constrained mobile or embedded computing system). Hence, there is a need for a platform that can aid interdisciplinary research. More specifically, we need a research platform that can benchmark each of the layers individually (for depth), as well as end-to-end execution for capturing the interactions across the layers (for breadth). Fig. 1 Aerial robotics is a cross-layer, interdisciplinary field. Designing an autonomous aerial robot to perform a task involves interactions between various boundaries, spanning from environment modeling down to the choice of hardware for the onboard compute Full size image In this paper, we present Air Learning (Sect. 4 )—an open source deep RL research simulation suite and benchmark for autonomous UAVs. As a simulation suite of tools, Air Learning provides a scalable and cost-effective applied reinforcement learning research. It augments existing frameworks such as AirSim (Shah et al., 2017 ) with capabilities that make it suitable for deep RL experimentation. As a gym, Air Learning enables RL research for resource constrained systems. Air Learning addresses the simulator level challenge, by providing domain randomization. We develop a configurable environment generator with a range of knobs to generate different environments with varying difficulty levels. The knobs (randomly) tune the number of static and dynamic obstacles, their speed (if relevant), texture and color, arena size, etc. In the context of our benchmarking autonomous UAV navigation task, the knobs help the learning algorithms generalize well without overfiting to a specific instance of an environment. Footnote 2 Air Learning addresses the learning challenges (RL algorithm, policy design, and reward optimization) by exposing the environment generator as an OpenAI gym (Brockman et al., 2016 ) interface and integrating it with Baselines (Hill et al., 2018 ), which has high-quality implementations of the state-of-the-art RL algorithms. We provide templates which the researchers can use for building multi-modal input policies based on Keras/Tensorflow. And as a DRL benchmark, the OpenAI gym interface enables easy addition of new deep RL algorithms. At the point of writing this paper, we provide two different reinforcement learning algorithms Deep Q-Networks (DQN) (Mnih et al., 2013 ) and Proximal Policy Optimization (PPO) (Schulman et al., 2017 ). DQN is an off-policy, discrete action RL algorithm, and PPO is an on-policy, continuous action control of UAVs. Both come ready with curriculum learning (Bengio et al., 2009 ) support. To address the resource-constrained challenge early on in the design and development of deep RL algorithms and policies, Air Learning uses a “hardware-in-the-loop” (HIL) (Adiprawita et al., 2008 ) method to enable robust hardware evaluation without risking real UAV platform. Hardware in the loop, which requires plugging in the computing platform used in the UAV into the software simulation, is a form of real-time simulation that allows us to understand how the UAV responds to simulated stimuli on a target hardware platform. Footnote 3 HIL simulation helps us quantify the real-time performance of reinforcement learning policies on various compute platforms, without risking experiments on real robot platforms before they are ready. We use HIL simulation to understand how a policy performs on an embedded compute platform that might potentially be the onboard computer of the UAVs. To enable systematic HIL evaluation, we use a variety of Quality-of-Flight (QoF) metrics, such as the total energy consumed by the UAV, the average length of its trajectory and endurance, to compare the different reinforcement learning policies. To demonstrate that Air Learning’s HIL simulation is essential and that it can reveal interesting insights, we take the best performing policy from our policy exploration stage and evaluate the performance of the policy on a resource-constrained low-performance platform (Ras-Pi 4) and compare it with a high-performance desktop counterpart (Intel Core-i9). The difference between the Ras-Pi 4 and the Core-i9 based performance for the policy is startling. The Ras-Pi 4 sometimes takes trajectories that are nearly 40% longer in some environments. We investigate the reason for the difference in the performance of the policy on Ras-Pi 4 versus Intel Core-i9 and show that the choice of onboard compute platform directly affects the policy processing latency, and hence the trajectory lengths. The discrepancy in the policy behavior from training to deployment hardware is a challenge that must be taken into account when designing the DRL algorithm for a resource-constrained robot. We define this behavior as ‘Hardware induced gap’ because of the performance gap in training machine versus deployment machine. We use a variety of metrics to quantify the hardware gap, such as percentage change between the QoF metrics that include flight time, success rate, the energy of flight, and trajectory distance. In summary, we present an open-source gym environment and research platform for deep RL research for autonomous aerial vehicles. The contributions within this context include: We present an open source benchmark to develop and train different RL algorithms, policies, and reward optimizations using regular and curriculum learning. We present a UAV mapless navigation task benchmark for RL research on resource constrained systems. We present a random environment generator for domain randomization to enable RL generalization. We introduce and show ‘Hardware induced gap’ – that the policy’s behavior depends on a computing platform it is running on, and that the same policy can result in a very different behavior if the target deployment platform is very different from the training platform. We describe the significance of taking energy consumption and the platform’s processing capabilities into account when evaluating policy success rates. To alleviate the hardware-induced gap, we train a policy using HIL to match the target platform’s latencies. Using this mitigation technique, we minimized the hardware gap between the training platform and resource-constrained target platform from 38% to less than 0.5% on flight time, 16.03% to 1.37% on the trajectory length, and 15.49% to 0.1% on the energy of flight metric. Air Learning will be of interest to both fundamental and applied RL research community. The point to point UAV navigation benchmark can yield to progress on fundamental RL algorithm development for resource-constrained systems where training and deployment platforms are different. From that point of view, Air Learning is another OpenAI Gym environment. For the applied RL researchers, interested in RL applications for UAV domains such as source seeking, search and reuse, etc., Air Learning serves as a simulation platform and toolset for full-stack research and development. 2 Real world challenges We describe the real-world challenges associated with developing deep RL algorithms on resource-constrained UAVs. We consolidate the challenges into four categories, namely Environment simulator, challenges related to the learning algorithm, policy selection challenges, and hardware-level challenges. Environment Simulator Challenges: The first challenge is that deep RL algorithms targeted for robotics need simulator. Collecting large amounts of real-world data is challenging because most commercial and off-the-shelf UAVs operate for less than 30 mins. To put this into perspective, creating a dataset as large as the latest “ImageNet” by Tencent for ML Images (Wu et al., 2019 ) would take close to 8000 flights (assuming a standard 30 FPS camera), thus making it a logistically challenging issue. But perhaps an even more critical and difficult aspect of this data collection is that there is a need for negative experiences, such as obstacle collisions, which can severely drive up the cost and logistics of collecting data (Gandhi et al., 2017 ). More importantly, it has been shown the environment simulator having high fidelity and ability to perform domain randomization aids in the better generalization of reinforcement learning algorithms (Tobin et al., 2017 ). Hence, any infrastructure for deep RL must have features to address these challenges to deploy RL policies in real-world robotics applications. Learning algorithm challenges: The second challenge is associated with reinforcement learning algorithms. Choosing the right variant of a reinforcement learning algorithm for a given task requires fairly exhaustive exploration. Furthermore, since the performance and efficiency of a particular reinforcement learning algorithm are greatly influenced by its reward function, to get good performance, there is a need to perform design exploration between the reinforcement learning algorithms and its reward function. Though these challenges are innate to the deep RL domain, having an environment simulator exposed as a simple interface (Brockman et al., 2016 ) can allow us to efficiently automate the RL algorithm selection, rewards shaping, hyperparameter tuning (Chiang et al., 2019 ). Policy selection challenges: The third challenge is associated with the selection of policies for robot control. Choosing the right policy architecture is a fairly exhaustive task. Depending upon the available sensor suite on the robot, the policy can be uni-modal or multi-modal in nature. Also, for effective learning, the hyperparameters associated with the policy architecture have to be appropriately tuned. Hyperparameter tuning and policy architecture search is still an active area of research, which has lead to techniques such as AutoML (Zoph et al., 2017 ) to determine the optimal neural network architecture. In the context of DRL policy selection, having a standard machine learning back-end tool such as Tensorflow/Keras (Abadi et al., 2015 ) can allow DRL researchers (or roboticist) to automate the policy architecture search. Hardware-level challenges: The fourth challenge is regarding the deployment of deep RL policies on the resource-constrained UAVs. Since UAVs are mobile machines, they need to accomplish their tasks with a limited amount of onboard energy. Because onboard compute is a scarce resource and RL policies are computationally intensive, we need to carefully co-design the policies with the underlying hardware so that the compute platform can meet the real-time requirements under power constraints. As the UAV size decreases, the problem exacerbates because battery capacity (i.e., size) decreases, which reduces the total onboard energy (even though the level of intelligence required remains the same). For instance, a nano-UAV such as a CrazyFlie ( 2018 ) must have the same autonomous navigation capabilities as compared to its larger mini counterpart, e.g., DJI-Mavic Pro ( 2018 ) while the CrazyFlie’s onboard energy is \(\frac{1}{15}\) th that of the Mavic Pro. Typically in deep RL research for robotics, the system and onboard computers are based on commercial off the shelf hardware platforms. However, whether the selection of these compute platforms is optimal is mostly unknown. Hence, having the ability to characterize the onboard computing platform early on can lead to resource-friendly deep RL policies. Air Learning is built with features to overcome the challenges listed above. Due to the interdisciplinary nature of the tool, it provides flexibility to researchers to focus on a given layer (e.g., policy architecture design) while also understanding its impact on the subsequent layer (e.g., hardware performance). In the next section, we describe the related work and list of features that Air Learning supports out of the box. 3 Related work Related work in deep RL toolset and benchmarks can be divided into three categories. The first category of related work includes environments for designing and benchmarking new deep RL algorithms. The second category of related work includes tools used specifically for deep RL based aerial robots. In the third category, we include other learning-based toolsets that support features that are important for deep RL training. The feature list and comparison of related work to Air Learning are tabulated in Table 1 . Table 1 Comparison of features commonly present in deep RL research infrastructures Full size table Benchmarking environments: The first category of related work includes benchmarking environments such as OpenAI Gym (Brockman et al., 2016 ), Arcade Learning Environments (Bellemare et al., 2015 ), and MujoCo (Todorov et al., 2012 ). These environments are simple by design and allow designing and benchmarking of new deep RL algorithms. However, using these environments for real-life applications such as robotics is challenging because they do not address the hardware-level challenges (Sect. 2 ) for transferring trained RL policies to real robots. Air Learning addresses these limitations by introducing Hardware-in-the-Loop (HIL), which allows end-user to benchmark and characterize the RL policy performance on a given onboard computing platform. UAV specific deep RL benchmarks: The second category of related work includes benchmarks that focus on UAVs. For example, AirSim (Shah et al., 2017 ) provides a high-fidelity simulation and dynamics for the UAVs in the form of a plugin that can be imported in any UE4 (Unreal Engine 4) (Valcasara, 2015 ) project. However, there are three AirSim limitations that AirLeaning addresses. First, the generation of the environment that includes domain randomization for the UAV task is left to the end-user to either develop or source it from the UE4 market place. The domain randomizations (Tobin et al., 2017 ) are very critical for generalization of the learning algorithm, and we address this limitation in AirSim using the Air Learning environment generator. Second, AirSim does not model UAV energy consumption. Energy is a scarce resource in UAVs that affects overall mission capability. Hence, learning algorithms need to be evaluated for energy efficiency. Air Learning uses energy model (Boroujerdian et al., 2018 ) within AirSim to evaluate learned policies. Air Learning also allows studying the impact of the performance of the onboard compute platform on the overall energy of UAVs, allowing us to estimate in the simulation how many missions UAV can do, without running in the simulation. Third, AirSim does not offer interfaces with OpenAI gym or other reinforcement learning framework such as stable baselines (Hill et al., 2018 ). We address this drawback by exposing the Air Learning random environment generator with OpenAI gym interfaces and integrate it with a high-quality implementation of reinforcement learning algorithms available in the framework such as baselines (Hill et al., 2018 ) and Keras-RL (Plappert, 2016 ). Using Air Learning, we can quickly explore and evaluate different RL algorithms for various UAV tasks. Another related work that uses a simulator and OpenAI gym interface in the context of UAVs is GYMFC (Koch et al., 2018 ). GYMFC uses Gazebo (Koenig & Howard, 2004 ) simulator and OpenAI gym interfaces for training an attitude controller for UAVs using reinforcement learning. The work primarily focuses on replacing the conventional flight controller with a real-time controller based on a neural network. This is a highly specific, low-level task. We focus more on high-level tasks, such as point-to-point UAV navigation in an environment with static and dynamic obstacles, and we provide the necessary infrastructure to carry research to enable on-edge autonomous navigation in UAVs. Adapting this work to support a high-level task such as navigation will involve overcoming the limitations of Gazebo, specifically in the context of photorealism. One of the motivations of building AirSim is to overcome the limitations of Gazebo by using state-of-the-art rendering techniques for modeling the environment, which is achieved using robust game engines such as Unreal Engine 4 (Valcasara, 2015 ) and Unity (Menard & Wagstaff, 2015 ). UAV agnostic deep RL benchmarks: The third category of related work includes deep RL benchmarks used for other robot tasks, such as grasping by a robotic arm or self-driving car. These related works are highly relevant to Air Learning because it contains essential features that improve the utility/performance of deep RL algorithms. The most prominent work in learning-based approaches for self-driving cars is CARLA (Dosovitskiy et al., 2017 ). It supports a photorealistic environment built on top of a game engine. It also exposes the environment as an OpenAI gym interface, which allows researchers to experiment with different deep RL algorithms. The physics is based on the game engine, and they do not model energy or focus on the compute hardware performance. Since the CARLA was built explicitly for self-driving cars, porting these features to UAVs will require significant engineering effort. For the robotic arm grasping/manipulation task, prior work (Ahn et al., 2020 ; Gu et al., 2016 ; Kalashnikov et al., 2018 ; Quillen et al., 2018 ) include infrastructure support to train and deploy deep RL algorithms on these robots. In Yahya et al. ( 2016 ), they introduce collective learning where they provide distributed infrastructure to collect large amounts of data with real platform experiments. They introduce an asynchronous variant of guided policy search to maximize the utilization (computer and synchronization between different agents), where each agent trains a local policy while a single global policy is trained based on the data collected from individual agents. However, these kinds of robots are fixed in a place; hence, they are not limited by energy or by onboard compute capability. So the inability to process or calculate the policy’s outcome in real-time only slows down the grasping rate. It does not cause instability. In UAVs, which have a higher control loop rate, uncertainty due to slow processing latency can cause fatal crashes (Giusti et al., 2016 ; Hwangbo et al., 2017 ). For mobile robots with/without grasping such as LocoBot (Locobot, 2018 ), PyRobot (Murali et al., 2019 ), ROBEL (Ahn et al., 2020 ) provides open-source tools and benchmarks for training and deploying deep RL policies on the LocoBot. The simulation infrastructure is based on Gazebo or MuJoCo, and hence it lacks photorealism in the environment and other domain randomization features. Similar to CARLA and robot grasping benchmarks, PyRobot does not model energy or focus on computing hardware performance. In soft learning (Haarnoja et al., 2018 ), the authors apply a soft-actor critic algorithm for the quadrupedal robot. They use Nvidia TX2 on the robot for data collection and also running the policy. The data collected is then used to train the global policy, which is then periodically updated to the robot. In contrast, in our work, we show that training policy on a high-end machine can result in a discrepancy in performance for aerial robot platform. Aerial robots are much more complex to control and unstable compared to ground-based quadrupedal robots. Hence small differences in processing time can hinder its safety. We propose training a policy using the HIL technique with the target platform’s latency distribution to mitigate the difference. Effect of action time in RL agents: Prior works (Riedmiller, 2012 ; Travnik et al., 2018 ) have studied the relationship between decision making time (i.e., time taken to decide an action) and task performance in RL agents. The authors propose reactive reinforcement learning algorithms propose a new “reactive SARSA” algorithm that orders computational components without affecting the training convergence to make decision making faster. In Air Learning, we expose a similar effect where differences in training hardware (high-end CPU/GPU) and deployment hardware (embedded CPUs) can result in entirely different agent behavior. To that end, we propose a novel action scaling technique based on Hardware-in-the-loop that minimizes the differences between training and deployment of the agent on resource-constrained hardware. Unlike “reactive SARSA” (Travnik et al., 2018 ), we do not make any changes to the RL algorithm. Another related work (Mahmood et al., 2018 ) studies the impact of delays in the action time in the robotic arm’s context. The authors use previously computed action until a new action is processed. We study the same problem in aerial robots, where we show that the differences in training and deployment hardware are another source of introducing processing delays and often overlooked. Since drones are deployed in a more dynamic environment, delayed action reduces the drones’ reactivity and can severely hinder their safety. To mitigate the performance gaps (hardware gap), we use the HIL methodology to model the target hardware delays and use them for training the policy. In summary, Air Learning provides an open source toolset and benchmark loaded with the features to develop deep RL based applications for UAVs. It helps design effective policies, and also characterize them on an onboard computer using the HIL methodology and quality-of-flight metrics. With that in mind, it is possible to start optimizing algorithms for UAVs, treating the entire UAV and its operation as a system. 4 Air Learning In this section, we describe the various Air Learning components. The different stages are shown in Fig. 2 , which allows researchers to develop and benchmark learning algorithms for autonomous UAVs. Air Learning consists of six keys components: an environment generator, an algorithm exploration framework, closed-loop real-time hardware in the loop setup, an energy and power model for UAVs, quality of flight metrics that are conscious of the UAV’s resource constraints, and a runtime system that orchestrates all of these components. By using all these components in unison, Air Learning allows us to fine-tune algorithms for the underlying hardware carefully. Fig. 2 Air Learning toolset for deep RL benchmarking in autonomous aerial machines. Our toolset consists of four main components. First, it has a configurable random environment generator built on top of UE4, a photo-realistic game engine that can be used to create a variety of different randomized environments. Second, the random environment generators are integrated with AirSim, OpenAI gym, and baselines for agile development and prototyping different state of the art reinforcement learning algorithms and policies for autonomous aerial vehicles. Third, its backend uses tools like Keras/Tensorflow that allow the design and exploration of different policies. Lastly, Air Learning uses the “hardware in the loop” methodology for characterizing the performance of the learned policies on real embedded hardware platforms. In short, it is an interdisciplinary tool that allows researchers to work from algorithm to hardware with the intent of enabling intra- and inter-layer understanding of execution. It also outputs a set of “Quality-of-Flight” metrics to understand execution Full size image 4.1 Environment generator Learning algorithms are data hungry, and the availability of high-quality data is vital for the learning process. Also, an environment that is good to learn from should include different scenarios that are challenging for the robot. By adding these challenging situations, they learn to solve those challenges. For instance, for teaching a robot to navigate obstacles, the data set should have a wide variety of obstacles (materials, textures, speeds, etc.) during the training process. We designed an environment generator specifically targeted for autonomous UAVs. Air Learning ’s environment generator creates high fidelity photo-realistic environments for the UAVs to fly in. The environment generator is built on top of UE4 and uses the AirSim UE4 (Shah et al., 2017 ) plugin for the UAV model and flight physics. The environment generator with the AirSim plugin is exposed as OpenAI gym interface. The environment generator has different configuration knobs for generating challenging environments. The configuration knobs available in the current version can be classified into two categories. The first category includes the parameters that can be controlled via a game configuration file. The second category consists of the parameters that can be controlled outside the game configuration file. The full list of parameters that can be controlled are tabulated in Table 2 . Figure 3 shows some examples of a randomly generated arena using the environment generator. For more information on these parameters, please refer “ Appendix ” section. Table 2 List of configurations available in current version of Air Learning environment generator Full size table Fig. 3 The environment generator generates different arena sizes with configurable wall texture colors, obstacles, obstacle materials etc. a arena with crimson colored walls with dimensions 50 m x 50 m x 5 m . The arena can be small or several miles long. The wall texture color is specified as an [R, G, B] tuple, which allows the generator to create any color in the visible spectrum. b Some of the UE4 asset used in Air Learning. Any UE4 asset can be imported and Air Learning environment generator will randomly select and spawn it in the arena. c Arena with random obstacles. The positions of the obstacles can be changed every episode or a rate specified by the user (Color figure online) Full size image 4.2 Algorithm exploration Deep reinforcement learning is still a nascent field that is rapidly evolving. Hence, there is significant infrastructure overhead to integrate random environment generator and evaluate new deep reinforcement learning algorithms for UAVs. So, we expose our random environment generator and AirSim UE4 plugin as an OpenAI gym interface and integrate it popular reinforcement learning framework with stable baselines (Hill et al., 2018 ), which is based on OpenAI baselines. Footnote 4 To expose our random environment generator into an OpenAI gym interface, we extend the work of AirGym (Kjell, 2018 ) to add support for environment randomization, a wide range of sensors (Depth image, Inertial Measurement Unit (IMU) data, RGB image, etc.) from AirSim and support exploring multimodal policies. We seed the Air Learning algorithm suite with two popular and commonly used reinforcement learning algorithms. The first is Deep Q Network (DQN) (Mnih et al., 2013 ) and the second is Proximal Policy Optimization (PPO) (Schulman et al., 2017 ). DQN falls into the discrete action algorithms where the action space is high-level commands (‘move forward,’ ‘move left’ e.t.c.,) and Proximal Policy Optimization falls into the continuous action algorithms (e.g., policy predicts the continuous value of velocity vector). For each of the algorithm variants, we also support an option to train the agent using curriculum learning (Bengio et al., 2009 ). For both these algorithms, we keep the observation space, policy architecture and reward structure same and compare agent performance. The environment configuration used in the training of PPO/DQN, the policy architecture, the reward function, is described in the appendix (“Appendix B ” section). Figure 4 a shows the normalized reward of the DQN agent (DQN-NC) and PPO agent (PPO-NC) trained using non-curriculum learning. One of the observations is that the PPO agent trained using non-curriculum learning consistently accrues negative reward throughout the training duration. In contrast, the DQN agent trained using non-curriculum learning starts at the same as the PPO agent but the DQN agent accrues more reward beginning in the 2000 th episode. Fig. 4 a Normalized reward during training for algorithm exploration between PPO-NC and DQN-NC. b Normalized reward during training for algorithm exploration between PPO-C and DQN-C. We find that the DQN agent performs better than the PPO agent irrespective of whether the agent was trained using curriculum learning or non-curriculum learning. The rewards are averaged over five runs with random seeds Full size image Figure 4 b shows the normalized episodic reward for the DQN (DQN-C) and PPO (PPO-C) agents trained using curriculum learning. We observe a similar trend as we saw with the agents trained using non-curriculum learning where the DQN agent outperforms the PPO agent. However, in this case, the PPO agent has a positive total reward. But the DQN agent starts to accrue more reward starting from the 1000 th episode. Also, the slight dip in the reward at 3800 th is due to the curriculum’s change (increased difficulty). Reflecting on the results, we gathered in Fig. 4 a, b, continuous action reinforcement learning algorithms such as PPO have generally been known to show promising results for low-level flight controller tasks that are used for stabilizing UAVs (Hwangbo et al., 2017 ). However, as our results indicate, applying these algorithms for a complex task, such as end-to-end navigation in a photo-realistic simulator, can be challenging for a couple of reasons. First, we believe that the action space for the PPO agent limits the exploration compared to the DQN agent. For the PPO agent, the action space is the components of velocity vector \(\texttt {{v}}_\texttt {{x}}\) and \(\texttt {{v}}_\texttt {{y}}\) whose value can vary from [-5 m / s , 5 m / s ]. Having such an action space can be a constraining factor for PPO. For instance, if the agent observes an obstacle at the front, it needs to take action such that it moves right or left. Now for PPO agent, since the action space is continuous values of [ \(\texttt {{V}}_\texttt {{x}}, \texttt {{V}}_\texttt {{y}}\) ], for it to move forward in the x -direction, the \(\texttt {{V}}_\texttt {{x}}\) can be any positive number while the \(\texttt {{V}}_\texttt {{{y}}}\) component has to be ‘0’. It can be quite challenging for the PPO agent (or continuous action algorithm) to learn this behavior, and it might require a much more sophisticated reward function that identifies these scenarios and rewards or penalizes these behaviors accordingly. In contrast, for the DQN agent, the action space is much simpler since it has to only yaw (i.e., move left or right) and then move forward or vice versa. Second, in our evaluation, we keep the reward function, input observation and the policy architecture same for DQN and PPO agent. We choose to fix these because we want to focus on showcasing the capability of the Air Learning infrastructure. Since RL algorithms are sensitive to hyperparameters and the choice of the reward function, it could be possible that our reward function, policy architecture could have inadvertently favored the DQN agent compared to the PPO agent. The sensitivity of the RL algorithms to the policy and reward is still an open research problem (Judah et al., 2014 ; Su et al., 2015 ). The takeaway is that we can do algorithm exploratory studies with Air Learning. For high-level task like point-to-point navigation, discrete action reinforcement learning algorithms like DQN allows more flexibility compared to continuous action reinforcement learning algorithms like PPO. We also demonstrate that incorporating techniques such as curriculum learning can be beneficial to the overall learning. 4.3 Policy exploration Another essential aspect of deep reinforcement learning is the policy, which determines the best action to take. Given a particular state the policy needs to maximize the reward. A neural network approximates the policies. To assist the researchers in exploring effective policies, we use Keras/TensorFlow (Chollet, 2015 ) as the machine learning back-end tool. Later on, we demonstrate how one can do algorithm and policy explorations for tasks like autonomous navigation though Air Learning is by no means limited to this task alone. 4.4 Hardware exploration Often aerial roboticists port the algorithm onto UAVs to validate the functionality of the algorithms. These UAVs can be custom built (NVIDAA-AI-IOT, 2015 ) or commercially available off-the-shelf (COTS) UAVs (Hummingbird, 2018 ; Intel, 2018 ) but mostly have fixed hardware that can be used as onboard compute. A critical shortcoming of this approach is that the roboticist cannot experiment with hardware changes. More powerful hardware may (or may not) unlock additional capabilities during flight, but there is no way to know until the hardware is available on a real UAV so that the roboticist can physically experiment with the platform. Reasons for wanting to do such exploration includes understanding the computational requirements of the system, quantifying the energy consumption implications as a result of interactions between the algorithm and the hardware, and so forth. Such evaluation is crucial to determine whether an algorithm is, in fact, feasible when ported to a real UAV with a specific hardware configuration and battery constraints. For instance, a Parrot Bepop (Parrot, 2019 ) comes with a P7 dual-core CPU Cortex A9 and a Quad core GPU. It is not possible to fly the UAV assuming a different piece of hardware, such as the NVIDIA Xavier (NVIDIA, 2019 ) processor that is significantly more powerful; at the time of this writing there is no COTS UAV that contains the Xavier platform. So, one would have to wait until a commercially viable platform is available. However, using Air Learning, one can experiment how the UAV would behave with a Xavier since the UAV is flying virtually. Hardware exploration in Air Learning allows for evaluation of the best reinforcement learning algorithm and its policy on different hardware. It is not limited by the onboard compute available on the real robot. Once the best algorithm and policy are determined, Air Learning allows for characterizing the performance of these algorithms and policies on different types of hardware platforms. It also enables to carefully fine-tune and co-design algorithms and policy while being mindful of the resource constraints and other limitation of the hardware. A HIL simulation combines the benefits of the real design and the simulation by allowing them to interact with one another as shown in Fig. 5 . There are three core components in Air Learning ’s HIL methodology: (1) a high-end desktop that simulates a virtual environment flying the UAV ( top ); (2) an embedded system that runs the operating system, the deep reinforcement learning algorithms, policies and associated software stack ( left ); and (3) a flight controller that controls the flight of the UAV in the simulated environment ( right ). Fig. 5 Hardware-in-the-loop (HIL) simulation in Air Learning Full size image The simulated environment models the various sensors (RGB/Depth Cameras), actuators (rotors), and the physical world surrounding the agent (Obstacles). This data is fed into the reinforcement learning algorithms that are running on the embedded companion computer, which processes the input and outputs flight commands to the flight controller. The controller then communicates those commands into the virtual UAV flying inside the simulated game environment. The interaction between the three components is what allows us to evaluate the algorithms and policy on various embedded computing platforms. The HIL setup we present allows for the swap-ability of the embedded platform under test. The methodology enables us to effectively measure both the performance and energy of the agent holistically and more accurately, since one can evaluate how well an algorithm performs on a variety of different platforms. In our evaluation, which we discuss later, we use a Raspberry Pi (Ras-Pi 4) as the embedded hardware platform to evaluate the best performing deep reinforcement learning algorithm and its associated policy. The HIL setup includes running the environment generator on a high-end desktop with a GPU. The reinforcement learning algorithm and its associated policy run on the Ras-Pi 4. The state information (Depth image, RGB image, IMU) are requested by Ras-Pi 3 using AirSim Plugins APIs which involves an RPC (remote procedural calls) over TCP/IP network (both high-end desktop and Ras-Pi 4 are connected by ethernet). The policy evaluates the actions based on the state information it received from the high-end desktop. The actions are relayed back to the high-end desktop through AirSim flight controller API’s. 4.5 Energy model in AirSim plugin In Air Learning, we use the energy simulator we developed in our prior work (Boroujerdian et al., 2018 ). The AirSim plugin is extended with a battery and energy model. The energy model is a function of UAVs velocity, acceleration. The values of velocity and acceleration are continuously sampled and using these we estimate the power as proposed in this work (Tseng et al., 2017 ). The power is calculated using the following formula: $$\begin{aligned} \begin{aligned} P&= \begin{bmatrix} \beta _{1} \\ \beta _{2} \\ \beta _{3} \end{bmatrix}^{T} \begin{bmatrix} \left\Vert \vec {v}_{xy}\right\Vert \\ \left\Vert \vec {a}_{xy}\right\Vert \\ \left\Vert \vec {v}_{xy}\right\Vert \left\Vert \vec {a}_{xy}\right\Vert \end{bmatrix} + \begin{bmatrix} \beta _{4} \\ \beta _{5} \\ \beta _{6} \end{bmatrix}^{T} \begin{bmatrix} \left\Vert \vec {v}_{z}\right\Vert \\ \left\Vert \vec {a}_{z}\right\Vert \\ \left\Vert \vec {v}_{z}\right\Vert \left\Vert \vec {a}_{z}\right\Vert \end{bmatrix} \\&\quad + \begin{bmatrix} \beta _{7} \\ \beta _{8} \\ \beta _{9} \end{bmatrix}^{T} \begin{bmatrix} m \\ \vec {v}_{xy} \cdot \vec {w}_{xy} \\ 1 \end{bmatrix} \end{aligned} \end{aligned}$$ (1) In Eq. 1 , v \(_{xy}\) and a \(_{xy}\) are the velocity and acceleration in the horizontal direction. v \(_{z}\) and a \(_{z}\) denotes the velocity and acceleration in the z direction. m denotes the mass of the payload. \(\beta _1\) to \(\beta _9\) are the coefficients based on the model of the UAV used in the simulation. For the energy calculation model, we use the columb counter technique as described in prior work (Kumar et al., 2016 ). The simulator computes the total number of columb that has passed over the battery over every cycle. Using the energy model Air Learning allows us to monitor the energy continuously during training or during the evaluation of the reinforcement learning algorithm. 4.6 Quality of flight metrics Reinforcement learning algorithms are often evaluated based on success rate where the success rate is based on whether the algorithm completed the mission. This metric only captures the functionality of the algorithm and grossly ignores how well the algorithm performs in the real world. In the real world, there are additional constraints for a UAV, such as the limited onboard compute capability and battery capacity. Hence, we need additional metrics that can quantify the performance of learning algorithms more holistically. To this end, Air Learning introduces Quality-of-Flight (QoF) metrics that not only captures the functionality of the algorithm but also how well they perform when ported to onboard compute in real UAVs. For instance, the algorithm and policies are only useful if they accomplish the goals within finite energy available in the UAVs. Hence, algorithms and policies need to be evaluated on the metrics that describe the quality of flight such as mission time, distance flown, etc. In the first version of Air Learning, we consider the following metrics. Success rate: The percentage of time the UAV reaches the goal state without collisions and running out of battery. Ideally, this number will be close to 100% as it reflects the algorithms’ functionality, taking into account resource constraints. Time to completion: The total time UAV spends finishing a mission within the simulated world. Energy consumed: The total energy spent while carrying out the mission. Limited battery available onboard constrains the mission time. Hence, monitoring energy usage is of utmost importance for autonomous aerial vehicles, and therefore should be a measure of policy’s efficiency. Distance traveled: Total distance flown while carrying out the mission. This metric is the average length of the trajectory that can be used to measure how well the policy did. 4.7 Runtime system The final part is the runtime system that orchestrates the overall execution. The runtime system starts the game engine with the correct configuration of the environment before the agent starts. It also monitors the episodic progress of the reinforcement learning algorithm and ensures that before starting a new episode that it randomizes the different parameters, so the agent statistically gets a new environment. It also has resiliency built into it to resume the training in case any one of the components (for example UE4 engine) crashes. In summary, using Air Learning environment generator, researchers can develop various challenging scenarios to design better learning algorithms. Using Air Learning interfaces to OpenAI gym, stable-baselines and TensorFlow backend, they can rapidly evaluate different reinforcement learning algorithms and their associated policies. Using Air Learning HIL methodology and QoF metrics, they can benchmark the performance of learning algorithms and policies on resource-constrained onboard compute platforms. 5 Experimental evaluation prelude The next few sections focus heavily on how Air Learning can be used to demonstrate its value. As a prelude, this section presents the highlights to focus on the big picture. Policy evaluation (Sect. 6 ): We show how Air Learning can be used to explore different reinforcement learning based policies. We use the best algorithm determined during the algorithm exploration step and use that algorithm to explore the best policy. In this work, we use Air Learning environment generator to generate three environments, namely, No Obstacles , Static Obstacles , and Dynamic Obstacles . These three environments create a varying level of difficulties by changing the number of static and dynamic obstacles in the environments for the autonomous navigation task. We also show how Air Learning allows end users to perform benchmarking of the policies by showing two examples. In the first example, we show how well the policies trained in one environment generalize to the other environments. In the second example, we show to which of the sensor inputs the policy is most sensitive towards. This insight can be used while designing the network architecture of the policy. For instance, we show that image input has the highest sensitivity amongst other inputs. Hence a future iteration of the policy can have more feature extractors (increasing the depth of filters) dedicated to the image input. System evaluation (Sect. 7 ): We show the importance of benchmarking algorithm performance on resource-constrained hardware such as what is typical of a UAV compute platform. In this work, we use a Raspberry Pi 4 (Ras-Pi 4) as an example of resource-constrained hardware. We use the best policies determined in the policy exploration step (Sect. 6 ) and use that to compare the performance between Intel Core-i9 and Ras-Pi 4 using HIL and the QoF metrics available in Air Learning. We also show how to artificially degrade the performance of the Intel Core-i9 to show how compute performance can potentially affect the behavior of a policy when it is ported over to a real aerial robot. In summary, using these focused studies, we demonstrate how Air Learning can be used by researchers to design and benchmark algorithm-hardware interactions in autonomous aerial vehicles, as shown previously in Fig. 2 . 6 Policy exploration In this section, we perform policy exploration for the DQN agent with curriculum learning (Bengio et al., 2009 ). The policy exploration phase aims to determine the best neural network policy architecture for each of the tasks (i.e., autonomous navigation) in different environments with and without obstacles. We start with a basic template architecture, as shown in Fig. 6 . The architecture is multi-modal and takes depth image, velocity, and position data as its input. Using this template, we sweep two parameters, namely # Layers and # Filters (making the policy wider and deeper). To simplify the search, for convolution layers, we restrict filter sizes to 3 \(\times\) 3 with stride 1. This choice ensures that there is no loss of pixel information. Likewise, for fully-connected layers, # Filter parameter denotes the number of hidden neurons in that layer. The choice of using #Layers and # Filters parameters to control both the convolution and fully-connected layers is to manage the complexity of searching over large NN hyperparameters design space. Fig. 6 The network architecture template for the policies used in DQN agents. We sweep the # Layers and # Filters parameters in the network architecture template. Both the agents take a depth image, velocity vector, and position vector as inputs. The depth image is passed through # Layers of convolutions layers with # Filters each. # Layers and # Filters are variables what we sweep. We also use a uniform filter size of (3 \(\times\) 3) with stride of 1. The combined vector space is passed to the # Layers of fully connected network, each with # Filters hidden units. The choice of using #Layers and # Filters parameters to control both the convolution and fully-connected layers is to manage the complexity of searching over large NN hyperparameters design space. The action space determines the number of hidden units in the last fully connected layer. For the DQN agent, we have twenty-five actions Full size image The # Layers and # Filters and the template policy architecture can be used to construct a variety of different policies. For example, a tuple of (# Filters \(=\) 32, # Layers \(=\) 5) will result in a policy architecture where there five convolution layers with 32 filters (with 3 \(\times\) 3 filters) followed by five fully-connected layers with 32 hidden neurons each. For each of the navigation tasks (in different environments), we sweep the template parameters (# Layers and # Filters) to explore multiple policy architectures for the DQN agent. 6.1 Training and testing methodology The training and testing methodology for the DQN agent running in the different environments is described below. Environments: For the point-to-point autonomous navigation task for UAVs, we create three randomly generated environments, namely, No Obstacles , Static Obstacles , and Dynamic Obstacles , with varying levels of static obstacles and dynamic obstacles. The environment size for all three levels is 50 m \(\times\) 50 m. For the No Obstacles environment, there are no obstacles in the main arena, but the goal position is changed every episode. For Static Obstacles , the number of obstacles varies from five to ten, and it is changed every four episodes. The end goal and position of the obstacles are changed every episode. For Dynamic Obstacles , along with five static obstacles, we introduce up to five dynamic obstacles of whose velocities range from 1 to 2.5 m/s. The obstacles and goals are placed in random locations every episode to ensure that the policy does not over-fit. Training methodology: We train the DQN agent using curriculum learning in the environments described above. We use the same methodology described in “Appendix B ” section, where we checkpoint policy in each zone for the three environments. The hardware used in training is an Intel Core-i9 CPU with an Nvidia GTX 2080-TI GPU. Testing methodology: For testing the policies, we evaluate the checkpoints saved in the final zone. Each policy is evaluated on 100 randomly generated goal/obstacle configuration (controlled by the ‘Seed’ parameter in Table 2 ). The same 100 randomly generated environment configurations are used across different policy evaluations. The hardware we use for testing the policies is the same as the hardware used for training them (Intel Core-i9 with Nvidia GTX 2080-TI). 6.2 Policy selection The policy architecture search for No Obstacles , Static Obstacles , and Dynamic Obstacles are shown in Fig. 8 . Figure 7 a–c show the success rate for different policy architecture searched for the DQN agent trained using curriculum learning on No Obstacles , Static Obstacles , and Dynamic Obstacles environments, respectively. In the figures, the x-axis corresponds # Filter sizes (32, 48, or 64) and the y-axis corresponds to the # Layers (2, 3, 4, 5, and 6) for No Obstacles / Static Obstacles environments and # Layers (5, 6, 7, 8, 9) for Dynamic Obstacles environment. The reason for sweeping different (larger) policies is because ”Dynamic Obstacles” will be a harder task, and a deeper policy might help improve the success rate compared to a shallow policy. Each cell corresponds to a unique policy architecture based on the template defined in Fig. 7 . The value in each cell corresponds to the success rate for the best policy architecture. The ± denotes the standard deviation (error bounds) across five seeds. For instance, in Fig. 7 a, the best performing policy architecture with # Filters of 32 and # Layers of 2 results in a 72% success rate. The success rate across five seeds results in a standard deviation of ± of 8% error. For evaluation, we always choose the best performing policy (i.e., the policy that achieves best success rate). Fig. 7 a , b , and c show the policy architecture search for the No Obstacles , Static Obstacles , and Dynamic Obstacles environments. Each cell shows the success rate for the policies for # Layers and # Filters’ corresponding values. The success rate is evaluated in Zone 3 , which is the region that is not used during training. Each policy is evaluated on the same 100 randomly generated environment configuration (controlled by the ‘Seed’ parameter described in Table 2 ). The policy architecture with the highest success rate is chosen as the best policy for DQN agents in the environment with no obstacles, static obstacles, and dynamic obstacles. The standard deviation error across multiple seeds are denoted by (±) sign. For the No Obstacles environment, the policy with # Layers of five and # Filters of 32 is chosen as the best performing policy. Likewise, for the Dynamic Obstacles environment, the policy architecture with # Layers of 7 and # Filter of 32 is chosen as the best policy Full size image Based on the policy architecture search, we notice that as the task complexity increases (obstacle density increases), a larger policy improves the task success rate. For instance, in the No Obstacles case (Fig. 7 a), the policy with # Filters of 32 and # Layers of 5 achieves the highest success rate of 91%. Even though we name the environment No Obstacles, the UAV agent can still collide with the arena walls, which lowers the success rate. For Static Obstacles case (Fig. 7 b), the policy with # Filters of 48 and # Layers of 4 achieves the best success rate of 84%. Likewise, for Dynamic Obstacles case (Fig. 7 c), the policy architecture with # Filters of 32 and # Layers of 7 achieves the best success rate of 61%. The success rate loss in Static Obstacles and Dynamic Obstacles cases can be attributed to an increase in the possibility of collisions with static and dynamic obstacles. 6.3 Success rate across the different environments To study how a policy trained in one environment performs in other environments, we take the best policy trained in the No Obstacles environment and evaluate it on the Static Obstacles and Dynamic Obstacles environments. We do the same for the best policy trained on Dynamic Obstacles and assess it on the No Obstacles and Static Obstacles environments. The results for the generalization study are tabulated in Table 3 . We see that the policy trained in the No Obstacles environment has a steep drop in success rate from 91 to 53% in Static Obstacles and 32% in Dynamic Obstacles environment, respectively. In contrast, we observe that the policy trained in the Dynamic Obstacles environment has an increased success rate from 61 to 89% in the No Obstacles and 74% in the Static Obstacles environment, respectively. Table 3 Evaluation of the best-performing policies trained in one environment tested in another environment Full size table The drop in the success rate for the policy trained in the No Obstacles environment is expected because, during its training, the agent might not have encountered a variety of obstacles (static and dynamic obstacles) to learn from as it might have encountered in the other two environments. The same reasoning can also apply to the improvement in the success rate observed in the policy trained in the Dynamic Obstacles environment when it is evaluated on the No Obstacles and Static Obstacles environments. In general, the agent performs best in the environment where it is trained, which is expected. But we also observe that training an agent in a more challenging environment can yield good results when evaluating in a much less challenging environment. Hence, having a random environment generator, such as what we have enabled in Air Learning, can help the policy generalize well by creating a wide variety of different experiences for the agent to experience during training. 6.4 Success rate sensitivity to sensor input ablation In doing policy exploration, one is also interested in studying the policy’s sensitivity towards a particular sensor input. So we ablate the sensor inputs to the policy to understand the effects. We ablate the policy’s inputs one by one and see the impact of various ablation and its success rate. It is important to note that we do not re-train the policy with ablated inputs. This is to perform reliability study and simulate the real-world scenario if a particular sensing modality is corrupted. The policy architecture we used for the DQN agent in this work is multi-modal in nature which receives depth image, velocity measurement \({V}_t\) and position vector \({X}_t\) as inputs. The \({V}_t\) is a 1-dimensional vector of the form [ \({v}_x\) , \({v}_y\) , \({v}_z\) ] where \({v}_x\) , \({v}_y\) , \({v}_z\) are the components of velocity vector in x, y and z directions at time ‘t’ . The \({X}_t\) is a 1-dimensional vector of the form [ \({X}_{{goal}}\) , \({Y}_{{goal}}\) , \({D}_{{goal}}\) ], where \({X}_{{goal}}\) and \({Y}_{{goal}}\) are the relative ‘x’ and ‘y’ distance with respect to the goal position and \({D}_{{goal}}\) is the euclidean distance to the goal from the agent’s current position. The baseline success rate we use in this study is when all the three inputs are fed to the policy. The velocity ablation study refers to removing the velocity input measurements from policy inputs. Likewise, the position ablation study and depth image ablation study refer to removing the position vector and depth image from the policy’s input stream. The results of various input ablation studies are plotted in Fig. 8 . Fig. 8 The effect of ablating the sensor inputs on the success rate. We observe that the depth image contributes the most to the policy’s success, whereas velocity input affects the least in the success. All the policy evaluations are in Zone3 on the Intel Core-i9 platform Full size image For the No Obstacles environment, the policy success rate drops from 91% to 53% when velocity measurements are ablated. When the depth image is ablated, we find that the success rate drops to 7%, and when the position vector is ablated, the success rate drops to 42%. Similarly, for Static Obstacles , we find that if the depth image input is ablated, it fails to reach the destination. Likewise, when the velocity and position inputs are ablated, we observe the success rate drops from 84% to 33%. Similarly, we see a similar observation in a Dynamic Obstacles environment where the success rate drops to 0% when the depth image is ablated. The depth image is the highest contributor to the policy’s success, whereas the velocity input is significant but least among the other two inputs. The drop in the policy success rate due to depth image ablation is evident from policy architecture since maximum features in the flatten layer are contributed by the depth image than velocity and position (both 1 \(\times\) 3 vectors). Another interesting observation is that when the position input is ablated, the agent also loses the information about its goal. The lack of goal position results in an exploration policy capable of avoiding obstacles (due to depth image input). In No Obstacles environment (where there are no obstacles except walls), the agent is free to explore unless it collides with the walls or exhaust maximum allowed steps. Due to the exploration, the agent reaches the goal position 42 out of 100 times. Our results are in line with prior work (Duisterhof et al., 2019 ; Palacin et al., 2005 ) where such random action-based exploration yields some amount of success. However, in a cluttered environment, random exploration may result in sub-optimal performance due to a higher probability of collision or exhausting maximum allowed steps (a proxy for limited battery energy). Using Air Learning, researchers can gain better insights into how reliable a particular set of inputs in the case of sensor failures. The reliability studies and its impact on learning algorithms are essential given the kind of application the autonomous aerial vehicles are targeted. Also, understanding the sensitivity of a particular input towards success can lead to better policies where more feature extraction can be assigned to those inputs. 7 System evaluation This section demonstrates how Air Learning can benchmark the algorithm and policy’s performance on a resource-constrained onboard compute platform, post-training. We use the HIL methodology (Sect. 4.4 ) and QoF metrics (Sect. 4.6 ) for benchmarking the DQN agent and its policy. We evaluate them on the three different randomly generated environments described in Sect. 6 . 7.1 Experimental setup The experimental setup has two components, namely, the server and System Under Test (SUT), as shown in Fig. 9 . The server component is responsible for rendering the environment (for example, No Obstacles ). The server consists of an 18 core Intel Core-i9 processor with an Nvidia RTX-2080. The SUT component is the system on which we want to evaluate the policy. The SUT is the proxy for the onboard compute system used in UAVs. In this work, we compare the policies’ performance on two systems, namely Intel Core-i9 and Ras-Pi 4. The key differences between the Intel Core-i9 and Ras-Pi 4 platform are tabulated in Table 4 . The systems are vastly different in their performance capabilities and represent ends of the performance spectrum. Fig. 9 The Experimental setup for policy evaluation on two different platforms. The platform under test is called the System Under Test (SUT). The environments are rendered on a server with Intel Core-i9 with Nvidia RTX 2080. Clock speed is a function in the AirSim plugin, which speeds up the environment time relative to the real world clock. In our evaluation, we set the clock speed to 2X. Time \(\hbox {t}_{{1}}\) is the time it takes to get the state information from the environment to the SUT. We use an Intel Core-i9 and a Ras-Pi 4 as the two SUTs. Time \(\hbox {t}_{{2}}\) is the time it takes to evaluate the forward pass of the neural network policy. This latency depends on the SUT. It is different for the Intel Core-i9 and the Ras-Pi 4. Time \(\hbox {t}_{{3}}\) is the actuation time for which the control is applied Full size image Table 4 The most pertinent System Under Test (SUT) specifications for the Intel Core-i9 and Ras-Pi 4 systems Full size table Three latencies affect the overall processing time. The first is \({t}_{{1}}\) , which is the latency to extract the state information (Depth Image, RGB Image, etc.) from the server. The state information is fetched from the server to the SUT. The communication protocol used between the server and the SUT is TCP/IP. Initially, we found that ethernet adapter on Intel Core-i9 faster compared to the ethernet adapter on Ras-Pi 4. We make the \({t}_{{1}}\) latencies between Intel Core-i9 and Ras-Pi 4 same by adding artificial sleep for Intel Core-i9 platform. Footnote 5 The second latency is \({t}_{{2}}\) , which is the policy evaluation time for the SUT (i.e., the Intel Core-i9 or the Ras-Pi 4). The policies are evaluated on the SUT, which predicts the output actions based on the input state information received from the server. The policy architecture used in this work has 40.3 Million ( No Obstacles and Static Obstacles ) and 161.77 Million parameters ( Dynamic Obstacles . The \({t}_{{2}}\) latency for No Obstacles policy on Ras-Pi 4 is 396 ms, while on the desktop, equipped with GTX 2080 Ti GPU and Intel Core i9 CPU, it is 11 ms. The desktop is 36 \(\times\) times faster. The third latency is \({t}_{{3}}\) . Once the policies are evaluated, it predicts actions. These actions are converted to the low-level actuation using the AirSim flight controller APIs. Footnote 6 These APIs have a duration parameter that controls the duration of a particular action must be applied. This duration parameter is denoted by \({t}_{{3}}\) , and it is kept the same for both SUTs. To evaluate the impact of the SUT performance on the overall learning behavior, we keep the \({t}_{{1}}\) and \({t}_{{3}}\) latencies constant for both Intel Core-i9 and Ras-Pi 4 systems. We focus only on the difference in the policy evaluation time (i.e., \({t}_{{2}}\) ) and study how it affects the overall performance time. Using this setup, we evaluate the best policy determined in Sect. 6 for environments with no obstacles, static obstacles, and dynamic obstacles. 7.2 Desktop vs. embedded SUT performance In Table 5 , we compare the performance of the policy on a Intel Core-i9 (high-end desktop) and the Ras-Pi 4. We evaluate the best policy on the No Obstacles , Static Obstacles and Dynamic Obstacles environments described previously in Sect. 6 . Table 5 Inference time, success rate, and Quality of Flight (QoF) metrics between Intel Core i9 desktop and Ras-Pi 4 for No Obstacles , Static Obstacles , and Dynamic Obstacles Full size table In the No Obstacles case, the policy running on the high-end desktop is 11% more successful compared to the policy running on the Ras-Pi 4. The flight time to reach the goal on the desktop is, on average, 25.29 s, whereas on the Ras-Pi 4, it is 37.37 s, which yields a performance gap of around 47.76%. The distance flown for the same policy on the desktop is 27.59 m, whereas on the Ras-Pi 4, it is 33.06 m, which contributes to a difference of 19.82%. Finally, the desktop consumes an average of 20 kJ of energy, while the Ras-Pi 4 consumes 25.4 kJ, which is 29.48% more energy. In the Static Obstacles case, the policy running on the desktop is 13% more successful than the policy running on Ras-Pi 4. The flight time to reach the goal on the high-end desktop is, on average, 30.25 s, whereas on the Ras-Pi 4, it is 34.44 s. That yields a performance gap of around 13.85%. For the distance flown, the policy running on the desktop has a trajectory length of 28.7 m, whereas the same policy on the Ras-Pi 4 has a trajectory length of 32.57 m. This contributes to a difference in 13.4%. For energy, the policy running on the desktop on an average consumes 19.2 kJ of energy, while policy running on Ras-Pi 4 on an average consumes 23.90 KJ of energy, which is about 32% more energy. In the Dynamic Obstacles case, the success rate between the desktop and the Ras-Pi 4 is 6%. The flight time to reach the goal on the desktop is, on average, 21.48 s, whereas on the Ras-Pi 4, it is 35.36 s, yielding a performance gap of around 64.61%. For the distance flown, the policy running on the desktop has a trajectory length of 23.51 m, whereas the same policy running on Ras-Pi 4 has a trajectory length of 32.86 m. This contributes to a difference in 40%. For energy, the policy running on the desktop, on average, consumes 18.76 kJ of energy while policy running on Ras-Pi 4 consumes 24.31 KJ of energy, which is about 30% more energy. Overall, across the three different environments, the policy evaluated on the Ras-Pi 4 achieves a success rate that is within 13% compared to the policy assessed on the desktop. While some degradation in performance is expected, the magnitude of the deterioration is more severe for the other QoF metrics, such as flight time, energy, and distance flown. This difference is significant to note because when the policies are ported to resource-constrained compute like the Ras-Pi 4 (a proxy for onboard compute in real UAVs), they could perform worse, such as being unable to finish the mission due to low battery. In summary, the takeaway is that evaluations of policies solely on a high-end machine do not accurately reflect the real-time performance on an embedded compute system such as those available on UAVs. Hence, relying on success rate as the sole metric is insufficient, though this is by and a large state of the art means to report success. Using Air Learning and its HIL methodology and QoF metrics, we can understand to what extent the choice of onboard compute affects the performance of the algorithm. 7.3 Root-cause analysis of SUT performance differences It is important to understand why the policy performs differently on the Intel Core i9 versus the Ras-Pi 4. So, we perform two experiments. First, we plot the policy trajectories on the Ras-Pi 4 and compare it to the Intel Core-i9 to understand if there is a flight path difference. Visualizing the trajectories helps us build intuition about the variations between the two platforms. Second, we take an Intel Core-i9 platform and degrade its performance by adding artificial sleep such that the policy evaluation times are similar to that of Ras-Pi 4. This helps us validate if the processing time is giving rise to the QoF metric discrepancy. To plot the trajectories, we fix the end goal’s position, obstacles, and evaluate 100 trajectories with the same configuration in the No Obstacles , Static Obstacles , and Dynamic Obstacles environments. The trajectories are shown in Fig. 10 a–c. They are representative of repeated trajectories between the start and end goals. The trajectories between the desktop and Ras-Pi 4 are very different—the desktop trajectory orients towards the goal and the proceeds directly. The Ras-Pi 4 trajectory starts toward the goal, but then drifts, resulting in a longer trajectory. This is likely a result of the actions taken because of stale sensory information, due to the longer inference time; recall there is a 20 \(\times\) difference in the inference time between the desktop and Ras-Pi 4 (Sect. 7.1 and Table 5 ). Fig. 10 Figures a , b , c compare the trajectories of Ras-Pi 4 and Intel Core-i9. The red columns in b and c denotes the position of static obstacles (Color figure online) Full size image To further root-cause and test whether the (slower) processing time ( \({t}_{{2}}\) ) is giving rise to the long trajectories, we take the best performing policy trained on the high-end desktop in the Static Obstacles environment and gradually degrade the policy’s evaluation time by introducing artificial sleep times into the program. Footnote 7 Sleep time injection allows us to model the big differences in the behavior of the same policy and its sensitivity to the onboard compute performance. Table 6 shows the effect of degrading the compute performance on policy evaluation. The baseline is the performance on the high-end Intel Core-i9 desktop. Intel Core-i9 (150 ms) and Intel Core-i9 (300 ms) are the scenarios where the performance of Intel Core-i9 is degraded by 150 ms and 300 ms, respectively. As performance deteriorates from 3 ms to 300 ms, the flight time degrades by 97%, the trajectory distance degrades by 21%, and energy degrades by 43%. Table 6 Degradation in policy evaluation using artificially injected program sleep (proxy for performance degradation) Full size table We visualize degradation impact by plotting the same policy’s trajectories on the baseline Intel Core-i9 system and the degraded versions of Intel Core-i9 systems (150 ms and 300 ms). The trajectory results are shown in Fig. 11 . As we artificially degrade, the drift in trajectories gets wider, which increases the trajectory length to reach the goal position, thus degrading the QoF metrics. We also see that the trajectory of the degraded Intel Core i9 closely resembles the Ras-Pi 4 trajectory. Fig. 11 Trajectory visualization of the best-performing policy on Intel Core i9 and artificially degraded versions of Intel Core i9 (150 ms) and Intel Core i9 (300 ms) Full size image In summary, the onboard compute choice and algorithm profoundly affect the resulting UAV behavior and shape of the trajectory. Additional quality of flight metrics (energy, distance, etc.) captures the differences better than just the success rate. Moreover, evaluations done purely on a high-end desktop might show lower energy consumption in a mission, but when the solution is ported to real robots, the solution might consume more energy due to the sub-par performance of the onboard compute. Using the hardware-in-the-loop (HIL) methodology allows us to identify these differences and other performance bottlenecks that arise due to the onboard compute without having to port things to the real robots. Hence, a tool like Air Learning with its HIL methodology helps identify such differences at the early stage. In the next section, we show how Air Learning HIL can mitigate the hardware gap and characterize the end-to-end learning algorithms and model these characteristics to create robust and performance-aware policies. 8 Mitigating the hardware gap In this section, we demonstrate how Air Learning HIL technique can be used to minimize the hardware gap due to differences in the training hardware and deployment hardware (onboard compute). To that end, we propose a general methodology where we train a policy on the high-end machine with added latencies to mimic the onboard compute’s performance. Using this method, we show that it minimizes the hardware gap from 38% to less than 0.5% on flight time metric, 16.03% to 1.37% on the trajectory length metric, and 15.49% to 0.1% on the energy of flight metric. One way to mitigate the hardware gap is to directly train the policy on the onboard computer available in the robot (Ahn et al., 2020 ; Ha et al., 2018 ; Kalashnikov et al., 2018 ). Though on-device RL training is practical for ground-based or fixed robots to overcome the ‘sim2real gap’ (Boeing & Bräunl, 2012 ; Koos et al., 2010 ), in the context of UAVs, training the RL policy on-device during flight has logistical limitations and not scalable (As explained in Sect. 2 ). Moreover, some of the onboard computers on these UAVs don’t have the necessary hardware resources required for on-device RL training. For instance, most hobbyist drones and research UAV platforms (e.g., CrazyFlie) are typically powered by microcontrollers and have a total of 1 MB memory. For most vision based navigation, its storage space is insufficient for the policy weights. Hence these resource constraints make RL training on-device extremely challenging. To overcome these resource constraint limitations and enable on-device RL training for UAVs, we introduce a methodology that uses HIL for training the RL policy. This methodology allows us to train the RL policy on a high-end machine (e.g., Intel Core-i9 with GPUs) while capturing the latencies incurred in processing the policy in the onboard computer. We describe the details of the methodology below. 8.1 Methodology The methodology is divided into three phases, namely ‘Phase 1’, ‘Phase 2’, and ‘Phase 3’ as shown in Fig. 12 . In Phase 1 (Fig. 12 a), we use the HIL to determine three specific latencies namely \(\hbox {t}_{{1}}\) , \(\hbox {t}_{{2}}\) , and \(\hbox {t}_{{3}}\) defined in Sect. 7.1 . We capture the latency distribution when the policies are run on-device (e.g., Ras-Pi). The distribution captures the variation in the decision-making times when the policy is deployed in the onboard computer. Fig. 12 A three-phase methodology for mitigating the hardware gap using hardware-in-the-loop training. a In phase 1, we use the hardware-in-the-loop methodology on a candidate policy to get the policy’s latency distribution on target hardware (Ras-Pi 4). We use prior work (Krishnan et al., 2020 ) as the cyber-physical model to determine the upper bound for maximum velocity. b In phase 2, we use the latency distribution to randomly sample the delay that needs to be added to the policy’s training. c In phase 3, the HIL trained policy is deployed on the target hardware for evaluation Full size image Once the latency distribution is captured, we calculate the maximum achievable velocities for safe navigation based on the decision-making time (Liu et al., 2016 ). This is to ensure that the drone can navigate safely without colliding with an obstacle. We evaluate the maximum safe velocity the aerial robot can travel based on the visual performance model proposed in this work (Krishnan et al., 2020 ). The model considers the time to action latency, drones’ physics (e.g., thrust-to-weight ratio, sensing distance, etc.) to determine the drone’s maximum safe velocity. In phase 2 (Fig. 12 b), we train the policy by adding extra delays sampled from the latency distribution determined in phase 1. The decision-making loop’s added delays mimic the typical processing delay when the policy is deployed on the resource-constrained onboard computer. The policy’s action space is also scaled based on the maximum velocity achievable based on the decision-making time (Krishnan et al., 2020 ; Liu et al., 2016 ). Once the policy is trained, in phase 3 (Fig. 12 c), we deployed the trained policy on the onboard compute (Ras-Pi) and evaluate its performance and quality of flight metrics. 8.2 Experimental setup and evaluation To validate the methodology, we train a policy on the Static Obstacle environment with at most two to three obstacles. The candidate architecture policy is 5 Layers with 32 Filters based on the template defined in Fig. 7 . We use the HIL setup described in Fig. 9 to evaluate the decision making latency on Ras-Pi 4, which is our target resource-constrained hardware platform. The simulation environment is rendered on the Intel Core-i9 server. We deploy a randomly initialized policy on the Ras-Pi 4 at this stage to benchmark the latencies. We do a rollout of 1000 steps using HIL to capture the variations in decision-making times. On the high-end server (Intel core-i9 with GTX 2080 TI), we train the candidate policy for the same task (i.e., Static Obstacles) with added delay element in the decision-making loop. The delay element’s actual value is randomly sampled from the latency distribution obtained for the candidate policy (5 Layers with 32 Filters) running on the Ras-Pi 4. Also, based on the maximum value of the latency from the distribution, we estimate the upper limit for the safe velocity for drone (Krishnan et al., 2020 ; Liu et al., 2016 ). This upper limit in safe velocity is then used to scale the action space such that at any point, the drone’s velocity at each step does not exceed the maximum safe velocity. Once the candidate policy’s training with added latency is complete, we deploy the policy on the Ras-Pi 4 platform (target resource-constrained onboard compute). We use the HIL methodology to evaluate the quality of flight metrics on Ras-Pi 4. The comparison in trajectories between Core-i9 and Ras-Pi 4 is shown in Fig. 13 . The two trajectories are very similar to each other and do not suffer from larger drifts that were seen before. Table 7 compares the quality of flight metric. The performance gap (denoted by “Perf Gap”) is reduced from 38% to less than 0.5% on the flight time metric, 16.03% to 1.37% on the trajectory length metric, and 15.49% to 0.1% on the energy of flight metric. Fig. 13 Comparison of trajectory for a policy that uses mitigation technique (denoted by the label “With mitigation”) with the policy that does not use mitigation technique (represented by the label “Without mitigation”). The policy trained on the training machine (denoted by the label (“HIL”) is also plotted for comparison. Using the mitigation technique, we reduced the trajectory length degradation from 34.15 to 29.03 m (to within 1.37%) Full size image Table 7 Evaluation of quality of flight between Ras-Pi 4 and Intel Core-i9 with and without mitigation Full size table In summary, we show that training policy with added delay to mimic the target platform can be used to minimize the hardware gap and performance difference between the training machine and the resource-constrained onboard compute. 9 Future work The Air Learning toolset and benchmark that we built can be used for solving several open problems related to UAVs which spans multiple disciplines. The goal of this work was to demonstrate the breadth of Air Learning as an interdisciplinary tool. In the future, Air Learning can be used to address numerous other questions, including but not limited to the following. Environments: In this work, we focus primarily on UAV navigation for indoor applications (Khosiawan & Nielsen, 2016 ). Future work can extend Air Learning ’s environment generator to explore new robust reinforcement learning policies for UAV control under harsh environmental conditions. For instance, AirSim weather APIs can be coupled with Air Learning environment generator to explore new reinforcement learning algorithms for UAV control with different weather conditions. Footnote 8 Algorithm design: Reinforcement algorithms are susceptible to hyperparameter tuning, policy architecture, and reward function. Future work could involve using techniques such as AutoML (Zoph et al., 2017 ) and AutoRL (Chiang et al., 2019 ) to determine the best hyperparameters, and explore new policy architectures for different UAV tasks. Policy exploration: We designed a simple multi-modal policy and kept the policy architecture same across DQN and PPO agent. In future work, one could explore other types of policy architectures, such as LSTM (Bakker, 2002 ) and recurrent reinforcement learning (Li et al., 2015 ). Another future work could expand our work by exploring energy efficient policies by using the capability available in Air Learning to monitor energy consumption continuously. Energy-aware policies can be associated with open problems in mobile robots, such as charging station problem (Kundu & Saha, 2018 ). System optimization studies: Future work on the system optimization can be classified into two categories. First, one can perform a thorough workload characterization for reducing the reinforcement learning training time. System optimizations will speed up the training process, thus allowing us to build more complex policies and strategies (OpenAI, 2018 ) for solving open problems in UAVs. Second, path to building custom hardware accelerators to improve the onboard compute performance can be explored. Having specialized hardware onboard would allow better real-time performance for UAVs. 10 Conclusion We present Air Learning, a deep RL gym and cross-disciplinary toolset, which enables deep RL research for resource constraint systems, and an end-to-end holistic applied RL research for autonomous aerial vehicles. We use Air Learning to compare the performance of two reinforcement learning algorithms namely DQN and PPO on a configurable environment with varying static and dynamic obstacles. We show that for an end to end autonomous navigation task, DQN performs better than PPO for a fixed observation inputs, policy architecture and reward function. We show that the curriculum learning based DQN agent has a better success rate compared to non-curriculum learning based DQN agent with the same number of experience (steps). We then use the best policy trained using curriculum learning and expose the difference in the behavior of aerial robot by quantifying the performance of the policy using HIL methodology on a resource-constrained Ras-Pi 4. We evaluate the performance of the best policy using quality of flight metrics such as flight time, energy consumed and total distance traveled. We show that there is a non-trivial behavior change and up to 40% difference in the performance of policy evaluated in high-end desktop and resource-constrained Ras-Pi 4. We also artificially degrade the performance of the high-end desktop where we trained the policy. We observe a similar variation in the trajectory as well as other QoF metrics as observed in Ras-Pi 4 thereby showing how the onboard compute performance can affect the behavior of policies when ported to real UAVs. We also show the impact of energy QoF on the success rate of the mission. Finally, we propose a mitigation technique using the HIL technique that minimizes the hardware gap from 38% to less than 0.5% on the flight time metric, 16.03% to 1.37% on the trajectory length metric, and 15.49% to 0.1% on the energy of flight metric. Notes For example, an FPV hobbyist drone can be built under $100: . The environment generator can be applied to other challenges in aerial robots, such as detecting thin wires and coping with translucent objects. Demonstration of HIL: . We also support Keras-RL, another widely used RL framework. The sleep latency value that was added to Intel Core-i9 was determined by doing a ping test with the packet size equal to the size of the data (Depth Image) we fetch from the server and averaged it over 50 iterations. . Adding artificial sleep into the high-end desktop is a simple first-order approximation of the Ras-Pi 4 system. In reality, we cannot fully equate the high-end desktop to the Ras-Pi 4 since there are other differences (e.g., system architecture, memory sub-system, and power). AirSim plugin weather APIs can be found here: . . .

Roboticists worldwide have been trying to develop autonomous unmanned aerial vehicles (UAVs) that could be deployed during search and rescue missions or that could be used to map geographical areas and for source-seeking. To operate autonomously, however, drones should be able to move safely and efficiently in their environment. In recent years, reinforcement learning (RL) algorithms have achieved highly promising results in enabling greater autonomy in robots. However, most existing RL techniques primarily focus on the algorithm's design without considering its actual implications. As a result, when the algorithms are applied on real UAVs, their performance can be different or disappointing. For instance, as many drones have limited onboard computing capabilities, RL algorithms trained in simulations can take longer to make predictions when they are applied on real robots. These longer computation times can make a UAV slower and less responsive, which could in turn affect the outcome of a mission or result in accidents and collisions. Researchers at Harvard University and Google Research recently developed Air Learning, an open-source simulator and gym environment where researchers can train RL algorithms for UAV navigation. This unique environment, introduced in a paper published in Springer Link's Special Issue on Reinforcement Learning for Real Life, could help to improve the performance of autonomous UAVs in real-world settings. "To achieve true autonomy in UAVs, there is a need to look at system-level aspects such as the choice of the onboard computer," Srivatsan Krishnan, one of the researchers who carried out the study, told TechXplore. "Therefore, the primary objective of our study was to provide the foundational blocks that will allow researchers to evaluate these autonomy algorithms holistically." In Air Learning, UAV agents can be exposed to and trained on challenging navigation scenarios. More specifically, they can be trained on point-to-point obstacle avoidance tasks in three key environments, using two training techniques called deep Q networks (DQN) and proximal policy optimization (PPO) algorithms. "Air Learning provides foundational building blocks to design and evaluate autonomy algorithms in a holistic fashion," Krishnan said. "It provides OpenAI gym-compatible environment generators that will allow researchers to train several reinforcement learning algorithms and neural network-based policies." On the platform developed by Krishnan and his colleagues, researchers can assess the performance of the algorithms they developed under various quality-of-flight (QoF) metrics. For instance, they can assess the energy consumed by drones when using their algorithms, as well as their endurance and average trajectory length when utilizing resource-constrained hardware, such as a Raspberry Pi. "Once their algorithms are designed, researchers can use the hardware-in-the-loop to plug in an embedded computer and evaluate how the autonomy algorithm performs as if it's running on an actual UAV with that onboard computer," Krishnan said. "Using these techniques, various system-level performance bottlenecks can be identified early on in the design process." When running tests on Air Learning, the researchers found that there usually is a discrepancy between predicted performances and the actual functioning of onboard computers. This discrepancy can affect the overall performance of UAVs, potentially affecting their deployment, mission outcomes and safety. "Though we specifically focus on UAVs, we believe that the methodologies we used can be applied to other autonomous systems, such as self-driving cars," Krishnan said. "Given these onboard computers are the brain of the autonomous systems, there is a lack of systematic methodology on how to design them. To design onboard computers efficiently, we first need to understand the performance bottlenecks, and Air Learning provides the foundational blocks to understand what the performance bottlenecks are." In the future, Air Learning could prove to be a valuable platform for the evaluation of RL algorithms designed to enable the autonomous operation of UAVs and other robotic systems. Krishnan and his colleagues are now using the platform they created to tackle a variety of research problems, ranging from the development of drones designed to complete specific missions to the creation of specialized onboard computers. "Reinforcement learning is known to be notoriously slow to train," Krishnan said. "People generally speed up RL training by throwing more computing resources, which can be expensive and lower entry barriers for many researchers. Our work QuaRL (Quantized reinforcement learning) uses quantization to speed up RL training and inference. We used Air Learning to show the real-world application of QuaRL in deploying larger RL policies on memory-constrained UAVs." Onboard computers act as the "brains" of autonomous systems, thus they should be able to efficiently run a variety of algorithms. Designing these computers, however, can be highly expensive and lacks a systematic design methodology. In their next studies, therefore, Krishnan and his colleagues also plan to explore how they could automate the design of onboard computers for autonomous UAVs, to lower their cost and maximize UAV performance. "We already used Air Learning to train and test several navigation policies for different deployment scenarios," Krishnan said. "In addition, as part of our research on autonomous applications, we created a fully autonomous UAV to seek light sources. The work used Air Learning to train and deploy a light-seeking policy to run on a tiny microcontroller-powered UAV."

10.1007/s10994-021-06006-6

Medicine

Researchers find potential map to more effective HIV vaccine

H. Liao et al. Co-evolution of a broadly neutralizing HIV-1 antibody and founder virus. Nature DOI: 10.1038/nature12053 (2013). dx.doi.org/10.1038/nature12053 In the current study, scientists identified one of the roughly 20 percent of HIV-infected individuals who naturally develop broadly neutralizing antibodies to the virus after several years of infection. This person in Africa was a volunteer in a study in which participants gave weekly blood samples beginning early in the course of infection. This individual had joined the study just 4 weeks after infection and was followed for more than 3 years. Having blood samples from such an early stage enabled researchers to pinpoint the particular "founder" virus that triggered the immune system to make an immature broadly neutralizing antibody against HIV, as well as the cell from which that antibody emerged. Analyses of the weekly samples also enabled the scientists to see the series of changes that the virus and antibody underwent over 2.5 years until the antibody matured to a form capable of potently neutralizing the virus. Scientists are now attempting to create a vaccine that harmlessly mimics the virus at key points in the observed process to generate broadly neutralizing HIV antibodies, first in uninfected animals and then in uninfected people. Journal information: Nature

http://dx.doi.org/10.1038/nature12053

https://medicalxpress.com/news/2013-04-potential-effective-hiv-vaccine.html

Abstract Current human immunodeficiency virus-1 (HIV-1) vaccines elicit strain-specific neutralizing antibodies. However, cross-reactive neutralizing antibodies arise in approximately 20% of HIV-1-infected individuals, and details of their generation could provide a blueprint for effective vaccination. Here we report the isolation, evolution and structure of a broadly neutralizing antibody from an African donor followed from the time of infection. The mature antibody, CH103, neutralized approximately 55% of HIV-1 isolates, and its co-crystal structure with the HIV-1 envelope protein gp120 revealed a new loop-based mechanism of CD4-binding-site recognition. Virus and antibody gene sequencing revealed concomitant virus evolution and antibody maturation. Notably, the unmutated common ancestor of the CH103 lineage avidly bound the transmitted/founder HIV-1 envelope glycoprotein, and evolution of antibody neutralization breadth was preceded by extensive viral diversification in and near the CH103 epitope. These data determine the viral and antibody evolution leading to induction of a lineage of HIV-1 broadly neutralizing antibodies, and provide insights into strategies to elicit similar antibodies by vaccination. Main Induction of HIV-1 envelope (Env) broadly neutralizing antibodies (BnAbs) is a key goal of HIV-1 vaccine development. BnAbs can target conserved regions that include conformational glycans, the gp41 membrane proximal region, the V1/V2 region, glycan-associated C3/V3 on gp120, and the CD4-binding site 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 . Most mature BnAbs have one or more unusual features (long third complementarity-determining region of the heavy chain (HCDR), polyreactivity for non-HIV-1 antigens, and high levels of somatic mutations), suggesting substantial barriers to their elicitation 4 , 10 , 11 , 12 , 13 . In particular, CD4-binding site BnAbs have extremely high levels of somatic mutation, suggesting complex or prolonged maturation pathways 4 , 5 , 6 , 7 . Moreover, it has been difficult to find Env proteins that bind with high affinity to BnAb germline or unmutated common ancestors (UCAs), a trait that would be desirable for candidate immunogens for induction of BnAbs 7 , 14 , 15 , 16 , 17 , 18 . Although it has been shown that Env proteins bind to UCAs of BnAbs targeting the gp41 membrane proximal region 16 , 19 , and to UCAs of some V1/V2 BnAbs 20 , so far, heterologous Env proteins have not been identified that bind the UCAs of CD4-binding site BnAb lineages 7 , 18 , 21 , 22 , 23 , although they should exist 21 . Eighty per cent of heterosexual HIV-1 infections are established by one transmitted/founder virus 24 . The initial neutralizing antibody response to this virus arises approximately 3 months after transmission and is strain-specific 25 , 26 . The antibody response to the transmitted/founder virus drives viral escape, such that virus mutants become resistant to neutralization by autologous plasma 25 , 26 . This antibody–virus race leads to poor or restricted specificities of neutralizing antibodies in ∼ 80% of patients; however in ∼ 20% of patients, evolved variants of the transmitted/founder virus induce antibodies with considerable neutralization breadth, such as BnAbs 2 , 20 , 27 , 28 , 29 , 30 , 31 , 32 , 33 . There are several potential molecular routes by which antibodies to HIV-1 may evolve, and indeed, types of antibody with different neutralizing specificities may follow different routes 6 , 11 , 15 , 34 . Because the initial autologous neutralizing antibody response is specific for the transmitted/founder virus 31 , some transmitted/founder Env proteins might be predisposed to binding the germ line or UCA of the observed BnAb in those rare patients that make BnAbs. Thus, although neutralizing breadth generally is not observed until chronic infection, a precise understanding of the interaction between virus evolution and maturing BnAb lineages in early infection may provide insight into events that ultimately lead to BnAb development. BnAbs studied so far have only been isolated from individuals who were sampled during chronic infection 1 , 3 , 4 , 5 , 6 , 7 , 20 , 27 , 29 . Thus, the evolutionary trajectories of virus and antibody from the time of virus transmission to the development of broad neutralization remain unknown. We and others have proposed vaccine strategies that begin by targeting UCAs, the putative naive B-cell receptors of BnAbs with relevant Env immunogens to trigger antibody lineages with potential ultimately to develop breadth 6 , 11 , 13 , 14 , 15 , 16 , 18 , 19 , 21 . This would be followed by vaccination with Env proteins specifically selected to stimulate somatic mutation pathways that give rise to BnAbs. Both aspects of this strategy have proved challenging owing to a lack of knowledge of specific Env proteins capable of interacting with UCAs and early intermediate antibodies of BnAbs. Here we report the isolation of the CH103 CD4-binding site BnAb clonal lineage from an African patient, CH505, who was followed from acute HIV-1 infection to BnAb development. We show that the CH103 BnAb lineage is less mutated than most other CD4-binding site BnAbs, and may be first detectable as early as 14 weeks after HIV-1 infection. Early autologous neutralization by antibodies in this lineage triggered virus escape, but rapid and extensive Env evolution in and near the epitope region preceded the acquisition of plasma antibody neutralization breadth defined as neutralization of heterologous viruses. Analysis of the co-crystal structure of the CH103 Fab fragment and a gp120 core demonstrated a new loop-binding mode of antibody neutralization. Isolation of the CH103 BnAb lineage The CH505 donor was enrolled in the CHAVI001 acute HIV-1 infection cohort 35 approximately 4 weeks after HIV-1 infection ( Supplementary Fig. 1 ) and followed for more than 3 years. Single genome amplification of 53 plasma viral Env gp160 RNAs 24 from 4 weeks after transmission identified a single clade C transmitted/founder virus. Serological analysis demonstrated the development of autologous neutralizing antibodies at 14 weeks, CD4-binding site antibodies that bound to a recombinant Env protein (resurfaced stabilized core 3 (RSC3)) 5 at 53 weeks, and evolution of plasma cross-reactive neutralizing activity from 41–92 weeks after transmission 30 ( Fig. 1 , Supplementary Table 1 and Supplementary Fig. 2 ). The natural variable regions of heavy-chain ( V h DJ h ) and light-chain ( V l J l ) gene pairs of antibodies CH103, CH104 and CH106 were isolated from peripheral blood mononuclear cells (PBMCs) at 136 weeks after transmission by flow sorting of memory B cells that bound RSC3 Env protein 5 , 13 , 36 ( Fig. 1b ). The V h DJ h gene of antibody CH105 was similarly isolated, but no V l J l gene was identified from the same cell. Analysis of characteristics of V h DJ h (V h 4–59, posterior probability (PP) = 0.99; D3–16, PP = 0.74; J h 4, PP = 1.00) and V l J l (Vλ3–1, PP = 1.00; Jλ1, PP = 1.00) rearrangements in monoclonal antibodies CH103, CH104, CH105 and CH106 demonstrated that these antibodies were representatives of a single clonal lineage that we designated as the CH103 clonal lineage ( Fig. 2 and Supplementary Table 2 ). Figure 1: Development of neutralization breadth in donor CH505 and isolation of antibodies. a , Shown are HIV-1 viral RNA copies and reactivity of longitudinal plasmas samples with HIV-1 YU2 gp120 core, RSC3 and negative control RSC3Δ371Ile (ΔRSC3) proteins. b , PBMCs from week 136 were used for sorting CD19 + , CD20 + , IgG + , RSC3 + and ΔRSC3 − memory B cells (0.198%). Individual cells indicated as orange, blue and green dots yielded monoclonal antibodies CH103, CH104 and CH106, respectively, as identified by index sorting. c , The neutralization potency and breadth of the CH103 antibody are displayed using a neighbour-joining tree created with the PHYLIP package. The individual tree branches for 196 HIV-1 Env proteins representing major circulating clades are coloured according to the neutralization IC 50 values as indicated. d , Cross competition of CH103 binding to YU2 gp120 by the indicated HIV-1 antibodies, and soluble CD4-Ig was determined by ELISA. mAbs, monoclonal antibodies. PowerPoint slide Full size image Figure 2: CH103 clonal family with time of appearance, V H DJ H mutations and HIV-1 Env reactivity. a , b , Phylogenies of V h DJ h ( a ) and V l J l ( b ) sequences from sorted single memory B cells and pyrosequencing. The ancestral reconstructions for each were performed as described in the Methods. The phylogenetic trees were subsequently computed using neighbour-joining on the complete set of DNA sequences (see Methods) to illustrate the correspondence of sampling date and read abundance in the context of the clonal history. Within time-point V H monophyletic clades are collapsed to single branches; variant frequencies are indicated on the right. Isolated mature antibodies are red, pyrosequencing-derived sequences are black. The inferred evolutionary paths to observed matured antibodies are bold. c , Maximum-likelihood phylogram showing the CH103 lineage with the inferred intermediates (circles, I1–4, I7 and I8), and percentage mutated V H sites and timing (blue), indicated. d , Binding affinities ( K d , nM) of antibodies to autologous subtype C CH505 (C.CH505; left box) and heterologous B.63521 (right box) were measured by surface plasmon reasonance. PowerPoint slide Full size image Neutralization assays using a previously described 5 , 37 panel of 196 geographically and genetically diverse Env-pseudoviruses representing the major circulated genetic subtypes and circulating recombinant forms demonstrated that CH103 neutralized 55% of viral isolates, with a geometric mean half-maximum inhibitory concentration (IC 50 ) of 4.54 mg ml −1 among sensitive isolates ( Fig. 1c and Supplementary Table 3 ). Enzyme-linked immunosorbent assay (ELISA) cross-competition analysis demonstrated that CH103 binding to gp120 was competed by known CD4-binding site ligands such as monoclonal antibody VRC01 and the chimaeric protein CD4-Ig ( Fig. 1d ); CH103 binding to RSC3 Env was also substantially diminished by gp120, with Pro363Asn and Δ371Ile mutations known to reduce the binding of most CD4-binding site monoclonal antibodies 5 , 30 ( Supplementary Fig. 3 ). Molecular characterization of the CH103 BnAb lineage The RSC3 probe isolated CH103, CH104, CH105 and CH106 BnAbs by single-cell flow sorting. The CH103 clonal lineage was enriched by V h DJ h and V l J l sequences identified by pyrosequencing PBMC DNA 34 , 38 obtained 66 and 140 weeks after transmission, and complementary DNA antibody transcripts 6 obtained 6, 14, 53, 92 and 144 weeks after transmission. From pyrosequencing of antibody gene transcripts, we found 457 unique heavy- and 171 unique light-chain clonal members ( Fig. 2a, b ). For comprehensive study, a representative 14-member BnAb pathway was reconstructed from V h DJ h sequences (1AH92U, 1AZCET and 1A102R) recovered by pyrosequencing, and V h DJ h genes of the inferred intermediate (I) antibodies (I1–I4, I7, I8) 11 , 16 , 34 (T. B. Kepler, manuscript submitted; ) that were paired and expressed with either the UCA or I2 V l J l depending on the genetic distance of the V h DJ h to either the UCA or mature antibodies ( Fig. 2c and Supplementary Table 2 ). The mature CH103, CH104 and CH106 antibodies were paired with their natural V l J l . The CH105 natural V h DJ h isolated from RSC3 memory B-cell sorting was paired with the V l J l of I2. Whereas the V h DJ h mutation frequencies (calculated as described in the Methods) of the published CD4-binding site BnAbs VRC01, CH31 and NIH45-46 are 30–36% (refs 5 , 6 , 7 , 22 , 39 ), the V h DJ h frequencies of CH103 lineage CH103, CH104, CH105 and CH106 are 13–17% ( Fig. 2c ). Furthermore, antibodies in CH103 clonal lineage do not contain the large (>3 nucleotides) insertion or deletion mutations common in the VRC01 class of BnAbs 1 , 2 , 3 , with the exception of the V L J L of CH103, which contained a three amino-acid light-chain complementarity-determining region 1 (LCDR1) deletion. It has been proposed that one reason that CD4-binding site BnAbs are difficult to induce is because heterologous HIV- 7 , 18 , 22 . We wondered, however, whether the CH505 transmitted/founder Env, the initial driving antigen for the CH103 BnAb lineage, would preferentially bind to early CH103 clonal lineage members and the UCA compared to heterologous Env proteins. Indeed, a heterologous gp120 transmitted/founder Env, subtype B 63521 (B.63521), did not bind to the CH103 UCA ( Fig. 2d ) but did bind to later members of the clonal lineage. Affinity for this heterologous Env protein increased four orders of magnitude during somatic evolution of the CH103 lineage, with maximal dissociation constant ( K d ) values of 2.4–7.0 nM in the mature CH103–CH106 monoclonal antibodies ( Fig. 2d ). The CH103 UCA monoclonal antibody did not bind to heterologous transmitted/founder Env proteins AE.427299, B.9021 and C.1086 ( Supplementary Table 4 ), confirming lack of heterologous Env binding to CD4-binding site UCAs. Moreover, the gp120 Env RSC3 protein was also not bound by the CH103 UCA and earlier members of the clonal lineage ( Supplementary Fig. 3a ), and no binding was seen with RSC3 mutant proteins known to disrupt CD4-binding site BnAb binding ( Supplementary Fig. 3b ). In contrast to heterologous Env proteins, the CH505 transmitted/founder Env gp140 bound well to all of the candidate UCAs ( Supplementary Table 5 ), with the highest UCA affinity of K d = 37.5 nM. In addition, the CH505 transmitted/founder Env gp140 was recognized by all members of the CH103 clonal lineage ( Fig. 2d ). Whereas affinity to the heterologous transmitted/founder Env B.63521 increased by more than four orders of magnitude as the CH103 lineage matured, affinity for the CH505 transmitted/founder Env increased by no more than tenfold ( Fig. 2d ). To demonstrate Env escape from CH103 lineage members directly, autologous recombinant gp140 Env proteins isolated at weeks 30, 53 and 78 after infection were expressed and compared with the CH505 transmitted/founder Env for binding to the BnAb arm of the CH103 clonal lineage ( Supplementary Table 6 and Supplementary Fig. 4 ). Escape-mutant Env proteins could be isolated that were progressively less reactive with the CH103 clonal lineage members. Env proteins isolated at weeks 30, 53 and 78 lost UCA reactivity and only bound intermediate antibodies 3, 2 and 1, as well as BnAbs CH103, CH104, CH105 and CH106 ( Supplementary Table 6 ). In addition, two Env escape mutants from week-78 viruses also lost either strong reactivity to all intermediate antibodies or all lineage members ( Supplementary Table 6 ). To quantify CH103 clonal variants from initial generation to induction of broad and potent neutralization, we used pyrosequencing of antibody cDNA transcripts from five time points, weeks 6, 14, 53, 92 and 144 after transmission ( Supplementary Table 7 ). We found two V h DJ h chains closely related to, and possibly members of, the CH103 clonal lineage ( Fig. 2a , Supplementary Table 7 ). Moreover, one of these V h DJ h chains when reconstituted in a full IgG1 backbone and expressed with the UCA V l J l weakly bound the CH505 transmitted/founder Env gp140 at an end-point titre of 11 μg ml −1 ( Fig. 2a ). These reconstructed antibodies were present concomitant with CH505 plasma autologous neutralizing activity at 14 weeks after transmission ( Supplementary Fig. 2 ). Antibodies that bound the CH505 transmitted/founder Env were present in plasma as early as 4 weeks after transmission (data not shown). Both CH103 lineage V h DJ h and V l J l sequences peaked at week 53, with 230 and 83 unique transcripts, respectively. V h DJ h clonal members fell to 46 at week 144, and V l J l members dropped to 76 at week 144. Polyreactivity is a common trait of BnAbs, suggesting that the generation of some BnAbs may be controlled by tolerance mechanisms 10 , 21 , 40 . Conversely, polyreactivity can arise during the somatic evolution of B cells in germinal centres as a normal component of B-cell development 41 . The CH103 clonal lineage was evaluated for polyreactivity as measured by HEp-2 cell reactivity and binding to a panel of autoantigens 10 . Although earlier members of the CH103 clonal lineage were not polyreactive by these measures, polyreactivity was acquired together with BnAb activity by the intermediate antibody I2, I1 and clonal members CH103, CH104 and CH106 ( Supplementary Fig. 5a, b ). The BnAbs CH106 and intermediate antibody I1 also demonstrated polyreactivity in protein arrays with specific reactivity to several human autoantigens, including elongation factor-2 kinase and ubiquitin-protein ligase E3A ( Supplementary Fig. 5c, d ). Structure of CH103 in complex with HIV-1 gp120 Crystals of the complex between the CH103 Fab fragment and the ZM176.66 strain of HIV diffracted to 3.25 Å resolution, and molecular replacement identified solutions for CH103 Fab and for the outer domain of gp120 ( Fig. 3a ). Inspection of the CH103–gp120 crystal lattice ( Supplementary Fig. 6 ) indicated that the absence of the gp120 inner domain was probably related to proteolytic degradation of the extended gp120 core to an outer domain fragment. Refinement to a R work / R free ratio of 19.6%/25.6% ( Supplementary Table 8 ) confirmed a lack of electron density for gp120 residues amino-terminal to gp120 residue Val 255 or carboxy-terminal to Gly 472 (gp120 residues are numbered according to standard HXB2 nomenclature), and no electron density was observed for gp120 residues 301–324 (V3), 398–411 (V4) and 421–439 (β20–21). Superposition of the ordered portions of gp120 in complex with CH103 with the fully extended gp120 core bound by antibody VRC01 (ref. 7 ) indicated a highly similar structure (Cα root mean squared deviation (r.m.s.d.) 1.16 Å) ( Fig. 3b ). Despite missing portions of core gp120, the entire CH103 epitope seemed to be present in the electron density for the experimentally observed gp120 outer domain. Figure 3: Structure of antibody CH103 in complex with the outer domain of HIV-1 gp120. a , Overall structure of the CH103–gp120 complex, with gp120 polypeptide depicted in red ribbon and CH103 shown as a molecular surface (heavy chain in green, light chain in blue). Major CH103-binding regions on gp120 are coloured orange for loop D, yellow for the CD4-binding site, and purple for loop V5. b , Superposition of the outer domain of gp120 bound by CH103 (red), and core gp120 bound by VRC01 (grey), with polypeptide shown in ribbon representation. c , CH103 epitope (green) on gp120 outer domain (red), with the initial CD4-binding site superposed (yellow boundaries) in surface representation. d , Sequence alignment of outer domains of the crystallized gp120 shown on the first line, and diverse HIV-1 Env proteins recognized by CH103. Secondary structure elements are labelled above the alignment, with grey dashed lines indicating disordered regions. Symbols in yellow or green denote gp120 outer domain contacts for CD4 and CH103, respectively, with open circles representing main-chain contacts, open circles with rays representing side-chain contacts, and filled circles representing both main-chain and side-chain contacts. PowerPoint slide Full size image The surface bound by CH103 formed an elongated patch with dimensions of ∼ 40 × 10 Å, which stretched across the site of initial CD4 contact on the outer domain of gp120 ( Fig. 3c ). The gp120 surface recognized by CH103 correlated well with the initial site of CD4 contact; of the residues contacted by CH103, only eight were not predicted to interact with CD4. CH103 interacted with these gp120 residues through side-chain contact with Ser 256 in loop D, main- and side-chain contacts with His 364 and Leu 369 in the CD4-binding loop, and main- and side-chain contacts with Asn 463 and Asp 464 in the V5 loop ( Fig. 3d ). Notably, residue 463 is a predicted site of N -linked glycosylation in strain ZM176.66 as well as in the autologous CH505 virus, but electron density for an N -linked glycan was not observed. Overall, of the 22 residues that monoclonal antibody CH103 was observed to contact on gp120, 14 were expected to interact with CD4 (16 of these residues with antibody VRC01), providing a structural basis for the CD4-epitope specificity of CH103 and its broad recognition ( Supplementary Table 9 ). Residues 1–215 on the antibody heavy chain and 1–209 on the light chain showed well-defined backbone densities. Overall, CH103 uses a CDR H3 dominated mode of interaction, although all six of the complementarity-determining regions (CDRs) interacted with gp120 as well as the light-chain framework region 3 (FWR3) ( Supplementary Fig. 7a, b and Supplementary Tables 10 and 11 ). It is important to note that ∼ 40% of the antibody contact surface was altered by somatic mutation in the HCDR2, LCDR1, LCDR2 and FWR3. In particular, residues 56 on the heavy chain, and residues 50, 51 and 66 on the light chain are altered by somatic mutation to form hydrogen bonds with the CD4-binding loop, loop D and loop V5 of gp120. Nevertheless, 88% of the CH103 V h DJ h and 44% of the V l J l contact areas were with amino acids unmutated in the CH103 germ line, potentially providing an explanation for the robust binding of the transmitted/founder Env to the CH103 UCA ( Supplementary Fig. 7c, d and Supplementary Table 12 ). Evolution of transmitted/founder Env sequences Using single genome amplification and sequencing 24 we tracked the evolution of CH505 env genes longitudinally from the transmitted/founder virus to 160 weeks after transmission ( Fig. 4 and Supplementary Fig. 8 ). The earliest recurrent mutation in Env, Asn279Lys (HIV-1 HXB2 numbering), was found 4 weeks after infection, and was in Env loop D in a CH103 contact residue. By week 14, additional mutations in loop D appeared, followed by mutations and insertions in the V1 loop at week 20. Insertions and mutations in the V5 loop began to accumulate by week 30 ( Fig. 4 ). Thus, the transmitted/founder virus began to diversify in key CD4 contact regions starting within 3 months of infection ( Supplementary Figs 8 and 9 ). Loop D and V5 mutations were directly in or adjacent to CH103/Env contact residues. Although the V1 region was not included in the CH103–Env co-crystal, the observed V1 CH505 Env mutations were adjacent to contact residues for CD4 and VRC01 so are likely to be relevant. It is also possible that early V1 insertions ( Fig. 4 ) were selected by inhibiting access to the CD4-binding site in the trimer or that they arose in response to early T-cell pressure. CD4-binding-loop mutations were present by week 78. Once regions that could directly affect CH103-lineage binding began to evolve (loop D, V5, the CD4-binding loop, and possibly loop V1), they were under sustained positive selective pressure throughout the study period ( Fig. 4 , Supplementary Figs 8 and 9 and Supplementary Table 13 ). Figure 4: Sequence logo displaying variation in key regions of CH505 Env proteins. The frequency of each amino acid variant per site is indicated by its height, deletions are indicated by grey bars. The first recurring mutation, Asn279Lys, appears at week 4 (open arrow). The timing of BnAb activity development (from Supplementary Fig. 2 and Supplementary Table 1 ) is on the left. Viral diversification, which precedes acquisition of breadth, is highlighted by vertical arrows to the right of each region. CD4 and CH103 contact residues, and amino acid position numbers based on HIV-1 HXB2, are shown along the base of each logo column. PowerPoint slide Full size image Considerable within-sample virus variability was evident in Env regions that could affect CH103-lineage antibody binding, and diversification in these regions preceded neutralization breadth. Expanding diversification early in viral evolution (4–22 weeks after transmission; Supplementary Figs 8 and 9 ) coincided with autologous neutralizing antibody development, consistent with autologous neutralizing antibody escape mutations. Mutations that accumulated from weeks 41 to 78 in CH505 Env contact regions immediately preceded development of neutralizing antibody breadth ( Fig. 4 and Supplementary Figs 8 and 9 ). By weeks 30–53, extensive within-sample diversity resulted from both point mutations in and around CH103 contact residues, and to several insertions and deletions in V1 and V5 ( Supplementary Fig. 9 ). A strong selective pressure seems to have come into play between weeks 30 and 53, perhaps due to autologous neutralization escape, and neutralization breadth developed after this point ( Fig. 4 and Supplementary Figs 8 and 9 ). Importantly, owing to apparent strong positive selective pressure between weeks 30 and 53, there was a marked shift in the viral population that is evident in the phylogenetic tree, such that only viruses carrying multiple mutations relative to the transmitted/founder, particularly in CH103 contact regions, persisted after week 30. This was followed by extreme and increasing within time-point diversification in key epitope regions, beginning at week 53 ( Supplementary Fig. 9 ). Emergence of antibodies with neutralization breadth occurred during this time ( Supplementary Fig. 2 and Supplementary Table 1 ). Thus, plasma breadth evolved in the presence of highly diverse forms of the CH103 epitope contact regions ( Fig. 4 and Supplementary Fig. 2 ). To evaluate and compare the immune pressure on amino acids in the region of CH103 and CD4 contacts, we compared the frequency of mutations in evolving transmitted/founder sequences of patient CH505 during the first year of infection and in 16 other acutely infected subjects followed over time ( Supplementary Fig. 10 ). The accumulation of mutations in the CH505 viral population was concentrated in regions likely to be associated with escape from the CH103 lineage ( Supplementary Fig. 10a ), and diversification of these regions was far more extensive during the first six months of infection in CH505 than in other subjects ( Supplementary Fig. 10b ). However, by one year into their infections, viruses from the other subjects had also begun to acquire mutations in these regions. Thus, the early and continuing accumulation of mutations in CH103 contact regions may have potentiated the early development of neutralizing antibody breadth in patient CH505. Neutralization of viruses and the CH103 lineage Heterologous BnAb activity was confined to the later members (I3 and later) of the BnAb arm of the CH103 lineage, as manifested by their neutralization capacity of pseudoviruses carrying tier 2 Env proteins A.Q842 and B.BG1168 ( Fig. 5a ). Similar results were seen with Env proteins A.Q168, B.JRFL, B.SF162 and C.ZM106 ( Supplementary Tables 14 and 15 ). By contrast, neutralizing activity of clonal lineage members against the autologous transmitted/founder Env pseudovirus appeared earlier, with measurable neutralization of the CH505 transmitted/founder virus by all members of the lineage after the UCA except monoclonal antibody 1AH92U ( Fig. 5a ). Thus, within the CH103 lineage, early intermediate antibodies only neutralized the transmitted/founder virus, whereas later intermediate antibodies gained neutralization breadth, indicating evolution of neutralization breadth with affinity maturation, and CH103–CH106 BnAbs evolved from an early autologous neutralizing antibody response. Moreover, the clonal lineage was heterogeneous, with an arm of the lineage represented in Fig. 5a evolving neutralization breadth and another antibody arm capable of mediating only autologous transmitted/founder virus neutralization. Although some escape-mutant viruses are clearly emerging over time ( Supplementary Table 4 ), it is important to point out that, although the escape-mutant viruses are driving BnAb evolution, the BnAbs remained capable of neutralizing the CH505 transmitted/founder virus ( Fig. 5a ). Of note, the earliest mutations in the heavy-chain lineage clustered near the contact points with gp120, and these remained fixed throughout the period of study, whereas mutations that accumulated later tended to be further from the binding site and may be affecting binding less directly ( Supplementary Fig. 11 ). Thus, stimulation of the CH103 BnAbs occurs in a manner to retain reactivity with the core CD4-binding site epitope present on the transmitted/founder Env. One possibility that might explain this is that the footprint of UCA binding contracts to the central core binding site of the CH103 mature antibody. Obtaining a crystal structure of the UCA with the transmitted/founder Env should inform this notion. Another possibility is that because affinity maturation is occurring in the presence of highly diverse forms of the CD4-binding site epitope, antibodies that favour tolerance of variation in and near the epitope are selected instead of those antibodies that acquire increased affinity for particular escape Env proteins. In both scenarios, persistence of activity to the transmitted/founder form and early viral variants would be expected. Figure 5b and Supplementary Fig. 11 show views of accumulations of mutations or entropy during the parallel evolution of the antibody paratope and the Env epitope bound by monoclonal antibody CH103. Figure 5: Development of neutralization breadth in the CH103 clonal lineage. a , Phylogenetic CH103 clonal lineage tree showing the IC 50 (μg ml −1 ) of neutralization of the autologous transmitted/founder (C.CH505), heterologous tier clades A (A.Q842) and B (B.BG1168) viruses as indicated. b , Interaction between evolving virus and developing clonal lineage mapped on to models of CH103 developmental variants and contemporaneous virus. The outer domain of HIV gp120 is depicted in worm representation, with worm thickness and colour (white to red) mapping the degree of per-site sequence diversity at each time point. Models of antibody intermediates are shown in cartoon diagram, with somatic mutations at each time point highlighted in spheres and coloured red for mutations carried over from I8 to mature antibody, cyan for mutations carried over from I4 to mature antibody, green for mutations carried over from I3 to mature antibody, blue for mutations carried over from I2 to mature antibody, orange for mutations carried over from II to mature antibody, and magenta for CH103 mutations from I1. Transient mutations that did not carry all the way to mature antibody are coloured in deep olive. The antibody (paratope) residues are shown in surface representation and coloured by their chemical types as indicated. PowerPoint slide Full size image Vaccine implications In this study, we demonstrate that the binding of a transmitted/founder Env to a UCA B-cell receptor of a BnAb lineage was responsible for the induction of broad neutralizing antibodies, thus providing a logical starting place for vaccine-induced CD4-binding site BnAb clonal activation and expansion. Importantly, the number of mutations required to achieve neutralization breadth was reduced in the CH103 lineage compared to most CD4-binding site BnAbs, although the CH103 lineage had reduced neutralization breadth compared to more mutated CD4-binding site BnAbs. Thus, this type of BnAb lineage may be less challenging to attempt to recapitulate by vaccination. By tracking viral evolution through early infection we found that intense selection and epitope diversification in the transmitted/founder virus preceded the acquisition of neutralizing antibody breadth in this individual—thus demonstrating the viral variants associated with development of BnAbs directly from autologous neutralizing antibodies and illuminating a pathway for induction of similar B-cell lineages. These data have implications for understanding the B-cell maturation pathways of the CH103 lineage and for replicating similar pathways in a vaccine setting. First, we demonstrate in CH505 that BnAbs were driven by sequential Env evolution beginning as early as 14 weeks after transmission, a time period compatible with induction of this type of BnAb lineage with a vaccine given the correct set of immunogens. Second, whereas heterologous Env proteins did not bind with UCAs or early intermediate antibodies of this lineage, the CH505 transmitted/founder Env bound remarkably well to the CH103 UCA, and subsequent Env proteins bound with increased affinity to later clonal lineage members. This suggests that immunizations with similar sequences of Env or Env subunits may drive similar lineages. Third, the CH103 lineage is less complicated than those of the VRC01 class of antibodies because antibodies in this lineage have fewer somatic mutations, and no indels, except CH103 V l, which has a deletion of three amino acid residues in the LCDR1 region. It should also be noted that our study is in one patient. Nonetheless, in each BnAb patient, analysis of viral evolution should determine a similar pathway of evolved Env proteins that induce BnAb breadth. The observation that rhesus macaques infected with the CCR5-tropic simian/human immunodeficiency virus (SHIV)-AD8 virus frequently develop neutralization breadth 42 suggests that certain Env proteins may be more likely to induce breadth and potency than others. Polyreactivity to host molecules in the CH103 lineage arose during affinity maturation in the periphery coincident with BnAb activity. This finding is compatible with the hypothesis that BnAbs may be derived from an inherently polyreactive pool of B cells, with polyreactivity providing a neutralization advantage via heteroligation of Env and host molecules 21 , 43 . Alternatively, as CH103 affinity maturation involves adapting to the simultaneous presence of diverse co-circulating forms of the epitope 44 , the selection of antibodies that can interact with extensive escape-generated epitope diversification may be an evolutionary force that also drives incidental acquisition of polyreactivity. Thus, a candidate vaccine concept could be to use the CH505 transmitted/founder Env or Env subunits (to avoid dominant Env non-neutralizing epitopes) to initially activate an appropriate naive B-cell response, followed by boosting with subsequently evolved CH505 Env variants either given in combination, to mimic the high diversity observed in vivo during affinity maturation, or in series, using vaccine immunogens specifically selected to trigger the appropriate maturation pathway by high-affinity binding to UCA and antibody intermediates 11 . These data demonstrate the power of studying subjects followed from the transmission event to the development of plasma BnAb activity for concomitant isolation of both transmitted/founder viruses and their evolved quasispecies along with the clonal lineage of induced BnAbs. The finding that the transmitted/founder Env can be the stimulator of a potent BnAb and bind optimally to that BnAb UCA is a crucial insight for vaccine design, and could allow the induction of BnAbs by targeting UCAs and intermediate ancestors of BnAb clonal lineage trees 11 . Methods Summary Serial blood samples were collected from an HIV-1-infected subject CH505 from 4 to 236 weeks after infection. Monoclonal antibodies CH103, CH104 and CH106 were generated by the isolation, amplification and cloning of single RSC3-specific memory B cells as described 5 , 6 , 7 , 22 , 36 . V h DJ h and V l J l 454 pyrosequencing was performed on samples from five time points after transmission 6 . Inference of UCA, and identification and production of clone members were performed as described in the Methods (see also Kepler, T. B., manuscript submitted; ). Additional V h DJ h and V l J l genes were identified by 454 pyrosequencing 6 , 34 , 38 and select V h DJ h and V l J l genes were used to produce recombinant antibodies as reported previously 34 and described in the Methods. Binding of patient plasma antibodies and CH103 clonal lineage antibody members to autologous and heterologous HIV-1 Env proteins was measured by ELISA and surface plasmon resonance 19 , 34 , 43 , 45 , and neutralizing activity of patient plasma and CH103 antibody clonal lineage members was determined in a TZM-bl-based pseudovirus neutralization assay 5 , 37 , 46 . Crystallographic analysis of CH103 bound to the HIV-1 outer domain was performed as previously reported 7 , and as described in the Methods. Online Methods Study subject Plasma and PBMCs were isolated from serial blood samples that were collected from an HIV-1-infected subject CH505 starting 6 weeks after infection up to 236 weeks after infection ( Supplementary Table 1 ) and frozen at −80 °C and liquid nitrogen tanks, respectively. During this time, no antiretroviral therapy was administered. All work related to human subjects was in compliance with Institutional Review Board protocols approved by the Duke University Health System Institutional Review Board. Antibodies isolated from PBMCs were tested in binding 45 and neutralization assays 46 . Inference of UCA and identification of clone members The inference of the UCA from a set of clonally related genes is described elsewhere (Kepler, T. B., manuscript submitted; ). In brief, we parameterize the VDJ rearrangement process in terms of its gene segments, recombination points, and n-regions sequences (non-templated nucleotides polymerized in the recombination junctions by the action of terminal deoxynucleotide transferase). Given any multiple sequence alignment (A) for the set of clonally related genes and any tree (T) describing a purported history, we can compute the likelihood for all parameter values, and thus the posterior probabilities on the rearrangement parameters conditional on A and T. We can then find the unmutated ancestor with the greatest posterior probability, and compute the maximum likelihood alignment A* and tree T* given this unmutated ancestor, and then recompute the posterior probabilities on rearrangement parameters conditional on A* and T*. We iterate the alternating conditional maximizations until convergence is reached. We use ClustalW 47 for the multiple sequence alignment, dnaml (PHYLIP) to infer the maximum likelihood tree, and our own software for the computation of the likelihood over the rearrangement parameters. The variable regions of heavy- and light-chain ( V h DJ h and V l J l ) gene segments were inferred from the natural pairs themselves. The posterior probabilities for these two gene segments are 0.999 and 0.993, respectively. We first inferred the unmutated ancestor from the natural pairs as described above. We identified additional clonally related variable region sequences from deep sequencing and refine the estimate of the UCA iteratively. We identified all variable region sequences inferred to have been rearranged to the same V h DJ h and J h , and to have the correct CDR3 length. For each sequence, we counted the number of mismatches between the sequence and the presumed V h DJ h gene up to the codon for the second invariant cysteine. Each iteration was based on the CDR3 of the current posterior modal unmutated ancestor. For each candidate sequence, we computed the number of nucleotide mismatches between its CDR3 and the unmutated ancestor CDR3. The sequence was rejected as a potential clone member if the z -statistic in a test for difference between proportion is greater than two (ref. 48 ). Once the set of candidates has been thus filtered by CDR3 distance, the unmutated ancestor was inferred on that larger set of sequences as described above. If the new posterior modal unmutated ancestor differed from the previous one, the process was repeated until convergence was reached. Owing to the inherent uncertainty in unmutated ancestor inference, we inferred the six most likely V h UCA sequences resulting in four unique amino acid sequences that were all produced and assayed for reactivity with the transmitted/founder envelope gp140 ( Supplementary Table 5 ). Phylogenetic trees Maximum-likelihood phylograms were generated using the dnaml program of the PHYLIP package (version 3.69) using the inferred ancestor as the outgroup root, ‘speedy/rough’ disabled, and default values for the remaining parameters. For the large antibody data sets, neighbour-joining phylogenetic trees were generated using the EBI bioinformatics server ( ) using default parameter values. All neighbour-joining trees were generated subsequent to the inference of the unmutated ancestors. Isolation and expression of V h DJ h and V l J l genes The V h DJ h and V l J l gene-segment pairs of the observed CH103, CH104 and CH106 antibodies, and the V h DJ h gene segment of CH105 were amplified by reverse transcription followed by PCR (RT–PCR) of flow-sorted HIV-1 Env RSC3-specific memory B cells using the methods described previously 5 , 6 , 7 , 22 , 36 . To compare V h mutation frequency of CH103, CH104, CH105 and CH106 antibodies with that of previously published of CD4-binding site BnAbs VRC01, CH31 and NIH45-46, V h sequences of these antibodies were aligned to the closest V h gene segment from the IMGT reference sequence set, and differences between the target sequence and the V h gene segment up to and including the second invariant cysteine were counted. The comparison 3′ of Cys 2 is omitted because the unmutated form of the ancestral sequence is not as well known. Additional V h DJ h and V l J l genes were identified by 454 pyrosequencing. Clonally related V h DJ h and V l J l sequences derived from either sorted single B cells or 454 pyrosequencing were combined and used to generate neighbour-joining phylogenetic trees ( Fig. 2a, b ). Antibodies that were recovered from single memory B cells are noted in the figure in red, and bold lines show the inferred evolutionary paths from the UCA to mature BnAbs. For clarity, related V h variants that grouped within monophyletic clades from the same time point were collapsed to single branches, condensing 457 V h DJ h and 174 V l J l variants to 119 and 46 branches, respectively, via the ‘nw_condense’ function from the Newick Utilities package (v. 1.6) 49 . The frequencies of V h DJ h variants in each B-cell sample are shown to the right of the V h DJ h tree in Fig. 2a , and were computed from sample sizes of 188,793, 186,626 and 211,901 sequences from weeks 53, 92 and 144, respectively. Two V h DJ h genes (IZ95W and 02IV4) were found at 14 weeks after transmission and paired with UCA V l J l for expression as IgG1 monoclonal antibodies. The IZ95W monoclonal antibody weakly bound the CH505 transmitted/founder Env gp140 with an end-point titre of 11 μg ml −1 . Among heavy-chain sequences in the tree, the mean distance of each to its nearest neighbour was calculated to be 8.1 nucleotides. The cumulative distribution function shows that, although there are pairs that are very close together (nearly 30% of sequences are 1 nucleotide from its neighbour), 45% of all sequences differ by 6 nucleotides or more from its nearest neighbour. The probability of generating a sequence that differs by 6 or more nucleotides from the starting sequence by PCR and sequencing is very small. The numbers of sequences obtained from a total of 100 million PBMCs were within the expected range of 50 to 500 antigen-specific B cells. We have analysed the number of unique V h DJ h and V l J l genes that we have isolated in several ways. First, we have clarified the calculations for the possible number of antigen-specific CD4-binding site memory B cells that could have been isolated from the samples studied. We studied five patient CH505 time points with pyrosequencing, with ∼ 20 million PBMCs per time point for a total of 100 million PBMCs studied. In chronic HIV infection, there is a mean of 145 total B cells per microlitre of blood, and 60 memory B cells per microlitre of blood 50 . This high percentage of memory B cells of ∼ 40% of the total B cells in chronic HIV infection is due to selective loss of naive B cells in HIV infection. Thus, in 100 ml of blood, there will be approximately 6 million memory B cells. If 0.1–1.0% are antigen specific, that would be 6,000–60,000 antigen-specific B cells sampled, and if, of these, 5% were CD4-binding site antibodies, then from 300 to 3000 CD4-binding site B cells would have been sampled in 100 million PBMCs studied. We studied 100 million PBMCs, therefore there should, by these calculations, be 1,000 CD4-binding site B cells sampled. This calculation therefore yields estimates that are completely compatible with the 474 V h DJ h genes amplified. To study the plausibility of sequences isolated further, the second method of analysis we used was as follows. Among heavy-chain sequences in the tree, one can compute the distance of each to its nearest neighbour. The mean distance to the nearest neighbour is 8.1 nucleotides. The cumulative distribution function shows that, although there are pairs that are very close together (nearly 30% of sequences are 1nt from its neighbour), 45% of all sequences differ by 6 nucleotides or more from its nearest neighbour. The probability of generating a sequence that differs by 6 or more nucleotides from the starting sequence by PCR and sequencing is very small. We believe the number of genes represented in our sample is closer to 200 than to 50, and most likely is larger than 200. The third analysis we performed was to compute the distance of each heavy-chain sequences in the tree to its nearest neighbour. The mean distance to the nearest neighbour is 8.1 nucleotides. We used agglomerative clustering to prune the sequence alignment. At the stage where no pairs of sequences were 3 nucleotides apart or closer, there were 335 out of 452 sequences remaining; when no pairs are 6 nucleotides apart or closer, there are still 288 sequences remaining. Therefore, with this analysis, we believe the number of genes represented in our sample is closer to 300 than to 50, and may be larger. Thus, by the sum of these re-analyses, we believe that the number of genes in the trees in Fig. 2 is plausible. The isolated Ig V h DJ h and V l J l gene pairs, the inferred UCA and intermediate V h DJ h and V l J l sequences, and select V h DJ h gene sequences identified by pyrosequencing were studied experimentally ( Supplementary Table 2 ), and used to generate a phylogenetic tree showing the percentage of mutated V h sites and time of appearance after transmission ( Fig. 2c ) and binding affinity ( Fig. 2d ). The isolated four mature antibodies are indicated in red, antibodies derived from 454 pyrosequencing are indicated in black, and inferred-intermediate antibodies (I1–I4, I7 and I8) are indicated by circles at ancestral nodes. The deep clades in this tree had modest bootstrap support, and the branching order and UCA inference were altered when more sequences were added to the phylogenetic analysis (compare the branching order of Fig. 2a and c ). The tree depicted in Fig. 2c, d was used to derive the ancestral intermediates of the representative lineage early in our study, and marked an important step in our analysis of antibody affinity maturation. The V h DJ h and V l J l genes were synthesized (GenScript) and cloned into a pcDNA3.1 plasmid (Invitrogen) for production of purified recombinant IgG1 antibodies as described previously 51 , 52 . The V h DJ h genes of I1–I4, I7 and I8 as well as the V h DJ h of CH105 were paired with either the V l gene of the inferred UCA or I2 depending on the genetic distance of the V h DJ h to either the UCA or mature antibodies for expressing as full-length IgG1 antibodies as described 51 ( Supplementary Table 2 ). Recombinant HIV-1 proteins HIV-1 Env genes for subtype B, 63521, subtype C, 1086, and subtype CRF_01, 427299, as well as subtype C, CH505 autologous transmitted/founder Env were obtained from acutely infected HIV-1 subjects by single genome amplification 24 , codon-optimized by using the codon usage of highly expressed human housekeeping genes 53 , de novo synthesized (GeneScript) as gp140 or gp120 (AE.427299) and cloned into a mammalian expression plasmid pcDNA3.1/hygromycin (Invitrogen). Recombinant Env glycoproteins were produced in 293F cells cultured in serum-free medium and transfected with the HIV-1 gp140- or gp120-expressing pcDNA3.1 plasmids, purified from the supernatants of transfected 293F cells by using Galanthus nivalis lectin-agarose (Vector Labs) column chromatography 16 , 52 , 54 , and stored at −80 °C. Select Env proteins made as CH505 transmitted/founder Env were further purified by superose 6 column chromatography to trimeric forms, and used in binding assays that showed similar results as with the lectin-purified oligomers. ELISA Binding of patient plasma antibodies and CH103 clonal lineage antibodies to autologous and heterologous HIV-1 Env proteins was measured by ELISA as described previously 34 , 52 . Plasma samples in serial threefold dilutions starting at 1:30 to 1:521,4470 or purified monoclonal antibodies in serial threefold dilutions starting at 100 μg ml −1 to 0.000 μg ml −1 diluted in PBS were assayed for binding to autologous and heterologous HIV-1 Env proteins. Binding of biotin-labelled CH103 at the subsaturating concentration was assayed for cross-competition by unlabelled HIV-1 antibodies and soluble CD4-Ig in serial fourfold dilutions starting at 10 μg ml −1 . The half-maximal effective concentration (EC 50 ) of plasma samples and monoclonal antibodies to HIV-1 Env proteins were determined and expressed as either the reciprocal dilution of the plasma samples or concentration of monoclonal antibodies. Surface plasmon resonance affinity and kinetics measurements Binding K d and rate constant (association rate ( K a )) measurements of monoclonal antibodies and all candidate UCAs to the autologous Env C. CH05 gp140 and/or the heterologous Env B.63521 gp120 were carried out on BIAcore 3000 instruments as described previously 19 , 43 , 45 . Anti-human IgG Fc antibody (Sigma Chemicals) was immobilized on a CM5 sensor chip to about 15,000 response units and each antibody was captured to about 50–200 response units on three individual flow cells for replicate analysis, in addition to having one flow cell captured with the control Synagis (anti-RSV) monoclonal antibody on the same sensor chip. Double referencing for each monoclonal antibody–HIV-1 Env binding interactions was used to subtract nonspecific binding and signal drift of the Env proteins to the control surface and blank buffer flow, respectively. Antibody capture level on the sensor surface was optimized for each monoclonal antibody to minimize rebinding and any associated avidity effects. C.CH505 Env gp140 protein was injected at concentrations ranging from 2 to 25 μg ml −1 , and B.63521 gp120 was injected at 50–400 μg ml −1 for UCAs and early intermediates IA8 and IA4, 10–100 μg ml −1 for intermediate IA3, and 1–25 μg ml −1 for the distal and mature monoclonal antibodies. All curve-fitting analyses were performed using global fit of to the 1:1 Langmuir model and are representative of at least three measurements. All data analysis was performed using the BIAevaluation 4.1 analysis software (GE Healthcare). Neutralization assays Neutralizing antibody assays in TZM-bl cells were performed as described previously 55 . Neutralizing activity of plasma samples in eight serial threefold dilutions starting at 1:20 dilution and for recombinant monoclonal antibodies in eight serial threefold dilutions starting at 50 μg ml −1 were tested against autologous and herologous HIV-1 Env-pseudotyped viruses in TZM-bl-based neutralization assays using the methods as described 5 , 37 , 55 . Neutralization breadth of CH103 was determined using a previously described 5 , 37 panel of 196 of geographically and genetically diverse Env-pseudoviruses representing the major circulated genetic subtypes and circulating recombinant forms. The subtypes shown in Fig. 1c are consistent with previous publications 5 , 56 , and the clades described in Los Alamos database ( ). HIV-1 subtype robustness is derived from the analysis of HIV-1 clades over time 57 . The data were calculated as a reduction in luminescence units compared with control wells, and reported as IC 50 in either reciprocal dilution for plasma samples or in micrograms per microlitre for monoclonal antibodies. Crystallization of antibody CH103 and its gp120 complex The antigen binding fragment (Fab) of CH103 was generated by LyS-C (Roche) digestion of IgG1 CH103 and purified as previously described 7 . The extended gp120 core of HIV-1 clade C ZM176.66 was used to form a complex with Fab CH103 by using previously described methods 58 . In brief, deglycosylated ZM176.66, constructed as an extended gp120 core 59 , that was produced using the method as described previously 7 and Fab CH103 were mixed at a 1:1.2 molar ratio at room temperature and purified by size-exclusion chromatography (Hiload 26/60 Superdex S200 prep grade, GE Healthcare) with buffer containing 0.35 M NaCl, 2.5 mM Tris, pH 7.0 and 0.02% NaN 3 . Fractions of the Fab or gp120–CH103 complex were concentrated to ∼ 10 mg ml −1 , flash frozen with liquid nitrogen before storing at −80 °C and used for crystallization screening experiments. Commercially available screens, Hampton Crystal Screen (Hampton Research), Precipitant Synergy Screen (Emerald BioSystems), Wizard Screen (Emerald BioSystems), PACT Suite and JCSG+ (Qiagen) were used for initial crystallization screening of both Fab CH103 and its gp120 complex. Vapour-diffusion sitting drops were set up robotically by mixing 0.2 μl of protein with an equal volume of precipitant solutions (Honeybee 963, DigiLab). The screen plates were stored at 20 °C and imaged at scheduled times with RockImager (Formulatrix.). The Fab CH103 crystals appeared in a condition from the JCSG+ kit containing 170 mM ammonium sulphate, 15% glycerol and 25.5% PEG 4000. For the gp120–CH103 complex ( Supplementary Table 8 ), crystals were obtained after 21 days of incubation in a fungi-contaminated 60 , 61 droplet of the PACT suite that contained 200 mM sodium formate, 20% PEG 3350 and 100 mM bistrispropane, pH 7.5. X-ray data collection and structure determination for gp120–CH103 Diffraction data were collected under cryogenic conditions. Optimal cryo-protectant conditions were obtained by screening several commonly used cryo-protectants as described previously 7 . X-ray diffraction data were collected at beam-line ID-22 (SER-CAT) at the Advanced Photon Source, Argonne National Laboratory, with 1.0000 Å radiation, processed and reduced with HKL2000 (ref. 62 ). For the Fab CH103 crystal, a data set at 1.65 Å resolution was collected with a cryo-solution containing 20% ethylene glycol, 300 mM ammonium sulphate, 15% glycerol and 25% PEG 4000 ( Supplementary Table 8 ). For the gp120–CH103 crystals, a data set at 3.20 Å resolution was collected using a cryo-solution containing 30% glycerol, 200 mM sodium formate, 30% PEG 3350 and 100 mM bistrispropane, pH 7.5 ( Supplementary Table 8 ). The Fab CH103 crystal was in the P 2 1 space group with cell dimensions at a = 43.0, b = 146.4, c = 66.3, α = 90.0, β = 97.7 and γ = 90.0, and contained two Fab molecules per asymmetric unit ( Supplementary Table 8 ). The crystal structures of Fab CH103 were solved by molecular replacement using Phaser 63 in the CCP 4 program suite 64 with published antibody structures as searching models. The gp120–CH103 crystal also belonged to the P 2 1 space group with cell dimensions at a = 48.9, b = 208.7, c = 69.4, α = 90, β = 107.2 and γ = 90.0, and contained two gp120–CH103 complexes per asymmetric unit ( Supplementary Table 8 ). The high-resolution CH103 Fab structure was used as an initial model to place the CH103 Fab component in the complex. With the CH103 Fab position fixed, searching with the extended gp120 core of ZM176.66 in the VRC01-bound form as an initial model failed to place the gp120 component in the complex. After trimming the inner domain and bridging sheet regions from the gp120 search model, Phaser was able to place correctly the remaining outer domain of gp120 into the complex without considerable clashes. Analysis of the packing of the crystallographic lattice indicated a lack of space to accommodate the inner domain of gp120, suggesting possible protease cleavage of gp120 by the containing fungi during crystallization 60 , 61 . Structural refinements were carried out with PHENIX 65 . Starting with torsion-angle simulated annealing with slow cooling, iterative manual model building was carried out on COOT 66 with maps generated from combinations of standard positional, individual B -factor, TLS (translation/libration/screw) refinement algorithms and non-crystallographic symmetry (NCS) restraints. Ordered solvents were added during each macro cycle. Throughout the refinement processes, a cross validation ( R free ) test set consisting of 5% of the data was used and hydrogen atoms were included in the refinement model. Structure validations were performed periodically during the model building/refinement process with MolProbity 67 and pdb-care 68 . X-ray crystallographic data and refinement statistics are summarized in Supplementary Table 8 . The Kabat nomenclature 69 was used for numbering of amino acid residues in amino acid sequences in antibodies. Protein structure analysis and graphical representations PISA 70 was used to perform protein–protein interfaces analysis. CCP 4 (ref. 66 ) was used for structural alignments. All graphical representation with protein crystal structures were made with Pymol 71 . Polyreactivity analysis of antibodies All antibodies in CH103 clonal lineage were assayed at 50 μg ml −1 for autoreactivity to HEp-2 cells (Inverness Medical Professional Diagnostics) by indirect immunofluorescence staining and a panel of autogens by antinuclear antibody assays using the methods as reported previously 10 . The intermediate antibody IA1 and CH106 were identified as reactive with HEp-2 cells and then selected for further testing for reactivity with human host cellular antigens using ProtoArray 5 microchip (Invitrogen) according to the instructions of the microchip manufacturer. In brief ProtoArray 5 microchips were blocked and exposed to 2 μg ml −1 IA1, CH106 or an isotype-matched (IgG1, k) human myeloma protein, 151 K (Southern Biotech) for 90 min at 4 °C. Protein–antibody interactions were detected by 1 μg ml −1 Alexa Fluor 647-conjugated anti-human IgG. The arrays were scanned at 635 nm with 10-μm resolution, using 100% power and 600 gain (GenePix 4000B scanner, Molecular Devices). Fluorescence intensities were quantified using GenePix Pro 5.0 (Molecular Devices). Lot-specific protein spot definitions were provided by the microchip manufacturer and aligned to the image. Accession codes Accessions GenBank/EMBL/DDBJ KC247375 KC247667 KC575845 KC576303 KC576304 KC576477 Protein Data Bank 4JAM 4JAN Data deposits The GenBank accession numbers for 292 CH505 Env proteins are KC247375 – KC247667 , and accessions for 459 V H DJ H and 174 V L J L sequences of antibody members in the CH103 clonal lineage are KC575845 – KC576303 and KC576304 – KC576477 , respectively. Atomic coordinates and structure factors for unbound CH103 Fab as well as CH103 Fab in complex with the ZM176.66 outer domain have been deposited with the Protein Data Bank under accession codes 4JAM for CH103 Fab, and 4JAN for the CH103–gp120 complex.

By tracking the very earliest days of one person's robust immune response to HIV, researchers have charted a new route for developing a long-sought vaccine that could boost the body's ability to neutralize the virus. The research team, led by Barton F. Haynes, M.D., director of the Duke Human Vaccine Institute, and John Mascola, M.D., acting director of the NIH Vaccine Research Center, have for the first time described the co-evolution of antibodies and virus in a person with HIV whose immune system mounted a broad attack against the pathogen. Findings are published April 3, 2013, in the journal Nature. Most vaccines work by inducing this antibody response, but the HIV virus has proved to be a difficult vaccine target. When HIV antibodies are produced, they typically have a limited range, and the virus changes rapidly to escape harm, leading to an arms race that the virus usually wins. The current research was aided by new technologies that can detect early infection and track the subsequent immune response and virus evolution. It fills gaps in knowledge that have impeded development of an effective vaccine for a virus that has killed more than 30 million people worldwide. "This project could only have been carried out by a multidisciplinary team working closely together," said Haynes, who led the work as a project of the Duke Center for HIV/AIDS Vaccine Immunology-Immunogen Discovery (CHAVI-ID) consortium, which is funded by the National Institute of Allergy and Infectious Diseases. "For the first time, we have mapped not only the evolutionary pathway of the antibody, but also the evolutionary pathway of the virus, defining the sequence of events involved that induce the broadly neutralizing antibodies." The evolution of the viral protein (green) from 14 weeks through 100 weeks post-transmission is compared to the maturation of the human antibody. Credit: Los Alamos National Laboratory The key to this finding was a person in Africa whose HIV infection was detected so early that the virus had not yet mutated to avoid the immune assault. The individual also exhibited a fortuitous trait that occurs in only about 20 percent of people infected with HIV – an immune system that produces broadly neutralizing antibodies. These immune weapons attack vulnerable sites of the virus that are conserved despite mutations. In identifying the early viral infection, the team found the outer envelope, the viral surface glycoprotein, which triggered the start of the broadly neutralizing antibody development. By tracking the precise virus and antibody pathways involved, the Duke CHAVI-ID and NIH teams now have a detailed road map for development of a potential vaccine, which involves immunogens with an outer envelope specifically selected to stimulate the production of broadly neutralizing antibodies. "The next step is to use that information to make sequential viral envelopes and test them as experimental vaccines," Haynes said. "This is a process of discovery and we've come a long way with regard to understanding what the problem has been."

dx.doi.org/10.1038/nature12053

Chemistry

Researchers discover a way to observe chromatin interaction changes in cancer-associated genetic mutations

Ciaran P. Seath et al, Tracking chromatin state changes using nanoscale photo-proximity labelling, Nature (2023). DOI: 10.1038/s41586-023-05914-y Journal information: Nature

https://dx.doi.org/10.1038/s41586-023-05914-y

https://phys.org/news/2023-04-chromatin-interaction-cancer-associated-genetic-mutations.html

Abstract Interactions between biomolecules underlie all cellular processes and ultimately control cell fate. Perturbation of native interactions through mutation, changes in expression levels or external stimuli leads to altered cellular physiology and can result in either disease or therapeutic effects 1 , 2 . Mapping these interactions and determining how they respond to stimulus is the genesis of many drug development efforts, leading to new therapeutic targets and improvements in human health 1 . However, in the complex environment of the nucleus, it is challenging to determine protein–protein interactions owing to low abundance, transient or multivalent binding and a lack of technologies that are able to interrogate these interactions without disrupting the protein-binding surface under study 3 . Here, we describe a method for the traceless incorporation of iridium-photosensitizers into the nuclear micro-environment using engineered split inteins. These Ir-catalysts can activate diazirine warheads through Dexter energy transfer to form reactive carbenes within an approximately 10 nm radius, cross-linking with proteins in the immediate micro-environment (a process termed µMap) for analysis using quantitative chemoproteomics 4 . We show that this nanoscale proximity-labelling method can reveal the critical changes in interactomes in the presence of cancer-associated mutations, as well as treatment with small-molecule inhibitors. µMap improves our fundamental understanding of nuclear protein–protein interactions and, in doing so, is expected to have a significant effect on the field of epigenetic drug discovery in both academia and industry. Main Mapping protein–protein interactions (PPIs) is central to our understanding of cellular biology 5 . The enormous challenges associated with this undertaking are magnified in the nucleus in which transient and multivalent interactions, fine-tuned by posttranslational modifications (PTMs), combine to choreograph DNA-templated processes such as transcription 6 . Perturbations to these regulatory mechanisms often lead to disease 7 , for example, somatic mutations that alter the composition and activity of chromatin-associated protein complexes are implicated in many human cancers and developmental disorders 8 , 9 , 10 . Moreover, recent studies have shown that histone proteins are themselves frequently mutated in cancers 11 , 12 , 13 . Understanding how these mutations lead to, or perpetuate, disease is the focus of intense investigation 11 , 12 , 13 , work that necessitates accurate comparative mapping of chromatin-associated PPI networks as a function of altered cell states 1 . The elucidation of nuclear PPIs has typically been performed by immunoprecipitation–mass spectrometry (IP–MS) workflows, in which antibodies recognizing selected proteins are used to enrich their target along with direct interactors 14 . However, IP–MS approaches rely on nuclear lysates as input, which may not be ideal for every system 15 , 16 , especially when the interactions are transient in nature (for example, driven by posttranslational modification) or require multiprotein complexes that bridge native chromatin 17 , 18 , 19 . This has fuelled the development of chemoproteomics approaches such as those based on photocrosslinking 20 , 21 or proximity-labelling 22 , 23 , 24 technologies that seek to capture PPIs in a native-like environment. Despite these ongoing advances, no single method exists to map chromatin interactomes in a general and unbiased manner and particularly to discern how such interactions are affected by perturbations such as mutation or drug treatment (Fig. 1a ). To address this need we considered the union of technologies recently disclosed by our two laboratories, µMap 4 and in nucleo protein trans -splicing 20 , to enable a traceless, short-range proximity-labelling method that can be readily deployed to any nuclear protein target. Protein trans -splicing using ultrafast split inteins facilitates the installation of iridium (Ir)-photocatalysts onto the N or C termini of target proteins. On irradiation with blue light-emitting diodes (LEDs) in the presence of a biotin–diazirine probe, localized carbene generation (through Dexter energy transfer from photoexcited iridium to aryl diazirines) allows interactomes to be determined, specifically those within about 10 nm of the iridium-centred photocatalyst (Fig. 1b ). Fig. 1: Development of a catalytic labelling platform using inteins. a , Chromatin interactomes can be perturbed through mutation leading to oncogenic phenotypes. Epigenetic drugs alter the chromatin interactome for therapeutic benefit. b , Cartoon showing strategy for nuclear photo-proximity labelling. Ir-photocatalysts can be incorporated into nuclear proteins via protein trans -splicing. The N-terminal fragment of Cfa N is fused to a target nuclear protein, whereas Cfa C is linked to the photocatalyst. In nucleo splicing provides Ir-conjugated nuclear proteins. Following excitation under blue light irradiation, the photocatalyst activates diazirines (to form carbenes via loss of molecular nitrogen) through a process called Dexter energy transfer. These carbenes are then poised to insert into neighbouring proteins that can be enriched using streptavidin beads. Comparative chemoproteomic analysis reveals how a given perturbation (for example, a mutation or drug treatment) affects local interactomes, providing insights into the mechanistic basis for disease or therapy. The mock volcano plot in Fig. 1 was created using BioRender . Full size image Our proposed workflow offers distinct advantages for the elucidation of subtly perturbed chromatin interactomes through mutation or ligand binding. Principally, the short radius afforded by this technology limits labelling to the close vicinity of a designated chromatin factor or nucleosome, only identifying proteins that are affected by a particular mutation or pharmacological intervention. This is critically important, given the structural and functional heterogeneity of chromatin 25 . Furthermore, the incorporation of the µMap catalyst is designed to be almost traceless and its small size is expected to minimize disruption to the native environment (for example, in comparison to fusion proteins), allowing the study of modifications that only minimally change the native interactome. We began our studies by preparing an Ir-conjugated intein fragment (complementary C-terminal fragment Cfa C -Ir), using a combination of solid-phase peptide synthesis and click chemistry (Fig. 2a,b ). We next fused the N-terminal fragment of the engineered Cfa split intein (Cfa N ) 26 , 27 to the C terminus of histone H3.1. For analytical convenience, we also included HA and FLAG epitope tags flanking the intein (Fig. 2c ). We then treated nuclei isolated from these transfected cells with the complementary split intein fragment (Cfa C ) linked to the Ir photocatalyst. This resulted in the site-selective incorporation of the photocatalyst onto the C terminus of H3.1 through in nucleo protein trans -splicing (Fig. 2e ). Irradiation of these nuclei in the presence of the diazirine–biotin probe (Fig. 2d ) led to dramatically enhanced protein labelling compared to control reactions that excluded Cfa C -Ir or diazirine–biotin (Fig. 2f ). Elution from streptavidin beads showed strong enrichment of histone H3 by western blot (Fig. 2f ). Fig. 2: Development of a chromatin-localized µMap proximity-labelling platform. a , Structure of the Cfa C -Ir construct used in this study. b , High-performance liquid chromatography trace (top) and electrospray ionization MS spectrum (bottom) for purified Cfa C -Ir (calculated [M] + = 6,514.6 Da; observed [M] + = 6,514.1 Da). c , General design of constructs used for µMap. d , Structure of diazirine-PEG3-biotin probe used in this study. e , Validation of the approach using histone H3. In nucleo protein trans -splicing with 0.5 μM Cfa C -Ir for 1 h results in the generation of the H3 spliced product (green) as shown by western blotting. f , Biotinylation of nuclear proteins, as detected by streptavidin blotting, is dependent on irradiation, Cfa C -Ir and the diazirine probe. H3 is enriched following labelling and streptavidin enrichment. g , Graphic representing differential localization between histone H3.1 and centromeric variant CENP-A. h , Volcano plot derived from a two-sided t -test showing proteins enriched in a comparative proteomics study between H3.1 and CENP-A using the µMap workflow. Selected proteins are highlighted that are discussed in the text. Dotted lines indicate cut-offs used. FDR values were calculated using the Benjamini–Hochberg procedure, as described in the Methods . For blot source data, see Supplementary Fig. 1 . Full size image We established a tandem mass tag (TMT)-based quantitative chemoproteomics workflow to determine the interactome for H3.1 versus the centromere-specific H3 variant CENP-A (Fig. 2g,h and Extended Data Fig. 1 ) to assess whether our workflow could delineate between nucleosomal interactomes at specific regions of chromatin. We used cut-offs of more than 0.5 log fold change (log 2 FC) and false discovery rate (FDR)-corrected P < 0.05 and hits were found in all replicates. Note, further analyses of the proteomic datasets in this study are provided in Extended Data Fig. 10 . Comparison of the H3.1 versus CENP-A interactomes returned established H3.1-modifying enzymes (for example, EHMT2 and SUV39H1) and reader proteins (for example, HP1 isoforms) (Fig. 2b and Extended Data Fig. 10 ; see Supplementary Information for full data table) 28 , 29 . CENP-A interacting proteins included both members of the FACT complex (SSRP1 and SPT16) involved in the deposition of CENP-A into chromatin (Fig. 2b ) 30 . Consistent with the centromeric localization of CENP-A, we see enrichment of transcriptional regulators (CBX6, DNMT3A, CHD4, RBBP4, KAT7 and KMT2A) only in the H3.1 samples. Furthermore, gene ontology (GO) analysis of the H3.1 hits showed strong enrichment of chromatin organization (−log 10 P = 56), histone lysine methylation (−log 10 P = 11) and epigenetic regulation of gene expression (−log 10 P = 25) (Extended Data Fig. 1 ), consistent with the known role of H3.1 as a platform for transcriptional regulation. With the basic chemoproteomics workflow established we questioned whether our method would sense subtle changes in nucleosomal structure, such as somatic mutations. Recent sequencing of patient tumour samples identified more than 4,000 histone mutations associated with a wide range of cancers 11 , 12 , 13 , 31 . Determining whether such mutations drive oncogenesis and cancer progression or are simply passenger mutations is critical to identifying new opportunities for therapeutic intervention. We sought to apply our method to the cancer-associated histone mutation H2A E92K, which is correlated with a range of cancers 13 . This mutation introduces a charge swap in the critical acidic patch interaction motif on the nucleosome, into which arginine residues of interacting proteins are known to anchor 32 , 33 , 34 , 35 . We hoped that our methodology could shed light on to what extent chromatin PPIs are perturbed by this oncohistone mutation. We found that the local chromatin micro-environment is indeed sensitive to the H2A E92K mutation (Fig. 3a and Extended Data Fig. 1 ; see Supplementary Information for full data table). GO analysis revealed diminished enrichment of proteins related to chromatin modification and re-organization as a function of the perturbed acidic patch (Fig. 3b ). Comparison of the dataset to the interactome of a construct with a triply mutated acidic patch (E61A, E90A and E92A) suggests that E92K affects a specific subset of acidic patch-binding proteins (Extended Data Fig. 2 ). Fig. 3: µMap as a method to uncover oncogenic function of the somatic mutation H2A E92K. a , A volcano plot derived from a two-sided t -test showing protein interactors from the µMap method comparing H2A to H2A E92K. Select proteins are highlighted. Dotted lines indicate cut-offs used. FDR values were calculated using the Benjamini–Hochberg procedure, as described in the Methods . b , Comparative GO analysis for wild-type versus E92K hits. GO terms consistent with the role of the acidic patch are de-enriched in the E92K mutant. c , Increased local acetylation of histone H4 is observed in the presence of H2A E92K, as determined by mononucleosome immunoprecipitation experiments. Western blot analysis showed an increase (2.1 ± 0.33) in acetylation in the E92K mutant compared to wild type. Bar plot shows change in acetylation levels on H4 normalized to wild type as determined by densitometry analysis of western blots (mean with s.d., n = 4 independent biological replicates, P = 0.0156). MN Pan-Ac, mononucleosome pan-acetylation levels. d , Volcano plot showing the concentration of significant (two-sided t -test, P < 0.05) ORFs in a comparative ATAC-seq experiment between HEK 293T cells stably expressing wild-type H2A versus H2A E92K. e , Violin plot comparing the overall concentration of ORFs between HEK 293T cells stably expressing wild-type H2A versus H2A E92K from ATAC-seq analysis ( P = <0.0001, two-sided t -test). The data show that the E92K population has an increased concentration of ORFs when compared to the wild-type sample. f , Left, graphic showing experimental design for measuring local DNA methylation. Right, box-and-whisker plots showing local concentration of 5-methyl deoxycytosine (Me-dC) normalized to global Me-dC levels. E92K mutant shows a 5% decrease in methylation ( n = 6 independent biological replicates, whiskers represent minimum to maximum, P = 0.0003, two-sided t -test). g , Western blot analysis comparing binding of selected proteins with either biotinylated wild-type nucleosomes or with biotinylated nucleosomes containing the H2A mutation E92K. * P < 0.05, *** P < 0.001, **** P < 0.0001. For blot source data, see Supplementary Fig. 1 . Full size image Notably, SIRT6, a deacetylase, was highly enriched in the wild-type sample, whereas the transcriptional activators, BRD2/3/4, were all enriched by E92K. Taken together, this suggests that the mutation acts to block histone lysine deacetylase activity, resulting in increased local acetylation and binding of associated reader proteins. To support this idea, we transiently expressed both wild-type and H2A E92K histones in HEK 293T cells and isolated intact mononucleosomes by anti-FLAG immunoprecipitation (Fig. 3c ). We found global lysine acetylation was increased twofold in the mutant nucleosomes above wild type, with a particular increase observed on histone H4. This effect is striking given that we are probably only enriching one copy of the mutant histone per nucleosome because of presumed stochastic incorporation (Extended Data Fig. 3 ) 36 , 37 . Following this, we questioned whether this local increase in acetylation would lead to a global effect by weakening DNA–nucleosome interactions, leading to a more accessible chromatin environment and consequently increasing the concentration of accessible open reading frames (ORFs). To probe this, we performed ATAC-seq analysis 38 on HEK 293T cell lines stably expressing either wild-type H2A or the E92K mutant (Extended Data Fig. 4 ). Consistent with our idea, we found 5,699 ORFs which showed a significant (FDR ≤ 0.05) change in concentration across four replicates. Of these, 90% indicate regions of higher accessibility in E92K in comparison to the wild-type H2A (Fig. 3d,e and Extended Data Fig. 5 ), consistent with our proposed model. We also noted that installation of E92K had a negative effect on nucleosomal DNMT3A binding (Fig. 3a ). Recent structural and biochemical work has demonstrated that DNMT3 binds to the acidic patch through two arginine residues (R740 and R743 in DNMT3B) 39 . Charge reversal mutants (for example, R740E) of these residues were shown to diminish nucleosomal binding and de novo methylation. Our data indicate that E92K may disrupt binding in the same manner. To verify this, we performed a mononucleosome FLAG-IP from HEK 293T cells stably expressing H2A or H2A E92K and analysed the precipitated DNA for MeC (Fig. 3f ). Consistent with our proposal, the data show an approximately 5% decrease in local DNA methylation in the E92K samples, supporting a model in which DNMT3 binding is impaired by E92K, leading to fewer de novo methylation events. We were able to validate loss of function binding interactions with SIRT6, DNMT3A, INCENP (part of the CEN complex 40 ) and the chromatin remodeller SMARCA4 (refs. 8 , 41 ) (BRG1) using the recombinant mononucleosome assay developed by McGinty 35 . Here, biotinylated mononucleosomes were immobilized on streptavidin beads before incubation with HEK 293T nuclear lysate. The eluent was subsequently blotted against hits from our photo-proximity labelling experiment to assess the effect of the E92K mutation on binding. In all cases, this assay supported the data presented by our µMap experiment (Fig. 3g ), with comparable enrichment values between western blot and proteomics datasets, further demonstrating that our method can accurately capture interacting proteins in an unbiased manner. Interestingly, performing the same comparative proteomics experiment with H2A/H2A(E92K)-APEX2 constructs did not reveal any differences in the interactome of mutant and wild-type samples when applying the same fold change and significance cut-offs used previously (Extended Data Figs. 6 and 10 ), presumably because of the increased labelling radius of the peroxidase-based platform (Extended Data Fig. 6 ). Together, these data demonstrate that the resolution provided by our proximity-labelling method can be used to uncover molecular level details for gain-of-function or loss-of-function interactions in the nucleus in a single experiment. We posited that our method could be used to determine the roles of small-molecule ligands in the chromatin micro-environment (Fig. 4a ). Epigenetic drug discovery has become a critical focus for therapeutic intervention, encompassing dozens of targets across a range of therapeutic areas 3 . We began by examining the effects that the bromodomain inhibitor JQ-1 has on chromatin (Fig. 4b ). JQ-1 is known to target the BET family of bromodomain containing readers (BRD2/3/4), with analogues progressing through the clinic 42 , 43 . We theorized that µMap would be sensitive enough to measure the effects that BRD inhibition would have on the chromosomal micro-environment. To probe this, we compared the interactomes of Ir-conjugated H2A in the presence or absence of 1 and 5 µM JQ-1 (Fig. 4b and Extended Data Figs. 4 , 7 and 10 ; see Supplementary Information for full data table). BRD2/3/4 were all enriched in the untreated sample (log 2 FC > 0.5, FDR < 0.05), consistent with blocking bromodomain–nucleosome interactions. The known JQ-1 off-target SOAT1 (ref. 44 ) was also enriched in the untreated sample. RNA-seq analysis of the ligand-treated cells showed that the identified interactors are not enriched on the basis of transcriptional changes associated with drug treatment (Extended Data Fig. 8 ). Fig. 4: Interactome mapping of nuclear proteins to assess ligand interactions. a , Performing the µMap workflow in the presence of a bioactive ligand unveils PPIs that are disrupted or promoted by ligand treatment. This method can provide nuclear target identification data. b , Volcano plot derived from a two-sided t -test showing the H2A interactome versus H2A + 5 µM JQ-1 (structure shown at left). This method identifies BRD2/3/4 as JQ-1 target proteins in addition to known off-target SOAT1. c , Volcano plot derived from a two-sided t -test showing the H2A interactome versus H2A + 2.5 µM pinometostat. Pinometostat target DOT1L is enriched in untreated cells, in addition to several proteins associated with loss of H3K79me. d , Summary of enriched proteins and their biological relevance is shown. e , RNA Pol II transcription is controlled by phosphorylation of the CTD. AT7519 inhibits CTD phosphorylation, stalling transcription. f , Volcano plot derived from a two-sided t -test showing RPB1-Ir versus RPB1-Ir + 2 µM AT7519. Dotted lines indicate cut-offs used. FDR values were calculated using the Benjamini–Hochberg procedure, as described in the Methods . Full size image Performing the same experiment using existing methods (APEX2, as above) showed enrichment of BRD2 and no enrichment of BRD3/BRD4 (Extended Data Fig. 7 ). We then performed a similar experiment with the DOT1L methyltransferase inhibitor, pinometostat, to assess both the selectivity of this ligand and the effect that depletion of H3K79 methylation has on the chromatin micro-environment 45 , 46 . DOT1L was enriched (log 2 FC > 0.5, FDR < 0.05) in the untreated sample, as were several proteins related to transcriptional activation, including BRD2/3/4 and POLR2E (Fig. 4c,d and Extended Data Fig. 7 ; see Supplementary Information for full data table). This observation is consistent with previous reports demonstrating that H3K79 methylation leads to recruitment of the acetyltransferase P300, subsequent BRD recruitment and transcriptional activation 46 . It was gratifying to see that in a single experiment we can extract both target-ID data for a small-molecule ligand along with downstream transcriptional effects. Finally, we applied our method to RNA polymerase II (POL II), which is responsible for transcribing protein-encoding genes in a highly regulated process involving the sequential association of multiple protein complexes 47 , 48 , 49 . Release of promoter-proximal paused POL II is primarily achieved through phosphorylation of the disordered C-terminal domain (CTD) of the RPB1 subunit, comprised of a repeating unit of Y-S-P-T-S-P-S, by cyclin dependent kinases (CDKs). Inhibition of CDKs has therefore become an attractive therapeutic strategy to stall POL II and downregulate transcription and therefore cancer proliferation. The small-molecule ligand AT7519, a pan-CDK inhibitor, has been developed to probe this strategy in the context of multiple myeloma and is now in clinical trials 50 . We questioned if we could determine how this ligand affects the POL II interactome and whether our method could unveil at what stage in its transcriptional cycle POL II is arrested by AT7519 treatment. To assess this, we expressed RPB1-Cfa N in HEK 293T cells, in which the intein is fused to the C terminus of the CTD (Extended Data Fig. 9 ). The µMap workflow was then performed in the presence and absence of AT7519 (Fig. 4e ). Comparing our treated and untreated datasets revealed CDK11, 9 and 1 to be enriched (log 2 FC > 0.5, FDR < 0.05) in the untreated sample, consistent with wide-ranging CDK inhibition by AT7519 (Fig. 4d ; see Supplementary Information for full data table) 50 . Mediator complex subunits in addition to several transcription factors were enriched in the treated sample, suggesting that CDK inhibition halts the progression of POL II at the pre-initiation complex, before CTD phosphorylation and association of the NELF and DSIF complexes 51 . Conclusion In summary, we have developed a photocatalytic proximity-labelling technology that can be deployed across the nuclear proteome. The short-range diazirine activation mechanism allows the collection of precise interactomics data that are sensitive to single amino acid mutations and can be used to detect changes caused by external stimuli such as ligand incubation. We believe this method will be broadly applicable across nuclear biology for the study of disease-associated mutations. Also, this method is an effective tool for ligand target ID in chromatin, identifying on- and off-target proteins and revealing how treatment with these molecules affects local chromatin interactomes. Methods Solid-phase peptide synthesis Boc-N α -Cfa C -CFGSGK(alloc)G-NH 2 was synthesized on a 0.1 mmol scale by standard Fmoc solid-phase peptide synthesis using DIC-Oxyma activation on a CEM Liberty Blue microwave-assisted peptide synthesizer on ChemMatrix Rink amide resin. Each residue was double coupled during the synthesis. Fmoc deprotection was performed at room temperature by the addition of 20% piperidine in dimethylformamide (DMF) with 0.1 M hydroxybenzotriazole. Alloc deprotection was performed by the addition of 0.1 equivalent (equiv.) Pd(PPh 3 ) 4 and 2.5 equiv. N , N ʹ-dimethylbarbituric acid in dichloromethane (DCM) with nitrogen agitation. Treatment was performed twice at room temperature for 30 min. The resin was sequentially washed with 3× DCM, 3× DCM:DMF (1:1 v/v), 3× DMF and 1× 5% w/v sodium diethyldithiocarbamate in DMF. The resin was then split into 0.02 mmol aliquots and treated as follows. Cfa C -biotin: biotin (5 equiv.) was coupled to the deprotected lysine side chain with PyAOP (4.95 equiv.) and N , N -diisopropylethylamine (DIEA) (10 equiv.) activation in NMP for 4 h with nitrogen agitation. The resin was washed with 3× NMP, 3× DMF and 3× DCM. Side-chain deprotection and cleavage from the resin was affected by addition of a 92.5:2.5:2.5:2.5 v/v/v solution of TFA:TIPS:EDT:H 2 O for 130 min at room temperature. The cleavage solution was reduced to less than 5 ml volume under a positive pressure of N 2 and the crude peptide was precipitated using cold diethyl ether. The crude peptide was isolated by refrigerated centrifugation, resuspended in 50/50 v/v H 2 O:MeCN with 0.1% TFA and lyophilized to yield a white solid. Cfa C -Ir: the iridium photocatalyst 52 was conjugated to the deprotected lysine side chain by treatment with NHS-Ir (1.2 equiv.) and DIEA (2 equiv.) in DMF for 2 h with nitrogen agitation in the absence of light. The resin was washed with 3× DMF and 3× DCM. Side-chain deprotection and cleavage from the resin was affected by addition of a 95:2.5:2.5 v/v/v solution of TFA:TIPS:H 2 O for 130 min at room temperature in the absence of light. The cleavage solution was reduced to less than 5 ml volume under a positive pressure of N 2 and the crude peptide was precipitated using cold diethyl ether. The crude peptide was isolated by refrigerated centrifugation, resuspended in 50/50 v/v H 2 O:MeCN with 0.1% TFA and lyophilized to yield a pale yellow solid. Crude lyophilized peptides were purified by preparative scale reversed-phase high-performance liquid chromatography and characterized by mass spectrometry. Cell culture HEK 293T cells were cultured as a monolayer in DMEM (Thermo Fisher), supplemented with 10% v/v FBS (Thermo Fisher), 100 U ml −1 of penicillin (Thermo Fisher) and 100 µg ml −1 of streptomycin (Thermo Fisher). Cells were maintained in an incubator at 37 °C with 5% CO 2 . Each 10 cm plate of HEK 293T cells at 70% confluency was transfected with a plasmid encoding POI-HA-Cfa N -FLAG (5 μg per plate) with lipofectamine 2000 (12 μl per plate) following the manufacturer’s instructions. After 6 h, the medium was aspirated and replaced with fresh medium. Transfection was performed for 24 or 48 h in an incubator at 37 °C with 5% CO 2 . Protein trans -splicing for TMT-10-plex experiments A total of 3 × 10 7 HEK 293T cells transfected with POI-HA-Cfa-N-FLAG were lysed by hypotonic lysis in 3 ml of RSB buffer (10 mM Tris buffer, 15 mM NaCl, 1.5 mM MgCl 2 , Roche cOmplete EDTA-free protease inhibitors, pH 7.6) for 10 min on ice. The crude nuclei were isolated by centrifugation at 400 g for 5 min at 4 °C. The nuclei were resuspended in 3 ml of RSB buffer and homogenized with ten strokes of a loose pestle Dounce homogenizer and pelleted at 400 g for 5 min at 4 °C. The nuclei were resuspended in cross-linking buffer (20 mM HEPES, 1.5 mM MgCl 2 , 150 mM KCl, Roche cOmplete EDTA-free protease inhibitors, pH 7.6) and centrifuged at 400 g for 5 min at 4 °C. Finally, the nuclei were resuspended in 300 μl of cross-linking buffer per 1 × 10 7 cells. To the isolated nuclei was added Cfa C -Ir in cross-linking buffer (0.5 µM final concentration). The nuclei were incubated at 37 °C for 1 h. The nuclei were isolated by centrifugation at 400 g for 5 min at 4 °C and washed twice with cross-linking buffer (500 μl) to remove excess peptide. The pellets were then resuspended in 3 ml of cross-linking buffer containing diazirine–biotin conjugate (200 µM) and irradiated with blue light for between 45 and 90 s in the Penn PhD Photoreactor M2 at 100% light intensity at 4 °C. The nuclei were re-isolated by centrifugation at 400 g for 5 min at 4 °C and washed once with cross-linking buffer to remove excess biotin–diazirine. The washed pellets were then resuspended in 2 ml of LB3 buffer (10 mM Tris, 100 mM NaCl, 1 mM EDTA, 0.5 mM EGTA, 0.1% sodium deoxycholate, 0.5% sodium lauroyl sarcosinate, pH 7.5) and sonicated using a Branson probe tip sonicator (12 cycles at 25% amplitude, 15 s on 15 s off on ice). The lysed nuclei were then clarified through centrifugation at 15,000 g for 20 min at 4 °C and the protein concentration of the supernatant was determined by BCA assay. Protein concentration was normalized across all experimental replicates and diluted to 1 mg ml −1 with binding buffer (25 mM Tris, 150 mM NaCl, 0.25% v/v NP-40, pH 7.5). Then 1 ml of each sample was incubated with 125 µl of prewashed magnetic Sepharose streptavidin beads ( , no. 28985738) for 2 h at room temperature with end-over-end rotation. The beads were subsequently washed twice with 1% w/v SDS in PBS, twice with 1 M NaCl in PBS and 10% ethyl alcohol in PBS ×3. Experiments with ligand treatment Following transfection, plated cells expressing POI-HA-Cfa-N-FLAG were treated with the small-molecule ligand for the specified amount of time (Supplementary Table 2 ). Following this, cells were scraped and pelleted as described previously. All buffers used in cell processing before irradiation (RSB, cross-linking) contained the ligand. Following irradiation, samples were treated as normal. For proteomic analysis Following streptavidin enrichment, the beads were resuspended in PBS (300 µl) and transferred to a new 1.5 ml LoBind tube. The supernatant was removed and the beads were washed with 3× PBS (0.5 ml) and 3× ammonium bicarbonate (100 mM). The beads were resuspended in 500 µl of 3 M urea in PBS and 25 µl of 200 mM dithiothreitol in 25 mM NH 4 HCO 3 was added. The beads were incubated at 55 °C for 30 min. Subsequently, 30 µl of 500 mM iodoacetamide in 25 mM NH 4 HCO 3 was added and incubated for 30 min at room temperature in the dark. The supernatant was removed and the beads washed with 3× 0.5 ml of DPBS and 6× 0.5 ml of triethyl ammonium bicarbonate (TEAB, 50 mM). The beads were resuspended in 0.5 ml of TEAB (50 mM) and transferred to a new protein LoBind tube. The beads were resuspended in 40 µl of TEAB (50 mM), 1.2 µl of trypsin (1 mg ml −1 in 50 mM acetic acid) was added and the beads incubated overnight with end-over-end rotation at 37 °C. After 16 h, a further 0.8 µl of trypsin was added and the beads were incubated for an extra 1 h at 37 °C. Meanwhile, TMT-10-plex label reagents (0.8 mg) (Thermo) were equilibrated to room temperature and diluted with 40 µl of anhydrous acetonitrile (Optima grade; 5 min with vortexing) and centrifuged. A total of 40 µl of each TMT reagent was added to the appropriate sample. The reaction was incubated for 2 h at room temperature. The samples were quenched with 8 µl of 5% hydroxylamine and incubated for 15 min. The samples were pooled in a new Protein LoBind tube and acidified with TFA (16 µl, Optima). Mass spectra were obtained using an Orbitrap Fusion Lumos mass spectrometer at Princeton Proteomics Facility and analysed using MaxQuant 53 . TMT-labelled peptides were dried down in a SpeedVac, redissolved in 300 µl of 0.1% TFA in water and fractionated into eight fractions using the Pierce High pH Reversed-Phase Peptide Fractionation Kit (no. 84868). Resulting MS/MS/MS data were searched in Maxquant against the Uniprot human protein database containing common contaminants. The proteinGroups.txt file was subsequently imported into Perseus ( ) 54 . The data were then filtered on the basis of the following criteria, ‘only identified by site’, ‘reverse’ and ‘potential contaminant’. The resulting data were log 2 -transformed and median normalization was performed. FDR-corrected P values were determined by a two-sample t -test following the Benjamini–Hochberg procedure. The data were visualized by plotting as a volcano plot. Procedure for mononucleosome IP Two 10 cm plates of HEK 293T cells transfected with H2A-E92K-HA-Cfa N -FLAG (approximately 30 million cells) and two 10 cm plates of HEK 293T cells transfected with H2A-HA-Cfa N -FLAG (approximately 30 million cells) were lysed in 1 ml of hypotonic lysis buffer (10 mM Tris, 15 mM NaCl, 1.5 mM MgCl 2 , Roche cOmplete EDTA-free protease inhibitors, 1 mM dithiothreitol, 5 mM sodium butyrate, pH 7.6) for 10 min on ice and the nuclei were pelleted at 400 g for 5 min at 4 °C. The nuclei were then resuspended in RSB + 1 mM dithiothreitol + 0.2% v/v Triton X-100 + 5 mM Na butyrate (500 µl per condition) and incubated on ice for 5 min. Nuclei were then pelleted at 400 g for 5 min at 4 °C. The nuclei were washed once more with RSB + 1 mM dithiothreitol + 5 mM Na butyrate and centrifuged at 600 g for 5 min at 4 °C. The nuclei were then resuspended in 500 μl of MNase digestion buffer (10 mM Tris, 60 mM KCl, 15 mM NaCl, 2 mM CaCl 2 , pH 7.5) and were incubated at 37 °C for 10 min. MNase (5 µl, NEB) was then added to each condition for 10 min at 37 °C (digestion time varies by enzyme batch and must be determined for each experiment. A total of 2 µl of the digest was removed every 2 min and quenched by addition of 20 mM EGTA. These aliquots were run on a 1.2% agarose gel and the digestion efficiency visualized with ethidium bromide staining). Once digestion to mononucleosomes was complete, the reaction was quenched with the addition of 20 mM EGTA on ice for 5 min. The sample was spun at 1,300 g for 5 min at 4 °C and the supernatant was collected (fraction S1). A total of 500 µl of buffer TE + 5 mM Na butyrate (10 mM Tris, 1 mM EDTA, pH 8.0) was added to the pellet and the sample was rotated end-over-end at 4 °C for 30 min. The sample was spun at 13,000 g for 5 min at 4 °C and the supernatant was collected (fraction S2). To 475 µl of S1 was added 475 µl of 2× buffer E (30 mM HEPES, 225 mM NaCl, 3 mM MgCl 2 , 0.4% Triton X-100, 20% v/v glycerol, pH 7.5) with constant vortexing dropwise for 1 min. To 475 μl of S2 was added 237 µl of 3× buffer D (60 mM HEPES, 450 mM NaCl, 4.5 mM MgCl 2 , 0.6 mM EGTA, 0.6% Triton X-100, 30% v/v glycerol, pH 7.5) with constant vortexing dropwise for 1 min. The samples were spun at 13,000 g for 5 min at 4 °C. The two fractions were combined in a 5 ml LoBind Eppendorf tube and 30 μΛ of magnetic FLAG beads (Sigma M2 anti-FLAG magnetic beads; prewashed with 1× buffer D) were added per condition. The FLAG-IP was performed overnight with end-over-end rotation at 4 °C. The beads were washed sequentially with 1× buffer D and 1× buffer D + 0.5% v/v Triton X-100 for 2 min each with rotation. A total of 40 µl of 1× SDS loading buffer was added and the beads were boiled for 10 min. Samples were run on a 15% Tris gel for western blotting with appropriate antibodies. Procedure for assaying local DNA methylation Mononucleosome isolation was performed as above from HEK 293T cells stably expressing H2A-HA-Cfa N -FLAG or H2A-E92K-HA-Cfa N -FLAG. Following this, mononucleosomes were enriched using 30 µl of magnetic FLAG beads (Sigma M2 anti-FLAG magnetic beads; prewashed with 1× buffer D) per condition. The FLAG-IP was performed overnight with end-over-end rotation at 4 °C. The beads were washed sequentially with 1× buffer D and 1× buffer D + 0.5% v/v Triton X-100 for 2 min each with rotation. The beads were then washed twice with TE and resuspended in 200 µl of elution buffer (50 mM Tris-Cl (pH 8.0), 10 mM EDTA, 1% SDS) supplanted with 1 µl of proteinase K (20 µg µl −1 ). The samples were rotated at 55 °C for 2 h. Following this, the samples were centrifuged at 10× 1,000 g for 1 min before being pelleted on a magnetic rack. The supernatants were transferred to clean LoBind tubes. The beads were then resuspended in 100 µl of elution buffer supplanted with 1 µl of proteinase K (20 µg µl −1 ) and rotated at 55 °C for a further 2 h. Again, the samples were centrifuged at 10,000 g for 1 min before being pelleted on a magnetic rack and the supernatants transferred to clean LoBind tubes. The supernatants were combined (in each replicate) and the eluted DNA was purified using a Monarch PCR & DNA Cleanup Kit according to the manufacturer’s instructions. The eluted DNA was then digested using the Nucleoside Digestion Mix (NEB: M0649S) following the manufacturer’s instructions. Digested nucleotide mixtures were then analysed by MS and the ratio of dC to methylated dC was calculated and normalized to global methylation levels. Briefly, LC–QQQ–MS quantitation of digested deoxynucleosides was performed following literature precedent 55 using a dynamic multiple reaction monitoring method on an Agilent 1260 LC Infinity II system coupled to an Agilent 6470 triple quadrupole mass spectrometer in positive ion mode. An InfinityLab Poroshell 120 SB C18 Column (Agilent, 683775-906(T), 2.7 µm particle size, 2.1 × 150 mm 2 ) was used for all analyses with a gradient composed of 0.1% formic acid in water (A) and acetonitrile (B) at 0.4 ml min −1 flow rate. The following mass spectrometer operating parameters were used: gas temperature 325 °C, gas flow rate 12 l min −1 , nebulizer pressure 20 p.s.i., capillary voltage 2,500 V and fragmentor voltage 70 V; collision energy was set to 14 for dC and 7 for m5dC. The MS1 (parent ion) to MS2 (deglycosylated base ion) transition for dC was set to m / z 228.1→ 112 and m / z 242.1→126 for m5dC. Commercially available deoxyribonucleosides were used to generate standard curves and the concentration of m5dC was normalized to dC concentration. Data quantification was performed with Agilent MassHunter Workstation Data Acquisition v.10.0.6. Subcellular fractionation One 10 cm plate of HEK 293T cells was transfected with the denoted plasmids and these were collected as previously described. Cells were lysed in hypotonic lysis buffer (10 mM Tris, 15 mM NaCl, 1.5 mM MgCl 2 , PI, pH 7.6) on ice for 10 min, followed by centrifugation at 400 g for 5 min at 4 °C. The supernatant was removed (cytosolic fraction) and the pellet was resuspended in hypotonic lysis buffer + 1% v/v Triton X-100. Nuclei were lysed with ten strokes of a tight pastel homogenizer followed by centrifugation at 10,000 g for 10 min at 4 °C. The supernatant was removed (nucleoplasmic fraction) and RIPA buffer was added to the pellet. The sample was sheared by probe sonication (2 × 10 s total, 25% amplitude, 1 s on, 1 s off) to yield the chromatin fraction. SDS loading buffer was added and the samples were analysed by western blotting using the indicated antibodies. Recombinant nucleosome immunoprecipitation Nuclear lysate was prepared as before. Following this, biotinylated nucleosomes ( ) were immobilized on streptavidin T1 magnetic dynabeads (MyOne, Thermo Fisher) in BB150 (20 mM HEPES pH 7.5, 150 mM NaCl, 10% glycerol, 1 mM dithiothreitol, 0.1% NP-40) by rotation for 2 h at 4 °C. A total of 60 μg of nucleosomes were incubated with 150 μl of resuspended resin in 1 ml of BB150. Streptavidin-bound nucleosomes were rinsed twice and washed for 30 min by rotating with 1 ml of BB150 at 4 °C and divided equally into three tubes. Nuclear lysate (700 μg) was mixed with 300 µl of BB150 and centrifuged at 10,000 g for 10 min at 4 °C before adding to nucleosome-bound streptavidin beads. The mixture was rotated for 2 h at 4 °C after which the beads were rinsed twice with 0.5 ml of BB150 and then rotated for 1 h at 4 °C with 0.5 ml of BB150. Washed dynabeads were moved to new tubes with about 500 μl of BB150, the beads were pelleted on a magnetic rack and the buffer was aspirated, washed beads were then centrifuged at 800 g for 2 min at 4 °C and aspirated again to remove residual buffer. Beads were then resuspended in 15 μl of 2× gel loading buffer (Bio-Rad) and boiled for 5 min before being pelleted on a magnetic rack. Eluted proteins were then analysed by western blotting using the indicated antibodies. ATAC-seq analysis ATAC-seq analysis was performed by Genewiz. Briefly, cells stably expressing H2A-HA-Cfa N -FLAG or H2A(E92K)-HA-Cfa N -FLAG were collected and cryopreserved before delivery to Genewiz for analysis. Both conditions were analysed over four biological replicates. RNA-seq analysis RNA-seq analysis was performed by Genewiz. Briefly, cells stably expressing H2A-HA-Cfa N -FLAG were treated with either JQ-1 (1 µM for 3 h) or pinometostat (2.5 µM for 24 h) collected and cryopreserved before delivery to Genewiz for analysis. Four biological replicates were performed for each condition. Generation of stable cell lines Constructs were cloned into a pCDH-CMV-MCS-EF1-Puro expression vector using standard restriction cloning procedures. Lentiviral cell lines were generated using established methods. Statistics and reproducibility Statistical analyses were performed using Prism (GraphPad). P values were determined by paired t -tests as appropriate and as listed in the figure legends. The statistical significances of differences (* P < 0.05, ** P < 0.01, *** P < 0.001, **** P < 0.0001) are specified throughout the figures and legends. We analysed GO using Metascape 56 . Heatmap was generated using Heatmapper ( Heatmapper.ca ) 57 . Expression and splicing blots were repeated successfully three times. Proteomics experiments in Figs. 2h , 3a , 4b , 4c and 4f have been performed once. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Data availability All relevant data are included in the manuscript and Supplementary Information . Mass spectrometry data files have been uploaded to the MassIVE proteomics database PXD038956 ( ftp://massive.ucsd.edu/MSV000090929/ ). Sequencing data are available on the GEO database ( GSE221674 ).

Studying the microbiology of any entity, be it a molecule or a dolphin, ideally means putting a spotlight as close to the source material as possible. That can be especially challenging when you're investigating the Rube Goldberg environment of a cell's nucleus. But in research published this week in Nature, Princeton chemists from the Muir Lab and the MacMillan Lab used two marquee technologies to shine a light right where they wanted it. In the process, they discovered critical, unexpected changes in interactions surrounding a DNA-protein complex called chromatin—essentially, an architecture that allows for the compaction of DNA—in the presence of genetic mutations widely associated with cancer. Interactions between biomolecules control every biological function. Mapping their activity leads to a deeper understanding of cell fate. So researchers paired the strengths of µMap, the MacMillan Lab's proximity-labeling system introduced three years ago, with in nucleo protein trans-splicing, a protein engineering technology introduced in 2016 and optimized since by the Muir Lab. "The whole point of µMap is to understand biology, broadly defined, in ways that you couldn't beforehand, because µMap gives you such incredibly precise information. This research is an example of doing precisely that," said David MacMillan, the James S. McDonnell Distinguished University Professor of Chemistry and a recipient of the 2021 Nobel Prize in Chemistry. "With this research, we saw that the biology is happening because of these mutations," he said. "It was impossible to see that before, so we were driving in the dark. It's another piece in this whole wall, a brick of what µMap's going to be able to do. It's still just the beginning, but this is a true collaboration." The combination of these technologies allowed researchers to tether an iridium photocatalyst to a protein of interest to study these minute interactions and how they change in the presence of mutations—all without impacting the complex microenvironment within the cell nucleus. The photocatalyst highlighted a radius of focus just one nucleosome wide, allowing researchers to peer into this microenvironment with unprecedented specificity. "So many things in biology and disease come down to how chromatin moves and changes," said Ciaran Seath, a former postdoc in the MacMillan Lab and co-lead author on the paper with Antony Burton, a former postdoc in the Muir Lab. "Viruses, aging, cancers, all the things we looked at, change how chromatin can move and react. We figured, if you could watch that, you could learn about all of these different problems in a modular way. "Sometimes you can't see or predict all the other things that are likely to happen in the machine of the nucleus. Now that we have paired these tools, we can measure these unforeseen consequences. It's like touching a point on a spider's web," he said. "You can see the whole thing move." Burton added, "What the synergy of these technologies gets you, in a minimally perturbative way, is the ability to install a small catalyst onto a protein of interest and then map what's nearby. We were able to illuminate protein interactions and complex, downstream effects at a level of detail that is very hard to do with other methods. "Most importantly, we can chart how these change as a function of mutation or drug treatment, presenting opportunities for academic and industrial application." The research was authored by Seath, now an assistant professor of chemistry at the Herbert Wertheim UF Scripps Institute for Biomedical Innovation & Technology in Jupiter, Florida; Burton, now a senior scientist in chemical biology at AstraZeneca in Boston, Massachusetts; and senior scientists Thomas Muir, the Van Zandt Williams Jr. Class of 1965 Professor, and MacMillan. Advancing the field of epigenetics The research has important implications for epigenetics, the branch of biology that explores changes in gene expression. At the heart of epigenetics are histone proteins, which package DNA and restrict access to the genome, thereby playing a key role in regulating transcription. Recently, mutations to these histone proteins have been discovered and linked to a wide range of cancers. "One of the things we looked at was adding a mutation that occurs in a histone that's linked to cancer," said Muir. "What we wanted to know is, what happens when that mutation is present; what can no longer be recruited, what's no longer in the neighborhood, what new things are brought in that shouldn't normally be there? "We were able to use these technologies to find all sort of things that get changed when we put this mutation on the chromatin, things related to gene regulation," he added. "We were able to infer mechanistic insights that relate to how genes get misregulated with a mutation." Beginning their research three years ago, the team hypothesized that the mutation associated with cancer precipitates some sort of loss of function. Biomolecules find chromatin and bind with it to leave a transcriptional mark. Researchers believed that the mutation blocked some of that action, thus leading to cell dysfunction. Their hypothesis was borne out with this research, which generated molecular detail on how a tiny change in the genome can lead to major impacts.

10.1038/s41586-023-05914-y

Medicine

New research could help in the fight against infection, cancer and allergies

Alistair Bailey et al. Selector function of MHC I molecules is determined by protein plasticity, Scientific Reports (2015). DOI: 10.1038/srep14928 Journal information: Scientific Reports

http://dx.doi.org/10.1038/srep14928

https://medicalxpress.com/news/2015-11-infection-cancer-allergies.html

Abstract The selection of peptides for presentation at the surface of most nucleated cells by major histocompatibility complex class I molecules (MHC I) is crucial to the immune response in vertebrates. However, the mechanisms of the rapid selection of high affinity peptides by MHC I from amongst thousands of mostly low affinity peptides are not well understood. We developed computational systems models encoding distinct mechanistic hypotheses for two molecules, HLA-B*44:02 (B*4402) and HLA-B*44:05 (B*4405), which differ by a single residue yet lie at opposite ends of the spectrum in their intrinsic ability to select high affinity peptides. We used in vivo biochemical data to infer that a conformational intermediate of MHC I is significant for peptide selection. We used molecular dynamics simulations to show that peptide selector function correlates with protein plasticity and confirmed this experimentally by altering the plasticity of MHC I with a single point mutation, which altered in vivo selector function in a predictable way. Finally, we investigated the mechanisms by which the co-factor tapasin influences MHC I plasticity. We propose that tapasin modulates MHC I plasticity by dynamically coupling the peptide binding region and α 3 domain of MHC I allosterically, resulting in enhanced peptide selector function. Introduction Peptides bound to Major histocompatibility complex class I (MHC I) molecules are displayed at the surface of most nucleated cells in jawed vertebrates for surveillance by cytotoxic T-lymphocytes (CTL) 1 , 2 , 3 . In transformed or diseased cells, peptides derived from viral or aberrantly expressed proteins are presented alongside peptides derived from native proteins. This is possible because MHC I has degenerate specificity: each molecule can bind a range of peptides of different lengths and sequences. In selecting peptides and presenting them at the cell surface, MHC I provides CTLs with a sample of the internal cell proteome. This property makes MHC I an attractive target for the development of immunotherapies that exploit the CTL response. Examples include therapies that modulate the overall cell surface presentation of peptides by MHC I, or that target MHC I at the cell surface with specific peptide vaccines to make cells more, or less, visible to CTLs. The selection of peptides by MHC I occurs in the endoplasmic reticulum (ER) and is modulated by a multi-protein complex into which the MHC I:β 2 m heterodimer is recruited. This peptide loading complex comprises of the peptide Transporter associated with Antigen Presentation (TAP), which transports peptides into the ER, chaperone proteins calreticulin and ERp57 and an MHC I-specific cofactor tapasin (reviewed in 4 ). Incorporation of MHC I molecules into the peptide loading complex locates them in close proximity to the peptide supply. Here, MHC I interacts with cofactor molecules to preferentially select peptides of high affinity from the large intracellular pool of many potential peptides of largely lower affinity 5 , 6 . Selection of high affinity peptides confers stability and immunogenicity to MHC I 7 , 8 and is one of the most important factors in establishing the specificity and intensity of a CTL response 9 . We refer to the peptide selector function of MHC I as its ability to preferentially select higher affinity peptides over lower affinity ones. This in turn profoundly influences the downstream MHC I function of presenting peptides to CTLs. MHC I molecules are highly polymorphic, but have a common tertiary structure ( Fig. 1 ) consisting of: the polymorphic heavy chain, monomorphic β 2 -microglobulin (β 2 m) and a peptide of generally 8–10 amino acids in length, non-covalently bound into a peptide binding groove. Intrinsic differences in the peptide selector function of different allelic variants of MHC I molecules become most apparent when the co-factor molecule tapasin is non-functional or absent. This is because tapasin masks these intrinsic differences by preferentially enhancing the selector function of MHC I molecules that are inefficient at selecting high affinity peptides 10 , 11 . When tapasin is absent, differences in the primary sequence of MHC I are sufficient to alter peptide selector function 12 . For example, two human alleles HLA-B*35:01 and HLA-B*52:01 differ by 12 residues in their primary sequence. In the absence of tapasin, HLA-B*35:01 molecules are expressed at a high level at the cell surface, whilst HLA-B*52:01 molecules are only observed at low levels 13 . Remarkably, even a single amino acid difference can alter the intrinsic peptide selector function of MHC I: HLA-B*44:02 (B*4402) and HLA-B*44:05 (B*4405) differ by a single residue at position 116 ( Fig. 1A ), yet they lie at opposite ends of the spectrum in their intrinsic ability to select high affinity peptides in the absence tapasin 10 . In tapasin-deficient cells, B*4402 is poor at sampling the peptidome and as a consequence is degraded in the endoplasmic reticulum (ER), while B*4405 is able to effectively select high affinity peptides and present them at the cell surface. Figure 1 Structure of the MHC I molecule (A) Ribbon representation of the MHC I molecule HLA-B*44:05 and its three components: a polymorphic heavy chain (yellow), non-covalently bound invariant β 2 m (yellow) and peptide (red). The polymorphic residue 116 between B*4402 and B*4405 is shown in blue (B) Comparison of B*4402 (PDB: 1M6O, green) and B*4405 (PDB: 1SYV, blue) structures. RMSD between them of 0.3 Å. (C) Combined ribbon and surface representation of the MHC I molecule peptide binding groove. Full size image These observations of allelic differences in intrinsic MHC I peptide selector function imply that tapasin normalizes the peptide selector function of MHC I alleles. However, a mechanistic explanation for how tapasin achieves this is lacking. Furthermore, despite there being many crystal structures of MHC I, the structural basis for intrinsic differences in selector function between MHC I alleles remains unknown. Importantly, what is not revealed by X-ray crystallography is the formation of a MHC I peptide complex and the significance or otherwise of the processes involved. For example, the structures of B*4402 and B*4405 reveal highly similar peptide bound conformations with a RMSD of 0.3 Å between their secondary structures 15 , 16 ( Fig. 1B ) and are therefore unable to provide insight into the known functional differences in peptide selection. However, the simple observation that the peptide is buried deep in the peptide binding groove of MHC I ( Fig. 1C ) suggests that MHC I is likely to explore intermediate conformational states during the formation of a stable complex, simply to allow peptides to enter and leave the groove. In other words, this observation alone suggests that MHC I molecules must be plastic in order to sample and bind a range of peptides. Different methods have provided indirect evidence that peptide binding to MHC I is associated with a conformational intermediate 17 , 18 , 19 , 20 , 21 , including molecular dynamics simulations 22 , 23 , 24 , 25 . Although these studies have helped to resolve the apparent paradox of how MHC I molecules of such degenerate specificity can bind peptides with such high affinity, they do not provide a framework for understanding the mechanism of peptide selection inside cells. Our aim therefore was to further elucidate the relationship between MHC I structure and MHC I function in the context of the complexity of the cellular environment. We sought to address this challenge for the MHC I presentation system, by adopting an interdisciplinary approach to infer mechanism from in vivo biochemical experiments. We use computational systems models that test several mechanistic hypotheses of MHC I peptide selection and quantify the uncertainty in each model. This provides us with a rigorous mathematical basis for comparing competing mechanistic hypotheses and for identifying which hypothesis best explains the experimental data. As becomes apparent below, in the case of antigen processing this centers on the plasticity of MHC I. We then relate observations at the cellular level to protein structure and function using molecular dynamics simulations of MHC I. As has been noted, combining three disciplines presents challenges to the reader in interpreting the results and assessing their validity 26 . However, this approach yields information about mechanism that only becomes apparent when considering all three disciplines in combination. Specifically, we test the hypothesis that conformational intermediates of MHC I molecules are directly relevant for the peptide selector function of MHC I. This work represent a first step towards establishing a general framework for inferring molecular mechanism from complex biological data in the context of MHC I peptide selection, with its subsequent importance for modulating the immune response. Results An enhanced intrinsic selector function for B*4405 over B*4402 To investigate the mechanisms of peptide selection by MHC I, we first quantified peptide selector function in vivo for two functionally distinct MHC I molecules, B*4402 and B*4405. We then developed computational systems models of MHC I peptide selection as sets of biochemical reactions to encode distinct mechanistic hypotheses and used Bayesian model selection to determine which hypotheses were most likely, given the experimental data. To quantify peptide selector function in vivo , we measured the fraction of a pulse-labelled cohort of MHC I molecules that were stable at 50 °C, 37 °C and 4 °C as they progressed through the secretory pathway over 120 minutes ( Fig. 2A,B , quantified in F,G). Since the thermal stability of MHC I has been shown to act as a surrogate for the affinity of peptide bound to MHC I 10 , this allows the quantification of the selection of high affinity peptides by MHC I. Figure 2A shows that, as demonstrated previously 10 , in tapasin-deficient cells B*4405 is able to select high affinity peptides. This is indicated by the relative proportions of thermostable MHC I complexes harvested from the cells at each time point. Figure 2 Computational systems models of the mechanisms of MHC I peptide selection, fit to in vivo peptide selection data for B*4402 and B*4405 in the absence of tapasin. ( A ) Time-dependent peptide selection measured with pulse-chase and thermostability in tapasin-deficient cells, as described in materials and methods. Tapasin deficient cultured 220 cells (721.220) were metabolically radiolabelled with 35 S-met for 5 min, then chased for the indicated times before being lysed and heated to the temperatures shown (data quantified in F). ( B ) Cell surface transit measured as percentage endoglycosidase-H (EndoH) resistance. MHC I was immunoprecipitated with W6/32 and treated at the different time points with (+) or without (−) EndoH, which distinguishes between sensitive, pre-cis-Golgi (ES) and resistant (ER) post-cis-Golgi MHC I (data quantified in G). Immunoprecipitations were performed in such a way as to record only MHC I bound to high affinity peptides ( 10 ). ( C ) A general computational systems model of the mechanisms of peptide selection in the absence of tapasin. Shapes represent molecular species and labelled boxes represent reactions and their rate parameters. ( D ) Different binding mechanisms are illustrated for o ne-conformation and two-conformation models, where peptide-dependent reactions are indicated by thick red symbols. ( E ) Comparison of model performance for different parameters taking allele-specific values ( allele parameters ). The most likely model (lowest BIC score) was the two-conformation model with MHC I opening as the peptide-dependent step and MHC I closing rate c as the allele-dependent parameter. ( F , G ) Comparison of the most likely model (solid lines) against experimental measurements (circles), with 95% confidence intervals (shaded regions). ( F ) Quantification of pulse chase experiments shown in panel A (circles), together with model simulations (lines). Red indicates the proportion of MHC I molecules that remained after heating to 50 °C (high affinity peptide-MHC complexes), green indicates heating to 37 °C (medium and high affinity complexes) and blue indicates heating to 4 °C (all complexes). ( G ) Quantification of cell surface transit experiments shown in panel B (circles), together with model simulations (lines). ( H ) Marginal posterior density of the allele-specific closing rate c , reflecting the probability of the parameter values, conditional on the measured data and underlying model. Full size image Over time, B*4405 ( Fig. 2A , quantified in F) has a larger proportion of complexes stable at higher temperatures. However, B*4402, with a lower proportion of stable complexes was therefore less able to select high affinity peptides. Furthermore, in the absence of tapasin, B*4402 is mostly sensitive to the enzyme Endoglycosidase H which indicates that B*4402 was less able to acquire high affinity peptides and progress to the cell surface ( Fig. 2B , quantified in G). A conformational intermediate of MHC I is significant for peptide selector function To investigate the mechanisms of MHC I peptide selection, we developed computational systems models of the antigen processing pathway as sets of biochemical reactions that encoded distinct mechanistic hypotheses. We defined our models as extensions to a previous computational systems model of MHC I assembly, peptide selection and cell surface presentation 27 . This modelling approach enabled us to observe the effects of individually or simultaneously varying all parameters over physiologically relevant ranges. It also allowed testing of different mechanistic hypotheses against biological data, thereby providing a springboard for further experimental investigation. Our previous analysis identified the rate of peptide binding to MHC I as the parameter differentiating HLA-B alleles B*4402, B*4405 and B*2705 27 . To determine how the rate of peptide binding might be controlled by MHC I itself, in addition to factors such as the concentration of peptide in the ER or the volume of the ER, we incorporated inherent plasticity of MHC I as a minimal extension to the original model 27 , by explicitly describing two conformational states of MHC I: a peptide-receptive open conformation to which peptides bind at rate b and a non-receptive closed conformation ( Fig. 2C,D ). We represented MHC I and peptide as a system of molecular species that interact with defined reaction rates, including rates of MHC I and peptide generation and degradation and transport from the ER to the cell surface. The reaction rates were considered as unknown parameters and were fitted to the experimental data (as described in the Supplementary Information ). In these extended models, the release of peptides from MHC I was described in two steps: MHC I opening, followed by peptide-unbinding. This gave rise to two variant models in which the interactions determined by the biochemical affinity of peptide for MHC I, the peptide-dependent step , could arise from one of two processes. One being that the rate of MHC I opening ( o i ) is peptide-dependent, the other being that the rate of peptide-unbinding from the open MHC I conformation ( u i ) is peptide-dependent. The subscript i indicates the opening or unbinding rate for a given peptide i . The descriptor “open” does not have a measured structural correlate, but we assumed that the structural correlate of “closed” was the common crystallographic structure of peptide bound MHC I complexes. This assumption includes MHC I containing very weakly bound peptides such as those that occupy only the F-pocket 28 . To determine which mechanistic hypotheses were most likely, we compared model simulations with experimental data using Bayesian parameter inference and model selection. To simulate the range of thermostability that was measured, representative peptides with high, medium or low affinities for MHC I were used in the model. In Fig. 2F,G , the experimental data from Fig. 2A,B is indicated by circles and the model simulations reproducing the data are indicated by lines. To assess the plausibility of each model, we determined the optimal parameter values using probabilistic inference techniques 27 . We examined a variety of hypotheses for allele-specific mechanisms, by enabling specific parameters (such as the rate of closing) to take on different values for each allele, while keeping the remaining parameters the same for all alleles 27 . For each hypothesis, we refer to the variable parameter as the allele parameter . Conclusions were based on the Bayesian Information Criterion (BIC) 29 , which minimizes the deviation between model simulation and experimental observation, while accounting for the effects of having different numbers of free parameters in the different models that were tested. Figure 2E shows the BIC statistic (shorter bars indicate statistical improvement) for a range of allele parameters for each of the three models, including no allele parameter. Comparing BIC statistics for the one conformational model ( Fig. 2E , top) indicated that an allele-specific rate of peptide binding to MHC I ( b ) explains the differences in peptide selection between the three alleles better than an allele-specific rate of peptide free MHC I degradation ( d M ). Importantly, we found that both of the two-conformation models outperformed the original one-conformation model in their ability to reproduce the experimental data, with improved performance sufficient for their additional parameters to be considered justifiable by the BIC statistic ( Fig. 2E ). In other words, this analysis clearly indicates that when testing model hypotheses of peptide selection, a two-conformation MHC I model provides a better explanation of the experimental observations than a one-conformation model. In particular, the one-conformation model was unable to reproduce a key experimental observation of slow conversion of MHC complexes to a mature measurable state, which had a maximal signal at 30 minutes post radiolabelling (see Figure S7 ). Moreover, the two-conformation model where the opening rate ( o i ) of the closed MHC I peptide complex (M c P i ) was peptide-dependent and the closing rate ( c ) of open peptide bound MHC I (M o P i ) was allele-dependent, provided the best description of the data ( Fig. 2E , bottom). Faster closing of MHC I corresponds with enhanced intrinsic selector function Interestingly, for the most likely model determined in Fig. 2E , B*4405 showed a relatively fast closing rate ( Fig. 2H , bottom) corresponding with faster progression to a stable conformation, consistent with its enhanced intrinsic selector function observed experimentally ( Fig. 2F , bottom). B*4402 showed a slower closing rate ( Fig. 2H , top), corresponding with a slower progression to a stable conformation, consistent with a diminished intrinsic selector function ( Fig. 2F,G , top). Our results therefore suggest that the intrinsic differences in selector function between these MHC I alleles arise from intrinsic differences in their ability to move from open to closed conformations. The closing rate c corresponds to the rate at which MHC I progresses to a stable conformation, from which it is able to exploit the peptide-dependent opening step. Specifically, frequent progression to a stable conformation (i.e. fast closing) enables the preferential exchange of low affinity peptides (characterized by fast o i ) for high affinity peptides (characterized by slow o i ), thus enhancing selector function. Put simply, once an MHC I molecule binds a peptide and closes, it is the peptide-dependent opening rate of the MHC I molecule that defines the affinity of the peptide-MHC interaction. Therefore, if one MHC I molecule closes faster than another, the faster closing MHC I molecule reaches the peptide-dependent part of the selection process more quickly, resulting in enhanced peptide selector function due to more rapid sampling of peptides. Peptide selector function of MHC I correlates with protein plasticity Following the identification of functionally relevant conformations of MHC I, we investigated the extent to which these conformations could be linked to protein plasticity. Specifically, we performed molecular dynamics simulations of MHC I alleles B*4402 16 and B*4405 15 and quantified their plasticity by computing the range and frequency of molecular conformations, as described by the relative motions of their backbone atoms, under conditions of thermodynamic equilibrium. We then compared the plasticity of these alleles to their corresponding peptide selector function. We performed molecular dynamics simulations using the GROMACS package 30 in the peptide bound and peptide free states (see Supplementary Information Table S1 for a summary of the simulations). The simulations were designed to capture the dynamic behavior of each MHC I molecule in a manner independent of timescale. Specifically, we used 10 ns block averaged root mean squared fluctuations to assess that there was adequate configurational sampling during the 420 ns simulations, to be consistent with equilibrium behavior ( Figure S1 ). To quantify the range and frequency of conformations adopted by MHC I molecules, we chose to focus on two representative distance measures, the F-pocket distance and the inter-domain distance . The F-pocket distance corresponds to the distance between the center of mass of helix residues 135–156 and helix residues 69–85 ( Fig. 3A , middle). This was chosen due to the hinge points created between the break in the α 2 helix near residue 156 and the end of the α 2 helix near residue 135, about which conformational change may occur 23 . Thus, the F-pocket distance was indicative of the range of conformations that could be adopted by the peptide binding groove. The inter-domain distance corresponds to the distance between peptide binding groove residues 96–100 and α 3 domain residues 220–227 ( Fig. 3A , right). This was chosen to represent the range of motion that could be exhibited between the heavy chain domains. From the molecular dynamics simulations, we quantified the frequency with which the MHC I molecules adopted conformations with a given F-pocket and inter-domain distance, represented as joint probability distributions ( Fig. 3B,C,E,F ). These distributions indicate the conformations that MHC I could adopt, together with the correlations between peptide binding groove and heavy chain conformations. Figure 3 Quantification of protein plasticity for MHC I alleles B*4402, B*4405 and B*4405 W147A from molecular dynamics simulations. ( A ) Left: Surface representation of peptide bound MHC I. Middle: Ribbon representations of peptide free MHC I. The polymorphism between B*4402 and B*4405 at position 116 in the peptide binding groove (brown) and mutation B*4405 W147A (blue). F-pocket distances were measured between the center of mass of helix residues 135–156 and 69–85 (red). Right: Inter-domain distances were measured between peptide binding groove residues 96–100 (red) and α 3 residues 220–227 (red). ( B–G ) Contour plots of the joint probability densities for the conformations of MHC I populated in each simulated condition, as defined by distances in ( A ). Black crosses indicate the initial structure conformation. Distributions for each individual distance are plotted on the outside of the adjacent axis. ( H–J ) The motion most correlated with the distance fluctuations across the F-pocket as defined in ( A ). Cones indicate the direction and amplitude of motion. The range of inter-domain twisting for each molecule is indicated by arrows (as depicted in Figure S5 ). See also Figures S1–S5 . Full size image To characterize how motion of the peptide binding groove relates to overall motion of MHC I, we performed Functional Mode Analysis (FMA) (See Supplementary Information and Figure S3 for further details) 31 . In correlating the F-pocket distance to the collective motions of the backbone atoms of the molecule, FMA determined a single collective motion most correlated with the distance fluctuations across the F-pocket, represented as porcupine plots ( Fig. 3H,I ). The direction of movement in the plots is represented by the orientation of the cones, while the amplitude of movement is represented by both the size and color of the cones. In this way, FMA provides a description of the conformational change undergone by the whole MHC I molecule corresponding to a conformational change of the peptide binding groove. We then compared the plasticity of MHC I alleles B*4402 and B*4405 ( Fig. 3 ) with their corresponding peptide selector function ( Fig. 2 ). In the peptide bound state ( Fig. 3B,C ), both molecules displayed similar plasticity and populated a single dominant conformation, with a sub-population of conformations arising from inter-domain motion for both alleles. Since these two alleles differ by only a single amino acid and exhibit almost identical crystal structures in their peptide bound state, we sought to investigate whether the observed differences in peptide selector function were due to differences in protein plasticity in the peptide free state. When we repeated the simulations following removal of the peptide, this revealed significant differences in plasticity ( Fig. 3E,F ). In the absence of peptide, B*4402 populated a single F-pocket conformation ( Fig. 3F ). In contrast, B*4405 populated several F-pocket conformations ( Fig. 3E ) and explored greater conformational space overall. In Fig. 3H,I , the direction and amplitude of the backbone atoms indicated a twisting motion that correlated with F-pocket opening, articulated around the domain linker region. This twisting motion was significantly larger for B*4405 than for B*4402 (see Figure S5 for more details of the twisting analysis). Taken together, the molecular conformations and movements of atoms indicated that, in the absence of peptide, the F-pocket of B*4402 was less plastic than that of B*4405, exploring a more limited conformational space and exhibiting a reduced twisting motion ( Fig. 3I ; Figure S5 ). Furthermore, these differences in plasticity in the absence of peptide correlated with differences in peptide selector function, with the more plastic B*4405 exhibiting significantly greater selector function than the less plastic B*4402 ( Fig. 2F,G ). Interestingly, this plasticity also correlated with the rate of conformational change from an open to a closed conformation in the computational systems model, given by the closing rate c , with B*4405 exhibiting a significantly greater closing rate than B*4402. Therefore, by comparing quantifications of B*4405 and B*4402 plasticity ( Fig. 3 ) with in vivo measurements of peptide selection ( Fig. 2F,G ), we identified mechanistic hypotheses linking protein plasticity to peptide selector function, as targets for further investigation. Plasticity of MHC I can predict peptide selector function in vivo To test our hypotheses experimentally, we performed molecular dynamics simulations of a site-directed mutant of B*4405 in which the highly conserved Tryptophan 147 was changed to Alanine (B*4405 W147A ).The identification of this novel mutant was somewhat fortuitous and stemmed from a previous study 32 describing the effect of 33 random single site mutations on the recognition of HLA-A*02:01 by alloreactive T cells. During this study, a significant effect of a W147 mutation on MHC I intracellular trafficking was observed, suggesting this site as a potential target for investigation. The B*4405 W147A novel mutant revealed an interesting phenotype that enabled an investigation into the link between sequence, plasticity and selector function. Significantly, in peptide free molecular dynamics simulations, B*4405 W147A explored fewer peptide binding groove conformations than wild-type B*4405 ( Fig. 3G ) and also showed a corresponding reduction in the range of the inter-domain twisting, similar to peptide-free B*4402 dynamics ( Fig. 3J ; Figure S5 ). Furthermore, B*4405 W147A showed mobility of the α 3 domain more like B*4402 than B*4405 ( Fig. 3J ), including a highly dynamic 220–227 α 3 loop. Together, these simulations predict that the intrinsic peptide selector function of B*4405 W147A should be qualitatively similar to that of B*4402. Having tested that B*4405 W147A molecules were able to select and bind to peptides normally ( Fig. 4A ), we tested our prediction by performing pulse chase analysis as in Fig. 2 for the B*4405 W147A mutant ( Fig. 4B,C ). We found that the W147A mutation of B*4405 profoundly affected the ability of MHC I to select stable peptides and present them at the cell surface in the absence of tapasin, as predicted. To mechanistically model the selector function of this mutant, we fit the allele parameter c (closing rate), using the optimal two-conformation model of Fig. 2 . The allele parameter for B*4405 W147A fell between the parameter values for B*4405 and B*4402 ( Fig. 4D ), resulting in a peptide selector function close to that of B*4402 (compare Fig. 4E,F with 2F,G), as predicted. Figure 4 Peptide selection of B*4405 W147A measured in vivo and compared with simulations of computational systems model. ( A ) Mutant B*4405 W147A has similar peptide binding ability to that of B*4402 and B*4405. This is demonstrated by performing a BFA decay assay with 220. tapasin cell lines expressing each allele. Stability of peptide loaded MHC I over time is measured with the conformation specific antibody W6/32. ( B,C ) Pulse-chase thermostability and EndoH assays in the absence of tapasin were carried out for B*4405 W147A as in Fig. 2A,B and as described in materials and methods. ( D ) Combined likelihood score against data for all three alleles in absence of tapasin, for different values of the allele parameter c. An optimum for B*4405 W147A is present between the mean posterior values for the other two alleles, as labelled. The areas indicate the 95% confidence intervals for those two parameters. (Middle, Right) simulation values for the maximum likelihood value of c. ( E , F ) Simulation (lines) of the two-conformation (opening) model with allele parameter c set to the value in panel ( D ) that best fits the experimental data (i.e. that optimizes the likelihood function). Plotted together with experimental measurements of B*4405 W147A (circles). ( E ) Experiments relate time-dependent peptide selection measured with pulse-chase and thermostability in tapasin-deficient cells (quantification of panel B). Red symbols/lines indicate heating to 50 °C (corresponding to high affinity peptide-MHC complexes), green indicates 37 °C (medium and high affinity complexes) and blue indicates 4 °C (all complexes). ( F ) Cell surface transit measured as percentage EndoH resistance (quantification of panel ( C )). Full size image Restraint of the tapasin binding site can modulate MHC I protein plasticity To further investigate our hypothesis that protein plasticity determines selector function, we sought to identify potential mechanisms that could be used by chaperones to regulate protein plasticity. Specifically, we focused on the MHC I allele B*4402 and the chaperone molecule tapasin, which has been shown to significantly alter peptide selector function. Previous work identified a C-terminus domain residue of tapasin, R333, which interacts with a residue of the 220–227 α 3 loop, E222, in docking simulations between HLA-B*08:01 and tapasin 4 . Within this region, mutation of position 222 has been shown to abrogate the interaction between MHC I and tapasin, leading to loss of function 33 . Furthermore, polymorphic differences in this region of chicken MHC I have been shown to affect peptide selector function when measured in vitro and correlate with changes in protein dynamics 34 . We therefore chose to focus on the α 3 loop comprising residues 220–227 as a means for tapasin to induce changes in MHC I selector function. Analysis of molecular dynamic simulations ( Fig. 3I ; Figure S5 ) revealed a dominant inter-domain twisting motion between the peptide binding domain and the α 3 domain. Furthermore, in peptide-free simulations of B*4402, only the 220–227 loop in the α 3 domain showed significant motions correlated with F-pocket opening ( Fig. 3I ). For B*4405, highly correlated motions were observed between the peptide binding domain and the α 3 domain ( Fig. 3H ), suggesting a communication between distant regions of MHC I best described as dynamic coupling (as illustrated by the covariance webs in Figure S2 ) 35 , 36 . To further investigate this coupling, we simulated peptide free B*4402 with weak restraints on the C α atoms of residues 220–227 in the α 3 domain, which also mimics the interaction of tapasin in this region. Figure 5A shows that, under these conditions, the plasticity of the F-pocket of B*4402 was shifted to more closely resemble that of peptide free B*4405, consistent with the existence of dynamic coupling of the α 3 and peptide binding domains. This effect was a specific property of the 220–227 region, as similar restraints applied to other α 3 regions (residues 188–194, 250–257) had no significant effect on F-pocket plasticity ( Fig. 5B,C ). Figure 5 Quantification of B*4402 plasticity with restrained residues at the tapasin binding site and measurement of tapasin binding for B*4402, B*4405 and B*4405 W147A . ( A ) Contour plots of the joint probability densities for the conformations populated by peptide free B*4402 with restrained α 3 domain residues 220–227, the location indicated in panel ( C ). Black crosses indicate the initial structure conformation. Distributions for each individual distance are plotted on the outside of the adjacent axis. Restraint of these residues increases B*4402 plasticity by modulating the peptide binding groove conformation. (B ) Contour plots of the joint probability densities for the control simulations, the locations are indicated in panel C. Black crosses indicate the initial structure conformation. Distributions for each individual distance are plotted on the outside of the adjacent axis. Restraint of control residues has little effect on B*4402 plasticity. ( C ) Sites of the restraints on MHC I corresponding with simulations in panels A and B. ( D ) B*4405 W147A exhibits sustained binding to tapasin, like B*4402, whereas B*4405 does not. Cells were lysed in digitonin to preserve the peptide loading complex, which was then immunoprecipitated with anti-tapasin antibody. Associated transporter associated with antigen processing (TAP) and MHC I (HC) were visualized by Western blot using specific antibodies. See also Figures S1-S5 . Full size image To determine whether differences in protein plasticity induced by restraints of the α 3 domain correlated with a physical interaction with tapasin, we measured the binding of tapasin to MHC I in vivo for the alleles B*4402, B*4405 and B*4405 W147A . We performed pull-down experiments to observe co-immunoprecipitation of MHC I with anti-tapasin antibodies from lysates of 220-tapasin-B*4405 W147A compared to B*4402 and B*4405. As shown previously, B*4402 had a sustained interaction with tapasin in vivo , but B*4405 did not 15 . Strikingly, The W147A mutation conferred sustained tapasin binding to the B*4405 molecule, to a level observed for B*4402 ( Fig. 5D ). Therefore, we reasoned that B*4405 was able to select high affinity peptides with significantly reduced tapasin interaction because the enhanced plasticity of B*4405 results in enhanced selector function. Whereas B*4402 required tapasin interaction to select high affinity peptides because the reduced plasticity of B*4402 resulted in reduced selector function. By using a restraint to mimic tapasin interaction, we observed B*4402 becoming more plastic. This is consistent with the possibility that tapasin binding to MHC I may be regulated by mobility of the 220–227 loop. Furthermore, the tapasin binding event may be communicated to the peptide binding groove in such a way as to induce plasticity that promotes peptide selection. Tapasin enhances peptide selector function by catalyzing transitions between conformational intermediates of MHC I Finally, we returned to computational systems modelling to infer the mechanisms by which tapasin enhances peptide selector function. We first quantified the effects of tapasin on peptide selection in vivo , for the three MHC I alleles B*4402, B*4405 and B*4405 W147A . As in the absence of tapasin, we developed computational systems models of MHC I peptide selection to encode distinct mechanistic hypotheses and used Bayesian model selection to determine which hypotheses were most likely, given the experimental data. To quantify peptide selector function in the presence of tapasin, we performed thermostability and EndoH pulse-chase experiments in tapasin-competent cells ( Fig. 6A,B ), similar to the experiments performed in tapasin-deficient cells ( Fig. 2 ). In the presence of tapasin, all three alleles progressed to the cell surface with the same rapid kinetics ( Fig. 6A , quantified in 6G) and loaded rapidly with stabilizing peptides ( Fig. 6B , quantified in 6H). This was consistent with the hypothesis that tapasin masked intrinsic differences in MHC I selector function, as previously identified. Figure 6 Computational systems models of the mechanisms of MHC I peptide selection, fit to in vivo peptide selection data for B*4402, B*4405 and B*4405 W147A in the presence and absence of tapasin. ( A ) Repeating the thermostability assay shown in Fig. 2A in the presence of tapasin indicates that B*4402 and B*4405 W147A now acquire thermostability equal to that of B*4405 (quantified in G). ( B ) Repeating the pulse chase assay shown in Fig. 2B in the presence of tapasin shows that all alleles select high affinity peptides in the presence of tapasin and traffic to the cell surface (quantified in H). ( C ) Graphical depiction of the general computational systems model of the mechanisms of peptide selection in the presence of tapasin, as in Fig. 2D . ( D ) Comparison of model performance for one-conformation and two-conformation models, as in Fig. 2E . The most likely model (with the lowest BIC score) was once again identified as the two-conformation model with MHC I opening as the peptide-dependent step and MHC I closing rate c as the allele-dependent parameter. ( E ) Flux analysis of the two-conformation model with peptide-dependent opening (including tapasin) reveals an anti-clockwise cycle of tapasin mediated peptide editing (the peptide-specific reactions are shown as thick red symbols and grey lines indicate unfavorable reactions). ( F – H ) Comparison of model behavior including a function for tapasin, analogous to Fig. 2F,G . Experimental measurements (circles) quantified from Figs 2A,B and 4A,B and panels ( A , B ) in this figure. In H solid black are + tapasin and dashed/open are – tapasin. Full size image To model potential mechanisms by which tapasin could enhance peptide selection, we introduced tapasin into the one-conformation and two-conformation models of Fig. 2 , as shown in Fig. 6C (see Supplementary Information and Figure S6 for details of the models). Since tapasin has been shown to elevate loading of stable peptides in vivo and to enhance peptide dissociation and association in vitro 37 , 38 , for the one-conformation model we assumed that tapasin enhanced peptide binding to and unbinding from MHC I, as previously suggested 27 . For the two-conformation models, we assumed that tapasin enhanced the peptide-dependent step, together with the rates of transition between conformational states. To determine which mechanistic hypotheses were most likely, we fit the parameters of each model to the experimental data, both in the presence ( Figs 2A,B and 4B,C ) and absence ( Fig. 6A,B ) of tapasin, for all three MHC I alleles simultaneously. As in the absence of tapasin ( Fig. 2E ), the model of best fit was identified as the two-conformation model, with MHC I opening as the peptide-dependent step and MHC I closing rate c as the allele-specific parameter ( Fig. 6D ). Furthermore, as in the absence of tapasin ( Fig. 4D ), the closing rate for B*4405 W147A was identified as lying between the values for B*4402 and B*4405 ( Table S2 ). Consistent with biochemical data 37 , 38 we observed a close fit of the model to the data when tapasin accelerated both the peptide-dependent rate of opening (factor q ) and the allele-dependent rate of closing ( c T > c ) of peptide-bound MHC I ( Table S2 ). Therefore, our modelling results suggest that tapasin confers enhanced selector function by acting as a catalyst to enhance MHC I plasticity and thus promote peptide exchange. To understand how specific interactions contribute to the overall behavior of the system, we performed a flux-analysis of the model ( Fig. 6E ; see Supplementary Information and Figure S6 for details of this analysis). Here, each transition in the model was removed systematically (i.e. set to zero) and the effect on the deviation of the model from the experimental data quantified using the maximum likelihood score. For example, we set the rate of the transition from TM o P i to TM o to zero and found only a small effect on model behavior, which suggests that this transition does not significantly affect peptide selection. This analysis was applied to all transitions and summarized in Fig. 6E , which shows essential transitions left in black, transitions with minor contributions in grey and non-essential transitions removed. The scheme illustrates some key features of the model: i) open MHC I preferentially binds to tapasin rather than peptide, ii) peptide binds preferentially to tapasin-bound MHC I and iii) tapasin is predicted to dissociate preferentially from closed, peptide occupied MHC molecules. This is consistent with recent demonstrations that MHC-binding peptides accelerate the dissociation of tapasin from MHC I, measured by surface plasmon resonance 38 . Discussion MHC class I antigen processing can be broadly divided into three phases: peptide generation, peptide selection and peptide presentation. The efficiency of peptide selection in the MHC class I pathway, as well as in the MHC class II pathway 39 , has great relevance for disease pathogenesis 40 , 41 and therapeutic developments 42 . Hence we sought a mechanism for MHC I peptide selection, by combining modelling and experimental techniques. We highlight three main areas of future investigation. Firstly, here we have simplified a complex system and excluded other factors that may potentially influence peptide selection, such as other chaperone molecules and the nature of the peptide supply 43 . Reducing the system makes the problem tractable in the first instance and creates a foundation upon which to build more complex systems. Secondly, we note that molecular dynamics simulations in the peptide free state do not have a known corresponding crystal structure: all known structures of MHC I are peptide bound. More generally, our observations from molecular dynamics are dependent upon simulation time and the reliability of the force field to reproduce physical behavior. The simulations presented here are of comparable length, or longer, than those previously performed for these MHC I molecules 22 , 23 , 24 . We anticipate that future NMR studies will test the validity of these observations, however there already exists direct structural evidence for MHC I populating more than one peptide bound conformation 24 , 44 . Thirdly, the differences in MHC I plasticity and corresponding selector function have been shown here for three HLA-B molecules and the generality of this observation remains to be explored, in particular by considering additional HLA alleles and point mutations. However, there is in vitro and computational evidence suggesting that plasticity is a common functional property of MHC I 19 , 20 , 22 , 24 , 45 , 46 and many other proteins 36 , 47 . In correlating our observations from the computational systems modelling with those from the molecular dynamics simulations, we can start to build a framework for understanding the mechanisms of MHC I peptide selection. This framework describes how the functional differences in conformational intermediates of distinct MHC I alleles can arise from intrinsic differences in the plasticity of the heavy chain structure encoded by a single amino acid polymorphism. Moreover, we have been able to distinguish between the intrinsic selector properties of MHC I molecules and the impact of the cofactor tapasin in modulating MHC I peptide selection. Our flux analysis identifies that tapasin binds preferentially to conformational intermediates of MHC I and accelerates the closing rate of MHC I, increasing the likelihood of selecting a high affinity peptide. These observations correlate with recent observations of the assembly efficiency of different HLA-B alleles in the absence and presence of tapasin 13 . The correlation with our model is that an MHC I allele with a slow closing rate progresses slowly to the peptide dependent step in the absence of tapasin, thus having a retarded selector function. The model assigns the function of tapasin as one of a catalyst-chaperone, accelerating the rate of closure of MHC I and thus the rate at which MHC I reaches the peptide dependent step. This is consistent with previous investigations indicating that peptide selection by MHC I depends upon conformational change 48 . Our analysis of the flux of MHC I through the model revealed an iterative cycle of tapasin catalyzed peptide exchange, which continues until a peptide of sufficiently high affinity enables MHC I to egress from the ER. This is not dissimilar to the model proposed for HLA-DM and HLA-DR in the MHC Class II pathway 49 , 50 , 51 . From examination of the crystal structure of DR and the DM-DR complex it has been postulated that DM modulates the conformation of peptide free DR such that peptides accessing the peptide binding groove must compete with DM for DR to reverse the conformational change, triggering DM dissociation. This is analogous to tapasin modulated conformational change of MHC I, leading to an increased rate at which MHC I reaches the peptide dependent step in our model, in turn enhancing both the rate and quantity, of high affinity peptides selected. A key finding in our model is that the difference in intrinsic peptide selection efficiency between alleles can be explained by intrinsic differences in their ability to undergo conformational changes. This is striking because B*4405 and B*4402 differ by only a single amino acid. Previously, it has been shown by molecular dynamics simulation 23 that the position 116 polymorphism alters the plasticity of B*4405 and B*4402. By extending these observations to include the B*4405 W147A allele we observed a correlation between plasticity and the rate of transition between an open and closed conformation of the peptide binding domain. This suggests a link between plasticity and closing rate such that more plastic MHC I alleles are able to close more quickly. Furthermore, by extending our observations to include the α 3 domain we can postulate two ways in which plasticity determines peptide selection at a structural level. Firstly, through our observation of coupled dynamics between the membrane proximal α 3 domain and peptide binding site of the MHC I heavy chain, we see how the 116 polymorphism or the W147A mutation can alter the intrinsic plasticity of these molecules. In the case of B*4405 W147A we observe a shift in plasticity towards that of the B*4402 phenotype. Secondly, in demonstrating that we can modulate the conformation of the peptide binding groove of B*4402 via the 220–227 region of the α 3 domain, we provide an explanation of how tapasin may catalyze peptide selection both directly via the peptide binding site and additionally allosterically: a phenomenon we also recently observed in chicken MHC 34 . It has also been reported that MHC I displays plasticity at the short 3 10 helix forming part of the peptide-binding groove interacting the N-terminus of bound peptides 20 and in the α 2-1 helix of HLA-A*02:01 crystallized with different peptide and T-cell receptor combinations 24 . Further molecular dynamics simulations report how different peptides can influence the plasticity of the MHC I binding groove of HLA-B*27 alleles 22 , 52 and recent NMR studies report conformational changes in T-cell receptor interaction of H2-L d bound to different peptides 45 . Notably, polymorphisms at position 156 in the α 2-1 helix in HLA-B*44 alleles, residing at a critical hinge-point important for articulating plasticity have been linked to long-term non-progression in HIV-1 infected individuals 41 and are associated with HLA-linked drug allergy 40 , 53 and non-permissive transplantation mismatches 42 . These correlations between protein dynamics and disease pathogenesis are suggestive of the importance of MHC I plasticity throughout the MHC I antigen processing and presentation pathway. Therefore, whilst there remains much to be understood regarding precisely what determines the generation of immunogenic MHC I molecules at the cell surface, the possibility of modulating protein plasticity as a therapeutic target provides an exciting avenue of investigation. Methods Molecular dynamics simulations The GROMACS version 4.5.3 30 molecular dynamics package was used for the all atom simulations with the Amber99SB-ILDN 54 force field. Further details are provided in the Supplementary Information . Specifically, we present block RMSF calculations in Figure S1 as our check for simulation stability and also provide a summary of the statistics of the MD simulations in Table S1 , including RMSD. Cell lines The 721.220 cell lines expressing tapasin, B*4402 and B*4405 are previously described 10 . RSV.5neo.B*44:05 was mutated at position 147 by site-directed mutagenesis. Mutation was confirmed by sequencing. 721.220 and 721.220.tapasin cells were transfected using Nucleofection and stable transfectants were grown under G418 and/or puromycin selection. Phenotype was confirmed by FACS and Western blot analysis. BFA Decay For cell surface decay experiments, cells were treated with Brefeldin A for the specified time points. Cells were stained with W6/32 and analyzed by FACS. Surface MHC I was expressed as percentage of mean channel fluorescence at time point 0. Pulse-chase and thermostability assays These were carried out as previously described 10 using 10 μCi/ml 35 S-Translabel and 500 U EndoH. Bands were detected using Personal Molecular Imager FX and quantified using Quantity One software. Immunoprecipitations Cells were lysed in PBS containing 1% Digitonin (WAKO), PMSF and IAA. Supernatants were pre-cleared with protein A sepharose, immunoprecipitated with anti-tapasin PaSta-1 antibody (kind gift from Peter Cresswell) and complexes were recovered with protein A. Beads were washed in 0.1% Digitonin and eluted with Laemmli sample buffer. Eluted proteins and samples of lysate, prior to immunoprecipitation were separated by SDS-PAGE, transferred by Western blot and detected by chemilluminescence using Fluor-S Multimager. Anti-human TAP1, anti-HLA-B (N-20), HRP-conjugated, anti-rabbit light chain and anti-goat antibodies were used. Computational modelling The computational models were constructed as systems of chemical reactions and simulated as ordinary differential equations assuming mass action kinetics. Numerical integration of the resulting equations and Bayesian parameter inference was performed as described previously 27 . Further details are provided in the Supplementary Information . Additional Information How to cite this article : Bailey, A. et al. Selector function of MHC I molecules is determined by protein plasticity. Sci. Rep. 5 , 14928; doi: 10.1038/srep14928 (2015).

New research has uncovered an important mechanism in the drive to understand immunological processes that protect us against infection, allergy and cancer. Researchers from Medicine, Chemistry and Biological Sciences in the University's Institute for Life Science (IfLS) have been collaborating with Microsoft Research UK to investigate the function of the antigen-presenting protein MHC1. Their research, which started more than a decade ago and has been part-funded by Cancer Research UK, has explored the pliability of the protein and how its ability to shape-shift dictates its function. Principal Investigator Tim Elliott, Professor of Experimental Oncology and Deputy Director for the IfLS, said: "This protein acts as a molecular interface between your body and your immune system. It alerts your immune system to the fact that the body has been infected by a virus or invaded by cancer, and guides white blood cells to kill them. What we have discovered is that it can only perform these vital functions if it is allowed to "wriggle" in a particular way. We also discovered that, because it is a pliable molecule, different parts of the protein communicate with one another - if we touch it in one place, the function in a distant part of the molecule changes. These findings are of real interest to both immunologists who are developing new immunotherapies for diseases and biology as a whole. They have generated real interest in the healthcare industry." The team's discoveries could have a major impact on the future of medical treatment and in the long term could see the development of cancer immunotherapies and vaccines against infection. It could also help to reduce allergies such as contact hypersensitivity by understanding how the additives used in healthcare products are detected by the immune system. The results of the study were recently published in the journal Scientific Reports and have led to the award of significant industry funding from the consumer goods giant Unilever. It has awarded the team £700,000 to work in collaboration with its scientists to continue exploring ways to minimise allergic reactions, involving MHC1, to additives in healthcare products. Professor Elliott said: "This is fantastic news. It is a strong indication of the quality of our research that consumer companies such as Unilever are interested in the fundamental science that we are exploring and can see the potential applications for industry even if they are still a long way ahead. By bringing together different disciplines under the umbrella of the IfLS we are making a step change in the way we solve biological problems."

10.1038/srep14928

Medicine

Study highlights possible Achilles' heel in key immune memory cells

Youdong Pan et al, Survival of tissue-resident memory T cells requires exogenous lipid uptake and metabolism, Nature (2017). DOI: 10.1038/nature21379 Journal information: Nature

http://dx.doi.org/10.1038/nature21379

https://medicalxpress.com/news/2017-03-highlights-achilles-heel-key-immune.html

Abstract Tissue-resident memory T (T RM ) cells persist indefinitely in epithelial barrier tissues and protect the host against pathogens 1 , 2 , 3 , 4 . However, the biological pathways that enable the long-term survival of T RM cells are obscure 4 , 5 . Here we show that mouse CD8 + T RM cells generated by viral infection of the skin differentially express high levels of several molecules that mediate lipid uptake and intracellular transport, including fatty-acid-binding proteins 4 and 5 (FABP4 and FABP5). We further show that T-cell-specific deficiency of Fabp4 and Fabp5 ( Fabp4 / Fabp5 ) impairs exogenous free fatty acid (FFA) uptake by CD8 + T RM cells and greatly reduces their long-term survival in vivo, while having no effect on the survival of central memory T (T CM ) cells in lymph nodes. In vitro , CD8 + T RM cells, but not CD8 + T CM cells, demonstrated increased mitochondrial oxidative metabolism in the presence of exogenous FFAs; this increase was not seen in Fabp4 / Fabp5 double-knockout CD8 + T RM cells. The persistence of CD8 + T RM cells in the skin was strongly diminished by inhibition of mitochondrial FFA β-oxidation in vivo . Moreover, skin CD8 + T RM cells that lacked Fabp4 / Fabp5 were less effective at protecting mice from cutaneous viral infection, and lung Fabp4 / Fabp5 double-knockout CD8 + T RM cells generated by skin vaccinia virus (VACV) infection were less effective at protecting mice from a lethal pulmonary challenge with VACV. Consistent with the mouse data, increased FABP4 and FABP5 expression and enhanced extracellular FFA uptake were also demonstrated in human CD8 + T RM cells in normal and psoriatic skin. These results suggest that FABP4 and FABP5 have a critical role in the maintenance, longevity and function of CD8 + T RM cells, and suggest that CD8 + T RM cells use exogenous FFAs and their oxidative metabolism to persist in tissue and to mediate protective immunity. Main Memory T cells protect the host through rapid recall responses to pathogens. A population of memory T cells that is vital for host defence, T RM cells, has recently been characterized 1 , 2 , 3 , 4 . T RM cells reside in epithelial barrier tissues and persist for long periods of time at the interface between host and environment 3 , 4 . Upon re-infection, CD8 + T RM cells provide a rapid antigen-specific immune response, creating an inflammatory and antiviral microenvironment that facilitates pathogen elimination 6 , 7 , 8 , 9 . Although previous studies have yielded clues 10 , 11 , 12 , 13 , little is known about the molecular program that regulates the long-term survival of these cells. To answer this question, we first evaluated skin T RM cell maturation by comparing gene expression patterns at different time points after infection. OT-I transgenic mouse T cells were transferred into recipient mice one day before immunization with a recombinant VACV that expresses chicken ovalbumin peptide (amino acids 257–264) under the control of an early gene promoter (rVACV OVA ) 14 . OT-I cells were readily found in the skin at day 5 after infection and reached their maximum level at day 10, before beginning to decrease in numbers ( Extended Data Fig. 1a ). Skin-infiltrating OT-I cells were sorted at different time points after infection and were analysed by transcriptional profiling. Principal-component analysis showed that transcriptomes of skin-infiltrating T cells clustered tightly from day 25 to day 90 after infection, suggesting that mouse skin CD8 + T RM cell maturation is largely completed by day 25 after infection ( Fig. 1a ). Transcriptomes of T RM cells are distinct from those of T CM cells and effector memory T (T EM ) cells ( Fig. 1a, b and Extended Data Fig. 1b ), consistent with previous reports 11 , 12 , 13 . Next, we directly compared T RM cells (day 30) and T CM cells ( Fig. 1c ). Notably, genes encoding FABP4 and FABP5 were among the most strongly upregulated genes in T RM cells, as was the gene that encodes CD36, a lipid-scavenger cell-surface receptor 15 ( Fig. 1c ). Quantitative real-time PCR (qPCR) confirmed the increased gene expression of Fabp4 and Fabp5 in CD8 + T RM cells ( Fig. 1d, e and Extended Data Fig. 1c ). Immunofluorescence staining of the skin showed expression of FABP4 and FABP5 in skin CD8 + T RM cells ( Fig. 1f ). To extend these observations to other peripheral tissues, mice with transferred OT-I cells were infected with VACV OVA by intratracheal infection and gene expression of Fabp4 and Fabp5 was measured 30 days later in lung CD8 + T RM cells. Consistently, increased Fabp4 and Fabp5 gene expression was observed ( Extended Data Fig. 1d ). Figure 1: Skin CD8 + T RM cells show increased expression of FABP4 and FABP5. a , Principal-component analysis (PCA) of gene-expression data for CD8 + T cell subtypes. Each time point represents an individual experiment wherein mRNA was pooled from 15–20 mice from 3–4 independent biological groups (5 mice per group). Numbered dots are for skin T cells derived after infection for the indicated number of days. b , Pearson correlation coefficients among CD8 + T cell subtypes. c , Heatmap of differentially expressed genes selected from a pair-wise comparison between OT-I T RM (day 30) and T CM cells. d , qPCR analysis of Fabp4 and Fabp5 expression in T N , T CM , T EM and T RM cells (day 30). e , qPCR analysis of Fabp4 and Fabp5 gene expression in skin CD103 − and CD103 + T RM cells (day 30). f , Immunofluorescence staining of FABP4 (top) and FABP5 (bottom) in OT-I T RM cells 30 days after infection. Scale bar, 20 μm. g , qPCR analysis of Pparg expression in T N , T CM , T EM and T RM (day 30). h , Effect of lentiviral Pparg siRNA knockdown (KD) on Fabp4 and Fabp5 expression in OT-I CD8 + T RM cells. Graphs in d , e , g , h show mean ± s.d. from triplicates. β-actin was used as internal control and mRNA was normalized to T N cells ( d , e , g ) or T RM cells transduced with a lentiviral vector encoding scrambled siRNA ( h ). T cells from 15–20 mice were pooled for each group. ** P < 0.01; NS, not significant. PowerPoint slide Source data Full size image Peroxisome proliferator-activated receptors (PPARs) are adipogenic regulators that have been reported to influence Fabp4 and Fabp5 gene expression 16 . Pparg , but not Ppara or Ppard , was selectively upregulated in T RM cells compared to naive T (T N ) cells, T CM and T EM cells ( Fig. 1g and data not shown). Knockdown of Pparg expression using short interfering (si)RNA lentiviruses, or treatment of mice with GW9662 (an irreversible PPARγ antagonist), inhibited Fabp4 and Fabp5 gene expression in CD8 + T RM cells ( Fig. 1h and Extended Data Fig. 1e, f ). These data indicate that PPARγ is an upstream regulator of Fabp4 and Fabp 5 gene expression. Upon activation, naive T cells undergo metabolic reprogramming as they proliferate and develop into different subsets of memory T cells 17 , 18 . The strongly upregulated T RM genes Fabp4 and Fabp5 encode lipid chaperone proteins that bind to hydrophobic ligands, thereby coordinating lipid uptake and intracellular trafficking 19 . Extracellular FFAs could be visualized in the mouse epidermis, where skin CD8 + T RM cells localize 3 ( Extended Data Fig. 2 ). Given the magnitude of their upregulation, we hypothesized that FABP4 and FABP5 might play a role in CD8 + T RM cell physiology in the skin. To test our hypothesis, we first compared the extracellular FFA uptake of OT-I memory T cell subtypes in vitro . Compared to T N , T CM and T EM cells, substantially more of a green fluorescent fatty acid, Bodipy FL C 16 , was internalized by OT-I T RM cells ( Fig. 2a ). FABP4 and FABP5 are highly homologous and bind to fatty acids with similar selectivity and affinity 19 . Given the compensatory and redundant role of these two molecules 20 , we used double-knockout mice that lacked both Fabp4 and Fabp5 ( Fabp4 −/− / Fabp5 −/− ) to analyse the contribution of these molecules to FFA uptake. FACS analysis showed that OT-I Fabp4 −/− / Fabp5 −/− T RM cells internalized substantially less fatty acids in vitro compared to wild-type cells ( Fig. 2b ), suggesting that FABP4 and FABP5 are important for the specific uptake of palmitate. We then investigated whether loss of FABP4 and FABP5 would impair T RM cell behaviour in vivo . OT-I wild-type and OT-1 Fabp4 −/− / Fabp5 −/− cells were mixed at a 1:1 ratio and transferred into congenic recipients. Mice were then infected with VACV OVA , and the number of OT-I wild-type and OT-I Fabp4 −/− / Fabp5 −/− cells in different anatomic compartments was assessed. No difference was observed between the number of spleen wild-type OT-I or OT-I Fabp4 −/− / Fabp5 −/− T CM cells at any time point, indicating that deficiency in FABP4 and FABP5 did not affect T CM cell survival ( Fig. 2c ). However, OT-I Fabp4 −/− / Fabp5 −/− cells in the skin displayed a marked reduction in persistence, beginning at 25 days after infection. The ratio of OT-I Fabp4 −/− / Fabp5 −/− to OT-1 wild-type T RM cells declined steadily over time thereafter ( Fig. 2c ). Deficiency of both FABP4 and FABP5 also decreased numbers of OT-I T RM cells detectable by immunofluorescence staining without affecting their recruitment or tissue localization ( Fig. 2d and Extended Data Fig. 3 ). By contrast, OT-1 Fabp4 −/− , Fabp5 −/− or Fabp4 +/− / Fabp 5 +/− T RM cells showed no defect in long-term survival ( Extended Data Fig. 4 ), consistent with the compensatory and redundant role of these two proteins 20 . Highly proliferative OT-I effector T (T eff ) cells had a higher uptake of FFAs than T N , T CM and T EM cells but less than T RM cells ( Extended Data Fig. 5a, b ). Notably, OT-I Fabp4 −/− / Fabp5 −/− T eff cells displayed a similar proliferative capacity and tissue-homing receptor expression as wild-type T eff cells at 60 h after infection ( Extended Data Fig. 5c ). These data indicate that the absence of FABP4 and FABP5 does not affect CD8 + T-cell proliferation or recruitment to the skin. Inhibition of Fabp4 / Fabp5 gene expression by knocking down Pparg expression in OT-I cells, or inhibition of PPARγ by GW9662 treatment, decreased the long-term persistence of CD8 + T RM cells in the skin ( Extended Data Fig. 6 ). These data suggest that FABP4 and FABP5 are essential for the long-term survival of CD8 + T RM cells in skin. Annexin-V staining of OT-I wild-type and Fabp4 −/− / Fabp5 −/− T RM cells shows that the latter had a higher rate of apoptosis ( Fig. 2e ). The gene-expression profile of Fabp4 −/− / Fabp5 −/− T RM cells revealed that immune-response genes were significantly downregulated, whereas genes involved in anti-inflammatory responses and apoptosis were upregulated compared to wild-type cells ( Extended Data Fig. 7 ). Figure 2: Loss of Fabp4 and Fabp5 decreases fatty-acid uptake and metabolism by CD8 + T RM cells and impairs their long-term maintenance. a , Representative histograms and average mean fluorescence intensity (MFI) of Bodipy FL C 16 uptake by T N , T CM , T EM or T RM cells (day 30). b , Representative histograms and average MFI of Bodipy FL C 16 uptake in OT-I wild-type and Fabp4 −/− / Fabp5 −/− T RM cells with or without pre-incubation with unlabelled palmitate (Palm). c , Number of OT-I wild-type and Fabp4 −/− / Fabp5 −/− T CM and T RM cells at different time points after infection in spleen and skin. d , Immunofluorescence staining and quantification of OT-I wild-type and Fabp4 −/− / Fabp5 −/− T RM cells at infected skin sites 45 days after infection. Arrowheads, T RM cells; FOV, field of view. Scale bar, 50 μm. e , Representative histograms and quantification of annexin V + cells in OT-I wild-type and Fabp4 −/− / Fabp5 −/− T RM cells 45 days after infection. Values are the % of annexin V + cells. f , OCR of OT-I T CM , T RM and Fabp4 −/− / Fabp5 −/− T RM cells (day 30) under basal conditions and in response to indicated mitochondria inhibitors. Results were normalized to control cells treated with bovine serum albumin (BSA) only. Eto, etomoxir. Graphs show mean ± s.d. of triplicates. g , Effect of lentiviral Cpt1a siRNA knockdown on OT-I CD8 + T RM survival in vivo in infected tissue. h , Representative dot plots and percentage of wild-type and Fabp4 −/− / Fabp5 −/− T CM and T RM cells 45 days after infection. Cells were gated on VACV-specific pentamer + CD8 + T cells. Graphs show mean ± s.d. of 5 ( a , b ) or 10 ( c ) mice, or 50 fields from 5 mice ( d , 10 fields per mouse), or symbols represent individual mice ( e , g , h ). * P < 0.05; ** P < 0.01; NS, not significant. PowerPoint slide Source data Full size image To determine the dependence of T RM cells on exogenous FFA uptake for oxidative metabolism, we used the Seahorse fatty-acid oxidation assay 21 . Addition of extracellular fatty acids induced a significantly higher basal and FCCP-stimulated maximum oxygen-consumption rate (OCR) in OT-I T RM cells ( Fig. 2f ). The increase in OCR was blocked by pre-treatment with etomoxir, an inhibitor of mitochondrial carnitine palmitoyltransferase 1a (CPT1A), an enzyme central to mitochondrial fatty acid β-oxidation 22 ( Fig. 2f ). By contrast, T CM or Fabp4 −/− / Fabp5 −/− T RM OT-I cells did not have an increased OCR when supplied with exogenous fatty acids, and the addition of etomoxir had no effect on their cellular respiration ( Fig. 2f ). In vivo knockdown of Cpt1a or treatment of mice with either etomoxir or trimetazidine 23 decreased the number of OT-I wild-type T RM cells to a similar extent as Fabp4 −/− / Fabp5 −/− cells ( Fig. 2g and Extended Data Fig. 8a–c ). These data suggest that skin CD8 + T RM use oxidative metabolism of exogenous FFAs to support their long-time survival. Early after infection, roughly equivalent numbers of wild-type and Fabp4 −/− / Fabp5 −/− CD8 + T eff cells were found in skin ( Fig. 2c ). Compared to skin T RM cells isolated at day 30, skin-infiltrating T eff cells isolated at days 10 and 15 had a lower OCR, but a higher basal extracellular acidification rate (ECAR), which corresponds to glycolysis 21 ( Extended Data Fig. 8d ). Deficiency in Fabp4 / Fabp5 decreased the OCR of T RM cells at day 30 after infection, but had no effect on the ECAR of skin-infiltrating T eff cells ( Extended Data Fig. 8d ). These data suggest that early skin-infiltrating T eff cells use glycolysis, which is unaffected by Fabp4 / Fabp5 gene expression. To establish the contribution of FABP4 and FABP5 to survival of non-transgenic CD8 + T RM cells, bone-marrow chimaeric mice, which have a 1:1 ratio of Thy1.1 + CD45.1 + wild-type and Thy1.1 + CD45.2 + Fabp4 −/− / Fabp5 −/− bone-marrow cells were infected with rVACV by skin scarification. After 45 days, infected skin tissue was collected and the number of VACV-pentamer + CD8 + T cells 3 was analysed by flow cytometry. Consistent with data from OT-I experiments, fewer Fabp4 −/− / Fabp5 −/− T RM cells were detected compared to wild-type T RM cells ( Fig. 2h ). T RM cells are more effective at clearing tissue VACV infections than T CM cells 3 . We evaluated the contribution of FABP4 and FABP5 to viral clearance of CD8 + T RM cells. Mice were adoptively transferred with OT-I wild-type or OT-I Fabp4 −/− / Fabp5 −/− cells and then infected by skin scarification with VACV OVA . Mice were re-challenged with VACV OVA 25 days later and skin viral load was measured 6 days later ( Fig. 3a ). FTY720, a sphingosine-1-phosphate receptor antagonist, was injected into mice to assess the contribution of circulating T CM cells to viral clearance ( Extended Data Fig. 9a ). Established OT-I wild-type T RM cells were highly effective at clearing virus from skin, as virus clearance was rapid and unaffected by FTY720 treatment ( Fig. 3b ). By contrast, OT-I Fabp4 −/− / Fabp5 −/− T RM cells were less effective at viral clearance, and least effective following FTY720 treatment ( Fig. 3b ). Treatment of mice with etomoxir reduced the VACV-clearing capacity of OT-I wild-type T RM cells to a level comparable to that of Fabp4 −/− / Fabp5 −/− cells ( Fig. 3b ), suggesting dependence on oxidative metabolism of FFAs. Upon stimulation in vitro , OT-I Fabp4 −/− / Fabp5 −/− T RM cells displayed impaired IFNγ production compared to wild-type cells ( Fig. 3c, d ). Similar results were obtained for OT-I Fabp4 −/− / Fabp5 −/− T RM cells residing at skin sites distant from the infection ( Extended Data Fig. 9b, c ). Figure 3: Skin Fabp4 −/− / Fabp5 −/− CD8 + T RM cells fail to protect mice against viral infectious challenge. a , Schematic of experimental design. i.p., intraperitoneal injection; s.s., skin scarification. b , qPCR analysis of skin viral load at six days after re-infection. U, un-immunized mice; I, immunized mice. c , d , Representative dot plots and quantification of IFNγ secretion by OT-I wild-type or Fabp4 −/− / Fabp5 −/− T RM cells. Symbols represent individual mice ( b , d ). e , Schematic of experimental design. f , g , Body weight (BW) ( f ) and survival measurements ( g ) of WR-VACV challenged mice, with or without FTY720 treatment. Data are representative of three independent experiments ( n = 10 mice per group). * P < 0.05; ** P < 0.01; NS, not significant. PowerPoint slide Source data Full size image We showed previously that lung CD8 + T RM cells generated by skin VACV vaccination could partially protect mice against a lethal respiratory challenge with VACV 9 . We therefore investigated the role of FABP4 and FABP5 in this protective capacity of lung CD8 + T RM cells generated by skin scarification. Rag1 −/− mice were reconstituted with transfer of CD4 + and CD8 + wild-type or Fabp4 −/− / Fabp5 −/− T cells one day before immunization with rVACV by skin scarification. At day 25, wild-type or Fabp4 −/− / Fabp5 −/− rVACV memory mice were challenged intranasally with lethal doses of the highly pathogenic Western Reserve (WR)-VACV ( Fig. 3e ). Mice with wild-type CD8 + T RM cells did not display symptoms associated with illness (for example, weight loss) ( Fig. 3f ). By contrast, mice with Fabp4 −/− / Fabp5 −/− CD8 + T RM cells showed marked weight loss after challenge and were only partially protected from virus-induced lethality ( Fig. 3f, g ). Treatment with FTY720 led to 100% mortality of mice that had received Fabp4 −/− / Fabp5 −/− CD8 + T RM cells, whereas animals with wild-type CD8 + T RM cells were partially, but significantly, protected ( Fig. 3f, g ). These data suggest that lung Fabp4 −/− / Fabp5 −/− CD8 + T RM cells were less protective against lethal respiratory VACV infection, and required the recruitment of circulating T CM cells. By contrast, wild-type CD8 + T RM cells alone protected 50% of mice from this lethal VACV infection, consistent with our previous data 9 . T RM cells in human skin have been implicated in the pathogenesis of several human skin diseases, including psoriasis 10 , 24 , 25 . We found that FABP4 and FABP5 were both strongly expressed in human skin CD8 + T RM cells compared to human blood T N , T CM and T EM cells using FACS analysis ( Fig. 4a, b ). Psoriasis is a chronic and recurring autoimmune disease ( Fig. 4c ) that is thought to be mediated by CD8 + T RM cells 5 , 26 . Immunofluorescence staining of lesional psoriatic skin showed a co-expression of CD8 and CD69 (ref. 5 ), indicating that CD8 + T cells in psoriasis tissue display a T RM -cell phenotype ( Fig. 4d ). Lipids could be visualized in lesional scalp skin ( Extended Data Fig. 10 ) and FABP4 and FABP5 protein expression could be detected in human psoriatic skin CD8 + T RM cells ( Fig. 4e ). In vitro incubation with exogenous Bodipy FL C 16 showed that human skin CD8 + T RM cells internalized more exogenous FFAs compared to blood T N , T CM and T EM cells ( Fig. 4f ), suggesting a role for FABP4 and FABP5 in the fatty-acid uptake of human CD8 + T RM cells, similar to that demonstrated here in mice. Figure 4: Human skin CD8 + T RM demonstrate increased expression of FABP4 and FABP5. a , b , Representative histograms and average MFI of intracellular staining of FABP4 ( a ) and FABP5 ( b ) in human peripheral blood mononuclear T N , T CM , and T EM cells compared to facial skin T RM . c , Haematoxylin and eosin staining of a human psoriatic lesion. Scale bar, 100 μm. d , Immunofluorescence staining of CD8 and CD69 expression in psoriatic human skin. Dashed line indicates epidermal–dermal junction. Scale bar, 50 μm. e , Immunofluorescence staining of FABP4 (top) and FABP5 (bottom) expression in skin CD8 + T RM cells from patients with psoriasis. Scale bar, 20 μm. f , Representative histograms and average MFI of Bodipy FL C 16 uptake by human T N , T CM , T EM or T RM cell subsets. a , b , f , Graphs show mean ± s.d. of 5 individuals; c – e , Representative images of n = 15 sections from 5 individuals. ** P < 0.01. PowerPoint slide Source data Full size image Skin and other epithelial tissues are lipid-rich but nutrient-poor microenvironments 15 , 27 , and CD8 + T RM cells appear to use mitochondrial β oxidation of exogenous FFAs or other lipids to support both their longevity and protective function. Although T CM cells depend in part on fatty-acid oxidation for cellular metabolism 17 , 28 , our data show that T CM cells cannot effectively internalize exogenous FFAs. Cell-intrinsic lipolysis and increased glycerol transport are used by T CM cells to support metabolic programming necessary for development 17 , 28 , 29 , but the dependence upon exogenous FFA uptake and metabolism for long-term survival is unique to T RM cells. Additionally, it is noteworthy that similar results were obtained from mice injected intradermally with etomoxir and mice with Cpt1a knockdown in OT-I cells ( Fig. 2g and Extended Data Fig. 8a, b ), suggesting that the etomoxir effects on CD8 + T RM cell persistence were mediated through CPT1A 30 . Given that generation of long-lived T RM cells are a goal of effective vaccination 4 , and that dysfunction of T RM cells underlies many auto-inflammatory tissue disorders 4 , 5 , a more detailed understanding of the unique lipid metabolic programs intrinsic to T RM cells and how these programs might be manipulated to increase or decrease T RM cell longevity and function, will be a subject of future investigation. Methods Mice Wide-type C57BL/6, CD45.1 + , Thy1.1 + , Rag1 −/− and μMT mice were purchased from Jackson Laboratory. Thy1.1 + Rag1 −/− OT-I mice were maintained through routine breeding in the animal facility of Harvard Institute of Medicine, Harvard Medical School. Fabp4 −/− , Fabp5 −/− , and Fabp4 −/− / Fabp5 −/− mice were provided by G.S.H. Mice were bred to generate Thy1.1 + CD45.1 + wild-type; Thy1.1 + CD45.2 + Fabp4 −/− / Fabp5 −/− ; Thy1.1 + CD45.1 + wild-type OT-I, Thy1.1 + CD45.2 + Fabp4 +/− / Fabp5 +/− OT-I, Thy1.1 + CD45.2 + Fabp4 −/− OT-I, Thy1.1 + CD45.2 + Fabp5 −/− OT-I and Thy1.1 + CD45.2 + Fabp4 −/− / Fabp5 −/− OT-I mice. Animal experiments were performed in accordance with the guidelines put forth by the Center for Animal Resources and Comparative Medicine at Harvard Medical School, and all protocols and experimental plans were approved by the HMS IACUC beforehand. Mice were randomly assigned to each group before the start and experiments were performed blinded with respect to treatment. For survival experiments, mice that had lost over 25% of their original body weight were euthanized. Viruses and infections Recombinant VACV expressing the OT-I T cell epitope OVA 257–264 and Western Reserve strain (WR-VACV) were originally obtained from B. Moss (NIH). Virus was expanded and titred by standard procedures as described previously 9 . 2 × 10 6 p.f.u. of VACV OVA was used for infection by either skin scarification or intratracheal infection. 2 × 10 6 p.f.u. WR-VACV was used at a lethal dose for intranasal infections, as described previously 9 . Antibodies and flow cytometry The following anti-mouse antibodies were obtained from BD PharMingen: PerCP-conjugated anti-CD3e (553067), PE-conjugated anti-CD8 (557654), PE-Cy7-conjugated anti-CD8 (552877), APC-Cy7-conjugated anti-CD8 (557654), PE-conjugated anti-Thy1.1 (561404), APC-conjugated anti-Thy1.1 (557266), Alexa Fluor 488-conjugated anti-KLRG1 (561619), PE-conjugated anti-integrin α4β7 (553811), PE-Cy7-conjugated anti-CD62L (560516), APC-Cy7-conjugated CD62L (560514), APC-conjugated anti-IFNγ (554413). Anti-mouse antibodies were also obtained from Biolegend: Alexa Fluor 488-conjugated anti-CD3e (100321), Alexa Fluor 647-conjugated anti-CD4 (100424), Alexa Fluor 594-conjugated anti-EpCAM (118222), FITC-conjugated anti-CD45.1 (110706), PE-conjugated anti-CD45.1 (110708), PE-Cy7-conjugated anti-CD45.2 (109830), APC-conjugated anti-CD45.2 (109814), PE-conjugated anti-CD45.2 (109808), APC-conjugated anti-KLRG1 (138412), PE-conjugated anti-CD44 (103008), PE-Cy7-conjugated anti-CD44 (103030), PE-Cy7-conjugated anti-CD69 (104512), APC-conjugated anti-CD103 (121414), APC-conjugated anti-integrin α4β7 (120608). The following anti-human antibodies were obtained from Biolegend: APC-Cy7-conjugated anti-CD3 (300425), PerCp-conjugated anti-CD4 (317431), APC-conjugated anti-CD8 (300911), Alexa Fluor 488-conjugated anti-CD69 (310916), PE-Cy7-conjugated anti-CD69 (310911), PE-conjugated anti-CD62L (304805), FITC-conjugated anti-CD45RO (304204). Antibodies were also obtained from Abcam: anti-human FABP4 (9B8D) and NOVUS: Alexa Fluor 405-conjugated anti-human FABP5 (FAB3077V). PE-conjugated B8R20–27/H-2Kb pentamers were obtained from ProImmune Ltd, and stained according to the manufacturer’s protocol. E- or P-selectin ligand expression was analysed by incubating cells with rmE-Selectin/Fc Chimera (575-ES; R&D System) or rmP-Selectin/Fc Chimera (737-PS; R&D System) in conjunction with PerCP-conjugated F(ab’)2 fragments of goat anti-human IgG Fc antibody (109-126-170; Jackson ImmunoResearch). To measure ex vivo uptake of Bodipy-conjugated palmitate (Bodipy FL C 16 ; D-3821; Thermo Fisher), cells were incubated for 30 min at 37 °C with 1 μM Bodipy FL C 16 in PBS with 20 μM FA-free BSA (A8806; Sigma-Aldrich). Bodipy uptake was quenched by adding 4× volume of ice-cold PBS with 2% FBS and then cells were washed twice before flow cytometry analysis. Annexin V staining was included to exclude dead/dying cells during FACS data acquisition. To block Bodipy uptake, cells were incubated with 100 μM palmitic acid (P0500; Sigma-Aldrich) for 10 min at 37 °C before Bodipy addition. Apoptosis was measured with the FITC Annexin V Apoptosis Detection Kit (640922; Biolegend) according to the manufacturer’s protocol. Flow cytometry data were acquired with a FACS Canto II flow cytometer (BD Biosciences) and data were analysed with Flowjo software (Tree Star). Preparation of cell suspensions Lymph nodes and spleen were collected and pressed through a 70-μm nylon cell strainer to prepare cell suspensions. Red blood cells (RBC) were lysed using RBC lysis buffer (00-4333-57; eBioscience). Skin tissue was excised after hair removal, separated into dorsal and ventral halves, minced, and then incubated in Hanks balanced salt solution (HBSS) supplemented with 1 mg ml −1 collagenase A (11088785103; Roche) and 40 μg ml −1 DNase I (10104159001; Roche) at 37 °C for 30 min. After filtration through a 70-μm nylon cell strainer, cells were collected and washed three times with cold PBS before staining. Mouse adoptive transfer and treatment Lymph nodes were collected from naive female donor mice at the age of 6–8 weeks. T cells were purified by magnetic cell sorting using a mouse CD8α + T-cell isolation kit (130-104-075; Miltenyi Biotec) or a mouse CD4 + T-cell isolation kit (130-104-454; Miltenyi Biotec), according to the manufacturer’s protocols. T cells were then transferred intravenously into female recipient mice at a total number of 5 × 10 5 or 2.5 × 10 5 cells per population in co-transfer experiments where cell types were transferred at a ratio of 1:1. To generate mixed bone-marrow chimaeras, T-cells- and NK-cell-depleted Thy1.1 + CD45.1 + wild-type and Thy1.1 + CD45.2 + Fabp4 −/− / Fabp5 −/− bone marrow was mixed in a 1:1 ratio and transferred at a number of 1 × 10 6 cells per population into sublethally irradiated recipient mice. Mice were rested for eight weeks before infection for full reconstitution of T cells and restoration of an intact immune system. Rag1 −/− T-cell reconstituted mice were generated by adoptive transfer of 3.5 × 10 6 CD4 + cells with 2 × 10 6 CD8 + wild-type or CD8 + Fabp4 −/− / Fabp5 −/− cells. T cells were labelled with carboxyfluorescein succinimidyl ester (CFSE, 65-0850; eBioscience) before co-transfer, where indicated. In some experiments, mice were treated daily with FTY720 (10006292; CAYMAN, 1 mg per kg) by intraperitoneal injection or with etomoxir (E1905; Sigma-Aldrich, 1 μg per site), GW9662 (M6191, Sigma-Aldrich, 1 mg per kg) or trimetazidine (653322, Sigma-Aldrich, 1 mg per kg) by intradermal injection. Microarray, data analysis and qPCR For each microarray dataset, OT-I cells from 15–20 mice were sorted with a FACSAria III (BD Biosciences) and pooled. RNA was extracted with an RNeasy Micro kit (74004; Qiagen). RNA quality and quantity was assessed with a Bioanalyzer 2100 (Agilent). RNA was then amplified and converted into cDNA by a linear amplification method with a WT-Ovation Pico System (3302-60; Nugen). Subsequently cDNA was labelled with the Encore Biotin module (4200-60; Nugen) and hybridized to GeneChip MouseGene 2.0 ST chips (Affymetrix) at the Translational Genomics Core of Partners Healthcare, Harvard Medical School. GeneChips were scanned using the Affymetrix GeneChip Scanner 3000 7G running Affymetrix Gene Command Console version 3.2. The data were analysed by using the Affymetrix Expression Console version 1.3.0.187 using the analysis algorithm RMA. To evaluate overall performance of microarray data, principal-component analysis and Pearson correlation coefficients among 12 diverse samples were applied by using 26,662 transcripts (R program). For relative qPCR, RNA was prepared as described above. A Bio-Rad iCycler iQ Real-Time PCR Detection System (Bio-Rad) was used with the following settings: 45 cycles of 15 s of denaturation at 95 °C, and 1 min of primer annealing and elongation at 60 °C. qPCR was performed with 1 μl cDNA plus 12.5 μl of 2× iQ SYBR Green Supermix (Bio-Rad) and 0.5 μl (10 μM) specific primers: mouse Fabp4 forward (5′-TTTCCTTCAAACTGGGCGTG-3′) and mouse Fabp4 reverse (5′-CATTCCACCACCAGCTTGTC-3′); mouse Fabp5 forward (5′-AACCGAGAGCACAGTGAAG-3′) and mouse Fabp5 reverse (5′-ACACTCCACGATCATCTTCC-3′); mouse Pparg forward (5′-TCGCTGATGCACTGCCTATG-3′) and mouse Pparg reverse (5′-GAGAGGTCCACAGAGCTG ATT-3′); mouse Actb forward (5′-CATTGCTGACAGGATGCAGAAGG-3′) and mouse Actb reverse (5′-TGCTGGAAGGTGGACAGTGAGG-3′). For absolute qPCR, each standard curve was constructed using tenfold serial dilutions of the target gene template ranging from 10 7 to 10 2 copies per ml and obtained by plotting values of the logarithm of their initial template copy numbers versus the mean C t values. The actual copy numbers of target genes were determine by relating the C t value to a standard curve. Immunofluorescence microscopy Mice were perfused with buffer A (0.2 M NaH 2 PO 4 , 0.2 M Na 2 HPO 4 , 0.2 M l -lysine and 0.1 M sodium periodate with 2% paraformaldehyde) and infected skin sites were collected and incubated for 30 min on ice in buffer A. Skin tissue was washed twice with PBS and incubated for 30 min at 4 °C in 20% sucrose. Fixed tissue was embedded in OCT (Tissue Tek IA018; Sakura) and frozen in liquid nitrogen. Skin sections were performed on a cryostat (Leica CM1850 UV) at 6-μm thickness and air-dried for 6–8 h. Sections were then fixed in −20 °C acetone for 5 min, rehydrated with PBS, and blocked with 2% FCS in PBS for 15 min at room temperature (20 °C). Sections were stained with rabbit anti-mouse/human FABP4 antibody (EPR3579; ab92501, Abcam), rabbit anti-mouse/human FABP5 (H-45, sc-50379, Santa Cruz) overnight at 4 °C in a semi-humid chamber. Sections were rinsed for 10 min in PBS, and labelled with donkey anti-rabbit Rhodamine Red-X (711-296-152; Jackson ImmunoResearch) for 1 h at room temperature (20 °C). Sections were rinsed for 10 min in PBS, and stained with Alexa Fluor 647-conjugated anti-mouse Thy1.1 (202508, Biolegend) in PBS for 1 h at room temperature. Then sections were rinsed three times (for 5 min each time) with TBS-Tween 20 by shaking and mounted with ProLong Diamond Antifade Mountant with DAPI (P36962; ThermoFisher). For tissue lipid visualization, sections were stained with BODIPY 493/503 (4,4-Difluoro-1,3,5,7,8-Pentamethyl-4-Bora-3a,4a-Diaza-s-Indacene) (D3922, Molecular Probes) before being mounted. Images were acquired with a Leica TCS SP8 confocal microscope (Harvard NeuroDiscovery Center Optical Imaging Core) and analysed with ImageJ. Lentiviral siRNA transduction Scrambled, Pparg and Cpt1a siRNA GFP lentiviruses were purchased from ABM (Applied Biological Materials Inc.) with sequences as follows. Scrambled siRNA: GGGTGAACTCACGTCAGAA; Pparg KD1: AATATGACCTGAAGCTCCAAGAATA; Pparg KD2: GTCTGCTGATCTGCGAGCC; Cpt1a KD1: GGAGCGACTCTTCAATACTTCCCGCATCC, Cpt1a KD2: GGTCATAGAGACATCCCTAAGCAGTGCCA. For siRNA lentivirus transduction, OT-I mice were infected with 2 × 10 6 VACV OVA by skin scarification. After 60 h, CD8 + T cells were collected from draining lymph nodes and incubated in medium with 10 μg ml −1 polybrene and 20 ng ml −1 hIL-2 at 37 °C for 30 min. Then cells were infected with scrambled, Pparg or Cpt1a siRNA GFP lentiviruses, in the presence of ViralPlus Transduction Enhancer G698 at 1:100 in order to enhance transduction efficiency. For adoptive transfer, 2.5 × 10 5 Pparg or Cpt1a siRNA-transduced OT-I cells (together with the same number of congenically scrambled-siRNA-transduced OT-I cells) were co-transferred into recipient mice that were previously infected with 2 × 10 6 VACV OVA by skin scarification. After 40 days, mice were euthanized and the number of siRNA-transduced OT-I cells in the infected skin tissue were isolated and enumerated by flow cytometry on the basis of the GFP marker. Recipient and both donor populations used for co-transfers differed in CD90 and CD45 alleles (CD90.2/45.2, CD90.1/45.1 or CD90.1/45.2—the combinations differed between experiments) to allow for identification of donor populations. Fatty-acid oxidation assay Oxidation of exogenous FFAs was measured using the XF Palmitate-BSA FAO substrate with the XF cell mito stress kit according to the manufacturer’s protocol (Seahorse Bioscience). Freshly isolated and sorted T cells (2.5 × 10 5 ) were incubated for 30 min with fatty-acid oxidation assay medium (111 mM NaCl, 4.7 mM KCl, 1.25 mM CaCl 2 , 2.0 mM MgSO 4 , 1.2 mM Na 2 HPO 4 , 2.5 mM glucose, 0.5 mM carnitine and 5 mM HEPES). When required, cells were pre-treated with etomoxir (40 μM) for 15 min. Afterwards, BSA (34 μM) or palmitate-BSA (200 μM palmitate conjugated with 34 μM BSA) was added to the medium, and the OCR was measured under basal conditions and in response to 1 μM oligomycin, 1.5 μM fluorocarbonyl cyanide phenylhydrazone (FCCP), and 100 nM rotenone + 1 μM antimycin A. Results were normalized to data from control cells in the presence of BSA. Determination of viral load VACV load was evaluated by qPCR as described previously 3 . In brief, 6 days after re-infection, inoculated skin samples were collected and DNA was purified with the DNeasy Mini Kit (51304; Qiagen) according to the manufacturer’s protocol. qPCR was performed with the Bio-Rad iCycler iQ Real-Time PCR Detection System (Bio-Rad Laboratories). The primers and TaqMan probe used in the qPCR assay are specific for the ribonucleotide reductase Vvl4L of VACV. The sequences are (forward) 5′-GACACTCTGGCAGCCGAAAT-3′; (reverse) 5′-CTGGCGGCTAGAATG GCATA-3′; (probe) 5′-AGCAGCCACTTGTACTACACAACATCCGGA-3′. The probe was 5′-labelled with FAM and 3′-labelled with TAMRA (Applied Biosystems). Amplification reactions were performed in a 96-well PCR plate (Bio-Rad Laboratory) in a 20 μl volume containing 2× TaqMan Master Mix (Applied Biosystems), 500 nM forward primer, 500 nM reverse primer, 150 nM probe, and the template DNA. Thermal cycling conditions were 50 °C for 2 min and 95 °C for 10 min for 1 cycle, followed by 45 cycles of amplification (94 °C for 15 s and 60 °C for 1 min). To calculate the viral load, a standard curve was generated from DNA of a VACV stock with a previously determined titre. Corresponding C t values obtained by the qPCR methods were plotted on the standard curve to estimate viral load in the skin samples. Intracellular cytokine detection Infected skin was collected 6 days after VACV OVA re-infection and single-cell suspensions were prepared as described above. Cells were then incubated with 2 μg ml −1 SINFEKL peptide of ovalbumin (RP 10611; GenScript) in the presence of brefeldin A (00-4506-51; eBioscience) for 7 h. Fc receptors were blocked with CD16/CD32 monoclonal antibodies (14-0161-82; eBioscience). Subsequently, intracellular IFNγ (554413; BD) as well as IFNγ isotype control (554686; BD) staining was performed using Intracellular Cytokine Detection Kits (BD Bioscience) according to the manufacturer’s instructions before data acquisition on a flow cytometer. Human tissue samples This is an experimental laboratory study performed on human tissue samples. All studies were performed in accordance with the Declaration of Helsinki. Blood from healthy individuals was obtained after leukapheresis, and normal skin was obtained from healthy individuals undergoing cosmetic surgery procedures. Lesional skin from patients with psoriasis was obtained from patients seen at the Brigham and Women’s Hospital or at Rockefeller University. All tissues were collected with informed consent (where applicable) and with prior approval from the Partners or Rockefeller Institutional Review Boards. Skin tissue was extensively minced and then incubated for 2 h at 37 °C in RPMI-1640 containing 0.2% collagenase type I (Invitrogen) and 30 Kunitz units per ml DNase I (Sigma Aldrich). Subsequently, cells were collected by filtering the collagenase-treated tissue through a 40-μm cell strainer (Fisher Scientific) followed by two washes with culture medium to remove any residual enzyme. Cells were stained with directly conjugated antibodies and analysed by flow cytometry. Gate strategy: T N cells, CD45RA + CD45RO − CD3 + CD8 + CD62L + ; T CM cells, CD45RO + CD3 + CD8 + CD62L + CCR7 + ; T EM cells, CD45RO + CD3 + CD8 + CD62L − CCR7 − ; T RM cells, CD45RO + CD3 + CD8 + CD62L − CD69 + . For immunofluorescent staining, skin tissues were embedded in OCT and frozen in liquid nitrogen immediately after surgery. Statistical analysis Comparisons between two groups were calculated using Student’s t -test (two tailed). Comparisons between more than two groups were calculated with one-way analysis of variance (ANOVA) followed by Bonferroni’s multiple comparison tests. Two-way ANOVA with Holm–Bonferroni post hoc analysis was used to compare weight loss between groups and log-rank (Mantel–Cox) test was used for survival curves. P < 0.05 was considered statistically significant. Data availability The microarray data that support the findings of this study are available in the Gene Expression Omnibus (accession number GSE79805 ); and Source Data are provided with the online version of the paper. Accession codes Primary accessions Gene Expression Omnibus GSE79805

The capacity for memory isn't exclusive to the brain. The immune system, with its sprawling network of diverse cell types, can recall the pathogens it meets, helping it to swiftly neutralize those intruders upon future encounters. For the last several years, BWH's Thomas Kupper, MD, Chair of the Department of Dermatology, and his colleagues have been studying a unique kind of immune memory cell, known as a tissue-resident memory T (TRM) cell. Discovered more than 10 years ago by Rachael Clark, MD, PhD (also in BWH Dermatology), these cells live in peripheral tissues, like the skin, gut and joints, and are thought to be a key source of the immune system's protective memory. Although much remains to be learned about the biology of these specialized memory T-cells, dysfunctional TRM cells are strongly implicated in autoimmune diseases, such as psoriasis, rheumatoid arthritis, inflammatory bowel disease and other conditions. To uncover the key genetic signals that guide the development of TRM cells, Kupper and his colleagues, led by postdoctoral fellow Youdong Pan, PhD, took an unbiased approach. They measured the level of gene activity for every gene in the genome as the cells developed in mice. That led the team to a remarkable finding, reported in a recent issue of the journal Nature: genes involved in lipid metabolism are highly active in TRM cells, roughly 20- to 30-fold more active than in other types of T-cells. Among these genes are two key mediators of lipid uptake, fatty-acid-binding proteins 4 and 5 (Fabp4 and Fabp5). Kupper and his colleagues teamed up with Gökhan Hotamisligil, PhD, an expert in lipid biology and metabolism at the Harvard T. H. Chan School of Public Health, to further dissect the roles of Fabp4 and Fabp5 in TRM cells. They found that TRM cells that lack both genes show a striking defect in their ability to import lipids. (Cells lacking just one of the genes are unaffected, likely because the two genes are highly related and perform overlapping functions.) Moreover, these Fabp4- and Fabp5-deficient TRM cells are significantly compromised both in their ability to protect against infection and their long-term survival in peripheral tissues. Based on his team's recent work, Kupper says the picture that is emerging of TRM cells highlights a unique dependence on fatty acids and other lipids as an energy source. Other types of T-cells can also metabolize lipids, but they cannot take them up from the environment, as TRM cells can. This could become an important Achilles' heel for investigators to target in the future. Drugs aimed at inhibiting lipid uptake could enable the selective removal of TRM cells from tissues, while leaving other types of T-cells intact. Current therapies for autoimmune disease are fairly broad in their activity—quieting down many types of immune cells, including TRM cells. But they work transiently, likely because TRM cells remain in place. "I think the real potential pay off of this story is to try and use this new information therapeutically," said Kupper. "While there are treatments for autoimmune diseases that impact pathogenic tissue-resident memory T-cells, none are able to effectively remove the cells from tissues. We've identified the first plausible mechanism for doing just that."

10.1038/nature21379

Physics

Teaching machines to spot essential information in physical systems

Koch-Janusz M, Ringel Z. Mutual information, neural networks and the renormalization group. Nature Physics, (2018). DOI: 10.1038/s41567-018-0081-4 Journal information: Nature Physics

http://dx.doi.org/10.1038/s41567-018-0081-4

https://phys.org/news/2018-03-machines-essential-physical.html

Abstract Physical systems differing in their microscopic details often display strikingly similar behaviour when probed at macroscopic scales. Those universal properties, largely determining their physical characteristics, are revealed by the powerful renormalization group (RG) procedure, which systematically retains ‘slow’ degrees of freedom and integrates out the rest. However, the important degrees of freedom may be difficult to identify. Here we demonstrate a machine-learning algorithm capable of identifying the relevant degrees of freedom and executing RG steps iteratively without any prior knowledge about the system. We introduce an artificial neural network based on a model-independent, information-theoretic characterization of a real-space RG procedure, which performs this task. We apply the algorithm to classical statistical physics problems in one and two dimensions. We demonstrate RG flow and extract the Ising critical exponent. Our results demonstrate that machine-learning techniques can extract abstract physical concepts and consequently become an integral part of theory- and model-building. Main Machine learning has been captivating public attention lately due to groundbreaking advances in automated translation, image and speech recognition 1 , game-playing 2 and achieving super-human performance in tasks in which humans excelled while more traditional algorithmic approaches struggled 3 . The applications of those techniques in physics are very recent, initially leveraging the trademark prowess of machine learning in classification and pattern recognition and applying them to classify phases of matter 4 , 5 , 6 , 7 , 8 , study amorphous materials 9 , 10 , or exploiting the neural networks’ potential as efficient nonlinear approximators of arbitrary functions 11 , 12 to introduce a new numerical simulation method for quantum systems 13 , 14 . However, the exciting possibility of employing machine learning not as a numerical simulator, or a hypothesis tester, but as an integral part of the physical reasoning process is still largely unexplored and, given the staggering pace of progress in the field of artificial intelligence, of fundamental importance and promise. The renormalization group (RG) approach has been one of the conceptually most profound tools of theoretical physics since its inception. It underlies the seminal work on critical phenomena 15 , and the discovery of asymptotic freedom in quantum chromodynamics 16 , and of the Kosterlitz–Thouless phase transition 17 , 18 . The RG is not a monolith, but rather a conceptual framework comprising different techniques: real-space RG 19 , functional RG 20 and density matrix RG 21 , among others. While all of those schemes differ quite substantially in their details, style and applicability, there is an underlying physical intuition that encompasses all of them—the essence of RG lies in identifying the ‘relevant’ degrees of freedom and integrating out the ‘irrelevant’ ones iteratively, thereby arriving at a universal, low-energy effective theory. However potent the RG idea, those relevant degrees of freedom need to be identified first 22 , 23 . This is often a challenging conceptual step, particularly for strongly interacting systems, and may involve a sequence of mathematical mappings to models, whose behaviour is better understood 24 , 25 . Here we introduce an artificial neural network algorithm iteratively identifying the physically relevant degrees of freedom in a spatial region and performing an RG coarse-graining step. The input data are samples of the system configurations drawn from a Boltzmann distribution; no further knowledge about the microscopic details of the system is provided. The internal parameters of the network, which ultimately encode the degrees of freedom of interest at each step, are optimized (‘learned’, in neural network parlance) by a training algorithm based on evaluating real-space mutual information (RSMI) between spatially separated regions. We validate our approach by studying the Ising and dimer models of classical statistical physics in two dimensions. We obtain the RG flow and extract the Ising critical exponent. The robustness of the RSMI algorithm to physically irrelevant noise is demonstrated. The identification of the important degrees of freedom, and the ability to execute a real-space RG procedure 19 , has not only quantitative but also conceptual significance: it allows one to gain insights into the correct way of thinking about the problem at hand, raising the prospect that machine-learning techniques may augment the scientific inquiry in a fundamental fashion. The RSMI algorithm Before going into more detail, let us provide a bird’s eye view of our method and results. We begin by phrasing the problem in probabilistic/information-theoretic terms, a language also used in refs 26 , 27 , 28 , 29 , 30 . To this end, we consider a small ‘visible’ spatial area \({\mathscr{V}}\) , which together with its environment \({\mathscr{E}}\) forms the system \({\mathscr{X}}\) , and we define a particular conditional probability distribution \({P}_{{\rm{\Lambda }}}({\mathscr{H}}| {\mathscr{V}})\) , which describes how the relevant degrees of freedom \({\mathscr{H}}\) (‘dubbed hiddens’) in \({\mathscr{V}}\) depend on both \({\mathscr{V}}\) and \({\mathscr{E}}\) . We then show that the sought-after conditional probability distribution is found by an algorithm maximizing an information-theoretic quantity, the mutual information, and that this algorithm lends itself to a natural implementation using artificial neural networks. We describe how RG is practically performed by coarse-graining with respect to \({P}_{{\rm{\Lambda }}}({\mathscr{H}}| {\mathscr{V}})\) and iterating the procedure. Finally, we provide a verification of our claims by considering two paradigmatic models of statistical physics: the Ising model—for which the RG procedure yields the famous Kadanoff block spins—and the dimer model, whose relevant degrees of freedom are much less trivial. We reconstruct the RG flow of the Ising model and extract the critical exponent. Consider then a classical system of local degrees of freedom \({\mathscr{X}}=\left\{{x}_{1},\ldots ,{x}_{N}\right\}\equiv \left\{{x}_{i}\right\}\) , defined by a Hamiltonian energy function H ({ x i }) and associated statistical probabilities \(P({\mathscr{X}})\propto {{\rm{e}}}^{-\beta {\rm{H}}\left(\left\{{x}_{i}\right\}\right)}\) , where β is the inverse temperature. Alternatively (and sufficiently for our purposes), the system is given by Monte Carlo samples of the equilibrium distribution \(P({\mathscr{X}})\) . We denote a small spatial region of interest by \({\mathscr{V}}\equiv \left\{{v}_{i}\right\}\) and the remainder of the system by \({\mathscr{E}}\equiv \left\{{e}_{i}\right\}\) , so that \({\mathscr{X}}=({\mathscr{V}},{\mathscr{E}})\) . We adopt a probabilistic point of view, and treat \({\mathscr{X}},{\mathscr{E}}\) and so on as random variables. Our goal is to extract the relevant degrees of freedom \({\mathscr{H}}\) from \({\mathscr{V}}\) . ‘Relevance’ is understood here in the following way: the degrees of freedom that RG captures govern the long-distance behaviour of the theory, and therefore the experimentally measurable physical properties; they carry the most information about the system at large, as opposed to local fluctuations. We thus formally define the random variable \({\mathscr{H}}\) as a composite function of degrees of freedom in \({\mathscr{V}}\) maximizing the ‘mutual information’ between \({\mathscr{H}}\) and the environment \({\mathscr{E}}\) . This definition, as we discuss in the Supplementary Information , is related to the requirement that the effective coarse-grained Hamiltonian be compact and short-ranged, which is a condition any successful standard RG scheme should satisfy. As we also show, it is supported by numerical results. Mutual information, denoted by I λ , measures the total amount of information about one random variable contained in the other 9 , 10 , 31 (thus, it is more general than correlation coefficients). It is given in our setting by: $${I}_{{\rm{\Lambda }}}({\mathscr{H}}:{\mathscr{E}})=\sum _{{\mathscr{H}},{\mathscr{E}}}{P}_{{\rm{\Lambda }}}({\mathscr{E}},{\mathscr{H}}){\rm{log}}\left(\frac{{P}_{{\rm{\Lambda }}}({\mathscr{E}},{\mathscr{H}})}{{P}_{{\rm{\Lambda }}}({\mathscr{H}})P({\mathscr{E}})}\right)$$ (1) The unknown distribution \({P}_{{\rm{\Lambda }}}({\mathscr{E}},{\mathscr{H}})\) and its marginalization \({P}_{{\rm{\Lambda }}}({\mathscr{H}})\) , depending on a set of parameters Λ (which we keep generic at this point), are functions of \(P({\mathscr{V}},{\mathscr{E}})\) and of \({P}_{{\rm{\Lambda }}}({\mathscr{H}}| {\mathscr{V}})\) , which is the central object of interest. Finding \({P}_{{\rm{\Lambda }}}({\mathscr{H}}| {\mathscr{V}})\) that maximizes I Λ under certain constraints is a well-posed mathematical question and has a formal solution 32 . However, since the space of probability distributions grows exponentially with the number of local degrees of freedom, it is, in practice, impossible to use without further assumptions for any but the smallest physical systems. Our approach is to exploit the remarkable dimensionality reduction properties of artificial neural networks 11 . We use restricted Boltzmann machines (RBMs), a class of probabilistic networks well adapted to approximating arbitrary data probability distributions. An RBM is composed of two layers of nodes, the ‘visible’ layer, corresponding to local degrees of freedom in our setting, and a ‘hidden’ layer. The interactions between the layers are defined by an energy function \({E}_{{\rm{\Theta }}}\equiv {E}_{a,b,\theta }({\mathscr{V}},{\mathscr{H}})\) = − \({\sum }_{j}{b}_{j}{h}_{j}\) − \({\sum }_{i}{a}_{i}{v}_{i}\) − \({\sum }_{ij}{v}_{i}{\theta }_{ij}{h}_{j}\) , such that the joint probability distribution for a particular configuration of visible and hidden degrees of freedom is given by a Boltzmann weight: $${P}_{{\rm{\Theta }}}({\mathscr{V}},{\mathscr{H}})=\frac{1}{{\mathscr{Z}}}{{\rm{e}}}^{-{E}_{a,b,\theta }({\mathscr{V}},{\mathscr{H}})}$$ (2) where \({\mathscr{Z}}\) is the normalization. The goal of the network training is to find parameters θ ij (‘weights’ or ‘filters’) and a i , b i optimizing a chosen objective function. Three distinct RBMs are used. Two are trained as efficient approximators of the probability distributions \(P({\mathscr{V}},{\mathscr{E}})\) and \(P({\mathscr{V}})\) , using the celebrated contrastive divergence (CD) algorithm 33 . Their trained parameters are used by the third network (see Fig. 1b ), which has a different objective: to find \({P}_{{\rm{\Lambda }}}({\mathscr{H}}| {\mathscr{V}})\) maximizing I Λ . To the end we introduce the real-space mutual information (RSMI) network, whose architecture is shown in Fig. 1a . The hidden units of RSMI correspond to coarse-grained variables \({\mathscr{H}}\) . Fig. 1: The RSMI algorithm. a , The RSMI neural network architecture. The hidden layer \({\mathscr{H}}\) is directly coupled to the visible layer \({\mathcal{V}}\) via the weights \({\lambda }_{i}^{j}\) (red arrows). However, the training algorithm for the weights estimates mutual information between \({\mathscr{H}}\) and the environment \({\mathcal{E}}\) . The buffer \({\mathscr{B}}\) is introduced to filter out local correlations within \({\mathcal{V}}\) (see Supplementary Information ). b , The workflow of the algorithm. The CD-algorithm-trained RBMs learn to approximate probability distributions \(P({\mathcal{V}},{\mathcal{E}})\) and \(P({\mathcal{V}})\) . Their final parameters, denoted collectively by \({{\rm{\Theta }}}^{({\mathcal{V}},{\mathcal{E}})}\) and \({{\rm{\Theta }}}^{({\mathcal{V}})}\) , are inputs for the main RSMI network learning to extract \({P}_{{\rm{\Lambda }}}({\mathscr{H}}| {\mathcal{V}})\) by maximizing I Λ . The final weights \({\lambda }_{i}^{j}\) of the RSMI network identify the relevant degrees of freedom. They are shown in Figs. 2 and 4 for Ising and dimer problems. Full size image The parameters \({\rm{\Lambda }}=({a}_{i},{b}_{j},{\lambda }_{i}^{j})\) of the RSMI network are trained by an iterative procedure. At each iteration, a Monte Carlo estimate of function \({I}_{{\rm{\Lambda }}}({\mathscr{H}}:{\mathscr{E}})\) and its gradients is performed for the current values of parameters Λ . The gradients are then used to improve the values of weights in the next step, using a stochastic gradient descent procedure. The trained weights Λ define the probability \({P}_{{\rm{\Lambda }}}({\mathscr{H}}| {\mathscr{V}})\) of a Boltzmann form, which is used to generate MC samples of the coarse-grained system. Those, in turn, become input to the next iteration of the RSMI algorithm. The estimates of mutual information, weights of the trained RBMs and sets of generated MC samples at every RG step can be used to extract quantitative information about the system in the form of correlation functions, critical exponents and so on, as we show below and in the Supplementary Information . We also emphasize that the parameters Λ identifying relevant degrees of freedom are re-computed at every RG step. This potentially allows RSMI to capture the evolution of the degrees of freedom along the RG flow 34 . Validation To validate our approach, we consider two important classical models of statistical physics: the Ising model, whose coarse-grained degrees of freedom resemble the original ones, and the fully packed dimer model, where they are entirely different. The Ising Hamiltonian on a two-dimensional (2D) square lattice is: $${H}_{{\rm{I}}}=\sum _{\left\langle i,j\right\rangle }{s}_{i}{s}_{j}$$ (3) with s i = ±1 and the summation over nearest neighbours. Real-space RG of the Ising model proceeds by the block-spin construction 19 , whereby each 2 × 2 block of spins is coarse-grained into a single effective spin, whose orientation is decided by a ‘majority’ rule. The results of the RSMI algorithm trained on Ising model samples are shown in Fig. 2 . We vary the number of both hidden neurons N h and the visible units, which are arranged in a 2D area \({\mathscr{V}}\) of size L × L (see Fig. 1a ). For a four-spin area, the network indeed rediscovers the famous Kadanoff block-spin: Fig. 2a shows a single hidden unit coupling uniformly to four visible spins (that is, the orientation of the hidden unit is decided by the average magnetization in the area). Figure 2b is a trivial but important sanity check: given four hidden units to extract relevant degrees of freedom from an area of four spins, the networks couples each hidden unit to a different spin, as expected. In the Supplementary Information we also compare the weights for areas \({\mathscr{V}}\) of different size, which are generalizations of the Kadanoff procedure to larger blocks. Fig. 2: The weights of the RSMI network trained on the Ising model. Visualization of the weights of the RSMI network trained on the Ising model for a visibile area \(\mathscr{V}\) of 2 × 2 spins. The ANN couples strongly to areas with large absolute value of the weights. a , The weights for N h = 1 hidden neuron: the ANN discovers Kadanoff blocking. b , The weights for N h = 4 hidden neurons: each neuron tracks one original spin. Full size image We next study the dimer model, given by an entropy-only partition function, which counts the number of dimer coverings of the lattice (that is, subsets of edges such that every vertex is the endpoint of exactly one edge). Figure 3a shows sample dimer configurations (and additional spin degrees of freedom added to generate noise). This deceptively simple description hides non-trivial physics 35 and correspondingly, the RG procedure for the dimer model is more subtle, since—in contrast to the Ising case—the correct degrees of freedom to perform RG on are not dimers, but rather look like effective local electric fields. This is revealed by a mathematical mapping to a ‘height field’ h (see Fig. 3a,b and ref. 36 ), whose gradients behave like electric fields. The continuum limit of the dimer model is given by the following action: $${S}_{{\rm{dim}}}[h]=\int {{\rm{d}}}^{2}x{\left(\nabla h({\bf{x}})\right)}^{2}\equiv \int {{\rm{d}}}^{2}x{{\bf{E}}}^{2}({\bf{x}})$$ (4) and therefore the coarse-grained degrees of freedom are low-momentum (Fourier) components of the electrical fields E x , E y in the x and y directions. They correspond to ‘staggered’ dimer configurations shown in Fig. 3a . Fig. 3: The dimer model. a , Two sample dimer configurations (blue links), corresponding to the E y and E x electrical fields, respectively. The coupled pairs of additional spin degrees of freedom on vertices and faces of the lattice (wiggly lines) are decoupled from the dimers and from each other. Their fluctuations constitute irrelevant noise. b , An example of mapping the dimer model to local electric fields. The so-called staggered configuration on the left maps to uniform non-vanishing field in the vertical direction: \(\left\langle {E}_{y}\right\rangle \ne 0\) . The ‘columnar’ configuration on the right produces both E x and E y that are zero on average (see ref. 36 for details of the mapping). Full size image Remarkably, the RSMI algorithm extracts the local electric fields from the dimer model samples without any knowledge of those mappings. In Fig. 4 , the weights for N h = 2 and N h = 4 hidden neurons, for an 8 × 8 area (similar to Fig. 3a ), are shown: the pattern of large negative (blue) weights couples strongly to a dimer pattern corresponding to local uniform E y field (see left panels of Fig. 3a,b ). The large positive (yellow) weights select an identical pattern, translated by one link. The remaining neurons extract linear superpositions E x + E y or E x − E y of the fields. Fig. 4: The weights of the RSMI network trained on dimer model data. a , N h = 2 hidden neurons for a visible area \({\mathscr{V}}\) of 8 × 8 spins. The two filters recognize E y and E x + E y electrical fields, respectively (compare with dimer patterns in Fig. 3a ). b , The trained weights for N h = 4 hidden neurons. Full size image To demonstrate the robustness of the RSMI, we added physically irrelevant noise, forming nevertheless a pronounced pattern, which we model by additional spin degrees of freedom, strongly coupled (ferromagnetically) in pairs (wiggly lines in Fig. 3a ). Decoupled from the dimers, and from other pairs, they form a trivial system, whose fluctuations are short-range noise on top of the dimer model. Vanishing weights (green in Fig. 4a,b ) on sites where pairs of spins reside prove that RSMI discards their fluctuations as irrelevant for long-range physics, despite their regular pattern. Notably, the filters obtained using our approach for the dimer model, which match the analytical expectation, are orthogonal to those obtained using Kullback–Leibler divergence. As expanded on in the Supplementary Information , this shows that standard RBMs minimizing the Kullback–Leibler divergence do not generally perform RG, thereby contradicting prior claims 37 . Finally, we demonstrate that by iterating the RSMI algorithm the qualitative insights into the nature of relevant degrees of freedom give rise to quantitative results. To this end, we revisit the 2D Ising model that (contrary to the dimer model) exhibits a non-trivial critical point at the temperature \({T}_{{\rm{c}}}={({\rm{log}}(1+\sqrt{2})/2)}^{-1}\) , separating the paramagnetic and ferromagnetic phases. We generate Monte Carlo samples of the system of size 128 × 128 at values T around the critical point, and for each one we perform up to four RG steps, by computing the Λ filters using RSMI, coarse-graining the system with respect to those filters (effectively halving the linear dimensions) and reiterating the procedure. In addition to the set of Monte Carlo configurations for the coarse-grained system, estimates of mutual information as well as the filters of the CD-trained RBMs are generated and stored. The effective temperature T of the system at each RG step can be evaluated entirely intrinsically either from correlations or the mutual information, as discussed in the Supplementary Information . Using the RBM filters, spin–spin correlations (for instance, next-nearest neighbour) can be computed. By comparing these with known analytical results 38 , an additional cross-check of the effective temperature can be obtained. In Fig. 5 , the effective T is plotted against \({{\rm{log}}}_{{\rm{2}}}({\xi }_{{\rm{128}}}{\rm{/}}\xi )\) , where ξ and ξ 128 are the current and 128 × 128 systems’ correlation lengths, respectively (this has the meaning of an RG step for integer values). The RG flow of the 2D Ising model is recovered: systems starting with T < T c flow towards ever-decreasing T (that is, an ordered state), while the ones with T > T c flow towards a paramagnet. In fact, the position of the critical point can be estimated with 1% accuracy just from the divergent flow. Furthermore, we evaluate the correlation length exponent ν , defined by ξ ∝ τ − ν . Using the finite-size data collapse (see Supplementary Fig. 4 ), its value, equal to the negative slope, is estimated to be ν ≈ 1.0 ± 0.15, consistent with the exact analytical result ν = 1. Fig. 5: RG flow for the 2D Ising model. The temperature T (in units of T c ) as a function of the RG step for systems with initial Monte Carlo temperatures (denoted by lines of different colour) below and above T c . See Supplementary Information for details. Full size image Future directions Artificial neural networks based on RSMI optimization have proved capable of extracting complex information about physically relevant degrees of freedom and using it to perform a real-space RG procedure. The RSMI algorithm we propose allows for the study of the existence and location of critical points, and RG flow in their vicinity, as well as estimation of correlations functions, critical exponents and so on. This approach is an example of a new paradigm in applying machine learning in physics: the internal data representations discovered by suitably designed algorithms are not just technical means to an end, but instead are a clear reflection of the underlying structure of the physical system (see also ref. 39 ). Thus, in spite of their ‘black box’ reputation, the innards of such architectures may teach us fundamental lessons. This raises the prospect of employing machine learning in science in a collaborative fashion, exploiting the machines’ power to distil subtle information from vast data, and human creativity and background knowledge 40 . Numerous further research directions can be pursued. Most directly, equilibrium systems with less understood relevant degrees of freedom—for example , disordered and glassy systems—can be investigated 9 , 10 . The ability of the RSMI algorithm to re-compute the relevant degrees of freedom at every RG step potentially allows one to study their evolution along the (more complicated) RG flow 34 . Furthermore, although we studied classical systems, the extension to the quantum domain is possible via the quantum-to-classical mapping of Euclidean path integral formalism. A more detailed analysis of the mutual-information-based RG procedure may prove fruitful from a theoretical perspective. Finally, applications of RSMI beyond physics are possible, since it offers a neural network implementation of a variant of the information bottleneck method 32 , successful in compression and clustering analyses 41 ; it can also be used as a local-noise-filtering pre-training stage for other machine-learning algorithms. Data availability The data that support the plots within this paper and other findings of this study are available from the corresponding author upon request.

Two physicists at ETH Zurich and the Hebrew University of Jerusalem have developed a novel machine-learning algorithm that analyses large data sets describing a physical system and extract from them the essential information needed to understand the underlying physics. Over the past decade, machine learning has enabled groundbreaking advances in computer vision, speech recognition and translation. More recently, machine learning has also been applied to physics problems, typically for the classification of physical phases and the numerical simulation of ground states. Maciej Koch-Janusz, a researcher at the Institute for Theoretical Physics at ETH Zurich, Switzerland, and Zohar Ringel of the Hebrew University of Jerusalem, Israel, have now explored the exciting possibility of harnessing machine learning not as a numerical simulator or a "hypothesis tester," but as an integral part of the physical reasoning process. One important step in understanding a physical system consisting of a large number of entities—for example, the atoms making up a magnetic material—is to identify among the many degrees of freedom of the system those that are most relevant for its physical behaviour. This is traditionally a step that relies heavily on human intuition and experience. But now, Koch-Janusz and Ringel demonstrate a machine-learning algorithm based on an artificial neural network that is capable of doing just that, as they report in the journal Nature Physics. Their algorithm takes data about a physical system without any prior knowledge about it and extracts those degrees of freedom that are most relevant to describe the system. Technically speaking, the machine performs one of the crucial steps of one of the conceptually most profound tools of modern theoretical physics, the so-called renormalization group. The algorithm of Koch-Janusz and Ringel provides a qualitatively new approach: the internal data representations discovered by suitably designed machine-learning systems are often considered to be obscure, but the results yielded by their algorithm provide fundamental physical insight, reflecting the underlying structure of physical system. This raises the prospect of employing machine learning in science in a collaborative fashion, combining the power of machines to distil information from vast data sets with human creativity and background knowledge.

10.1038/s41567-018-0081-4

Medicine

Research delves into link between test anxiety and poor sleep

Nancy Hamilton et al, Test Anxiety and Poor Sleep: A Vicious Cycle, International Journal of Behavioral Medicine (2021). DOI: 10.1007/s12529-021-09973-1

http://dx.doi.org/10.1007/s12529-021-09973-1

https://medicalxpress.com/news/2021-04-delves-link-anxiety-poor.html

Abstract Background Test anxiety may be better thought of as a biopsychosocial process affecting academic performance during the days leading up to an exam, rather than a static appraisal of attitudes related to test taking. This was a passive observational study following students 2 days before a midterm exam and was designed to test the Sleep Anxiety Performance Process (SAPP) model in the context of a psychology statistics exam. Methods Undergraduates ( N = 167) enrolled in a statistics class, January–November 2015. Participants completed an electronic battery of measures and Sleep Mood Study Diaries (SMS) during the mornings, 2 days before a statistics exam. Instructors confirmed exam scores. Results A path model showed a reciprocal bi-directional relationship between Sleep Quality and restfulness (Q&R) and test anxiety 2 days before a scheduled exam, with test anxiety measured in the morning, before the exam predicting exam performance. Prior exam performance, being a non-native English speaker (ESL), and class performance motivation also predicted exam performance. Conclusions These data support the SAPP model’s premise that that sleep and anxiety feed one another, as a reciprocal process, that collectively impairs academic performance, with direct effects on academic performance, but with implications for overall student health. Access provided by MPDL Services gGmbH c/o Max Planck Digital Library Working on a manuscript? Avoid the common mistakes Introduction Test anxiety has been defined as the emotional, physiological, and behavioral responses in the anticipation of negative evaluation from an upcoming test or exam [ 1 ]. Between 10 and 40% of students experience test anxiety and students with higher levels of test anxiety perform more poorly on tests and have lower Grade Point Averages (GPA: the accumulated average of point values assigned to final letter grades, typically ranging from 1 to 4, with an A = 4 points, a B = 3, etc.) [ 2 ]. However, consequences of test anxiety extend beyond the classroom and are correlated with poor health behaviors, including dysregulated sleep patterns and poor sleep quality [ 3 ]. Given the well-known bi-directional relationship between sleep and anxiety [ 4 ], it would be important to understand the specific pathways by which anxiety and sleep predicts exam performance and potentially unfold over the course of the several days prior to an exam. Characteristics of the relationship between test-taking and anxiety have been well documented. Although trait anxiety is related to test anxiety and GPA [ 5 ], state and test-specific anxiety measured immediately before an exam have been shown to have stronger relationships to performance than do more distal measures of state anxiety [ 6 , 7 ]. In one particularly well-designed study, researchers assessed trait anxiety, state anxiety, and statistics specific anxiety. Measured 2 weeks before the exam, trait anxiety was related to statistics specific anxiety and state anxiety. However, performance was predicted by attitudes about math and interest in the class, statistics specific anxiety, and state anxiety measured proximal to the test [ 6 ]. Test anxiety varies across time [ 7 ], across classes, and attitudes about subject matter are important. In particular, statistics classes are associated with high levels of anxiety [ 1 , 8 ], especially for female students [ 9 ], and class performance has been shown to be driven both by attitudes about the specific class and anxiety [ 6 ]. Although most test anxiety studies have only examined the relationship between test anxiety and performance, extant research suggests that sleep is related to both academic performance and anxiety. For instance, sleep deprivation has been found to compromise a broad array of memory processes [ 10 ]. Average sleep duration, sleep habits, and chronotype have been found to predict GPA [ 11 , 12 , 13 , 14 ] and medical students’ exam scores [ 15 ]. In an illustrative study, end of the semester undergraduate course performance was predicted by prior academic performance, study habits, chronotype, sleep duration, and sleep quality [ 16 ]. In any study of exam anxiety, it would be important to measure both anxiety and sleep as sleep problems such as insomnia and anxiety are highly correlated [ 17 ]. State and trait anxiety have documented relationships to shorter total sleep time and disrupted sleep [ 18 , 19 ], longer sleep onset latencies, more Wake Time after Sleep Onset (WASO [ 20 ]), and there is evidence of a bidirectional relationship between poor sleep and anxiety [ 4 ]. A few studies have examined exam-related anxiety, sleep, and exam performance. Among college students, test anxiety was related to sleep duration and quality the night before an exam, but only test anxiety predicted actual exam performance [ 18 ]. Stress (a construct related to anxiety) was found to increase, and sleep quality (but not duration) was found to decrease in the week before an exam period [ 21 ]. Stress and sleep duration were related to academic performance among medical students [ 22 , 23 ]. During the month before a high-stakes exam, stress increased and sleep duration decreased; however, the individual relationship of stress and sleep disruption were not examined [ 22 ]. Collectively, these studies indicate that anxiety (and stress) and poor sleep (shorter duration and poor quality) go hand in hand. However, the temporal relationship between sleep disruption and anxiety has not yet been explicitly examined. The Present Study Test anxiety is a serious academic problem for college students. However, the bidirectional relationship between sleep and anxiety suggests that it would be important to examine the temporal relationship between sleep (duration and quality) and test anxiety as predictors of academic performance. Because there is an abundance of research documenting a reciprocal relationship between sleep problems such as insomnia and anxiety, we proposed that the effects of test anxiety on academic performance would unfold as a reciprocal process, across several days, as represented by the Sleep Anxiety Performance Process (SAPP) model. Moreover, given the relationship of sleep problems to decrements in memory, we predicted that sleep would mediate the relationship between anxiety and academic performance. In order to understand the temporal relationship between test-taking anxiety and sleep (duration and quality), we used electronic diaries to examine the trajectory of anxiety and sleep during two days before a statistics exam. However, we were also mindful that there were other relevant predictors of achievement and potential confounds of the relationships in question; thus, we controlled for prior exam performance, and individual difference variables including trait anxiety and symptoms of depression, enduring sleep patterns (chronotype and insomnia), and baseline sleep quality, as well as class performance motivation and study habits. After controlling for these individual differences, the SAPP model predicted that the effect of anxiety on exam performance would be mediated by sleep quality and sleep duration. Method Design and Participants Undergraduate students were recruited from seven sections of introductory statistics (STATS) classes at a large university in the Midwestern USA. Data were gathered from STATS classes taught by five different instructors, gathered during a single calendar year, two spring sections ( n = 40), one summer section ( n = 9), and five fall sections ( n = 118). A total of 314 students were offered participation in the study ( N = 167) consented to participate. Power was analyzed via simulation, where relationships between the core variables of varying strength were evaluated. Consistent with demographics of psychology majors at this university, 75% of participants were female; the majority were Caucasian (74%); 8% identified as African-American, 5% Hispanic, 8% Asian, and 5% identified themselves as having multiple ethnic affiliations. English was a second language for 9% of participants. STATS is a class for second year college students and ages ranged from 18 to 40 ( M = 20.65, SD = 2.83). Procedure A research assistant who explained the nature of the study recruited students from classrooms. Recruitment and baseline assessments varied according to the semester schedule, with study entry following the first exam. Interested STATS students linked to an online baseline questionnaire (Qualtrics survey software), signed an electronic consent form, provided their email addresses, and were sent Sleep, Mood, Study daily diaries (SMS diaries) at 6:00 a.m. each morning for the 6 days leading up to a target exam. Participants were asked to respond to time-stamped diaries as quickly as possible after waking. In an effort to minimize response burden and capture the strongest anxiety-performance relationship [ 6 ], test anxiety was measured only twice: the morning before the exam and the morning of the exam. After completing the final SMS diary, instructors confirmed students’ self-reported exam scores. At that time students were also provided with a debriefing statement. Each student could receive up to 15 total extra credit points. The protocol for this study was approved by the University of Kansas Institutional Review Board (IRB). Measures Although all questionnaires were administered in an electronic format, psychometric properties (i.e., internal consistency and construct validity) remained similar to the paper and pencil versions of the questionnaires. Baseline Battery Student characteristics in the battery included items to identify participants’ age, sex, ethnicity, year in school, employment, school credits, previous exam score, and class performance motivation (“How important is it for you to do well in this class?”) (4-point Likert items). The Pittsburgh Sleep Quality Inventory (PSQI) [ 24 ] was used as an omnibus measure of sleep quality (Cronbach’s α = 0.66). PSQI Total Score range: 0–21. A score of 5 or greater indicates clinically relevant sleep disturbance. The Morning-Eveningness Questionnaire (MEQ) was used to measure chronotype. MEQ contains 19 Likert items (Cronbach’s α = 0.76-0.79 [ 25 ]:). Higher scores indicate greater “morningness” (70–86: definite; 59–69: moderate), lower scores indicate “eveningness” (16–20: definite; 31–41: moderate), and central scores indicate “intermediate” tendencies (42–58). The Insomnia Severity Index (ISI) was used to measure symptoms of insomnia (Cronbach’s α = 0.90–0.91 [ 26 ]. The ISI consists of 7 Likert items that measure clinical insomnia. Higher total scores denote higher severity of clinical insomnia, clinical cut-offs: sub-threshold insomnia (8–14), moderately severe insomnia (15–21), and severe clinical insomnia (22–28). The Beck Depression Inventory-II (BDI) was used to measure symptoms of depression (Chronbach’s α = 0.86–0.95; [ 27 ]). Scores exceeding 10–12 indicate risk for depression. The State Trait Anxiety Inventory, Trait form (STAI-T) was used to measure symptom of anxiety [ 28 ]. The STAI-T uses 20, 4-point Likert items, with higher scores indicating greater anxiety (Cronbach’s α = 0.86–0.95). Recent administrations of the STAI-T with college students show average scores ranging from 37 to 41.51[ 29 , 30 , 31 ]. Sleep Mood Study Diaries Sleep was measured each morning using a sleep diary adapted from the Consensus Sleep Diary (Qs 1–10) [ 32 ]. Questions asked for bedtime, time it took to fall asleep, number of awakenings, number of awakenings after sleep onset (WASO), time out of bed, sleep duration. Sleep Quality (Q), and Restfulness (R) (7-point Likert items) were highly correlated (0.81, p < 0.01) and thus were averaged (Q&R). From this measure, the following variables were calculated: sleep onset latency (SOL), total sleep time (TST), and sleep efficiency (SE). Study behavior was measured by asking the following questions: “How many minutes did you study for your statistics class yesterday?” “Did you go to (scheduled) class”? Test anxiety was assessed on the mornings before (day 5) and of the exam (day 6). Students rated test-taking anxiety using a version of the State Trait Anxiety Inventory-State Version [ 28 ], with an altered instruction set to capture anxiety about the upcoming exam (STAI-test). The STAI-test was selected because it is brief (20 4-point Likert items), reactive to changes in anxiety across successive days, but with high observed internal consistency using this instruction set (Cronbach’s α = 0.85) that is similar to established instruction sets (Cronbach’s α = 0.86–0.95). Although sleep diary and study behaviors were gathered for six mornings, STAI-Test was measured only twice, in close proximity to the test in order to minimize response burden. Procedure A research assistant who explained the nature of the study recruited students from classrooms. Recruitment and baseline assessments varied according to the semester schedule, with study entry following the first exam. Interested STATS students linked to an online baseline questionnaire (Qualtrics survey software), signed an electronic consent form, provided their email addresses, and were sent Sleep, Mood, Study daily diaries (SMS diaries) at 6:00 a.m. each morning for the 6 days leading up to a target exam. Participants were asked to respond to time-stamped diaries as quickly as possible after waking. In an effort to minimize response burden and capture the strongest anxiety-performance relationship [ 6 ], test anxiety was measured only twice: the morning before the exam and the morning of the exam. After completing the final SMS diary, instructors confirmed students’ self-reported exam scores. At that time, students were also provided with a debriefing statement. Each student could receive up to 15 total extra credit points. The protocol for this study was approved by the University of Kansas Institutional Review Board (IRB). Method of Analysis Data were analyzed in R using the “lavaan” package. The small amount of missing data was addressed using a saturated correlates model for full-information maximum likelihood (FIML). Because students were nested within classroom and semester, we investigated whether there it would be necessary to use a multilevel SEM model. In no instance did analyses reveal evidence of systemic between-cluster differences (composed either of instructor or semester). Simulation data were used to estimate power [ 33 ]. Given the plausible range of effect sizes and range of probable covariates; N = 150 conferred sufficient power to detect a moderate sized simple effect of the primary predictors (i.e., Test Anxiety, Sleep Q&R, and TST). The total fraction of missing information in the data used for the final analysis model was 7.8%, with 11–14% for individual SMS diary items. This fraction of missing data is small and within the range where FIML can recover lost information and point estimates on which to base unbiased parameters estimates [ 34 ]. Results Descriptive Data Table 1 includes descriptive data and correlations among individual difference and demographic data. BDI scores were below clinical range, and STAI-T scores only slightly higher than recently reported college student samples [ 30 , 35 , 36 ]. However, there were indications of sleep problems, with ISI scores in the moderately severe range and PSQI total scores that exceeded the clinical cut off for sleep disturbance. On average, students fell into the “intermediate” range in terms of morning-evening tendencies. As shown in Table 1 , correlations among individual difference variables were in expected directions (e.g., STAI-T, BDI, PSQI, and ISI were all positively correlated). Table 1 Descriptive data Full size table Primary Hypotheses Analyses were constructed in a sequential manner, meaning that a simpler model was constructed first, that explained variability in final exam scores using individual difference variables (i.e., BDI, STAI-T, PSQI, ISI, MEQ chronotype, age, sex, race/ethnicity, and English as a Second Language (ESL), class performance motivation, first exam score). Non-significant predictors were pruned from the model based on Wald tests for predictor significance (seeTable 2 for detailed output). Footnote 1 A model constraining non-significant predictor slopes to zero was then analyzed and was determined to have excellent fit (CFI = 0.996, TLI = 0.994). Table 2 Baseline model Full size table The modified baseline model with individual differences was then incorporated into the hypothesized path model, such that theorized sleep-anxiety-performance processes could be evaluated while simultaneously controlling for relevant individual difference variables. Variables pruned from the individual differences model that nevertheless could potentially impact steps in the sleep-anxiety-performance process were also included. The final model included the following individual difference variables: Age, Sex, Race/Ethnicity, ESL, First Exam Grade, and Class Performance Motivation; Class time (AM/PM); and ISI and Chronotype. To this, we added reports from the SMS diaries reflecting reports from night four (exam-2 days) and five (exam-1 day): Minutes studied, Sleep Q&R, WASO, TST, and STAI-Test (exam-1 only), and on the day of the exam STAI-test and the Final Exam Grade. The final model largely supports the hypothesized conceptualization of the SAPP model and fit the data well (CFI = 0.963, TLI = 0.935, RMSEA 0.029). Figure 1 can be broken down into its component features: Higher Q&R preceded lower levels of STAI-test for both the night prior to the exam ( χ 2 [1] = 7.17, p = 0.007) as well as two nights before ( χ 2 [1] = 6.789, p = 0.009). Increasing levels of STAI-test had an adverse effect on Q&R (for STAI-test, the day prior to the exam, χ 2 [1] = 5.172, p = 0.023) as well as exam performance itself (for STAI-test on exam day, χ 2 [1] = 16.15524, p ≤ 0.001). In fact, after controlling for prior exam performance and other predictors, test anxiety explained approximately 9.8% of variability in test performance, increasing the multiple R -squared for test performance from ~ 30 to ~ 40% significantly improving the fit of the model, p < 0.001. Although Sleep Q&R predicted test anxiety, our results did not indicate that sleep was directly related exam performance. In fact, none of the sleep parameters (e.g. Q&R, TST, SOL, SE) had significant direct or lagged effects on the final exam score. Fig. 1 The Sleep Anxiety Performance Process Model (SAPP). WASO number of awakenings after sleep onset, Q&R Sleep Quality and Restfulness, TST Total Sleep Time, STAI-Test, State Test Anxiety, ESL English as a Second Language, solid lines = positive path; dashed lines = negative path. * p < 0.05, ** p < 0.01, *** p < 0.001 Full size image Few individual difference variables predicted exam scores. Neither symptoms of psychopathology, nor general demographic differences, nor general sleep dysfunction as measured by the ISI, chronotype, nor baseline sleep quality as measured by the PSQI mediated the relationship between sleep quality and anxiety, nor had any direct effects on exam performance. Individual difference variables that did predict exam score include first exam score, class performance motivation, and ESL. Each of these variables significantly predicted final exam scores in expected ways: Class performance motivation and first exam scores were positive indicators of Final Exam scores, whereas ESL students had significantly lower final exam scores. Risk of Failure As a follow-up analysis, we examined whether significant predictors of exam scores could be used to predict whether students passed or failed the exam. Using logistic regression, first exam grade, STAI-Test, ESL status, and Sleep Q&R were entered into an equation to predict whether students passed (score ≥ 70) or failed the exam (score < 70). Nearly one-third of the students failed their exam (27.86%). Single-unit increases in first exam scores were associated with greater odds of passing the exam (1.067, Wald 11.956, p < 0.05) and higher STAI-test scores reduced the chance of a passing score on the exam (0.943, Wald = 7.932, p < 0.05), as did speaking a first language other than English (14.017, Wald = 4.043, p < 0.05). Although this model explained a significant portion of the variance, Nagelkerke R 2 = 0.33, and correctly classified 77% of students, it proved less likely to identify true failures 44.1% than true passers 89%. Discussion Universities are struggling with the high rate of drop-out. Almost 40% of freshmen fail to return to their original college or university after the first year, many for academic reasons [ 31 ]. One significant contributor to academic failure is anxiety about academic performance, specifically test-taking anxiety. Students who are at increased risk for ineffective or maladaptive coping with the stressors of college may be even more vulnerable to the negative effects of anxiety on their academic performance and health-behaviors including sleep [ 3 ]. Consistent with this observation, this study found that students who experienced even moderate amounts of anxiety had lower test scores. Rather than having separate effects on exam performance, over the course of two days and nights, poor sleep quality predicted increased anxiety, which in turn perpetuated disrupted sleep, which then predicted further increases in anxiety and ultimately poor academic performance. Although the snapshot of the sleep-anxiety relationship captured here was not consistent with the specific a-priori predicted relationship (poor sleep mediating the relationship between test anxiety and exam performance), the model is consistent with the overall precept of the SAPP model, that sleep and anxiety feed one another, creating a vicious cycle that collectively impairs academic performance. This was not a clinical sample. Nevertheless, test anxiety had a significant effect on test performance. A one-unit increase of test anxiety on the day of the test was associated with a 0.388-point lower test score. For students who were one standard deviation above the mean for anxiety, the effect size equates to a nearly 5-point lower (4.96) exam score. This could mean the difference between earning a D vs. a C-. Although the relationship between test anxiety and performance is fairly well understood, what was new in this study was documentation of the reciprocal relationship between sleep and anxiety. In particular, test anxiety and poor sleep quality appeared to operate as a positive feedback loop, with poor sleep disinhibiting anxiety, and those feelings and thoughts, in turn, disrupting sleep. Instead of poor sleep mediating the relationship between test anxiety and exam performance, anxiety, and sleep appeared to form a feed forward process, with test anxiety ultimately predicting exam performance. Although poor sleep predicted test anxiety, none of the other sleep parameters (e.g., SOL, WASO, TST, SE) were either directly related to academic performance or test anxiety. Instead, in this sample of young adults, the effects of WASO and TST appeared to have been captured in the appraisal of poor sleep quality and restorativeness. Had we chosen a different focus of this model, we could have modeled WASO and TST as mediators of the relationship of poor sleep quality to test anxiety. Other predictors of exam performance were primarily related to academic history. Prior exam performance was a strong predictor of future exam performance. However, this is not a “clean” academic covariate. Similar to what we observed during the 2 days prior to the statistics exam, we would expect that previous exams were also affected by the cyclical relationship between sleep and anxiety. Dispositional measures (e.g., Trait anxiety, depression, insomnia symptoms, circadian rhythms) were largely unrelated to final exam performance, underscoring the fact that this was a largely mentally healthy sample and that the effects of anxiety and sleep on performance were situational and not driven by general response biases or trait level effects. Non-native speakers had significantly lower scores than native speakers and were 14 times more likely to fail than a native speaker. It should be noted that this was a not a planned analysis, only a small proportion of our sample (roughly 10%) were non-native speakers, and the student’s level of English fluency was not measured. These caveats aside, the results of this study suggest the need for additional interventions to support non-native English speakers. It is worth commenting on the relationship of motivation and study habits to academic performance. Perhaps it is not surprising that class performance motivation to do well in a class was related to actual performance, but was surprising that study habits did not relate to exam score. In fact, the only relationship of study time was to shorter sleep time the night before the exam. Follow-up exploratory analyses on study time showed that academic performance was not correlated with total minutes studied, average or variability, minutes of study during any day during the week, nor a curvilinear function. Examination of the data revealed no obvious sources of bias in reporting of study times. Although surprising, this finding is consistent with a body of work that shows little relationship between homework and achievement in elementary and secondary education [ 29 ]. As this is a single study and this was not a hypothesis driven finding, replication would be critical. Test Anxiety and Broader Implications for Health It is tempting to conceptualize test anxiety as a psychological state that has effects that are limited to academic performance and success. However, in other studies, highly test anxious students were found to have higher levels of stress, increased levels of maladaptive health behaviors such as increased caffeine intake, smoking, and lower levels of adaptive behaviors such as healthy eating, physical activity, and hygienic sleep [ 3 ]. The SAPP model may help to explain Oaten’s findings by documenting that test anxious students may have poor sleep quality multiple days before an exam. Thus, it seems plausible that at least some of the effects of test anxiety on health behaviors (e.g., caffeine intake, healthy eating) observed by Oaten, were attempts to mitigate feelings of fatigue associated with poor sleep quality. Limitations We attempted to comprehensively measure the relationships among sleep, anxiety, and academic performance by using a combination of trait level measures and daily diary measures of sleep and anxiety. However, there were several notable limitations. First, non-invasive wrist worn actigraphs would have provided a more accurate measure of TST, WASO, and SOL. A wealth of literature has documented that sleep diary reports of sleep parameters do not correlate well with objective sleep measured via polysomnography [ 37 ]. Self-reported sleep often is more strongly related to measures of mood and anxiety and has a more modest association with biological and performance measures [ 38 ]. In addition, test anxiety and sleep were measured concurrently in order to minimize response burden for non-financially compensated research participants. However, morning and evening reports would reduce commonly recognized sources of bias (e.g., mood-dependent recall) [ 39 ]. Moreover, these were observational data, weakening our ability to make causal statements about the relationships observed in this study. And finally, generalizability is limited by class type (statistics) and should be made with care in universities with more racially and ethnically diverse undergraduate populations. Summary and Clinical Application The SAPP model offers a comprehensive way to understand the relationship between sleep and anxiety on academic performance. Test anxiety appears separate from trait anxiety and is thus likely to be more variable. One study that tracked the ebb and flow of test anxiety across the semester found that anxiety peaked at the beginning of the semester and dissipated as students become more familiar with the subject matter and test formats [ 7 ]. However, unremitting anxiety was among the most commonly identified mental health concern named by students leaving college [ 40 ]. Test anxiety has typically been thought of as an emotion that arises as students face a test. However, the SAPP model proposes that test anxiety does not start and stop on the day of an exam. Instead, the non-trait nature of test anxiety and the temporal relationship between sleep and text anxiety span at least 2 days leading up to the exam, which suggests a clear opportunity for intervention. As universities and high schools struggle to retain students, interventions to address anxiety-related performance deficits may be part of a broader strategy to prevent drop out because of academic failure. Research on test anxiety interventions has primarily focused on the cognitive–behavioral methods, skill-building, and relaxation interventions to reduce test anxiety and improve test performance (e.g., [ 41 ]. However, individual or even group therapy for all test anxious students is not likely to be a practical solution to this problem. Fortunately, lower-burden interventions show promise. Unlimited exam-time for math and statistics exams [ 9 ] and a seven-minute expressive writing intervention directing students to write about “thoughts and feelings” about an upcoming exam have been shown to alleviate or minimize the effect of anxiety on a quantitative reasoning test [ 42 ]. However, the good news for test anxious students and for universities who wish to retain test anxious students is that interventions such as these are effective, easily delivered via instructional change, and come at no additional cost to the university. Notes It should be noted that race/ethnicity was a significant predictor of exam score. However, the effect was entirely dependent upon a single student’s outlier (poor) test score. Because of the low probability for replication, this effect is not pictured in the model or reported in Table 2 .

College students across the country struggle with a vicious cycle: Test anxiety triggers poor sleep, which in turn reduces performance on the tests that caused the anxiety in the first place. New research from the University of Kansas just published in the International Journal of Behavioral Medicine is shedding light on this biopsychosocial process that can lead to poor grades, withdrawal from classes and even students who drop out. Indeed, about 40% of freshman don't return to their universities for a second year in the United States. "We were interested in finding out what predicted students' performance in statistics classes—stats classes are usually the most dreaded undergrad class," said lead author Nancy Hamilton, professor of psychology at KU. "It can be a particular problem that can be a sticking point for a lot of students. I'm interested in sleep, and sleep and anxiety are related. So, we wanted to find out what the relationship was between sleep, anxiety and test performance to find the correlation and how it unfolds over time." Hamilton and graduate student co-authors Ronald Freche and Ian Carroll and undergraduates Yichi Zhang and Gabriella Zeller surveyed the sleep quality, anxiety levels and test scores for 167 students enrolled in a statistics class at KU. Participants completed an electronic battery of measures and filled out Sleep Mood Study Diaries during the mornings in the days before a statistics exam. Instructors confirmed exam scores. The study showed "sleep and anxiety feed one another" and can hurt academic performance predictably. "We looked at test anxiety to determine whether that did predict who passed, and it was a predictor," Hamilton said. "It was a predictor even after controlling for students' past performance and increased the likelihood of students failing in class. When you look at students who are especially anxious, it was almost a five-point difference in their score over students who had average levels of anxiety. This is not small potatoes. It's the difference between a C-minus abd a D. It's the difference between a B-plus and an A-minus. It's real." Beyond falling grades, a student's overall health could suffer when test anxiety and poor sleep reinforce each other. "Studies have shown students tend to cope with anxiety through health behaviors," Hamilton said. "Students may use more caffeine to combat sleep problems associated with anxiety, and caffeine can actually enhance sleep problems, specifically if you're using caffeine in the afternoon or in the evening. Students sometimes self-medicate for anxiety by using alcohol or other sedating drugs. Those are things that we know are related." Hamilton said universities could do more to communicate to students the prevalence of test anxiety and provide them with resources. "What would be really helpful for a university to do is to talk about testing anxiety and to talk about the fact that it's very common and that there are things that can be done for students who have test anxiety," she said. "A university can also talk to instructors about doing things that they can do to help minimize the effect of testing anxiety." According to Hamilton, instructors are hindered by the phenomenon as well: Anxiety and associated sleep problems actually distort instructors' ability to measure student knowledge in a given subject. "As an instructor, my goal when I'm writing a test is to assess how much a student understands," she said. "So having a psychological or an emotional problem gets in the way of that. It actually impedes my ability to effectively assess learning. It's noise. It's unrelated to what they understand and what they know. So, I think it behooves all of us to see if we can figure out ways to help students minimize the effects of anxiety on their performance." The KU researcher said testing itself isn't the problem and suggested an increase in regular tests might reduce anxiety through regular exposure. However, she said a few small changes to how tests are administered also could calm student anxiety. "In classes that use performance-based measures like math or statistics, classes that tend to really induce a lot of anxiety for some students, encouraging those students to take five minutes right before an exam to physically write about what they're anxious about can help—that's cheap, that's easy," Hamilton said. "Also, eliminating a time limit on a test can help. There's just really nothing to be gained by telling students, 'You have an hour to complete a test and what you don't get done you just don't get done.' That's really not assessing what a student can do—it's only assessing what a student can do quickly." Hamilton said going forward she'd like research into the link between test anxiety and poor sleep broadened to include a more diverse group of students and also to include its influence on remote learning. "The students in this study were mostly middle-class, Caucasian students," she said. "So, I hesitate to say these results would generalize necessarily to universities that have a more heterogeneous student body. I also would hesitate to say how this would generalize into our current Zoom environment. I don't know how that shakes out because the demands of doing exams online are likely to be very different."

10.1007/s12529-021-09973-1